KDIR 2026 Abstracts

Full Papers Short Papers
Full Papers
Paper Nr: 35
Title:

From Reviews to Tweets: Multilingual and Cross-Domain Sentiment Analysis in Romance Languages
Authors:

Sergiu Limboi, Irina Savilencu and Laura Dioşan
Abstract: Sentiment analysis in multilingual settings remains challenging, especially when models are transferred across domains with different linguistic and structural properties. This paper investigates multilingual sentiment analysis and cross-domain generalization across three Romance languages: Spanish, French, and Romanian. We combine experiments on structured review datasets, namely Amazon Reviews Multilingual and LaRoSeDa, with evaluation on newly collected X (formerly Twitter) data. We compare TF–IDF with Logistic Regression against XLM-RoBERTa in binary and three-class classification settings. In addition, we analyze zero-shot predictions from a pretrained multilingual sentiment model on unlabeled X data. To support supervised cross-domain evaluation, we construct a manually annotated X benchmark of 1,000 tweets per language. Additionally, we apply a formal robustness analysis to quantify cross-domain degradation, revealing moderate Source Drop (8–9 points) but substantial Target Drop (12–24 points), indicating that target-domain supervision substantially improves performance on social media sentiment analysis. Results show that XLM-RoBERTa achieves higher macro-F1 than TF–IDF on the annotated X datasets, with language-dependent gains. Overall, our findings highlight the importance of language-specific evaluation and domain-adapted data for multilingual sentiment analysis.
Download
Paper Nr: 49
Title:

Denoising User-Item Graphs with Modality Alignment for Multimodal Recommendation
Authors:

Peilun Li, Yuzhuo Dang and Fei Cai
Abstract: Recommender systems (RS) can alleviate information overload by modeling users’ preferences from their historical interactions. As interaction data are often sparse and ID-only models offer limited semantics for cold-start and long-tail items, multimodal recommenders leverage the increasing availability of visual and textual item content for personalization in scenarios such as e-commerce and media recommendation, by incorporating such features into user-item and item-item graph structures. However, implicit user-item interactions are frequently noisy, so graph neural recommenders can amplify misleading edges and degrade embeddings, and heterogeneous visual and textual representations may remain poorly aligned when fused without explicit cross-modal regularization, which limits recommendation quality. To address these issues, we propose FUSION, extending the FREEDOM graph-based multimodal framework by replacing stochastic user-item edge dropout with learned soft pruning, adding an explicit alignment loss between projected image and text features, and retaining frozen multimodal item-item graphs for high-order item relation modeling. Extensive experiments on three public Amazon datasets show that FUSION outperforms representative graph-based multimodal baselines in terms of Recall and NDCG.
Download
Paper Nr: 57
Title:

Memory-Efficient Approximate Nearest-Neighbor Search via Two-Stage Filtering
Authors:

Naoya Higuchi, Yasunobu Imamura, Takeshi Shinohara, Kouichi Hirata and Tetsuji Kuboyama
Abstract: Approximate nearest-neighbor (ANN) search is considered under a practical setting using a standard desktop PC. We focus on large-scale datasets that do not fit in main memory, where feature vectors are stored in secondary storage such as SSDs and only compact indices are maintained in memory. This paper revisits a previously proposed lightweight ANN indexing framework based on two-stage filtering under a storage-aware setting. In particular, we investigate end-to-end query latency, including secondary-storage access costs, under strict memory constraints and cache-controlled execution environments. Experiments on datasets ranging from tens of millions to one billion vectors show that the method achieves stable millisecond-level query latency with practical recall while using only a modest amount of main memory. These results demonstrate that efficient ANN search is possible even when the memory footprint is severely limited.
Download
Paper Nr: 71
Title:

Trust-Aware Federated Knowledge Discovery under Non-IID Client Data
Authors:

Mahran Jazi and Irad Ben-Gal
Abstract: Federated learning enables collaborative model training without centralizing raw data, but standard aggregation may be unreliable when clients are statistically heterogeneous and differ in data quality. We propose Trust-Aware Federated Knowledge Discovery (TAF-KD), a graph-guided aggregation framework that constructs a dynamic client-relation graph from update-direction alignment, predictive reliability on an approved anchor set, and temporal update stability. In TAF-KD, “trust” is an operational, round-specific aggregation score rather than a cryptographic guarantee, a permanent reputation measure, or proof of client correctness. The graph is used to adapt client aggregation weights while preserving raw-data locality. We evaluate TAF-KD in a controlled stress test using the scikit-learn handwritten-digits dataset with ten clients, Dirichlet label skew, and 70% systematic label remapping in the three largest client partitions. Across three random seeds, TAF-KD achieves the highest mean peak macro-F1 among the compared methods, exceeding FedAvg by 16.25 and 18.93 percentage points at moderate and severe label skew, respectively. These results are specific to the evaluated benchmark and corruption protocol. They do not establish superiority under clean heterogeneity, robustness against adaptive adversaries, fairness to legitimate minority distributions, or update-level privacy.
Download
Paper Nr: 84
Title:

SLogic: Subgraph-Informed Logical Rule Learning for Knowledge Graph Completion
Authors:

Trung Hoang Le, Tran Cao Son, Ishtiaq Ahmed and Huiping Cao
Abstract: Logical rule-based methods offer an interpretable approach to knowledge graph completion (KGC) by capturing compositional relationships in the form of human-readable inference rules. While existing logical rule-based methods learn rule confidence scores, they typically assign a global weight to each rule schema, applied uniformly across the graph. This is a significant limitation, as a rule’s importance often varies depending on the specific query instance. To address this, we introduce SLogic (Subgraph-Informed Logical Rule learning), a novel framework that assigns query-dependent scores to logical rules. The core of SLogic is a context-aware scoring function. This framework is in line with the specificity principle in commonsense reasoning. This function determines the importance of a rule by analyzing the subgraph locally defined by the query’s head entity, thereby enabling a differentiated weighting of rules specific to their local query contexts. Extensive experiments on benchmark datasets show that SLogic achieves competitive performance among rule-based methods. It also generates query-dependent, human-readable logical rules that serve as explicit explanations for its inferences.
Download
Paper Nr: 101
Title:

A New Method of Linear Dimensionality Reduction for Improved Classification Accuracy
Authors:

Jonathan Bodine and Dorit S. Hochbaum
Abstract: Dimensionality reduction is a standard preprocessing step for classification. Two of the most widely used linear methods, Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA), each have well-known weaknesses: PCA ignores the class labels, while LDA assumes that all classes share a single covariance and yields at most c−1 informative directions given c classes. This paper introduces Class-Whitened PCA (CW-PCA), a supervised linear dimensionality reduction method that whitens each class separately and then keeps the directions of greatest total variance in the combined representation. This pools the class-specific covariances by their harmonic, rather than arithmetic, mean and removes the c −1 cap. It further introduces Blended Class-Whitened PCA (BCW-PCA), a two-parameter family that interpolates between CW-PCA and LDA, and CV-BCW-PCA, the method that selects the two parameters by cross-validation on each dataset. For both random forest and nearest-neighbor classifiers, across 11,100 synthetic datasets, CW-PCA gives better average classification accuracy than PCA and LDA, and across twenty real-world datasets, CV-BCW-PCA is on average more accurate than each of PCA, LDA, and CW-PCA.
Download
Paper Nr: 120
Title:

Toward Reproducible Federated Learning Intrusion Detection: A Reporting Protocol for Trustworthy Benchmark Knowledge Discovery
Authors:

Moad Ezzaibouh and Ali Idri
Abstract: Extracting trustworthy knowledge from federated machine-learning experiments depends on disciplined empirical reporting, yet the practices used to evaluate federated learning (FL) intrusion detection systems (FL-IDS) have not kept pace with reproducible-research standards. Surveying recent FL-IDS studies, we identify four recurring reporting deficiencies: single-seed evaluation, exclusive reliance on a metric at the fixed decision threshold of 0.5, the absence of a threshold-aware diagnostic, and the absence of any robustness account under increasing federation size. Using evidence from a completed 450-run benchmark (two datasets, three architectures, five aggregation strategies, three federation sizes, five seeds), the experimental findings indicate that these practices may lead to misleading conclusions under the evaluated conditions: strategy rankings shift across seeds, F1 at the default threshold can label a recoverable model as a failure, and ranking quality (AUC) can stay high while F1 collapses. We then propose a concise seven-element reporting protocol and demonstrate, through an illustrative case study, the conclusion each element protects. The protocol introduces no new algorithm: six elements are post-hoc analyses of standard training outputs, while the seventh, multi-seed replication, is a deliberate and necessary expense. It is offered as a practical checklist for authors, reviewers, and benchmark designers.
Download
Paper Nr: 127
Title:

Meta-Adaptive Representation Selection for Self-Supervised Time-Series Forecasting
Authors:

Ons Saadallah, Isabek Mamatkulov, Mátyás Andó and Tamás Gábor Orosz
Abstract: Self-Supervised learning (SSL) has emerged as an effective approach for learning transferable representations for time-series forecasting. Existing methods, however, typically rely on a single encoder, implicitly assuming that one latent representation is universally suitable for all time series and forecasting scenarios. In practice, different SSL objectives and network architectures capture complementary temporal characteristics, making the optimal representation dependent on the underlying signal dynamics. This paper introduces Meta-Adaptive Representation Selection (MARS), a meta-learning framework that dynamically selects the most suitable self-supervised representation for each input window. MARS first learns a diverse pool of heterogeneous representations using reconstruction-based and contrastive encoders. It then characterizes each input window through statistical descriptors and latent embeddings, enabling a meta-learner to predict which representation is expected to minimize the downstream forecasting error. By adapting the representation rather than relying on a single embedding space, MARS exploits the complementary strengths of heterogeneous self-supervised encoders. Experiments on three multivariate forecasting benchmarks demonstrate that no single representation consistently achieves the best forecasting performance across datasets. Instead, adaptive representation selection consistently matches or outperforms the strongest fixed encoder, highlighting that learning when to use a representation is as important as learning the representation itself. These results establish adaptive representation selection as a promising direction for self-supervised time-series forecasting.
Download
Paper Nr: 139
Title:

GL-3DViT: A Global-Local Feature Fusion Based Hybrid Network for Hyperspectral Land Cover Classification
Authors:

Shruti Pravin Khonde, Gaurav Mishra, Prabhat Dansena, Ashish Ranjan and Rani Kumari
Abstract: Classification of land use and land cover (LULC) hyper-spectral remote sensing data is a difficult task. This is due to the fact that, hyper-spectral data is high in dimension and contains complex texture patterns. Moreover, variability in the data increases the difficulty for the classification task. Convolutional Neural Networks (CNNs) are suitable for local spatial texture based classification. However, CNNs fails to capture contextual dependencies in long range. Though, Transformer architectures can handle these issues; however, they face difficulty to capture fine local texture features. Thus, in this work, we propose a hybrid architecture that integrates 3D Convolutional Neural Networks (3D-CNNs) with Vision Transformers (ViTs) for enhanced LULC classification. Initially, we apply volumetric convolutions in order to extract the fine texture features considering correlation between different spectral bands. The Vision Transformer based architecture captures both global contextual information and long-range dependencies. Moreover, due to the use of local feature extraction utilizing volumetric convolutions with global attention modeling, the proposed architecture achieves superior classification for the task at hand. Experimental results and comparative analysis on the EuroSAT dataset confirms the same. Proposed model achieves the best overall performance, with an accuracy of 98.20%, AUC-ROC of 99.93%, and weighted F1-score of 98.19%.
Download
Paper Nr: 144
Title:

Refining Prompts from Their Own Mistakes: A Case Study on Sustainability Reports
Authors:

Lars Wilhelmi and Matthias Schumann
Abstract: Reliable automatic extraction of sustainability metrics from corporate reports remains difficult. The values appear in tables, charts, and footnotes with selective disclosure. Extraction quality with large language models (LLMs) depends heavily on prompt design. Automatic prompt optimization promises to replace manual trial and error. Yet most methods are tuned on generic reasoning benchmarks, return opaque prompts, and react only to an aggregate score. We present a diagnosis-driven prompt refinement method. It grounds each edit in an empirically derived taxonomy of prompt-component requirements and in an analysis of why the current prompt fails. The diagnosis includes a vision step that inspects the source pages. Every revision is recorded as an auditable edit. On a public CO2-emission benchmark, the refined prompt lifts a small, inexpensive extractor 0.22 tolerance-F1 above four published expert prompts on unseen companies. The one-time refinement cost is about seven dollars. On this benchmark, the loop reconstructs most of the domain’s correctness criteria from its own errors, so an optional expert codebook yields only minor additional gains.
Download
Paper Nr: 150
Title:

Micro-Prompt Chaining for Relational Extraction: An Inference Orchestration Architecture for Complex Documents
Authors:

Gabriela Batista de Souza and Carina F. Dorneles
Abstract: Extracting related entities from long, unstructured documents remains a central challenge for Information Retrieval (IR) and Natural Language Processing (NLP), particularly in legal-administrative document collections. In conventional Retrieval-Augmented Generation (RAG) pipelines, semantic links between entities, such as companies and their shareholders, are often lost when retrieval and inference follow a linear, document-level design; we refer to this failure as contextual fragmentation. In this paper, we propose an LLM-driven inference-orchestration architecture based on Micro-Prompt Chaining, which decomposes relational extraction into dependent, schema-constrained inferences with task-scoped retrieval and Pydantic validation. We evaluate open-weight LLMs (including Llama 3.3 and DeepSeek-R1) on two Brazilian Portuguese corpora: (i) a confidential annotated corpus of documents provided by the Brazilian Public Ministry; (ii) a reproducible synthetic corpus. Measured by Global Document Accuracy (GDA), the architecture improves relational reconstruction. On the Public Ministry corpus, GDA rises from 40% to 60% for single-company documents and from 20% to 32% for multi-company documents. On the synthetic corpus, peak GDA rises from 97% to 100% (single-company) and from 88% to 100% (multi-company). These gains show that changing the inference workflow improves relational reconstruction on both corpora, including the harder Public Ministry setting.
Download
Paper Nr: 181
Title:

HSTN: A Hybrid Seasonal-Trend Network for Long-Term Time Series Forecasting via Selective State Space Models and Convolutional Networks with Dynamic Trust Gating
Authors:

Ilyes Gouider, Ilhem Ltifi, Hounaida Moalla, Naima Halouani and Andrei Doncescu
Abstract: Precise long-term time-series forecasting remains a significant challenge due to non-stationary data distributions and the difficulty of capturing both long-range dependencies and local temporal continuities. This paper proposes HSTN (Hybrid Seasonal-Trend Network), a novel architecture designed to address these challenges through dual-stream decomposition and adaptive fusion mechanisms. HSTN applies Reversible Instance Normalization (RevIN) to mitigate distribution shift, followed by moving-average series decomposition that explicitly decouples input into seasonal and trend components. The HSTN block employs a dual-pathway structure: a Selective State Space Model (SSM) captures global, long-term dependencies within seasonal signals, while multi-scale dilated depthwise-separable convolutions (MultiScaleConv) extract local continuity and sudden changes within trend components. An adaptive Dynamic Trust Gate fuses these representations by computing sample-specific weights, enabling prioritization of the most relevant pathway based on signal characteristics. Experiments on seven benchmark datasets demonstrate competitive performance with computational efficiency, leveraging parallelized selective scans with near-linear complexity. Complete implementation and execution scripts are publicly available.
Download
Paper Nr: 188
Title:

Forget Chunks: Explainable Retrieval via Normative Knowledge Graphs
Authors:

Eduardo de Souza, André Gomes Regino, Eric Tadeu Camacho de Oliveira, Cassia de Souza Carvalho and Julio Cesar dos Reis
Abstract: Normative documents such as laws and regulations present challenges for information retrieval due to their hierarchical structure, semantic density, cross-references, and temporal evolution. Traditional retrieval approaches often fail to capture the underlying normative logic. This study proposes a rule-centric solution for retrieving normative documents, in which a rule-oriented search retrieves normative Knowledge Graph triples (subject, action, condition) that serve as the harness governing contextual evidence and constraining the generative reasoning process and answer. Grounded in the principle that the correct unit of retrieval is the normative rule rather than the text chunk, our approach introduces a dedicated :Rule class as a symbolic anchor that connects entities, conditions, and legal provisions within structured subgraphs. Experimental findings from a legal GDPR QA dataset show that our rule-centric retrieval nearly doubles grounding recall over a vector-based baseline built on the same corpus (84.81% vs 52.42%), while achieving factuality comparable to it (90.86% vs 92.83%), indicating that the gains stem from more complete and traceable normative grounding rather than from raw answer factuality alone. All evaluated models, including a small local language model, outperformed the fine-tuned baseline reported for the original benchmark (68% factuality). These findings suggest broader implications for the economics and governance of legal AI, including lower operational costs and reduced dependency on frontier-scale models.
Download
Short Papers
Paper Nr: 20
Title:

Academic Collaborator Recommendation Based on Hyperlink Prediction
Authors:

Chen Wenjie
Abstract: Academic collaborator recommendation is widely recognized as an important way for researchers to broaden their academic perspectives and identify potential partners across diverse fields. However, existing research primarily focuses on recommending pairwise collaborations and struggles to discover group collaborations. To address this limitation, we propose an Academic Collaborator Recommendation Framework (ACRF) based on hypergraph structure and hyperlink prediction. In ACRF, a hypergraph-based collaboration network is first constructed to model higher-order group collaborations. Then, hyperlink prediction is performed by integrating scholars’ topic distribution features with structural features of the hypergraph. Finally, a two-stage strategy is used to generate recommendation lists. We validate the effectiveness of ACRF in the organ chips field, demonstrating its capability to efficiently identify high-value 3-ary and 4-ary co-authorships. Large-scale experiments on publication data show that ACRF outperforms leading methods in key metrics and can effectively predict diverse scholarly collaborations.
Download
Paper Nr: 22
Title:

Factors Influencing Knowledge Sharing among Public-Sector Herbal Researchers in Thailand: A Conceptual Framework for Platform Development
Authors:

Sukrutai Phansiripat, Mongkolchai Wiriyapinit and Artiwan Shotipruk
Abstract: Thailand's herbal research sector holds considerable strategic value within the national Bio-Circular-Green (BCG) Economic Model, yet a persistent knowledge sharing (KS) gap continues to slow the commercialization of research outputs across Technology Readiness Levels (TRL) 4 through 7, the transition phase widely referred to as the 'Valley of Death.' This study conducts an integrative review of twenty-five empirical and theoretical studies to identify and categorize factors influencing KS behavior among public-sector herbal researchers in Thailand, and proposes a conceptual framework. Four interdependent antecedent dimensions are identified-Individual, Organizational, Socio-Cultural, and Technological-and their effects on KS Behavior are explained through a serial mediation structure grounded in the Theory of Planned Behavior (TPB), where KS Attitude, Subjective Norms, and Perceived Behavioral Control mediate KS Intention. The dependent variable is operationalized using the four SECI knowledge conversion modes, with a counterfactual path linking negative KS Attitude to Knowledge Hiding. Two Thai-specific moderating conditions condition the strength of these relationships. The framework is compared against conventional public-sector KM models, and a general empirical validation approach is outlined to guide its transition toward implementation-ready platform design.
Download
Paper Nr: 44
Title:

An End-to-End Framework for Turkish Financial Report Summarization with Large Language Models and LSTM-Based Stock Classification on BIST100
Authors:

Barış Can, Zahide Gür Çembertaş, Ömer Yalçın, Fatih Serdar Aydemir and Bekir Cetintav
Abstract: Quick and accurate interpretation of financial reports is important for investment decisions. However, manual examination of reports published daily by the Public Disclosure Platform and investment institutions is not sustainable. In this study, an end-to-end pipeline encompassing data collection, summarization, key information extraction, analytical attribute generation, and sentence reconstruction was developed for automatic processing of Turkish financial reports. Analytical attributes containing directional information were extracted from summaries, and sentences suitable for Turkish language structure were generated to select the most appropriate candidates. System performance was evaluated on 1508 summaries using ROUGE, BLEU, semantic similarity, Sentence Mover Similarity, G-Eval, and Reason-Then-Score metrics. Results showed limited performance in overlap-based metrics but high performance in semantic similarity and LLM-based criteria. Additionally, to strengthen the investment decision support dimension, an LSTM-based classification model was developed for 30-day performance prediction based on historical price and technical indicator data of BIST100 stocks, and 12 labeling strategies were compared. The XU100 index-relative performance-based labeling strategy achieved the highest accuracy at 50.48%. The impact of sentiment signals from the LLM pipeline on the LSTM model was evaluated through a controlled ablation study.
Download
Paper Nr: 52
Title:

Transitivity-Aware Neuro-Symbolic Regularization for BERT-Based Natural Language Inference
Authors:

Alisher Kaziz, Bolatzhan Kumalakov and Richard Harrison
Abstract: This paper examines whether neuro-symbolic regularization can improve the reliability of BERT-based Natural Language Inference beyond aggregate classification accuracy. The proposed framework augments a BERT sentence-pair classifier with Logical Neural Network penalties and an additional entailment-transitivity constraint. Three model variants are compared on the all-NLI pair-class setting: a standard BERT baseline, a BERT model with logical regularization and a BERT model with both logical and transitivity-aware regularization. The transitivity-aware model shows superior results among comparative models, with Macro-F1 performance of 86.62%. It exhibits a lower misclassification rate for entailment, contradiction and neutral labels while also reducing false positives for entailment and contradiction relative to the baseline model. These gains are modest rather than transformative, but they indicate that logically structured relations between examples can improve selected error patterns that remain hidden in aggregate metrics. The study therefore positions transitivity-aware regularization as a preliminary but promising mechanism for improving consistency in transformer-based NLI systems.
Download
Paper Nr: 66
Title:

GraphDocGen: Context-Aware Documentation for Java Repositories Using Repository Knowledge Graphs
Authors:

Maria Gomes, Eva Maia and Isabel Praça
Abstract: Automatic software documentation often relies on isolated code fragments, although relevant knowledge is distributed across the repository. Existing context-aware approaches commonly provide this information as unstructured text, making relevance and provenance difficult to preserve. This paper presents GraphDocGen, a knowledge-graph-driven approach for generating contextualized documentation for Java repositories. Using CoReGraph, it retrieves technical, document-derived, and organizational relations associated with each target artifact and organizes them into ranked and provenance-preserving context packs. The approach was evaluated on Java methods and repository documents by comparing graph-enriched generation with a no-context condition. Results indicate that graph-derived context improves alignment with repository-aware references in selected configurations, although its effect varies across models, inputs, artifacts, and metrics. These findings provide preliminary evidence that repository knowledge graphs can support structured and traceable documentation generation.
Download
Paper Nr: 77
Title:

From Operational Deviations to Reliability-Oriented Evidence: A Two-Phase Data-Driven Approach for Heat Pump Systems under Weakly Labeled Field Conditions
Authors:

Ahmed Qarqour, Gernot Heisenberg, Sahil-Jai Arora and Hakan Akdag
Abstract: Heat-pump systems are becoming a cornerstone of sustainable heating, generating continuous multivariate operational data at scale. However, maintenance records in real deployments provide only weak labels – documenting service events without specifying degradation onset, fault duration, or remaining useful life. Existing data-driven reliability approaches often assume accurately labeled failures and well-defined degradation trajectories. This paper proposes a two-phase framework for deriving interpretable reliability-oriented evidence under weakly labeled field conditions. In Phase 1, mode-specific LSTM autoencoders identify deviations from healthy operational behavior and consolidate persistent deviations into anomaly episodes. In Phase 2, these episodes are characterized by their morphology, temporal evolution, and sensor contributions, organized into recurring operational signatures, and aggregated into component-level degradation evidence. Applied to field-installed heat pump systems, the approach identified repeated and persistent anomaly episodes before a documented maintenance event, while no further persistent episodes were observed after the intervention. Rather than directly estimating remaining useful life, a central prognostics and health management (PHM) objective, the proposed framework transforms system-level deviations into interpretable, maintenance-relevant evidence under limited fault information.
Download
Paper Nr: 79
Title:

Synonymy-Aware Masked Language Modeling for OCR Post-Correction: The Case of Illegible Mentions in French Citizen Contributions
Authors:

Sami Guembour and Catherine Dominguès
Abstract: In this paper, we address the problem of corpus correction when the target word is unknown or missing, but its position in the text is known. We apply this setting to the Cahiers citoyens corpus, which brings together citizen contributions collected as part of the Grand Débat National. The corpus contains both handwritten (in great majority) and typed contributions, resulting in diverse lexical forms, writing styles, and readability levels. After digitization, OCR processing, and human verification, some words could not be identified and were replaced with the token ILLISIBLE, which means illegible. We propose a method for improving corpus quality by predicting and correcting these missing words. Our approach relies on CamemBERT, fine-tuned on the Cahiers citoyens corpus, and uses masked language modeling to predict the most plausible word from its context. We evaluate several parameters, including the impact of fine-tuning, the choice of context unit, and the integration of synonymy into candidate selection, using a manually verified ground-truth dataset. Finally, we assess the impact of the correction on downstream semantic analysis by applying sentence clustering to the corrected corpus. The results show that the correction improves corpus usability, increases topic diversity, and slightly increases the number of sentences assigned to thematic clusters.
Download
Paper Nr: 83
Title:

Explainability-Aware Active Learning for Recommender Systems
Authors:

Sami Alshehri, Sami Khenissi, Kyle Spurlock, Behnoush Abdollahi, Wenlong Sun, Pegah Sagheb and Olfa Nasraoui
Abstract: Recommender systems often rely on accurate but opaque models, limiting recommendation transparency. This problem is especially acute for new users, whose few ratings constrain both recommendation quality and the neighborhood evidence available for explanations. We introduce Explainability-Aware Active Learning (ExAL), which incorporates neighborhood-based explainability into active query selection for matrix-factorization recommenders. ExAL includes Greedy and Explore variants, as well as Balanced variants that combine exploration and exploitation, and its query score approximates the effect of one online model update. In the MovieLens-1M results, the validation-adaptive balanced policy (ExAL–Balanced–Auto) achieves the highest observed MAP@10 and NDCG@10 among the compared methods while maintaining Mean Ex-plainability Precision (MEP) close to the highest observed MEP. Across the tested variants, the exploration– exploitation policy affected the balance between ranking quality and neighborhood-based explainability.
Download
Paper Nr: 87
Title:

A Reproducibility Study in Disaster Data Mining: The Hidden Cost of Data Leakage
Authors:

Nicholas Johnson, Kyle Spurlock and Olfa Nasraoui
Abstract: The reproducibility of machine learning models in disaster data mining is a prerequisite for drawing sound, actionable conclusions about their real-world reliability. A central but often overlooked threat to this reproducibility is data leakage, a subtle form of information contamination that arises when training and test sets share samples from the same disaster event. In this paper, we provide an evidence-based answer to the question: To what extent do the evaluation results of ML models reflect their true generalization ability in a realistic disaster scenario? We conduct a systematic comparative evaluation of several supervised learning models across four disaster-related classification tasks, contrasting the standard shuffled train-test split protocol against a disaster-specific evaluation strategy that strictly withholds all samples from a given disaster during training. Our findings reveal that conventional evaluation practices can produce significantly overoptimistic performance estimates due to implicit data leakage from event-specific signals such as location-based hashtags (e.g., #FloodPH). The magnitude of this effect is substantial: accuracy differences between the two protocols range from 10% to 50% depending on the task and model. These findings demonstrate the need for more rigorous, ecologically valid evaluation protocols in disaster data mining research if the results are to have any implications on real-world deployment for disaster relief. Our findings carry both methodological and vast humanitarian implications since the over-estimated performance results from a random (but non realistic) evaluation protocol can have severe consequences in real-world deployment.
Download
Paper Nr: 99
Title:

Efficient Variants of the Conjugate Gradient Algorithm for Neural Networks
Authors:

Götz-Henrik Wiegand, Tomas Hrycej, Massimo Pavone, Bernhard Bermeitinger and Siegfried Handschuh
Abstract: The use of second-order optimization methods has been limited by their high computational expense. While recent variants focus on reducing this cost, such simplifications often risk premature termination, a more severe issue than expense itself. We identify primary causes of premature causes: simplified line search procedures and non-descending conjugate search directions. Computing experiments on a reduced GPT-2 model and Resnet-18 show the superiority of (1) a safe Golden Section line search over the Wolfe method, and (2) the Dai-Yuan beta method of the search direction, which generalizes convergence guarantees to standard Wolfe conditions.
Download
Paper Nr: 108
Title:

Evaluating Embedding Models for Scientific Paper Titles: A Prelude to Temporal Topic Tracking
Authors:

Karlo Babić, Patrik Mandić, Slobodan Beliga and Ana Meštrović
Abstract: Tracking the evolution of scientific literature requires measuring the semantic similarity of ideas over time. To generate data for such temporal tracking, it is necessary to identify text embedding models that perform well on concise, domain-specific texts like scientific paper titles. We evaluate seven transformer-based embedding models on titles from the arXiv quantum physics (quant-ph) dataset. We use two evaluation methods. First, we measure intrinsic metadata alignment using category clustering and title-abstract correlations. Second, we use an LLM-based ranking framework where Gemini-3-Pro generates consensus similarity rankings for model comparison. Results show that mxbai-embed-large-v1 achieves the highest category-based alignment (P@5), while bge-large-en-v1.5 achieves the strongest title-abstract correlation. bge-small-en-v1.5 provides the best balance of speed and accuracy. We present a preliminary application of the chosen embeddings using a sliding-window temporal clustering algorithm. This pipeline will be used in future work to generate datasets of evolving topics, which are required to train parameters and evaluate evolutionary frameworks.
Download
Paper Nr: 113
Title:

Contextual Unsupervised Shapelet Learning: Towards Interpretable Time Series Clustering
Authors:

Matthieu Delahaye, Lina Fahed, Florent Castagnino and Philippe Lenca
Abstract: Explainable Artificial Intelligence (XAI) for time series remains challenging as relevant information is distributed across ordered observations and temporal relations. For example, in unsupervised settings, time series clustering can reveal unknown groups but the resulting clusters remain difficult to interpret. Shapelet learning methods address this problem by representing local temporal patterns that can be related back to observed subsequences. However, existing unsupervised shapelet-learning methods mainly rely on temporal information and provide limited support for integrating knowledge external to the series. To face this challenge, we propose a contextual unsupervised shapelet-learning framework that combines shapelet-based temporal similarities with external contextual knowledge. Our framework provides a free label evaluation based on cluster validity and stability, and connects global cluster descriptions with local shapelet patterns, which yields more interpretable clusters and greater explainability for domain experts. We evaluate our framework through an urban crime case study in Toronto, Canada, where weekly offence time series are associated with census-based socioeconomic and demographic neighbourhood profiles. Although using temporal information alone produced more clearly separated clusters according to the internal metrics, integrating the actual neighbourhood context led to more reproducible partitions: the mean agreement between independent runs, measured by the Adjusted Rand Index, increased from 0.524 to 0.702. The resulting four cluster solution obtained a mean cluster separation score of 0.619, measured by the Silhouette index, and supported complementary interpretations based on offence composition, neighbourhood profiles, spatial distribution and local temporal patterns.
Download
Paper Nr: 113
Title:

Contextual Unsupervised Shapelet Learning: Towards Interpretable Time Series Clustering
Authors:

Matthieu Delahaye, Lina Fahed, Florent Castagnino and Philippe Lenca
Abstract: Explainable Artificial Intelligence (XAI) for time series remains challenging as relevant information is distributed across ordered observations and temporal relations. For example, in unsupervised settings, time series clustering can reveal unknown groups but the resulting clusters remain difficult to interpret. Shapelet learning methods address this problem by representing local temporal patterns that can be related back to observed subsequences. However, existing unsupervised shapelet-learning methods mainly rely on temporal information and provide limited support for integrating knowledge external to the series. To face this challenge, we propose a contextual unsupervised shapelet-learning framework that combines shapelet-based temporal similarities with external contextual knowledge. Our framework provides a free label evaluation based on cluster validity and stability, and connects global cluster descriptions with local shapelet patterns, which yields more interpretable clusters and greater explainability for domain experts. We evaluate our framework through an urban crime case study in Toronto, Canada, where weekly offence time series are associated with census-based socioeconomic and demographic neighbourhood profiles. Although using temporal information alone produced more clearly separated clusters according to the internal metrics, integrating the actual neighbourhood context led to more reproducible partitions: the mean agreement between independent runs, measured by the Adjusted Rand Index, increased from 0.524 to 0.702. The resulting four cluster solution obtained a mean cluster separation score of 0.619, measured by the Silhouette index, and supported complementary interpretations based on offence composition, neighbourhood profiles, spatial distribution and local temporal patterns.
Download
Paper Nr: 122
Title:

Signs of Trouble: Diffusion Augmentation and Backdoor Risks in Traffic Sign Detection
Authors:

William Aiken, Paula Branco, Guy-Vincent Jourdan and Iosif-Viorel Onut
Abstract: Autonomous driving depends on fast and reliable detection across a wide range of objects and conditions. Synthetic data offers a promising way to augment rare traffic-sign classes, but the security consequences of trusting a generative model as a training data source remain poorly understood, particularly in federated learning (FL). In this work, we study benign and backdoored diffusion models that generate synthetic traffic-sign data for downstream YOLOv7 object detection. Our evaluation covers centralized and federated training with real-only data, benign synthetic augmentation, and different poisoning scenarios, including targeted single-class label-flipping as well as broader all-to-one label-flipping attacks. Clean synthetic augmentation substantially improves detection utility, but poisoned synthetic data can also transfer trigger-conditioned behaviour into downstream detectors. Across TT100K and CANDOR-99 datasets, one-to-one and all-to-one attacks produce class-specific backdoor attacks while clean mAP@.5:.95 remains close to benignly augmented baselines, and a single malicious client in FL can further amplify the effect. These results identify diffusion-based augmentation as both a useful source of training data and a consequential supply-chain attack surface.
Download
Paper Nr: 124
Title:

RAISE: Tokenizer-Aware Rule-Based Explanations for Transformer Log Anomaly Detection
Authors:

Paul K. Mvula, Paula Branco, Guy-Vincent Jourdan, Iosif-Viorel Onut and Herna L. Viktor
Abstract: System logs are a primary source of evidence for detecting anomalies and intrusions in computer infrastructure, yet transformer-based detectors increasingly used for this purpose remain opaque: operators need not only a detection decision but an interpretable rationale tied to specific log tokens. Comparing sparse symbolic explainers with dense attribution explainers requires resolving structural asymmetries that existing protocols often overlook: sparse methods may return no output, clause lengths differ from attribution budgets, and shared-output versus overall agreement capture different properties. We address this by introducing a unified evaluation protocol that separates native per-explainer configuration selection from final shared-budget comparison while explicitly retaining empty outputs. Within this protocol, we introduce RAISE, a tokenizer-aware, model-agnostic rule-based explainer that produces compact symbolic rationales and component-level summaries for individual log lines. RAISE and the protocol are evaluated across 28 cross-system scenarios spanning four transformer detector families and seven log datasets. Results show complementary behaviour rather than universal dominance: RAISE is compact and efficient; attribution methods provide broader sensitivity maps; and Anchors can be informative where available, but has substantially lower explanation availability. A controlled four-scenario cross–intra diagnostic isolates detector-regime effects. Masking diagnostics measure faithfulness to detector behaviour, not expert, semantic, or causal correctness.
Download
Paper Nr: 143
Title:

A Formal and Empirical Study on the Geometry of Multi-Hop Query Identification
Authors:

Tarik Hastor, Ahmed Haračić and Fedja Arnautović
Abstract: Agentic retrieval systems excel at multi-hop question answering but cost far more than conventional retrieval-augmented generation, motivating a lightweight way to identify multi-hop queries before invoking expensive workflows. We investigate the boundary hypothesis: that multi-hop queries can be identified by their proximity to Voronoi boundaries in retrieval embedding spaces, and should therefore exhibit smaller retrieval margins than single-hop queries. We develop a geometric framework relating retrieval margins, Voronoi geometry, embedding normalization, neighborhood statistics, and classifier performance, with five results formalized and machine-verified in Lean 4 using Mathlib. Evaluating on 2WikiMultiHopQA across four settings of increasing methodological control, we find the apparent effect weakens or reverses once title leakage, probe-construction artifacts, and normalization are addressed: controlled multi-hop queries exhibit larger margins and greater boundary separation. Thus multi-hop structure is not encoded as boundary proximity, yet hop count remains highly identifiable from other geometric features.
Download
Paper Nr: 147
Title:

Refusal Theater: Evaluation of Declared Refusal and Operational Disclosure in LLM Safety Responses
Authors:

Selin Şaşal and Ozgu Can
Abstract: Large Language Models (LLMs) are increasingly deployed in security-sensitive applications. However, the reliability of their refusal behavior as a safety mechanism remains poorly understood. This study investigates whether declared refusal reliably predicts the practical safety of a model response or whether refusal language and operational disclosure can diverge within the same output. Thus, we introduce a two-axis evaluation framework that independently assesses declared refusal and operational disclosure and apply it to Claude Sonnet 4.6 using a controlled sequence of semantically refined cybersecurity prompts. The results show that diagnostic reframing and multi-turn context accumulation can trigger medium-to-high operational disclosure without jailbreak language. Moreover, declared partial refusal can co-occur with high operational disclosure in a single response - a failure mode that we term refusal theater. On the contrary, the comparative risk-reporting setting produced controlled responses under equivalent conversational pressure, indicating that refusal reliability is content-category-dependent rather than a uniform model property. These findings challenge binary refusal scoring as a sufficient safety evaluation method for LLMs and demonstrate the need for evaluation frameworks that measure operational disclosure independently of declared refusal. As a result, the study provides a practical methodology for identifying failure modes that remain undetected by conventional refusal-based safety evaluations.
Paper Nr: 151
Title:

BFEL: Building Footprint Extraction from 3D Airborne LiDAR Point Clouds Using CNN and Geometric Methods
Authors:

Saqib Alam Ansari and Jaya Sreevalsan-Nair
Abstract: Extracting building footprints from airborne LiDAR point clouds is essential to urban mapping and GIS applications. There is a gap in the state of the art regarding generalized methodology across different datasets. We propose BFEL, a modular system for building footprint extraction from LiDAR, with five stages. The stages include binary semantic segmentation, building instance isolation, adaptive alpha-shape seed generation for boundary initialization, learning-based boundary refinement, and geometric regularization. Our novel component, GeoRefineNet, is a one-dimensional convolutional network that uses geometry-aware features to predict displacements in a local frame and refine the initial boundary. All processes preserve the closed-loop structure of the footprint. The refined boundary geometry is further optionally regularized to create clean and structurally consistent footprints. The boundary quality is further tested using experiments on the AHN3 dataset and cross-dataset evaluation on the ISPRS Vaihingen dataset, and compared to the initial alpha-shape boundary. Overall, our experimental results show that BFEL provides an effective and practical solution for producing accurate building footprints from airborne LiDAR point clouds.
Download
Paper Nr: 153
Title:

On the Limits of Triple-Only Retrieval for GraphRAG: Evidence from Zero-Shot Ontology-Guided Exploration
Authors:

Mateus Edival Rodrigues da Silveira, Julio Cesar dos Reis and Allan M. de Souza
Abstract: Retrieval-Augmented Generation (RAG) has emerged as a relevant approach to mitigate problems inherent to Large Language Models (LLMs), such as lack of up-to-date knowledge and hallucinations. Nevertheless, further addressing the multi-hop reasoning deficits of standard RAG, Knowledge Graphs (KGs) have been increasingly integrated as a reliable source of structured knowledge for reasoning. However, state-of-the-art GraphRAG approaches frequently impose expensive training phases that require task-specific, labelled data. Additionally, they lack zero-shot adaptability across different schemas. This study introduces ForesightRAG, an iterative, training-free pipeline that uses the graph ontology to ground LLM generation exclusively in structured Resource Description Framework (RDF) triples. Our framework provides a clear and verifiable reasoning path for the model’s conclusions. Our empirical evaluation across standard KGQA datasets reveals a critical finding. Although our proposed method grounds its responses in verifiable factual traces, combining semantic graph traversal with raw triples does not consistently outperform traditional parametric Chain-of-Thought (CoT) baselines. While not surpassing state-of-the-art performance, this work delivers a comprehensive analysis of ForesightRAG’s limitations and outlines a clear pathway to improvements that can overcome the strict limitations of purely structural context.
Download
Paper Nr: 156
Title:

Large Language Model-Based Autonomous Agents: A Survey of Architectures, Coordination, Evaluation, and Practical Guidance
Authors:

Adel Hidri, Noha Hamoudah, Marwa Torki and Minyar Sassi Hidri
Abstract: Large language models are increasingly used not only as text generators but also as the core reasoning component of autonomous and semi-autonomous agent systems. When coupled with memory, planning, tool use, and interaction with external environments, these systems can address tasks that are difficult to solve through single-turn prompting alone. This paper presents a survey of large language model-based autonomous agents, with emphasis on architectural design, coordination strategies, evaluation methodology, failure recovery, and practical system selection. The survey organizes prior work along three dimensions: agent cardinality, memory design, and planning style. It also compares representative systems such as ReAct, Reflexion, AutoGen, MetaGPT, CrewAI, LangGraph, DyLAN, and ToolMaker, while reviewing application domains including software engineering, research support, healthcare, education, finance, and customer service. In addition, it synthesizes current evaluation practices across AgentBench, SWE-bench, SWE-bench Verified, GAIA, and WebArena, highlighting the extent to which scaffold design influences reported performance. The survey also examines failure propagation, state synchronization, retry behavior, benchmark limitations, and reproducibility factors that influence agent-system evaluation. The objective is to provide a systems-oriented synthesis for researchers and a practical design reference for practitioners developing deployable agent systems.
Download
Paper Nr: 169
Title:

Assessing Pre-Trained LLMs for Topic-Modeling Based Sentiment Analysis
Authors:

Patrick A. M. de Araújo, Marcelo da Silva Reis and Julio Cesar dos Reis
Abstract: The rapid advancement of natural language processing (NLP) has significantly improved sentiment analysis, essential for understanding user emotions in text. Pre-trained large language models (LLMs) excel in zero-shot, one-shot, and few-shot learning, making them potentially powerful for sentiment analysis. However, integrating topic modeling presents challenges, particularly in enhancing classification accuracy and handling nuanced tasks. This study evaluates how in-context learning with LLMs favors sentiment analysis with and without topic modeling. Using zero-shot, one-shot, and few-shot learning, we compare it across three-label and five-label classifications on three relevant datasets. Experimental results with statistical test validation demonstrate that Meta Llama 3 8B performs better in simpler three-label tasks, whereas Mixtral 8x7B excels in complex five-label scenarios. Although topic modeling sometimes leads to improvements-particularly in few-shot settings with five labels-its impact is not universally positive across all datasets and configurations. These findings highlight both the opportunities and limitations of topic modeling when integrated into sentiment classification workflows.
Download
Paper Nr: 176
Title:

UM-MTE v1.0: An English–Maltese Corpus for Evaluating Open-Weight Translation Models
Authors:

Davis Poku Amankwa, David Ohiole Imolore, Chukwuebuka Sylvanus Ngwu and Joel Azzopardi
Abstract: Maltese remains comparatively under-represented in resources for training and evaluating machine translation. We introduce UM-MTE version 1.0, a corpus of 13,535 non-synthetic English–Maltese parallel documents and articles from the Malta Government Gazette and legislation, Newsbook local news, and the Constitution of Malta. It contains 18.28 million English and 15.56 million Maltese whitespace-delimited tokens and preserves document-, article- or notice-level alignment. We establish a reproducible bidirectional benchmark on a frozen, length-stratified sample of 500 bilingual units: 250 Gazette/legislation, 200 Newsbook and 50 Constitution units. A general 4B LLM, a translation-specialized 4B LLM and a dedicated 600M neural machine translation (NMT) system are compared with failure-preserving BLEU and chrF++, confidence intervals, paired tests and COMET on a common length-safe subset. NLLB-200 leads English-to-Maltese with 17.46 BLEU and 51.34 chrF++, whereas Gemma 3 4B leads Maltese-to-English with 32.38 BLEU and 57.97 chrF++. Source and failure analyses expose substantial domain, direction and system variation. Release-wide Unicode, duplication, length-ratio and language-ID checks, together with an exact OPUS-100 overlap audit, delimit the evidence supported by the corpus.
Download
Paper Nr: 178
Title:

Coupling Sentence-Transformer Embeddings and Self-Organizing Maps for Interactive Semantic Document Exploration
Authors:

Michel L. Charles, James Larouy and Andrei Doncescu
Abstract: The exponential growth of unstructured textual data makes the manual exploration of large corpora increasingly impractical. This paper presents a complete pipeline coupling Sentence-Transformer embeddings (allMiniLM-L6-v2) with Self-Organizing Maps (SOM) for interactive semantic document exploration, and integrating a generative language model (llama3.2) for real-time cluster labelling. Documents are encoded as dense 384-dimensional vectors and projected onto a two-dimensional Kohonen map that preserves topological structure, ensuring thematically related documents are mapped to neighbouring neurons. The pipeline is evaluated on the 20 Newsgroups benchmark (18,846 documents, 20 categories). A systematic study of four grid configurations shows all quality metrics improve monotonically with grid size. The retained 80×80 configuration achieves QE=0.759, TE=8.4%, Silhouette=+0.023, DB=2.01 (57% improvement over baseline). External metrics (NMI=0.52, ARI=0.31, purity=0.68) demonstrate competitive clustering performance. A web interface with real-time cluster labelling by a local language model operates at zero API cost. Generality is demonstrated on Shakespearean literary corpora with quantitative validation metrics.
Download
Paper Nr: 180
Title:

CLIP-CMA & NeuroClipBD: Competitive Image–Text Retrieval and Brain Decoding with Minimal Trainable Parameters
Authors:

Mohamed Hajji, Hounaida Moalla, Naima Halouani and Andrei Doncescu
Abstract: Language-supervised vision models such as CLIP align closely with high-level visual cortex, which makes them natural targets for fMRI-to-image retrieval. We study whether a CLIP space adapted with few trainable parameters on curated captions is also easier to decode from brain activity. Phase 1 (CLIP-CMA) fine-tunes OpenCLIP ViT-L/14 on MS-COCO 2017 with partial unfreezing, gated residual adapters and a training-time cross-modal fusion objective (16.1 % of 437.9 M parameters trainable). It reaches an image-to-text Recall@1 of 71.94 % (95 % CI 70.7–73.2) on the 5K val-2017 split, +6.8 points over the frozen model. Phase 2 (NeuroClipBD, BD-v2) trains a 67 M-parameter residual MLP that maps 15,724 cortical voxels of NSD subject 01 to the frozen CLIP-CMA image space on a single GPU. On the exhaustive 982-image test pool (chance 0.10 %) it obtains Recall@1 = 16.8 % (CI 14.6–19.3), Recall@5 = 40.6 % and Recall@10 = 55.3 %, above linear decoders trained on the same voxels. We show analytically that the pooled cross-modal attention reduces to a learned linear map, and we discuss which controls—in particular the same decoder trained against frozen CLIP targets—are required before the adapted space can be said to be more decodable.
Download
Paper Nr: 25
Title:

Guardian Crawler: Retrieval-First Knowledge Discovery with Bounded LLM Augmentation for Noisy Web Intelligence
Authors:

Joshua Castillo, Santosh Nukavarapu and Ravi Mukkamala
Abstract: Retrieving relevant evidence from noisy web data is challenging, particularly in sensitive domains containing incomplete reports, heterogeneous language, and irrelevant content. We present Guardian Crawler, a reproducible retrieval-first testbed for controlled experiments on knowledge discovery and evidence-grounded summarization over synthetic web-like corpora. The architecture combines BM25 retrieval with risk-aware, embedding-augmented, and hybrid reranking, followed by constrained retrieval-augmented generation with explicit document citations. Experiments on a synthetic 900-document corpus and 10 queries produced the highest descriptive retrieval scores under risk-based reranking, with P@10 = 1.00 and NDCG@10 = 0.94, compared with 0.94 and 0.81 for BM25. The best hybrid and BM25+Semantic configurations reached NDCG@10 values of 0.94 and 0.88, respectively. All 41 evaluable generated bullets passed the lexical coverage threshold; an automated LLM judge classified 36 as supported, one as partially supported, and four as unsupported. These results demonstrate the feasibility of Guardian Crawler as a controlled testbed but do not establish statistical superiority, human-validated faithfulness, or transfer to live-web investigative environments.
Download
Paper Nr: 27
Title:

LuminaCite: A Hybrid Knowledge Discovery and Information Retrieval System for Explainable Interdisciplinary Literature Discovery
Authors:

Müge Akbulut
Abstract: Scholarly literature discovery requires both focused retrieval and the identification of hidden, cross-disciplinary connections. This paper presents LuminaCite, an explainable hybrid knowledge discovery and information retrieval (KD+IR) platform for exploratory scholarly search in a large Turkish digital library corpus. The system combines Latent Dirichlet Allocation (LDA) as an interpretable topic-based discovery layer with pennant retrieval as a citation-based contextual ranking layer. A user-controlled weighting parameter exposes the relevance-diversity trade-off and displays LDA, pennant, and integrated scores at the result level. LuminaCite was implemented over 633,931 Turkish scholarly articles and 3,768,409 citation links from DergiPark. A scenario-based evaluation with two representative scholarly search tasks compares LDA, pennant retrieval, and the integrated model using NDCG@10, Precision@10, Shannon diversity, and unique topic counts. The results show that pennant retrieval yields the strongest local relevance scores, LDA yields the broadest topical diversity, and the integrated model provides a balanced, transparent compromise between the two objectives.
Download
Paper Nr: 36
Title:

A Legal Page-Indexed Navigation Engine (L-PINE) for Precise Information Retrieval in Intellectual Property Law
Authors:

Rishabh Dixit, Trupti Khodwe and Madhav Rao
Abstract: The deployment of Large Language Models (LLMs) within the legal domain is often hindered by two primary failures: factual hallucinations and the inability to maintain structural context within complex, table-heavy statutory documents. While traditional Retrieval-Augmented Generation (RAG) relies on semantic similarity via vector embeddings, these methods frequently suffer from domain collisions and context fragmentation when applied to the nuances of Intellectual Property (IP) law. This paper introduces L-PINE, a ”reasoning-first” retrieval architecture that utilizes the PageIndex technique, a page-centric indexing methodology that preserves the native structural boundaries of legal documents and replaces standard embedding-based retrieval with a multi-stage, rule-guided reasoning pipeline. The system employs a Librarian for domain-specific routing, a reasoning-based Navigator for precise page selection, and a Neighbor-Aware Fetcher to preserve the integrity of legal schedules and cross-page clauses. We evaluate our approach using a specialized 100-query stress test derived from Indian Patent, Trademark, and Design Rules. Our results demonstrate a factual accuracy rate of 93.16%, significantly outperforming baseline legal assistants in traceability and citation grounding. Furthermore, we report a routing hit rate of 85.0%, illustrating the efficacy of neighbor-aware stitching in mitigating retrieval near-misses. Unlike fine-tuned models, L-PINE exhibits near-instant adaptability to unseen legal datasets, establishing a new benchmark for verifiable and auditable AI in highly regulated legal environments.
Download
Paper Nr: 55
Title:

Leakage-Aware Short-Term Solar Photovoltaic Power Forecasting Using Machine Learning and Feature Ablation
Authors:

Hamda Khalifa Humaid Alremeithi and Ishtiaq Rasool Khan
Abstract: Accurate short-term solar photovoltaic (PV) power forecasting is important for renewable energy integration, plant operation, and grid stability. This study develops a leakage-aware machine learning approach for 15-minute-ahead plant-level AC power forecasting using Plant 1 from the Solar Power Generation Data dataset. Inverter-level generation records were aggregated into timestamp-level plant output and merged with weather sensor data. The target was total AC power output, while DC power, daily yield, and total yield were excluded from the main predictor set to avoid leakage-like operational variables. Weather, time-based, lagged, rolling, and interaction features were engineered and evaluated using a chronological train-validation-test split. Persistence and seasonal naïve baselines were compared with Linear Regression, Ridge Regression, Support Vector Regression, Decision Tree, Random Forest, Gradient Boosting, and XGBoost. Machine learning models substantially outperformed the naïve baselines. Linear Regression achieved the lowest overall RMSE, while Random Forest achieved the best daylight-period MAPE. Random Forest ablation showed that weather variables combined with AC power lags and rolling statistics produced the strongest daylight RMSE.
Download
Paper Nr: 61
Title:

Automatic Price Prediction from CAD Files
Authors:

Robin Sutter, Yannik Holenstein, Bella Fernández and Bernhard Bermeitinger
Abstract: The paper presents a proof of concept to automate quotation generation for modular construction, thereby replacing a slow, manual process. From 975 priced STEP models, about 40 geometric and structural features were extracted to make the data suitable for machine learning. Several models were tested: while LSTM on raw STEP data failed, tree-based models (random forest and especially XGBoost) performed best, with XGBoost using six key features. The deep-learning-based TabNet underperformed due to the limited dataset size. Due to the limited number of samples for expensive designs, a two-stage pipeline is proposed: a model handles typical cases, while a classifier flags high-priced cases for manual quotation.
Download
Paper Nr: 110
Title:

Interpretable Feature Selection for Regression via Associative Local Distribution Rule Mining
Authors:

Tomonori Obana and Kaoru Shimada
Abstract: Feature selection aims to improve machine learning model performance by selecting features effective for prediction. However, among existing feature selection methods, most filter methods have focused on classification, and research on regression is limited. In addition, although these methods provide a certain degree of interpretability regarding the feature selection mechanism, they evaluate the relationship between each feature and the target variable individually based on global statistics. Therefore, understanding the conditions in which the selected features are considered important is difficult. In this study, we propose an interpretable filter method built on the discovery of associative local distribution rules (ALDRs), which are itemsets associated with distinctive local distributions of a continuous target variable and, on the quantification of feature importance based on them. First, ALDRs are discovered using GNMiner, which employs an evolutionary process with accumulated outcomes across generations. Subsequently, the proposed method quantifies each feature’s contribution to characterizing the local distributions of the target variable. Finally, based on this quantification, a promising feature subset is identified. This approach also provides clues that support interpretability of the selection process. Experimental results suggest that our method helps interpret the selection process and indicate that it achieves performance comparable to four existing methods.
Download
Paper Nr: 115
Title:

A Multi-Layered Large Language Model Firewall: Defense-in-Depth against Prompt Injection
Authors:

Vesile Çelik and Özgü Can
Abstract: Large Language Models (LLMs) have rapidly proliferated in production systems such as conversational assistants, code-generation tools, and search interfaces; however, their instruction-following behaviour has introduced new attack surfaces, most notably prompt injection and jailbreaking. Existing defenses are often inadequate: keyword and regular-expression filters are easily bypassed by semantic variations of an attack, while alignment-based protections require costly model retraining and can still be circumvented. This work presents a multi-layered, defense-in-depth firewall that can be added as middleware to existing LLM applications without retraining the underlying model. The architecture comprises three complementary layers: Layer 1 (semantic analysis), based on vector embeddings and cosine similarity; Layer 2, a lightweight judge language model that performs contextual reasoning; and Layer 3, an output filter that masks personal data using regular expressions and named-entity recognition. The system is built on the OpenAI text-embedding-3-small and gpt-4o-mini models and FastAPI services. On a balanced set of one thousand samples, the full pipeline achieves an F1 of 0.986 and a false-positive rate of 0.8%; the hypotheses that the layered defense is superior to individual layers (H2) and that semantic analysis is superior to keyword filters (H1) were tested using the McNemar test and bootstrap confidence intervals. Although the Layer 1 semantic analyzer has strong discriminative power (ROC AUC = 0.979), it is shown that the default threshold (0.85) needs to be recalibrated. The Layer 3 output filter masks personal data with an F1 of 0.903.
Download
Paper Nr: 131
Title:

LBK: A Heuristic Approach to Meta-Based Automated Reasoning in Knowledge Graphs
Authors:

Simon von Oppenkowski, Benedikt Lerch and Klemens Schnattinger
Abstract: Based on the original premise addressing the challenge of uncertainty in Knowledge Graphs (KGs), this paper serves as an extension for a fully automated solution to the uncertainty problem. The fundamental challenge of uncertainty introduced by newly added nodes into a KG remains. The core concept of an abstraction layer in the form of a subgraph is continuously pursued. In the underlying LBK framework, this isolated graph validates vague or uncertain knowledge by analyzing the surrounding neighbors and abstracting their properties through Galois derivation. This process leads to the consideration of domain-specific knowledge belonging to the adjacent nodes of the newly introduced entity. There, safety validation is achieved through the application of Byzantine Fault Tolerance in graphs. Originally, a threshold of > 0.75 was established as a fully automated validation boundary, where any range between 0.5 and 0.75 required supervised classification, and any value below < 0.5 was deemed false. The calculation is derived from the set of derived features in relation to the total set of features introduced by the validation set of nodes. In this paper, we address the challenge of supervised classification, as it blocks not only human resources but also the execution of further dependent processes. To purposefully solve this problem and make the framework truly fully automated, we propose a heuristic acceptance corridor, an adaptive statistical reasoning mechanism based on previously learned supervised classifications, thereby approximating a machine-based heuristic.
Download
Paper Nr: 135
Title:

Comparative Study of Constraint-Aware AI Paradigms for Climate Monitoring
Authors:

Imen Gatfaoui, Zouhayra Ayadi and Imed Riadh Farah
Abstract: This paper presents a structured literature-based comparative study of three major AI paradigms for climate monitoring: Vision-Language Models (VLMs), Explainable Artificial Intelligence (XAI), and Knowledge-Guided Reasoning (KGR). Although these paradigms have significantly advanced multimodal perception, explainability, and scientific reasoning, none simultaneously ensures multimodal understanding, scientific consistency, and constraint-guided reasoning. We compare representative approaches according to five dimensions: multimodal representation, explainability, explicit constraint integration, scientific consistency, and scalability. The analysis highlights their complementary strengths and recurring limitations, particularly the separation between multimodal learning and constraint-aware reasoning. Based on these findings, we identify key scientific gaps and discuss constraint-guided neuro-symbolic reasoning as a promising integration perspective combining multimodal learning, semantic and physical constraints, and symbolic knowledge. This study highlights the potential of constraint-aware neuro-symbolic AI as a promising direction for developing interpretable, scientifically consistent, and scalable climate monitoring systems.
Download
Paper Nr: 146
Title:

Following the Promise: Retrieval-Augmented Analysis of Parliamentary Action
Authors:

Amalia Pons, Antoni Mestre, Manoli Albert and Vicente Pelechano
Abstract: Political accountability requires connecting electoral commitments with subsequent parliamentary activity. However, this task remains difficult because political documents are long, heterogeneous, and semantically complex. This paper presents a coherence-aware Retrieval-Augmented Generation framework for political information retrieval. The proposed system treats electoral manifestos as a reference corpus, retrieves relevant commitments, and uses a Large Language Model to generate evidence-grounded coherence judgements for parliamentary actions. Unlike generic document question-answering systems, the framework combines domain-specific preprocessing, metadata-enriched chunking, semantic retrieval, an explicit coherence rubric, and constrained prompting to support interpretable political analysis. We frame manifesto-to-action comparison as an information retrieval and knowledge discovery task: given a parliamentary document or action, the system must retrieve the most relevant electoral commitments and assess whether the action is coherent, partially coherent, incoherent, or indeterminate with respect to those commitments. The framework is evaluated through an expert-based parliamentary study involving five Members of Parliament and two parliamentary ad-visors. The evaluation assesses retrieval relevance, agreement with expert coherence judgements, explanation quality, perceived usefulness, and unsupported claim rate. The evaluation follows a lightweight but realistic parliamentary setting combining technical metrics and expert-based user assessment. The paper contributes a task formulation, a domain-specific RAG architecture, and an evaluation protocol for grounded political information retrieval.
Download
Paper Nr: 152
Title:

A Machine Learning-Based Web System for Predicting Patient No-Shows and Supporting Medical Appointment Rescheduling in Level II-2 Healthcare Facilities
Authors:

Bryan Anderson Llamocuro Hermosilla, Romina Lelia del Sol Sosa Saavedra and Alejandrina Nelly Huarcaya Junes
Abstract: Patient no-shows in outpatient healthcare services contribute to underutilized clinical resources, longer waiting times, lost care opportunities, and reduced service quality. Although machine learning models have demonstrated the ability to predict appointment non-attendance, many approaches focus primarily on predictive performance without integrating predictions into operational appointment-management workflows. This paper presents Citalyx, a machine learning-based web system designed to predict patient no-shows and support medical appointment rescheduling in Level II-2 healthcare facilities. The system integrates historical data processing, appointment management, medical slot availability, no-show risk prediction, operational dashboards, and recommendation generation. A LightGBM-based predictive component estimates non-attendance probability and classifies appointments by risk level to support preventive actions such as attendance confirmation, reminder prioritization, slot reassignment, and rescheduling. Evaluation using historical outpatient appointment data yielded an AUC-ROC of 0.734, recall of 0.83, precision of 0.31, F1-score of 0.45, and Brier score of 0.144. These results provide initial evidence that Citalyx can support proactive appointment management by connecting predictive analytics with operational decision-making.
Download
Paper Nr: 157
Title:

Three-Level Multi-Core CPU Optimization for Pneumonia Detection in Chest X-Ray Classification
Authors:

Adel Hidri, Marwa Torki, Noha Hamoudah and Minyar Sassi Hidri
Abstract: Deep learning techniques have demonstrated impressive success in the task of medical image classification. However, the process of model training is typically computationally demanding and cannot be conducted within limited computing facilities in many healthcare settings. This work evaluates whether a mobile multi-core CPU can accelerate chest X-ray classification training without a dedicated GPU. A three-level CPU optimization configuration—TensorFlow strategy-based synchronization, operator-level multithreading, and generator-based input preparation—is applied to the training of a user-defined convolutional neural network on an Intel Core i7-1065G7 CPU. Evaluation on the Chest X-Ray Pneumonia dataset indicates that training time can be reduced from 10.52 to 2.31 minutes per 10 epochs, yielding a 4.56× observed speedup using four physical CPU cores while maintaining an accuracy of 91.52%. The results are interpreted as platform-specific training improvements rather than general evidence of CPU superiority over GPUs.
Download
Paper Nr: 162
Title:

AI-Augmented Weak Signal Detection for Emerging Technologies
Authors:

Anderson Rezende, Daniel Martins, Carlos Eduardo Barbosa, Lucas Nóbrega, Luiz Felipe Martinez, Matheus Argôlo, Marcos Antonio de Almeida and Jano Moreira de Souza
Abstract: Anticipating emerging technologies requires the ability to detect weak signals–subtle, dispersed indications of potential structural change–within large volumes of heterogeneous textual data. However, their ambiguity and semantic instability make systematic identification difficult using traditional scanning approaches. This paper proposes a pipeline that integrates autonomous software agents for distributed data acquisition, Large Language Models (LLMs) for contextual semantic analysis, and the Precision, Importance, Novelty, Credibility (PINC) framework for structured prioritization. The architecture formalizes the interaction between large-scale retrieval, Artificial Intelligence (AI)-driven extraction, and multicriteria scoring into a reproducible detection process. The methodology was evaluated across four domains–transport, health, education, and finance. Results show that the combined use of LLM-based extraction and PINC scoring effectively reduces extensive document sets into a focused subset of high-priority signals, with expert validation confirming the strategic relevance of 85–90% of top-ranked cases. These findings support the use of integrated semantic and multicriteria frameworks to enhance structured foresight practices.
Download
Paper Nr: 168
Title:

Enhancing Natural Voice Communication with LLMs via Joint Emotion Prediction and 3-Dimensional Speech Filler Insertion
Authors:

Semih Yağcı and Senem Kumova Metin
Abstract: Traditional voice AI pipelines often separate semantic text generation from acoustic synthesis, causing prosodic and paralinguistic information to be lost before text-to-speech (TTS) generation. We present a text-conditioned framework in which Large Language Models (LLMs) jointly predict utterance-level emotion and insert three-dimensional speech-filler tags encoding filler identity, relative pitch, and duration. From 124,000 LibriHeavy rows, we retained filler-containing utterances and identified 9,747 filler occurrences after acoustic preprocessing and forced alignment. Mistral-7B-Instruct-v0.2, Qwen2.5-7B-Instruct, and Llama-3-8B-Instruct were fine-tuned using Quantized Low-Rank Adaptation (QLoRA) and evaluated using exact-match metrics and BERTScore. Qwen2.5-7B achieved the highest Emotion Accuracy (0.5291) and Weighted F1-score (0.4798), compared with 0.4492 and 0.3887 for the strongest classical baseline. Mistral-7B-v0.2 achieved the highest filler-word, pitch, and duration accuracies of 0.8458, 0.7681, and 0.7693, respectively. All three LLMs obtained a Full-Sentence BERTScore F1 of 0.9982. The results support the feasibility of generating structured control metadata for expressive speech synthesis, while also revealing substantial performance limitations for underrepresented emotions.
Download
Paper Nr: 183
Title:

Grounded and Guided: A Human-in-the-Loop Design Pattern for LLM-Based Evidence Synthesis, from Literature Reviews to Multi-Agent Report Generation
Authors:

Mauro Pelucchi, Rosanna La Gamba, Ettore Colombo, Simone Perego, Sophie Gvasalia, Rafeh Babar and Luca Agostoni
Abstract: Large language models (LLMs) promise to compress the most labour-intensive stages of evidence synthesis, yet naive automation fabricates references and hides scoping decisions that institutions must be able to audit. This paper distils a domain-independent design pattern for controllable LLM-based evidence synthesis-curated sources, editable retrieval strategies, mandatory human checkpoints, generation grounded in a user-approved evidential closure, substitutable models and exportable, traceable outputs-and presents it through two production-oriented use cases developed over three design cycles (2023–2026) in a European Union agency and a global labour-market analytics company, with the University of Bergamo. Use case I is a guided literature-review assistant: a four-step wizard over a serverless, operation-based API that orchestrates hosted LLMs and federates four curated repositories. Use case II generalises the pattern to data-grounded reporting through nine specialised agents under hybrid orchestration, with data access mediated by the Model Context Protocol. We document the model selections of each cycle, position the pattern against screening tools, commercial and deep-research assistants, survey generators and multi-agent frameworks, and discuss trade-offs and limitations.
Download
Paper Nr: 190
Title:

Knowledge Extraction from Clinical Trial Protocols Using Large Language Models for Automated Patient Matching
Authors:

Maria Clara Castro Higino de Sousa, André Gomes Regino and Julio Cesar dos Reis
Abstract: Clinical trial recruitment remains one of the main bottlenecks in clinical research due to the complexity of eligibility criteria and the heterogeneous nature of electronic health records (EHRs). Although recent advances in Large Language Models (LLMs) have demonstrated promising capabilities for processing clinical text, most existing approaches address either eligibility criteria extraction or patient matching as isolated tasks. This paper proposes a parameterized LLM-based framework that integrates automated clinical trial protocol acquisition, structured eligibility criteria extraction, and semantic patient matching within a unified workflow. Our framework is independent of specific language models or clinical databases, enabling adaptation to different recruitment scenarios with minimal modifications. The evaluation applies the framework to retrieve clinical trial protocols from \textit{ClinicalTrials.gov}, extract inclusion and exclusion criteria into a structured representation, and semantically compare these criteria against heterogeneous patient records obtained from EHRs. The framework was instantiated using Gemini 2.5 Flash and evaluated on $1,000$ cardiology clinical trial protocols together with $4,761$ patient records from the \textit{MIMIC-IV-Ext Cardiac Disease} dataset. To enable scalable evaluation, an independent LLM-based judge was validated against manual annotations, achieving competitive agreement results ($\kappa = 0.690$). The eligibility extraction module achieved a mean Recall of $0.976$, a mean Precision of $0.981$, and an average quality score of 4.39/5. Our experiments on the semantic patient matching further demonstrate the feasibility of combining structured eligibility representations with LLM-based semantic reasoning for automated candidate identification.
Download