Migrants traversing geographically distinct routes such as the Trans-Saharan and Balkan corridors often recount strikingly parallel lived experiences: police violence, smuggler exploitation, dangerous crossings, and family separation. We introduce the task of experiential intertextuality detection: automatically identifying shared experiential echoes across migration narratives without requiring annotated training data. From 108 French migration narratives spanning both corridors, we automatically generate sentence pairs and score them using annotation-free methods: lexical baselines, sentence embeddings, POS-based structural features, a migration-specific theme lexicon, context-aware narrative features, and zero-shot LLM scoring with Qwen2.5-7B and Mistral-7B under three prompting strategies. We validate all methods against 816 expertannotated intertextuality judgments (interannotator Krippendorff’s α=0.27). Our results reveal that all surface, structural, and embedding methods correlate only weakly with expert judgments (r≤0.30); Qwen2.5-7B zero-shot achieves the best single-method correlation (r=0.38); few-shot examples degrade Qwen but dramatically improve Mistral; narrative position significantly predicts intertextuality, with departure-phase pairs showing the highest experiential echoes; and a supervised hybrid combining all 31 features achieves r=0.45, a 21% improvement over the best individual method.

The curse of dimensionality presents a significant challenge in data mining, pattern recognition, computer vision, and machine learning applications. Feature selection is a primary approach to address this challenge. It aims to eliminate irrelevant and redundant features while preserving the relevant ones to reduce computation time, improve prediction performance, and enhance the understanding of data. In this paper, we introduce a new feature selection (FS) technique based on the Logical Analysis of Data (LAD), a pattern learning framework that combines optimization, Boolean functions, and combinatorial theory. One of its main objectives is to generate minimal support sets of features (subsets of features) that discriminate between different groups of data. To generate such subsets, we first reduce the complexity of the LAD optimization task by transforming it into the problem of enumerating minimal hitting sets in a hypergraph, for which efficient implementations exist. Those feature subsets are then ranked based on a scoring method before selecting the highest quality one. Moreover, we explore the relationship between optimal Decision Trees (DTs) and LAD-based FS, introducing new optimality criteria, namely DTs involving a minimum number of features. Finally, we conduct comparative evaluations of LAD-based approach against several state-of-the-art (SOTA) FS methods on benchmark datasets, including two-class binary datasets and numerical datasets with two and multiple classes. Experiments reveal that our approach is competitive with SOTA methods, selecting high-quality feature subsets that maintain or enhance the performance of DTs and other classifiers like SVM, KNN, and Naive Bayes.

On a worldwide scale, an increasing number of victims of human trafficking were observed these last years, covering a majority of countries and territories. Among them, a large portion of women and girls are recruited primarily for sexual exploitation. United Nations Office on Drugs and Crime (UNODC) highlights the difficulties of access to justice which deprive victims of protection, a central issue behind our work. Our contribution is part of an emerging research trend, combining Artificial Intelligence (AI), Humanities and Social Sciences (HSS). It makes an original use of legal database to identify Human Trafficking Networks (HTNs), involving both sexual abuse victims and exploiters. First, a reformulation of the legal database as a numerical database is proposed, using new features expressing relationships between people involved in the same court case, likely to better reveal HTNs. Secondly, six machine learning algorithms, including Decision Tree, Random Forest, Gradient Boosting, Logistic Regression, Support Vector Machine (SVM) and K-Nearest Neighbors (KNN) are used to train on numerical database and learn to classify the input court case into one of the three classes: Not suspicious, Suspicious, or Probably suspicious. We in details discuss knowledge-based feature engineering, dataset balancing, parameters tuning, and best models selection. The comparative empirical evaluations between those classification algorithms have been conducted in order to highlights the relevance of our HTNs detection approach. To help the end-users, to better understand the displayed HTNs, for Decision Tree and Random Forest, we also provide explanations of why such court case can be classified. Those results were finally discussed with experts in the field of human trafficking, providing us with interesting feedback shedding light to this multidimensional form of modern-day slavery problem.