Conception, implémentation et évaluation d’un système de recommandation musical robustes

Thesis Abstract

This study addresses the persistent challenge of delivering accurate and robust musical recommendations in environments with diverse user preferences, sparse feedback, and rapidly evolving catalogs. The problem centers on achieving high prediction accuracy while maintaining fairness, resilience to cold-start scenarios, and scalability across heterogeneous listening contexts. The aim is to design, implement, and empirically evaluate a robust music recommendation system that combines content-based features, collaborative signals, and context-aware modeling to improve user satisfaction and engagement. Specific objectives include (1) developing a hybrid recommender architecture that fuses metadata-based representations, latent factors from collaborative filtering, and inter-user similarity dynamics; (2) integrating robust anomaly detection and outlier handling to mitigate manipulation and noise in user interactions; (3) incorporating fairness-aware ranking to reduce algorithmic bias across genres, languages, and demographic subgroups; (4) evaluating robustness under cold-start and data-sparse conditions through systematic perturbations; and (5) assessing system performance across diverse playback contexts (mobile, desktop, offline mode) using real-world streaming traces. The methodology adopts a pragmatic mixed-methods design combining quantitative experimentation with qualitative validation. The research population comprises 2,500 anonymized user profiles drawn from a mid-sized music streaming service, with a stratified sample of 1,000 active users for evaluation and 500 new users for cold-start experiments. A further corpus of 200,000 tracks with rich metadata (genre, tempo, mood, language, release year) complements behavioral data. Data collection instruments include server-generated interaction logs (play, skip, like, add-to-playlist) over a 12-month period, user surveys (n=400) to gauge perceived recommendation quality, and A/B testing dashboards to compare model variants. Instrument validity is ensured through pilot testing and expert review, while reliability is established via test-retest checks on feature extraction pipelines. Analytical techniques center on a hybrid modeling framework and rigorous evaluation. The core recommender employs a multi-tower neural architecture that fuses content embeddings (e.g., MFCC-derived audio features and metadata embeddings) with user- and item-based collaborative signals. A Bayesian personalized ranking objective guides the optimization to balance accuracy and diversity. Robustness is addressed through anomaly-robust loss functions and an outlier-aware calibration layer. Fairness considerations are operationalized via disparate impact analysis and a constrained optimization routine enforcing equalized opportunity across demographic slices. The approach includes a graph-based session modeling component to capture temporal user-context interactions. Model evaluation uses offline metrics (NDCG@10, Recall@20, MAP) and online metrics (click-through rate, completion rate, average listening time) across five experimental conditions baseline collaborative filtering, content-enhanced hybrid, robustified hybrid with outlier handling, fairness-constrained hybrid, and context-aware hybrid. Statistical analyses include regression-based significance testing (ANOVA) to compare variants and effect size estimation, along with ablation studies to quantify the contribution of each module. Thematic analysis of user surveys supplements quantitative findings, focusing on perceived relevance, novelty, and trust in recommendations. Expected findings indicate that the robust, hybrid model improves top-N accuracy by 12–16% relative to a strong baseline, while maintaining 8–12% higher engagement in cold-start scenarios and reducing low-quality recommendations by 20–25% via anomaly filtering. Fairness-aware adjustments are anticipated to reduce within-genre and cross-language bias, with marginal trade-offs in overall ranking precision (<2%), deemed acceptable in user-centric evaluations. The study anticipates that context-aware components yield measurable gains in mobile and offline modes, where data sparsity is most pronounced. The theoretical contribution includes a formalized integration of robustness and fairness into hybrid recommender architectures, supported by empirical evidence across large-scale streaming data. The study contributes to knowledge by demonstrating a scalable, interpretable, and equitable music recommender that remains resilient under data sparsity and manipulation, while delivering consistent user-perceived relevance. Practical implications encompass guidelines for deploying hybrid recommender systems in production environments, strategies for balancing accuracy and fairness, and methodologies for continuous monitoring of robustness metrics. Limitations include potential generalizability constraints to markedly different catalog sizes and user bases, and dependence on high-quality metadata. Recommendations for future work emphasize extending the model to multilingual and cultural dimension-aware recommendations, investigating causal inference for user satisfaction, and exploring federated learning paradigms to preserve privacy without compromising performance.