Hybrid Architecture for Recommendation Systems: Fusion of Association Rules and Latent Semantics in a B2B Context

This thesis addresses the design and development of an advanced recommendation system for CNH Industrial’s B2B portal, a complex context characterized by anonymous transactional data, extensive catalogs and the need for technically relevant suggestions. The existing system, based on association rules (FP-Growth), although effective in identifying the most frequent co-occurrences, shows limits in terms of accuracy and ability to cover the vast long tail of the catalogue, highlighting a classic trade-off between precision and coverage. To overcome these limitations, a tailor-made hybrid architecture was proposed and validated. The model synergistically blends two components: the explicit knowledge extracted from FP-Growth, which provides a solid basis of interpretable rules, and the latent semantic knowledge learned through a Word2Vec distributional model (Prod2Vec), which represents each product as a vector in a contextual space. This architecture has been further enhanced by specific mechanisms to manage two critical domain challenges: the cold-start problem for new products, addressed with a content-based strategy (TF-IDF), and the seasonality of purchasing patterns, managed via a dynamic weighting system. The optimal system configuration was identified through a rigorous Grid Search procedure The experimental evaluation, conducted on a real Test Set, demonstrated a substantial and statistically significant improvement over the system in production. The final hybrid model achieved superior values on a complete metric framework, with particularly relevant increases in two key areas: ranking quality (nDCG), demonstrating the ability to reorder suggestions in a more useful way for the user, and practical utility (Success Rate), ensuring greater coverage and reducing cases of recommendation failure. In addition to the methodological contribution, the thesis includes the development of an interactive dashboard in Power BI to monitor the economic impact of the recommendations, offering a concrete link between academic research and a validated industrial solution.