An Interpretable Learning-Based Multi-Agent Approach for Recommender System

Abstract

Recommender system methods rely on finding correlations between users and items by analysing their data and interaction history. Using different approaches, recommender systems make recommendations based on the derived correlations between users and items. Besides, recommender systems may also provide justifications to explain why the recommendations are made to the user. Matrix factorisation is one of the most successful recommendation methods. However, despite their success, most of the existing matrix factorisation methods suffer from some limitations. For example, they require extra effort to find the optimal number of latent factors. Besides, the produced latent factors have no clear meaning, making them not helpful for explaining the recommendations. In addition, they resort to cross-domain data sources to enhance the recommendation accuracy and explainability, which both require an extra computational cost. In this thesis, we propose an explainable recommender system approach that utilises only the available data in the local dataset of the system. More precisely, we propose WAFE: a feature engineering method, WSLER: an explainable recommender system framework, and MARS: a multi-agent recommender system framework. WAFE produces effective user and item features using only the primary data available in the local system dataset (i.e., item data and the user-item rating history). In addition, WSLER utilises the features produced by WAFE to define the matrix factorisation latent factors. This utilisation is to enhance the recommendation accuracy and explainability. Moreover, MARS extends WSLER to combine multiple recommender agents collaboratively to enhance the system robustness. We have evaluated our approach in different aspects, including recommendation accuracy and explainability, using six benchmark datasets. The evaluation results show that our proposed approach enhances the performance of the matrix factorisation method and outperforms the baseline methods.