Diversity Aware Bundle Recommendation System

Abstract

In today’s digital world, recommendation systems are essential for helping users find relevant content among vast amounts of data. While traditional systems suggest individual items, bundle recommendation systems group related items together, offering users a more tailored and convenient experience, particularly in e-commerce. These bundles are designed to improve user satisfaction by providing items that complement each other. However, one of the major challenges in bundle recommendations is the lack of diversity. Many existing methods focus on relevance, leading to redundancy and reducing the variety within bundles, which may result in user fatigue. This thesis addresses this issue by introducing three bundling methods that integrate both diversity and relevance into the recommendation process. The methods, namely Bundle Partition and Bundle Function, are evaluated against traditional Similarity-Based approaches to ensure they maintain high relevance while introducing meaningful diversity. The study leverages real world datasets from Amazon’s Appliances, All_Beauty, and Luxury_Beauty categories. These datasets were pivotal in assessing the efficacy of the proposed methods across multiple domains. The results show that the Bundle Partition and Bundle Function methods successfully introduce higher levels of diversity, as measured by Intra-List Diversity (ILD), while maintaining high relevance scores through average ratings. Furthermore, by assessing novelty using Mean Inverse User Frequency (MIUF), it was demonstrated that these methods offer users a fresh experience by recommending items that are both relevant and new to them. This research highlights the importance of diversity in improving user satisfaction and engagement in bundle recommendations.