Improving Recommender System Diversity with Variational Autoencoders

Sheetal Borar, Hilde J. P. Weerts, Binyam Gebre, Mykola Pechenizkiy

Introduction

Modern recommender systems (RS) primarily optimize for relevance, aiming to predict the next item a user is most likely to engage with. While effective for short-term engagement, this focus often creates "filter bubbles" and popularity bias, leading to a homogeneous user experience. A significant technical hurdle is the standard representation of a user as a single latent vector, which fails to capture the multifaceted nature of human interests. To address this, we explore VAE-GUP (VAE-based Generation of User Profiles), a method that leverages the generative capabilities of Variational Autoencoders to enhance recommendation diversity.

Methodology

VAE-GUP: Stochastic User Profiling

Unlike traditional models that use a deterministic point estimate (the mean) of a VAE's latent distribution, VAE-GUP exploits its stochastic nature. During inference, the model samples $n$ different latent vectors for a single user, effectively creating multiple "sub-profiles" that represent different facets of their interests.

The recommendation pipeline: from user profile generation to candidate ranking and feedback collection.

The system follows a multi-stage process:

  1. Stochastic Inference: Sampling multiple latent profiles from the VAE distribution.
  2. Diverse Item Selection: Generating recommendation lists for each profile and selecting items that maximize inter-item distance (diversity rank).
  3. Relevance Re-Ranking: A final pass ensures that the diverse set of items remains highly relevant to the user's core interests.

TILD: A Semantic Diversity Metric

To better evaluate diversity, the authors introduce Temporal Inter-list Diversity (TILD). Traditional metrics often rely on simple item ID counts, which can be misled by near-duplicates. TILD utilizes SBERT embeddings to measure the semantic distance between items across different sessions, providing a more accurate reflection of representational diversity.

Experimental Evaluation

The proposed method was evaluated on two large-scale datasets:

  • MovieLens-20M: A benchmark dataset with 20 million movie ratings.
  • Bol.com Dataset: A real-world e-commerce dataset from the Benelux region, containing 6 million records across over 11,000 categories.

Key Results

The experiments compared VAE-GUP against standard Autoencoders (AE) and $\beta$-VAEs.

  • Diversity Gains: VAE-GUP significantly improved both Intra-list Diversity (ILD) and TILD. On the Bol.com dataset, it achieved a 36% improvement in TILD@5 compared to the $\beta$-VAE baseline.
  • The Relevance Trade-off: Increasing diversity often comes with a "diversity tax." On MovieLens, relevance (NDCG@5) decreased by approximately 13%. However, this trade-off is often acceptable for platforms seeking to break popularity bias and surface "long-tail" items.
  • Qualitative Impact: Qualitative analysis showed that VAE-GUP could successfully identify and recommend niche interests (e.g., "measuring equipment") that baseline models overlooked in favor of broader categories like "books."

Discussion and Future Work

VAE-GUP demonstrates that the variance in a VAE's latent space is a valuable resource for exploration. By surfacing a wider variety of items, the system not only improves user experience but also promotes fairness for smaller vendors by reducing popularity bias.

Future research will focus on optimizing the relevance/diversity balance to minimize the drop in NDCG. Additionally, addressing the computational overhead of multi-profile sampling and conducting real-world A/B testing are critical steps toward deploying these diverse recommender systems in production environments.

Summary updated

Abstract

Focusing only on relevance when training recommender systems can lead to suboptimal recommendations. On eCommerce platforms, this not only leads to poor user experience for customers but also limits opportunities for vendors due to popularity bias. Consequently, there is an increasing interest in systems that produce recommendations that are not only relevant but also diverse. In this work, we have shown how diversity can be measured and improved in the context of a large eCommerce recommendation platform. We first evaluate how diversity should be measured feasibly in real-world recommender systems and introduce TILD: a novel diversity metric that measures representational diversity across user sessions. Second, we introduce VAE-based Generation of User Profiles (VAE-GUP), wherein we use the generative nature of VAEs to learn a user profile distribution and sample multiple user profiles. In a simple post-processing step, we combine the recommendations generated by these profiles to generate a final list of more diverse recommendations. Through empirical analysis of benchmark and real-world datasets, we show that VAE-GUP can improve several diversity metrics without significantly compromising relevance. This research has been done in collaboration with Bol.com, the largest e-commerce retailer in the Benelux region.
BibTeX
@inproceedings{borar2023improving,
  author = {Sheetal Borar and Hilde J. P. Weerts and Binyam Gebre and Mykola Pechenizkiy},
  title = {{Improving Recommender System Diversity with Variational Autoencoders}},
  booktitle = {BIAS},
  year = {2023},
  doi = {10.1007/978-3-031-37249-0\_7},
}
Figure