Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
102 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Managing Popularity Bias in Recommender Systems with Personalized Re-ranking (1901.07555v4)

Published 22 Jan 2019 in cs.IR

Abstract: Many recommender systems suffer from popularity bias: popular items are recommended frequently while less popular, niche products, are recommended rarely or not at all. However, recommending the ignored products in the `long tail' is critical for businesses as they are less likely to be discovered. In this paper, we introduce a personalized diversification re-ranking approach to increase the representation of less popular items in recommendations while maintaining acceptable recommendation accuracy. Our approach is a post-processing step that can be applied to the output of any recommender system. We show that our approach is capable of managing popularity bias more effectively, compared with an existing method based on regularization. We also examine both new and existing metrics to measure the coverage of long-tail items in the recommendation.

Managing Popularity Bias in Recommender Systems with Personalized Re-ranking

The paper authored by Himan Abdollahpouri, Robin Burke, and Bamshad Mobasher addresses a prevalent issue in recommender systems: popularity bias. Popularity bias favors frequently rated items, thus overshadowing niche items, commonly referred to as "long-tail" items. This bias can limit the diversity of recommendations and hinder the discovery of new content by users with specialized preferences. The paper proposes a personalized diversification re-ranking strategy designed to ameliorate this bias without compromising recommendation accuracy.

Introduction to the Problem

Recommender systems operating in domains such as movies, music, or books often disproportionately promote popular items, leaving long-tail items less visible. The long tail describes a phenomenon where a small number of items have high popularity, while a large number of items are less popular. The objective of the proposed approach is to change this imbalance, thereby enhancing the representation of long-tail items in recommendation outputs.

Methodology Overview

The proposed solution employs a re-ranking technique inspired by the xQuAD (eXplicit Query Aspect Diversification) model. Instead of altering the underlying recommendation algorithm, the method applies a post-processing re-ranking step. This approach can adjust recommendations generated by any base algorithm by selecting items not just based on their predicted interest to the user, but also based on their category—whether they belong to the short head or the long tail.

The authors introduce two variants of this method: Binary xQuAD and Smooth xQuAD. Binary xQuAD adheres to the original xQuAD premise by making binary decisions about whether a recommendation set covers the necessary diversity. In contrast, Smooth xQuAD uses a gradient approach, considering the proportional representation of item categories to make selections.

Experimental Evaluation and Results

The effectiveness of the proposed algorithms was assessed using the Movielens and Epinions datasets. These datasets are well-recognized benchmarks consisting of sparse user ratings for movies and products, respectively.

In terms of performance metrics, four were principally considered:

  • Average Recommendation Popularity (ARP): Reflects the average popularity of items in recommendation lists.
  • Average Percentage of Long Tail Items (APLT): Measures the percentage of long-tail items included in recommendations.
  • Average Coverage of Long Tail items (ACLT): Indicates the proportion of the long tail exposed across the system's recommendations.
  • NDCG (Normalized Discounted Cumulative Gain): Evaluates the ranking quality on a normalized scale.

The experiments demonstrate that both re-ranking algorithms effectively enhance long-tail coverage—as shown by increased APLT and ACLT metrics—while managing the trade-off between diversity and accuracy. In particular, Smooth xQuAD exhibited superior performance in enhancing ACLT, especially in dense datasets like Movielens, indicating a more balanced exposure of long-tail items across recommendation lists.

Implications and Future Directions

The re-ranking strategy proposed in this research provides a flexible mechanism to manage popularity bias in recommendation systems, allowing businesses to better engage niche markets. By facilitating a more equitable exposure of less popular items, the paper contributes to broadening user content discovery and supporting content diversity.

Looking ahead, the approach could be further refined and extended to multi-stakeholder settings, where various stakeholders (e.g., content providers, advertisers) may have competing interests impacting content exposure decisions. Additionally, integrating user preference dynamics over time and considering the implications of user-item interactions can offer deeper insights for achieving personalized fairness and diversity.

In conclusion, managing popularity bias through personalized re-ranking represents a significant step toward optimizing the utility of recommendation systems in diverse and real-world applications.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (3)
  1. Himan Abdollahpouri (25 papers)
  2. Robin Burke (40 papers)
  3. Bamshad Mobasher (34 papers)
Citations (241)
Youtube Logo Streamline Icon: https://streamlinehq.com