[Paper Review] Reducing Popularity Bias in Recommendation Over Time

Himan Abdollahpouri, Robin Burke|arXiv (Cornell University)|Jun 27, 2019

Recommender Systems and Techniques21 references4 citations

TL;DR

This paper proposes a time-sensitive approach to reduce popularity bias in recommendation systems by dynamically adjusting long-tail item exposure based on historical recommendation coverage. Extending the xQuAD diversification algorithm, it uses cumulative recommendation history to prioritize underexposed long-tail items, achieving significantly better long-tail coverage than baseline methods while maintaining comparable accuracy on two public datasets.

ABSTRACT

Many recommendation algorithms suffer from popularity bias: a small number of popular items being recommended too frequently, while other items get insufficient exposure. Research in this area so far has concentrated on a one-shot representation of this bias, and on algorithms to improve the diversity of individual recommendation lists. In this work, we take a time-sensitive view of popularity bias, in which the algorithm assesses its long-tail coverage at regular intervals, and compensates in the present moment for omissions in the past. In particular, we present a temporal version of the well-known xQuAD diversification algorithm adapted for long-tail recommendation. Experimental results on two public datasets show that our method is more effective in terms of the long-tail coverage and accuracy tradeoff compared to some other existing approaches.

Motivation & Objective

• To address the limitation of existing methods that treat popularity bias as a one-shot problem, ignoring historical recommendation patterns.
• To improve long-tail item exposure by compensating for past omissions in recommendation history.
• To maintain high recommendation accuracy while increasing diversity and coverage across the item catalog over time.
• To develop a re-ranking framework that adapts to user-specific long-tail preferences using temporal feedback.

Proposed method

• Adapts the xQuAD diversification algorithm to incorporate cumulative recommendation history $C_{T_0:T_i}$, replacing per-list coverage with global coverage tracking.
• Uses a scoring function $s = P(v|u) + \lambda P(v,C_{T_0:T_i}^\prime|u)$, where the second term penalizes items from already-represented popularity categories.
• Computes marginal likelihood $P(v,C_{T_0:T_i}^\prime|u)$ over long-tail ($\Gamma$) and short-head ($\Gamma^\prime$) categories to guide diversity.
• Ranks items iteratively, selecting the highest-scoring item at each step and updating the coverage set $C_{T_0:T_i}$ to reflect past recommendations.
• Introduces a personalization factor by weighting user-specific interest in long-tail items in the diversity term.
• Employs a re-ranking mechanism that operates on top of any base recommendation algorithm, enabling broad applicability.

Experimental results

Research questions

• RQ1 Can a time-aware re-ranking strategy improve long-tail coverage compared to one-shot diversity methods?
• RQ2 How does incorporating historical recommendation coverage affect the balance between accuracy and long-tail exposure?
• RQ3 To what extent can a system compensate for past omissions in long-tail item exposure through dynamic re-ranking?
• RQ4 Does personalizing long-tail promotion based on user history lead to better overall coverage and fairness?
• RQ5 How does the performance of the temporal xQuAD method vary across different datasets and popularity distributions?

Key findings

• The proposed method achieved significantly higher long-tail coverage than baseline approaches, particularly in reducing over-reliance on a small set of long-tail items.
• On the Epinions dataset, the method maintained NDCG scores comparable to baselines, indicating no significant loss in ranking accuracy.
• The approach outperformed existing re-ranking methods in terms of unique long-tail item exposure, avoiding repeated recommendations of the same niche items.
• The algorithm effectively reduced the concentration of long-tail recommendations on a few items, leading to broader catalog coverage.
• Results showed that temporal feedback enables better compensation for past omissions, making long-tail promotion more systematic and less random.
• The method demonstrated robustness across two diverse datasets, suggesting generalizability beyond specific data characteristics.