PaperlayerView exact revision on arXiv

Source-linked AI summary

Managing Popularity Bias in Recommender Systems with Personalized Re-ranking

Himan Abdollahpouri, Robin Burke, Bamshad Mobasher

arXiv:1901.07555v4cs.IR

TL;DR

Popularity bias causes recommenders to overexpose popular items and underexpose important long-tail products. The paper introduces personalized xQuAD-based re-ranking as an algorithm-agnostic post-processing method, finding that it promotes long-tail items with small accuracy losses and outperforms regularization-based promotion on two datasets.

  • Problem

    Popularity-biased recommenders underrepresent long-tail items, limiting discovery even though long-tail coverage matters for businesses and information providers.

  • Method

    The paper adapts xQuAD into personalized re-ranking that uses user category preferences to balance base-recommender accuracy with long-tail exposure.

  • Results

    Across two datasets, re-ranking methods boosted long-tail items while keeping accuracy loss small compared with a model-based regularization technique.

  • Takeaways & Limitations

    Long-tail promotion should be evaluated with APLT and ACLT alongside ARP, because ARP alone can miss differences in popularity-bias performance.

  • Takeaways & Limitations

    The method assumes independence among remaining items conditional on short-head and long-tail categories, and future work considers extending it to multistakeholder recommendation.

Abstract

from arXiv · show

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.

INTRODUCTION

Popularity bias causes recommenders to favor well-rated short-head items, limiting discovery of long-tail products and potentially disadvantaging newer or niche producers. The paper proposes a flexible post-processing approach to balance exposure across catalog segments while trading off accuracy and long-tail coverage.

  • INTRODUCTION: Popularity bias makes collaborative filtering favor highly rated short-head items over niche long-tail items.This can limit new-item discovery, overlook users with niche tastes, and disadvantage less popular or newer producers.
  • INTRODUCTION: The popularity distribution separates a highly rated top 20% short head from the remaining 80% tail.The first vertical line in Figure 1 marks this division; short-head items cumulatively receive many more ratings.
  • INTRODUCTION: The long tail contains items suitable for collaborative recommendation, whereas the distant tail is too sparse for reliable cross-user comparison.The paper focuses on the long-tail segment and leaves distant-tail cold-start items to content-based or hybrid methods.
  • INTRODUCTION: The proposed approach controls item exposure as a post-processing phase for standard recommendation algorithms.It adapts xQuAD and lets system designers tune the trade-off between recommendation accuracy and long-tail coverage.

Related Work

Prior work treats long-tail recommendation as an accuracy problem, a source of discovery and preference information, or a broader diversity and fairness concern. This paper instead targets popularity bias directly and distinguishes its re-ranking method from model-specific regularization.

  • Related Work: Long-tail items support serendipitous discovery and can account for substantial sales that traditional stores would not normally carry.One cited study reports that 30-40% of Amazon book sales come from such titles.
  • Related Work: Long-tail recommendations can help active-learning systems uncover diverse user preferences.Users are less likely to know these items, making them useful for exploring profile areas where preferences may be diverse.
  • Related Work: Popularity-biased markets may become more homogeneous and provide fewer opportunities for discovering obscure products, innovation, and creativity.The paper frames long-tail recommendation as having a social-good dimension in addition to commercial value.
  • Related Work: Unlike earlier work focused on accuracy and precision, this paper focuses on reducing popularity bias and balancing representation across the popularity distribution.It positions long-tail promotion as a distinct objective from conventional recommender performance optimization.
  • Related Work: Regularization-based long-tail methods are restricted to factorization models and do not account for individual tolerance toward long-tail items.The paper contrasts this with re-ranking, which can operate on any algorithm’s output and personalize diversification.
  • Related Work: Related research also studies generic diversity, personalized diversity, xQuAD-based result diversification, and fairness-aware recommendation.These lines of work motivate the paper’s adaptation of diversification methods to popularity bias.

CONTROLLING POPULARITY BIAS

The paper adapts xQuAD to popularity bias by recognizing that users differ in their interest in long-tail items. It therefore gives under-represented long-tail items a user-specific bonus rather than applying uniform diversification.

  • CONTROLLING POPULARITY BIAS: xQuAD selects items iteratively by estimating how well each candidate satisfies an uncovered aspect of the query.The paper uses this result-diversification framework as the basis for popularity-aware recommendation re-ranking.
  • CONTROLLING POPULARITY BIAS: Uniformly increasing diversity across popularity levels may work poorly because users differ in their interest in long-tail items.The proposed adaptation is designed to recognize these differences rather than impose the same diversification pressure on every user.
  • CONTROLLING POPULARITY BIAS: The method adds a personalized bonus to under-represented long-tail items based on each user’s historical interest in them.This personalization factor is the mechanism used to balance long-tail exposure with user-specific preferences.

METHODOLOGY

The method re-ranks an existing recommendation list by combining base-recommender relevance with category-aware diversification. Its personalized xQuAD formulation uses user category preferences, iterative selection, and either binary or smooth coverage estimates.

  • METHODOLOGY: The modified xQuAD method takes a ranked list from any base recommender and produces a shorter re-ranked list balancing popularity bias and accuracy.The output list S is constructed from the candidate list R after the base algorithm has generated it.
  • METHODOLOGY: The selection criterion combines base relevance P(v|u) with a diversification term weighted by λ.The first term represents interest in the item independent of the current list, while the second promotes category diversity.
  • METHODOLOGY: The diversification term aggregates marginal likelihood over long-tail Γ and short-head Γ′ categories.The formulation follows xQuAD assumptions that remaining items are conditionally independent given their category.
  • METHODOLOGY: Binary xQuAD uses an indicator for whether a category is already represented, while Smooth xQuAD uses that category’s fraction of items in S.Binary xQuAD seeks minimal re-ranking by adding the best uncovered long-tail item without pursuing further diversity.
  • METHODOLOGY: User preference P(d|u) is estimated from the ratio of profile items belonging to category d.It measures whether a user is more interested in short-head or long-tail items.

EXPERIMENT

The experiments use two public datasets, with filtering designed to retain users and items suitable for evaluating long-tail recommendation. The resulting datasets differ substantially in scale and reduction after applying the criteria.

  • Datasets: Experiments use MovieLens 1M and Epinions, two public datasets with ratings from users across movie and consumer-opinion domains.MovieLens contains approximately 3,900 movies and 1,000,209 ratings; Epinions contains 139,736 items and 664,824 ratings before filtering.
  • Filtered datasets: After filtering, MovieLens retained 6,040 users, 3,043 movies, and 995,492 ratings, whereas Epinions retained 8,144 users, 5,195 items, and 220,117 ratings.The reductions were about 0.4% for MovieLens and around 66% for Epinions.
  • Popularity split: Both datasets divide items into long-tail and short-head categories, with the short head accounting for 80% of ratings and the long tail for 20%.The paper notes that other popularity-distribution divisions remain future work.

EVALUATION

The evaluation compares ranking and re-ranking algorithms using long-tail exposure, coverage, and accuracy metrics. It includes a coverage metric that addresses a limitation of average long-tail percentage and uses NDCG to assess accuracy trade-offs.

  • Algorithms: Four algorithms are compared: RankALS, LT-Reg, Binary xQuAD, and Smooth xQuAD, with the latter two re-ranking RankALS outputs.RankALS is the ranking baseline, while LT-Reg is the regularized long-tail diversification method.
  • Recommendation pipeline: Recommendations are generated by reranking lists of 100 RankALS items into final lists of 10 items per user.The re-ranking variants receive RankALS output as their input.
  • Popularity metric: Average Recommendation Popularity (ARP) measures the average number of training-set ratings received by recommended items.For each user, ARP averages item popularity across the recommended list and then across test users.
  • Long-tail exposure: Average Percentage of Long Tail Items (APLT) measures the average percentage of recommended items belonging to the long-tail set.APLT captures the proportion of long-tail recommendations in users’ lists.
  • Long-tail coverage: Average Coverage of Long Tail items (ACLT) measures the fraction of long-tail items covered across the entire recommendation output.It complements APLT because APLT can be high when users repeatedly receive the same long-tail items.
  • Accuracy metric: Normalized Discounted cumulative Gain (NDCG) evaluates ranking accuracy alongside the long-tail diversity metrics.This enables examination of diversity-accuracy trade-offs.

RESULTS

Across Epinions and MovieLens, personalized re-ranking improves long-tail exposure with accuracy trade-offs, generally outperforming regularization on catalog coverage. The results also show that metric choice and dataset sparsity materially affect the observed comparison.

  • Epinions: On Epinions, Smooth and Binary re-ranking expose more of the long-tail inventory than LT-Reg, while Binary preserves slightly better ranking accuracy.Diversity improves across algorithms as ranking accuracy declines; Smooth is especially effective on ACLT, while Binary makes minimal list adjustments.
  • Epinions: Relative to NDCG loss on Epinions, Binary and Smooth have similar diversity-accuracy trade-offs, whereas LT-Reg improves less as ranking accuracy is sacrificed.ARP is the only metric on which the algorithms are fairly similar, particularly at lower NDCG loss.
  • MovieLens: MovieLens shows a stronger benefit for Smooth re-ranking, while LT-Reg is somewhat more effective than in Epinions; ARP remains similar across algorithms.The relative comparison confirms similar ARP values despite differences on other long-tail metrics.
  • Cross-dataset comparison: With 10% NDCG loss, Epinions reaches around 15% long-tail-catalog exposure, while MovieLens achieves equivalent or greater benefit at 0.2% loss; LT-Reg is much less effective.Long-tail diversification is more difficult in the sparser Epinions dataset, and both datasets start from baseline exposure values very close to zero.
  • Cross-dataset comparison: Binary and Smooth perform similarly on sparse Epinions, but Smooth improves ACLT more strongly on MovieLens by promoting multiple higher-quality long-tail items.The paper attributes this difference to the greater availability of higher-quality long-tail candidates in MovieLens.
  • Metric interpretation: ARP alone is insufficient for evaluating long-tail diversity because algorithms can have similar ARP while differing substantially in catalog coverage and long-tail promotion.The authors recommend considering ARP together with the other metrics; ARP does not require a predefined long-tail threshold.

CONCLUSION AND FUTURE WORK

The paper argues that long-tail coverage matters for introducing users to new products despite recommender systems’ popularity bias. Its re-ranking methods boost long-tail items with small accuracy loss, while future work extends the model to multistakeholder recommendation.

  • Long-tail coverage can determine whether recommender systems introduce users to new products and experiences, because short-head items are already widely known.
  • The re-ranking methods boost long-tail items while keeping the accuracy loss small on two datasets compared with the model-based technique.
  • Average recommendation popularity alone is insufficient for evaluating long-tail promotion because similar ARP values can accompany different popularity-bias outcomes.
  • Future work could apply the model to multistakeholder recommendation and add a parameter controlling each stakeholder’s priority.
Loading 1901.07555v4…