Mitigating Popularity Bias in Recommendation with Global Listwise Learning and Progressive Bi-Weighting

• Long-tail feedback can amplify recommendations of popular items, while IPS methods are limited by local objectives and propensity misspecification.Popularity-only estimates can underestimate tail-item exposure probabilities; inaccurate estimates can introduce empirical bias.Source: §1 · p. 2
• Mult-IPS uses IPS-weighted multinomial likelihood to estimate global, unbiased user preferences across the item set.It is presented as addressing the limits of local preference learning in pointwise and pairwise debiasing methods.Source: §1 · pp. 2–3
• BiW smooths propensity estimates with a collection model, and the paper analyzes empirical-bias bounds and the collection model's optimal form.The smoothing is intended to improve robustness and increase tail-item weights while attenuating popular-item weights.Source: §1 · pp. 2–3
• Progressive Bi-Weighting gradually increases propensity-weight influence, shifting training from discriminative representation learning toward popularity debiasing.The strategy is motivated by the potential harm of aggressive reweighting early in training.Source: §1 · p. 3
• Mult-BiW consistently outperforms state-of-the-art debiasing baselines across six real-world datasets.Source: §1 · p. 3

2.1 Problem Formulation

• Top-N recommendation predicts scores for unobserved user-item pairs and recommends each user's N highest-scoring items.Source: §2.1 · p. 3
• Popularity-debiased recommendation trains on long-tail-biased interactions and evaluates on a test set with a uniform item distribution.The stated objective is maximizing performance on the test data despite non-IID popularity distributions between training and test.Source: §2.1 · pp. 3–4

2.2 IPS for Popularity Debiasing in Item Recommendation

• IPS reweights observed interactions by propensity to address the mismatch between observed interactions and user relevance under exposure and popularity effects.Propensity is the probability that a relevant item is observed through user interaction.Source: §2.2 · pp. 4–5
• The described IPS estimator is pointwise and treats items independently, which may limit ranking quality and capture of users' global preference structures.Source: §2.2 · p. 5

3.1 IPS-Weighted Multinomial Likelihood

• Mult-IPS uses an IPS-weighted multinomial likelihood to learn global preferences across the item set, and its estimator is unbiased for the ideal multinomial loss.The softmax probability constraint makes items compete for a fixed probability budget, unlike pointwise likelihoods; the framework is model-agnostic.Source: §3 · p. 5, §3.1 · p. 6, §3.2 · p. 7

3.2 Bi-Weighting

• BiW smooths individual propensity estimates with a collection-level prior, attenuating small tail-item estimates and moderating large popular-item estimates.The smoothing parameter α controls the interpolation; α→0 recovers Mult-IPS, while α→1 yields uniform item weights.Source: §3.2 · pp. 7–9
• BiW can reduce empirical Mult-IPS bias, and the analysis derives a bias upper bound and an optimal collection model by optimizing against that bound.The collection model is selected through a convex optimization problem.Source: §3.2 · pp. 8–9

3.3 Progressive Bi-Weighting

• Progressive BiW gradually increases propensity-based reweighting, shifting training from discriminative representation learning toward popularity debiasing.The pace parameter η can be adjusted for different backbones and their representational capacity and convergence behavior.Source: §3.3 · p. 10

3.4.1 Comparison with Propensity Clipping.

• Unlike clipping, which truncates small propensities below threshold M, BiW makes a soft, global adjustment across items.The paper compares Mult-CIPS and Mult-BiW empirically in Section 4.3.2.Source: §3.4.1 · p. 10

3.4.2 Applicability of Doubly Robust Estimation.

• Direct doubly robust estimation is challenging for this implicit-feedback setting because it lacks explicit observation indicators, and relevance-imputation extensions are left for future work.The paper defines propensities using interaction signals and notes that extending Mult-BiW would require strategies for imputing relevance probabilities.Source: §3.4.2 · p. 10

3.4.3 Model Complexity.

• Mult-BiW requires O(|U||I|d) operations per epoch, versus O(|D|(1+K)d) for negative-sampling pointwise or pairwise objectives.The latter is asymptotically more efficient when |D|(1+K) ≪ |U||I|, though the runtime gap is often modest with GPU-optimized dense matrix multiplication and softmax.Source: §3.4.3 · p. 11

4 Experiments

• The five research questions examine baseline comparisons, component contributions, hyperparameter sensitivity, debiasing recommendation policies, and performance across popularity groups.The experiments use six real-world datasets.Source: §4 · p. 11

4.1.1 Datasets.

• Six datasets include Amazon Books, Google Local, Gowalla, Yelp Challenge 2019, ML10M, and Netflix; Coat and Yahoo! R3 are also tested in the Appendix.The first four datasets use 20-core preprocessing; ML10M and Netflix test sets are not perfectly balanced, while Coat and Yahoo! R3 have naturally uniform test sets.Source: §4.1.1 · pp. 11–12

4.1.3 Evaluation Metrics.

• Models were trained on long-tailed, imbalanced interactions and evaluated on balanced validation and test sets with uniform item frequency, using a 70%/10%/20% training/validation/test split.Recall@N and NDCG@N were reported at N=10 and 20, averaging five random initializations. Tables 3 and 4 report MF performance across six datasets and training time for Mult-BiW-MF versus CPR-MF, respectively.Source: §4.1.1 · p. 12, §4.1.3 · pp. 12, 14, Table 3 · p. 13

4.1.2 Baselines.

• The comparison includes seven debiasing baselines and Mult-IPS and Mult-CIPS variants, with MF as the default backbone for fairness.The baselines are Expo-MF, Rel-MF, UBPR, DICE, MACR, CPR, and TTEN.Source: §4.1.2 · p. 12

4.1.4 Hyperparameter Settings.

• The embedding dimension was fixed at 64; learning rate was tuned over {1e-3, 5e-4, 2e-4, 1e-4}, and L2 regularization over {0, 1e-7, 1e-6, 1e-5, 1e-4}.Models used Adam with batch size 256; Mult-FBiW α was tuned over {0, 0.1, ..., 1.0}, and Mult-PBiW η over {0.5, 1.0, 2.0}.Source: §4.1.4 · p. 14

4.2.1 Effectiveness Comparison.

• CPR and TTEN are the strongest baseline methods, while UBPR consistently outperforms the pointwise Rel-MF among IPS-based baselines.The authors attribute Mult-BiW’s advantage over IPS-based methods to its multinomial listwise objective and Bi-Weighting’s smoothing of propensity estimates.Source: §4.2.1 · pp. 14–15
• 9.43% in Recall@20 and 14.80% in NDCG@20 are Mult-BiW’s average improvements over the strongest baseline on each dataset; it performs best across all datasets.Gains are especially pronounced on Amazon Books and Yelp, which have more severe popularity skew.Source: §4.2.1 · p. 15

4.2.2 Efficiency Comparison.

• Mult-BiW has comparable training time to CPR on ML10M and Netflix and is faster on the other four datasets.It adds no model parameters or sampling procedures; Bi-Weighting factors are precomputed from item popularity statistics with negligible training overhead.Source: §4.2.2 · p. 15

4.3.1 Loss Function Comparison.

• The multinomial likelihood consistently outperforms all alternative loss functions across datasets, supporting listwise unbiased learning.Pairwise loss ranks second, while pointwise squared and logistic losses perform substantially worse.Source: §4.3.1 · p. 16

4.3.2 Bi-Weighting vs. IPS.

• Fixed and progressive Bi-Weighting consistently outperform IPS and CIPS across datasets and backbones, with PBiW best in most cases.The comparison attributes improved robustness to smoothing estimated propensities.Source: §4.3.2 · p. 16
• Bi-Weighting methods improve faster and more stably over training epochs than IPS and CIPS, and PBiW consistently outperforms FBiW.Continuously decreasing α too aggressively may lead to overfitting.Source: §4.3.2 · pp. 16–17

4.4.1 Effect of Model Balance Coefficient

• FBiW significantly outperforms IPS across a wide range of intermediate α values, while PBiW surpasses FBiW even at its optimal α.α=0 reduces FBiW to vanilla IPS, whereas α=1 ignores propensities.Source: §4.4.1 · p. 17

4.4.2 Effect of Cumulative Learning Parameter

• Smaller η generally improves Mult-MF, whereas larger η improves LightGCN by allowing more representation-learning epochs before stronger debiasing.The passage attributes this difference to LightGCN's multilayer graph convolution and more complex representation learning.Source: §4.4.2 · p. 17, Figure 5 · p. 18

4.5 Debiasing Visualization (RQ4)

• Mult-BiW-MF has the flattest popularity-versus-top-10-frequency curves and lowest Pearson correlations among the compared debiasing methods.All debiasing methods reduce the strong positive correlation seen for Mult-MF.Source: §4.5 · p. 18

4.6 Heterogeneous Analysis of Item Groups (RQ5)

• Mult-IPS-MF and Mult-BiW-MF reduce head-item accuracy while improving tail-item performance over Mult-MF; Mult-BiW-MF gains most on less popular items.The comparison reports average NDCG@10 by item-popularity group.Source: §4.6 · p. 18, Figure 7 · p. 19

5.1 Collaborative Recommendation

• Collaborative recommendation spans memory-based similarity, matrix factorization, item-similarity, deep and graph models, self-supervised learning, and generative approaches; this method is model-agnostic and complementary.The paper frames its contribution as applicable across existing recommenders.Source: §5.1 · p. 19

5.2 Loss Functions for Implicit Feedback

• Pointwise losses model individual interactions, pairwise losses model relative preferences, and listwise losses optimize item lists; popularity-debiasing work has mostly used pointwise or pairwise objectives.The paper introduces a multinomial listwise framework for IPS-based debiasing.Source: §5.2 · p. 20

5.3 Popularity Bias and Debiasing Strategies

• Popularity debiasing includes pre-processing, in-processing, and post-processing approaches, while IPS methods are attractive but depend critically on accurate propensity estimation.Bi-Weighting improves propensity estimation and is integrated with a multinomial listwise IPS objective.Source: §5.3 · pp. 20–21

6 Conclusion

• Mult-BiW consistently outperforms state-of-the-art debiasing methods by a substantial margin across six real-world datasets while mitigating popularity bias.The paper says the method produces more balanced recommendation outcomes.Source: §6 · p. 21
• Mult-IPS combines IPS with multinomial likelihood to capture global, unbiased user preferences over the entire item set.BiW combines individual propensity estimates with a collection-level model and smoothing to improve propensity-estimation robustness.Source: §6 · p. 21
• The analysis bounds empirical bias from inaccurate propensities and derives the optimal form of BiW’s collection model.Source: §6 · p. 21
• Progressive Bi-Weighting uses cumulative learning to transition from representation learning to popularity debiasing, addressing adverse effects of aggressive reweighting on representation learning.Source: §6 · p. 21
• Future work could explore more advanced smoothing and incorporate user exposure signals, viewing histories, or click-through data into Bi-Weighting.Source: §6 · p. 21