ICPE: An Item Cluster-Wise Pareto-Efficient Framework for Recommendation Debiasing

TL;DR

This work targets biases in implicit-feedback recommender systems caused by a highly skewed item popularity distribution, where head items disproportionately influence learning. It introduces ICPE, a model-agnostic framework that reframes training as an item cluster-wise multi-objective optimization, using Popularity Discrepancy-based Clustering to form K clusters, a Pareto-efficient solver to set cluster weights, and counterfactual inference to remove global propensity during prediction. Theoretical results link item popularity to gradient influence and provide a generalization bound for ICPE under distribution shift. Empirical results across Last.Fm, Gowalla, and Yelp2018 show that ICPE improves overall accuracy and debiasing metrics across major backbones (MF, NeuMF, LightGCN), with substantial gains in niche-item performance and coverage. The approach offers a scalable, principled solution to long-tail bias in implicit-feedback recommendations and suggests potential extensions to other long-tail learning tasks.

Abstract

Recommender system based on historical user-item interactions is of vital importance for web-based services. However, the observed data used to train the recommender model suffers from severe bias issues. Practically, the item frequency distribution of the dataset is a highly skewed power-law distribution. Interactions of a small fraction of head items account for almost the whole training data. The normal training paradigm from such biased data tends to repetitively generate recommendations from the head items, which further exacerbates the biases and affects the exploration of potentially interesting items from the niche set. In this work, we innovatively explore the central theme of recommendation debiasing from an item cluster-wise multi-objective optimization perspective. Aiming to balance the learning on various item clusters that differ in popularity during the training process, we propose a model-agnostic framework namely Item Cluster-Wise Pareto-Efficient Recommendation (ICPE). In detail, we define our item cluster-wise optimization target as the recommender model should balance all item clusters that differ in popularity, thus we set the model learning on each item cluster as a unique optimization objective. To achieve this goal, we first explore items' popularity levels from a novel causal reasoning perspective. Then, we devise popularity discrepancy-based bisecting clustering to separate the item clusters. Next, we adaptively find the overall harmonious gradient direction for cluster-wise optimization objectives from a Pareto-efficient solver. Finally, in the prediction stage, we perform counterfactual inference to further eliminate the impact of global propensity. Extensive experimental results verify the superiorities of ICPE on overall recommendation performance and biases elimination.

Figures (10)

• Figure 1: Biases Occurrence and Amplification Loop.
• Figure 2: Empirical study on Gowalla dataset.
• Figure 3: The schematic overview of ICPE. Firstly, according to the training target that the recommender model should balance all clusters of items that differ in popularity, we explore items' actual popularity level (via a specific neural model) from a novel cause-effect perspective and then devise Popularity Discrepancy-based Clustering to separate the discriminated item clusters. Next, we acquire the non-conflict gradient direction for multiple item cluster-wise objectives from the Pareto-efficiency solver. Finally, we combine the above components in the training stage and perform counterfactual inference in the prediction stage to further eliminate the impact of global propensity.
• Figure 4: Causal graphs describing the recommender model. $U$: user, $I$: candidate item, $P$: global propensity, $Y$: interaction probability. We identify $P$ as a confounder between $I$ and $Y$, and propose counterfactual inference by reducing the natural direct effect from $P$ on $Y$.
• Figure 5: Recall-Head@20 and Recall-Niche@20 on LightGCN

Theorems & Definitions (6)

• Definition 3.1
• Definition 3.2
• Theorem 4.1
• proof
• Theorem 4.2
• proof