From Insight to Intervention: Interpretable Neuron Steering for Controlling Popularity Bias in Recommender Systems

TL;DR

This work targets popularity bias in recommender systems by proposing PopSteer, a post-hoc, interpretable intervention that leverages a Sparse Autoencoder (SAE) to reveal neuron-level popularity signals and steer activations. By generating synthetic profiles that emphasize popular vs. unpopular items, the method identifies bias-encoding neurons via Cohen’s d and adjusts their activations proportionally to their variability and bias strength, producing a steered user embedding without retraining the base model. Across three public sequential datasets, PopSteer improves exposure fairness (higher item coverage and lower Gini) with only minimal impact on ranking accuracy, while offering clear interpretability and controllability over the fairness-accuracy trade-off. The approach is scalable, requires no loss-function changes, and provides actionable insights into the neural basis of bias, with potential to extend to other biases and deployment scenarios.

Abstract

Popularity bias is a pervasive challenge in recommender systems, where a few popular items dominate attention while the majority of less popular items remain underexposed. This imbalance can reduce recommendation quality and lead to unfair item exposure. Although existing mitigation methods address this issue to some extent, they often lack transparency in how they operate. In this paper, we propose a post-hoc approach, PopSteer, that leverages a Sparse Autoencoder (SAE) to both interpret and mitigate popularity bias in recommendation models. The SAE is trained to replicate a trained model's behavior while enabling neuron-level interpretability. By introducing synthetic users with strong preferences for either popular or unpopular items, we identify neurons encoding popularity signals through their activation patterns. We then steer recommendations by adjusting the activations of the most biased neurons. Experiments on three public datasets with a sequential recommendation model demonstrate that PopSteer significantly enhances fairness with minimal impact on accuracy, while providing interpretable insights and fine-grained control over the fairness-accuracy trade-off.

Figures (8)

• Figure 1: Overview of PopSteer process, first run with synthetic data for interpreting SAE's neuron ($a \rightarrow b \rightarrow c \rightarrow d$), and second run with real-world dataset for steering SAE's neurons to mitigate bias and derive modified user's embedding ($e \rightarrow b \rightarrow c \rightarrow d \rightarrow f \rightarrow g$): (a) synthetic interaction data containing users’ profiles extremely representing interest towards either popular or unpopular items, (b) a pretrained recommendation model trained on real (not synthetic) data, (c) users' and items' embeddings learned from the pretrained recommendation model, (d) SAE used for interpreting popularity bias according to activation behavior of neurons in hidden dimension, (e) users' profiles in real data, (f) updated user embedding after steering SAE's neurons, and (g) predicting the relevance score for a target user-item pair using the updated user embedding.
• Figure 2: Performance comparison of PopSteer with baselines, in terms of nDCG versus (a) item coverage and (b) Gini Index.
• Figure 3: Head-item share $H$ over users for each dataset. Circles mark the mean $H$ of the top-10 users for the SAE neuron with the largest positive Cohen's $d$ (green, popularity-aligned) and the most negative Cohen's $d$ (orange, unpopularity-aligned).
• Figure 4: Effect of reducing activation of $K^\prime$ neurons linked to popularity bias ($\beta=1$).
• Figure 5: 2D UMAP projections of SAE activations for real, synthetic popular/unpopular, and steered user profiles. For ML-1M and BeerAdvocate, $\alpha^{\text{Unpop}} = 3$, $\alpha^{\text{Pop}} = 3$. For Yelp, $\alpha^{\text{Unpop}} = 1$, $\alpha^{\text{Pop}} = 1$.