Extracting Interaction-Aware Monosemantic Concepts in Recommender Systems

TL;DR

This paper introduces a novel method to extract interpretable, interaction-aware concepts from user and item embeddings in recommender systems using a sparse autoencoder with a prediction-aware training objective.

Contribution

It proposes a new approach to identify monosemantic neurons in recommendation models that preserve user-item interaction information and enable post hoc interpretability and control.

Findings

Neurons capture genre, popularity, and temporal trends.

Method generalizes across models and datasets.

Supports targeted filtering and content promotion.

Abstract

We present a method for extracting \emph{monosemantic} neurons, defined as latent dimensions that align with coherent and interpretable concepts, from user and item embeddings in recommender systems. Our approach employs a Sparse Autoencoder (SAE) to reveal semantic structure within pretrained representations. In contrast to work on language models, monosemanticity in recommendation must preserve the interactions between separate user and item embeddings. To achieve this, we introduce a \emph{prediction aware} training objective that backpropagates through a frozen recommender and aligns the learned latent structure with the model's user-item affinity predictions. The resulting neurons capture properties such as genre, popularity, and temporal trends, and support post hoc control operations including targeted filtering and content promotion without modifying the base model. Our method generalizes across different recommendation models and datasets, providing a practical tool for interpretable and controllable personalization. Code and evaluation resources are available at https://github.com/DeltaLabTLV/Monosemanticity4Rec.