# Dual-Tower Counterfactual Session-Aware Recommender System

## TL;DR

This paper introduces a new recommender system that uses a counterfactual approach to improve session-based recommendations by combining long-term and short-term user preferences.

## Contribution

The novel dual-tower architecture with counterfactual causal modeling enhances session-aware recommenders by reducing bias and spurious correlations.

## Key findings

- The proposed method outperforms existing benchmarks in session-based recommendation tasks.
- The counterfactual framework improves robustness by capturing causal influences of long-term preferences.
- Self-supervised training and data augmentation reduce inherent biases in recommendation models.

## Abstract

In the complex dynamics of modern information systems such as e-commerce and streaming services, managing uncertainty and leveraging information theory are crucial in enhancing session-aware recommender systems (SARSs). This paper presents an innovative approach to SARSs that combines static long-term and dynamic short-term preferences within a counterfactual causal framework. Our method addresses the shortcomings of current prediction models that tend to capture spurious correlations, leading to biased recommendations. By incorporating a counterfactual viewpoint, we aim to elucidate the causal influences of static long-term preferences on next-item selections and enhance the overall robustness of predictive models. We introduce a dual-tower architecture with a novel data augmentation process and a self-supervised training strategy, tailored to tackle inherent biases and unreliable correlations. Extensive experiments demonstrate the effectiveness of our approach, outperforming existing benchmarks and paving the way for more accurate and reliable session-based recommendations.