Multi-Head Auto-Correlation Attention Networks for Session-based Social Recommendation

Abstract

Session-based Social Recommendation (SSR) aims to improve next-item prediction by combining a user’s session activities with insights from their social networks. However, the brevity of sessions makes SSR models prone to noise, and many methods rely on complex Deep Neural Networks (DNNs), which often lead to long training times. To address these challenges, we introduce a novel SSR approach that leverages autocorrelation, a concept from stochastic processes, to model sequential dependencies more effectively. By applying Fast Fourier Transforms (FFT) to compute autocorrelation and integrating this with multi-head attention mechanisms, our method captures temporal patterns within session data more accurately. Extensive experiments on two public datasets demonstrate that our method outperforms existing state-of-the-art SSR models in both accuracy and efficiency. Our code is available at https://github.com/LUUUUUUZ/MACNet.