Attention-based Frequency-aware Multi-scale Network for Sequential Recommendation

The key problem of sequential recommendation is how to capture user sequential patterns and enrich user sequential representations from historical interactions, mainly due to the uncertainty of user behavior and the limited information. The RNN-based methods capture the long- and short-term level patterns. The CNN-based methods treat the representation of the user’s historical interaction as an “image”, and discover the point-level patterns, union-level patterns and union-level with skip once. The attention-based methods mine the focus-level patterns. However, all the previous methods have only studied how to capture users’ sequential patterns in the time domain. In many cases, if we only consider the time domain information, these methods will have trouble in mining the user’s sequential patterns. To solve this problem, we consider the frequency domain to capture frequency-level patterns for the first time. Because a non-periodic historical behavior sequence in the time domain may be brutal to reflect the user’s intention but much more accessible in the frequency domain. In light of this, we propose a novel Attention-based Frequency-aware Multi-scale Network (AFMN) for Sequential Recommendation. We introduce Fourier transform to decompose the simple embedding vector, the representation of the user’s historical interaction, into a multi-frequency embedding vector to enrich the user’s behavior sequence representation. The frequency-aware attention layers adaptively focus on the important frequency components and output a refined multi-frequency embedding vector. Given the multi-frequency embedding vector, we develop a non-local attention module to aggregate attribute-level and item-level features of the previous L items. Empirical results on four public benchmark datasets show that our method can achieve a significant improvement over the state-of-the-art baselines.