Deep Exponential Moving Average Learning Method for Sequential Recommendation

Sequential recommendation is a recommendation paradigm that models the user’s historical interaction sequence to predict the user􀆳s preference for the next item. Due to the diversity and complexity of user interaction history, interaction sequences often contain noisy and irrelevant noise terms, which affects the model to capture stable sequence patterns. At the same time, there are various sequence signals in the interactive sequence, such as trend signal, semantic signal and residual signal, which cannot be accurately expressed by a single type of modeling. To solve these two problems, this paper proposes a deep exponential moving average based learning method for sequential recommendation (DeepEMA). In order to solve the noise problem in user interaction sequences, DeepEMA uses the exponential moving average method in the field of time series to smooth the sequences, filter the noise and initially extract the sequence trend. In order to capture complex and diverse sequence signals in interaction sequences, DeepEMA models the sequential trend signal, dimensional semantic signal, and residual signal respectively through a multi-module architecture including trend module, dimensional semantic module, and residual extraction module. Experimental results on four publicly available datasets demonstrate that, the proposed model demonstrates superior recommendation accuracy in terms of hit ratio (HR) and normalized discounted cumulative gain (NDCG) metrics. For example, there is an average improvement of 5.67% on the HR@5 metric and 4.88% on the NDCG@5 metric compared with the most advanced baseline. And the sequence noise problem is significantly improved. The ablation experiments also verify the effectiveness of each module.