LSTM-UBI: a user behavior inertia based recommendation method

Practice shows that human behavior has certain patterns, also called as behavioral inertia. It often plays a crucial role in user’s decision-making. The current relevant research usually analyzes repetitive behaviors, and takes them as the basis for recommendation. However, in reality, human behavior depends on constraints in time sequence, cause and effect, state and other aspects. Therefore, we propose a more comprehensive user behavior inertia model and apply it to the recommendation system. We takes purchase recommendation as the research scenario. First, we summarize three patterns of user purchase behavior and model the time series data using LSTM. The purpose is to strengthen the differentiation of behaviors and improve the interpretability of the model. Second, we formal define, quantify and apply the user’s behavior inertia in the purchasing recommendation. Then, We adapt Newton’s first law to the recommendation problem by defining user interest retention as an “external force” that changes the user’s current state. Finally, in order to better simulate the human thinking mode, this paper calculates the user’s interest retention rate through the Ebbinghaus forgetting curve. The obtained results are compared with the user’s behavior inertia to determine whether the candidate product can change the user’s current purchase status. Extensive experiments based on real collected datasets show that the recommendation method based on user behavior inertia has better recommendation performance compared to the baseline method, where the Hit value is improved by about 2 % -7 % and the NDCG value is improved by about 1 % -3 % .