Session-Recommendation Based on Gated Neural Network to Extract Item Structure Feature Graph

The issue of inaccurate user portrait recommendations for short-term conversations is solved. The classical recommendation is based on an assumption: the user's historical behavior can represent the user's stable long-term preferences, and the result is generated based on the user's historical behavior. However, there is sparseness in the user's historical behavior information record in the actual data set, and the obtained data cannot accurately represent the user's preferences. Research on a session recommendation focuses on user behavior data in the short term, and aims at predict anonymous user behavior based on in-session behavior information. It plays a vital role in application scenarios where user behavior information is scarce. Feature extraction is performed on the user's anonymous behavior information graph generated in a fixed period of time as the embedding representation of user's interest preference; the graph structure of the same item in the global click interaction graph is combined with the input a gated graph neural network as the item attribute embedding representation. The prediction part adopts simulated interactive calculation to produce results. With the help of gated graph neural network, features extraction and analysis of data in short-term sessions: session-based user behavior data and in-session item attributes are obtained from the graph data structure, which reduces errors caused by inconsistent data processing methods. The user's current interest is captured based on the behavior information in the session, the global inter-item conversion graph is generated based on the sequence of click moments, and the feature vector is obtained through graph neural network processing. The generation of recommendation results is more time-sensitive, complete the anonymity problem due to the scarcity of user's information, and improves the accuracy of recommendations.