| CPC G06N 5/022 (2013.01) | 9 Claims |
|
1. A point-of-interest recommendation method based on a temporal knowledge graph, comprising:
S1. constructing a dynamic temporal knowledge graph and a static group knowledge graph based on complete historical behavior trajectories of all users; wherein the dynamic temporal knowledge graph is a graph set formed by dynamic relationship knowledge graphs of different historical time slices, each of the dynamic relationship knowledge graphs records a dynamic relationship between all of the users and points of interest in the historical time slices, the dynamic relationship comprises a visit relationship for recording a visit behavior of the user to the point of interest, and a follow-up relationship for recording a neighboring visit behavior of the user to different points of interest; wherein the static group knowledge graph records a static relationship between all of the users and the points of interest in all of the historical time slices, the static relationship comprises a social relationship for recording a friend relationship between the users, a location relationship for recording a spatial area where the point of interest is located, a neighboring relationship for recording whether the different points of interest are neighboring points, a category relationship for recording a point-of-interest category to which the point of interest belongs, and a group relationship for recording a user group grouped according to a visited point of interest and a visited spatial area;
S2. obtaining a historical behavior trajectory substring of a target user before a time to be predicted, extracting sequentially a user review text of each of the points of interest visited by the user therefrom, performing word embedding on the user review text using an aspect-level sentiment analysis module built based on a pre-trained model, and concatenating sentiment embeddings of all of the user review texts to obtain a user review sentiment embedding sequence;
S3. inputting the historical behavior trajectory substring, the dynamic temporal knowledge graph, the static group knowledge graph, and the user review sentiment embedding sequence into a point-of-interest recommendation model, wherein an embedding module first performs word embedding on the input data, then a multimodal knowledge fusion module fuses the dynamic temporal knowledge graph and the static group knowledge graph based on a heterogeneous mutual attention mechanism and fuses the point of interest, the user, and other multimodal information to obtain a point-of-interest fusion feature representation and a user fusion feature representation, and finally a decoding module concatenates the point-of-interest fusion feature representation and the user fusion feature representation and then inputs into a cascaded recurrent neural network and a multi-layer perceptron to predict a point of interest likely to be visited by the target user at a next moment, wherein
a processing flow in the multimodal knowledge fusion module is as follows:
S31. inputting respectively the dynamic temporal knowledge graph and the static group knowledge graph into a heterogeneous graph transformer network for information fusion, and obtaining the dynamic temporal knowledge graph and the static group knowledge graph fused;
S32. arranging all of the points of interest in the historical behavior trajectory substring according to an order of user visits, extracting sequentially a hidden layer vector corresponding to each of the points of interest from the dynamic temporal knowledge graph fused to form a user behavior trajectory embedding with global time slice information, extracting sequentially a hidden layer vector corresponding to each of the points of interest from the static group knowledge graph fused to form a user behavior trajectory embedding with global static information; extracting a point-of-interest level group feature and an area level group feature from the static group knowledge graph fused;
S33. using the user review sentiment embedding sequence as a query, fusing the user behavior trajectory embedding with the global time slice information and an original user behavior trajectory embedding through an attention mechanism to obtain the user behavior trajectory embedding fused; using the user behavior trajectory embedding fused as a value, the user behavior trajectory embedding with the global static information as the query, and the point-of-interest level group feature as a key, and inputting into an Encoder module of a Transformer model for fusion encoding to obtain a point-of-interest fusion feature representation; and
S34. concatenating and fusing the point-of-interest level group feature and the area level group feature, using the obtained group feature fused as a key, user embedding feature representations of all of the users in the static group knowledge graph fused as a query, the user embedding feature representations of all of the users in an original static group knowledge graph as a value, and inputting into the Encoder module of the Transformer model for fusion encoding to obtain a user fusion feature representation.
|