POI Recommendation Research Based On Heterogeneous Knowledge

With the advent of the mobile internet era,people’s lives are becoming more and more convenient.Without leaving their homes,they can satisfy their daily needs through numerous applications.The prosperity of the virtual world has created massive amounts of online data.Faced with this large amount of data,businesses have been considering how to provide consumers with matching content.This has led to the emergence of interest-based recommendation services.In the past,traditional sequential interest-based recommendation services were able to capture users’ spatiotemporal semantic features well but were unable to make a finegrained distinction among interest points.Although knowledge embedding could capture users’ characteristics well,the growth of data would put high demands on computing power.To address these issues,this article proposes the following work based on heterogeneous knowledge and deep learning technology:(1)In order to satisfy users’ more fine-grained preferences and solve the problem of knowledge graphs’ limited storage of dynamic user information,a multi-view heterogeneous knowledge learning method is proposed.First,to integrate POI attribute knowledge representation and user access sequence features,a heterogeneous knowledge embedding method is designed.Second,to mine the behavioral similarities among users,a user trajectory similarity graph is constructed.At the same time,to learn the potential attribute similarities among POIs,a POI attribute space distance calculation method is designed to construct the POI attribute similarity graph.Finally,a multi-view joint learning method is constructed to model various complex relationships in the POI recommendation scenario.The results on two real datasets indicate that this method is more effective than the baseline approach in terms of NDCG and Recall metrics.(2)In order to alleviate the problem of memory bottleneck caused by large vectors in the recommendation process and improve the ability of sequence interest point recommendation to distinguish among interest points,a self-attention-based interest point recommendation model SAHK(Self-attention Heterogenous Knowledge)is proposed.This model consists of a knowledge embedding layer and a self-attention layer.In the knowledge embedding layer,the structural semantic features and node semantic features in the POI attribute graph are extracted through biased random walks to obtain attribute vectors.Then,the heterogeneous knowledge vector is obtained by concatenating the attribute vector with the sequence vector.Subsequently,the heterogeneous knowledge vector is reduced in size using the quotient trick.In the self-attention layer,different weights are adaptively given to the interest point vectors using self-attention,so as to learn the features in the heterogeneous knowledge vector.Then,the implicit features in the vector are learned through a feedforward neural network.After being validated on two real datasets,this method proves to be more effective than the benchmark approach in terms of NDCG and Hr metrics.