Towards Heterogeneous Feature And Impilict Feedback Analysis For Social-based Recommender Model Study

Recently,with the development of location-based social networks(LBSNs),several online platforms reveal a substantial amount of image data which not only reveal visual contents of POIs but also users' preference.Thus the rapid development of computer vision provides an opportunity to investigate the potential of high level visual features on the recommender scenario considering the wide application of computer vision theory on other research fields.With the assumption that the combination of visual content and other side information(e.g.,social relations)can generate a more comprehensive feature learning process and improve the performance of the recommendation model.In this paper,we propose an improved recommendation model with the combination of visual content and other heterogeneous features and leverage the final fused framework on two real-world datasets.Experimental results show that our proposed model is effective in POI recommendation task.Our main contributions are listed as follows:1)We proposed a unified model named VCG,which incorporates visual contents and other heterogeneous features in LBSNs.Aiming to extract the useful information of the image data,we utilized state-of-art architecture Convolutional Neural Network(CNN)to discover the high level visual feature.Based on an overlapping community discover method,we capture two social relations of POIs(i.e.,POI-POI,POI-Community).Moreover,we design an objective function with social regularization terms based on weighted matrix factorization to learn latent vectors of users,POIs and communities for the reconstruction of the rating matrix.2)We introduced an improved power-law distribution based geographical influence model.For discover the pattern of POIs in the geographical distribution and incorporating the user feedback information,we leveraged the frequency of check-in records to build a power-law distribution model.Then we learned the parameters of the power-law model with records and estimated the visiting probabilities between users and POIs.3)Based on two real-world datasets,we designed a comprehensive experiment to verify our proposed model.First,we compare VGC with five state-of-art baselines and our model outperforms other methods both in precision and recall as evaluation metrics.Second,we also conduct several different parameter analysis experiments to verify the interpretability of the parameter chosen.Last,we utilize the component-wise study method to verify the component effective of every individual component compared with the integrated model.