Deep Belief Networks Considering WeMaps' Recommendation Factors

Objectives: In the cold start phase of the Wemaps' recommendation system, traditional deep belief networks have low efficiency in mining fine-grained information, resulting in a decrease in recommendation accuracy. Therefore, this paper proposes a D eep B elief network considering WeMaps' r ecommendation Fact ors (DBRFact). The structured information of clustering and rich semantic information in the deep belief networks are used to alleviate the problem of low recommendation accuracy during the cold start. In addition, the implicit semantic interaction information between the users and the WeMaps is mined to implement customized recommendations. Methods: First, a Density-Based Spatial Clustering of Application with Noise clustering algorithm is employed to cluster the historical samples of the users and the WeMaps. Secondly, the core points (i.e., recommendation factors), boundary points, and outliers are selected, and the recommendation factors are constructed into different visual and hidden layers. Finally, the visual neurons of the restricted Boltzmann machine are replaced with Gaussian units. Then, a deep belief network is used to fine-tune the input data from low level to high level to improve the recommendation accuracy. Results: The accuracies on the FoodMaps, DrivingMaps, TourMaps, and SchoolMaps datasets were 0.77532, 0.76818, 0.77518 and 0.77464, respectively, and the Root Mean Square Errors (RMSE) were 0.19078, 0.19476, 0.19033 and 0.19092, respectively. Conclusions: The experimental results have showed that the proposed algorithm can accurately recommend the interested WeMaps' information for users.