A Multi-Dimensional Context-Aware Recommendation Approach Based on Improved Random Forest Algorithm

Context-aware recommender systems focus on improving recommendation accuracy by adding contextual information and have been widely used in the real-world applications. However, conventional context-aware recommendation approaches have the drawbacks of giving the same weight to all context features, ignoring that users may have different preferences to different contexts, and the effects of context features on the process of recommendations may be different. In this paper, we propose a multi-dimensional context-aware recommendation approach based on improved random forest (MCRIRF) algorithm. The MCRIRF first improves the random forest algorithm by randomly selecting features from multiple feature subspaces that are classified by the importance of features. In addition, the MCRIRF uses the improved random forest algorithm to decompose and reduce the dimensions of context features of users, items, and contexts. Then, the MCRIRF calculates the weights of the 3-D user-item-context recommendation model. In the end, the MCRIRF recommends top-n items with high forecasting ratings to users with similar contexts. LDOS-CoMoDa data set and Cycle Share data set are used for simulation, and six other recommendation approaches are considered in comparison. The experimental results indicate that the MCRIRF can reduce the mean absolute error and root mean squared error of the two data sets by about 2%-16% and 2%-13%, respectively. Thus, the evaluation presents encouraging results, indicating that the MCRIRF would be useful in the context-aware recommendation.