Abstract: To solve the problems in transportation recommendation service,such as single recommended methods and ignoring the user travel preferences,a cascade gradient boosting decision tree (CaGBDT) model is proposed based on the multi-grained cascade forest structure. CaGBDT uses the cascade structure to increase the depth, and then realizes the deep-level representation learning of features. Meanwhile to solve the imbalance of sample class,an index optimization layer based on Powell algorithm is proposed. By searching a threshold for each class,the weight of the prediction results of the model is modified to maximize the evaluation indexes. In addition,a user travel global relationship graph can be constructed by the CaGBDT model via referring the user's travel record, and the spatial context relationship of the user's travel is extracted automatically by the graph embedding method to improve the efficiency of feature extraction.

Key words: traffic recommendation, graph embedding, feature engineering, cascade forest, gradient boosting decision tree