Joint Representation Learning For Transportation Modal Recommendation

With more and more users using map client navigation,how to recommend the most appropriate route based on user's preference and the environmental context is a hot topic not only in industry but also in academia.At present,people pay more attention to single mode transportation recommendation or route planning.The application of route planning is quite mature.For example,Baidu Map can accurately recommend the most suitable path under different travel modes,which can meet the needs of most users.However,there are relatively few studies on travel modal recommendation.When people arrive in a strange city and are not familiar with the local traffic conditions,it is confusing to choose which transportation mode is suitable.Therefore,this paper proposes a joint representation learning framework to recommend the transportation modes.The learned representation can not only represent user's preferences and OD's preferences in history records,but also represents user's demographics and OD's demographics.This paper considers that the user's preference and OD's preference of transportation mode in history directly affects the choice of travel mode at present,because individuals are unique and the people with the same demographics and contexts,which may have different preferences will have different choices.In order to model the user's preference,we propose a model of BTrans2 Vec,which establish the user-mode and OD-mode bipartite graph.Then use the graph embedding to learn the representation of user,OD and mode.We also propose an anchor-based method to reduce the impact of unbalanced distribution.In the end,we propose an online strategy to recommend the transportation mode in real time.We conduct extensive experimental evaluations based on real-world datasets,the results demonstrate that BTrans2 Vec outperforms other baselines.In addition,the travel preference may be distorted for users who have few map client records in history.How to solve the preference distortion problem by using user's demographic and OD's demographic is important.We argue that in the case of the distortion of travel preference,users with the same demographics should have the same embedding,and ODs with the same demographics should have the same embedding.Therefore,we propose a model of Trans2 Vec,we add the edge of user-user and od-od according to the demographics.We add a regularization term with the user demographic,and constrain the embedding with the demographics.Besides,in order to solve the problem that different portrait features have different weights for transportation recommendation,we use logistic regression model to learn the weight of different features.We compare the results of Trans2 Vec and BTrans2 Vec on different datasets,and find that Trans2 Vec is slightly better than BTrans2 Vec on all metrics.Indeed,our method has been deployed into one of the largest navigation Apps to serve hundreds of millions of users...