Research On Recommendation Algorithm For Driver Team Formation

In order to motivate Didi drivers' work enthusiasm and organizational identifications,Didi Chuxing has held a series of driver team competitions in different cities.In the process of team formation,Didi proposed a greedy recommendation algorithm based on spatiotemporal similarity to recommend team members for captains.The algorithm successfully connects the team leader and the team member,effectively improves the efficiency of the driver team,but it also has obvious defects and large room for improvement.In order to further improve the efficiency of driver team formation and reduce the cost of operation,this paper takes the success rate of team formation as the main evaluation indicator to further study and validate the recommendation algorithm of team formation.In this paper,the original recommendation algorithm is improved from two aspects of prediction and recommendation,and a genetic recommendation algorithm based on the drivers' intention probability of team formation is proposed.In the prediction stage,this paper collects the historical recommendation records of the platform and driver-to-driver contact behavior from November to December 2017,constructs a sample data set based on driver portrait information,and trains the Easy Ensemble model based on logistic regression on the data set.The modified posterior probability of the model is used to predict the formation intention probability among drivers.In the recommendation process,the recommendation matrix is re-coded as binary string,and the expected success rate of team formation under the recommendation matrix is taken as the fitness function of genetic algorithm.On the premise of satisfying the activity rules,the approximate global optimal recommendation matrix is searched by designing the generational genetic evolution process of selection,crossover and variation.In order to test the performance of the proposed recommendation algorithm in improving the success rate of team formation,an online A/B test was conducted in 26 cities.The drivers in each activity were randomly divided into random recommendation group,greedy recommendation group and genetic recommendation group.The results show that the average success rate of the random recommendation group is 52.6%,the greedy recommendation group is 69.3%,and the genetic recommendation group is 82.8%.The genetic recommendation algorithm based on the teaming intention probability proposed in this paper has higher success rate of team formation than traditional recommendation strategy.In addition,the advantages of the latter two recommendation strategies are more stable and significant in large-scale activities or cities.