Design And Implementation Of Aboard-Point Recommendation System Based On Spark

With the advent of the "Internet Plus" era,Internet + traffic has developed rapidly in recent years.As a fast-growing internet technology company,company D has changed people's traditional ways of traveling and greatly improved the efficiency of urban transportation.When online dating vehicles have become a common phenomenon,user travel data has grown exponentially;the accumulation of traffic data and the maturity of big data processing technologies have made it possible to provide deep learning and machine learning applications based on traffic big data.Through the analysis of the order,this paper finds that after the user calls the car through the application software of Company D,it usually needs to communicate with the network about one or more times to determine the position of the car.In order to reduce the cost of communication between drivers and passengers,and reduce the length of pick-up time,an in-depth investigation and analysis was conducted on the behavior data of the user network vehicles,aiming at excavating the real problems and pain points of current products through surface phenomena.Through the analysis of a large amount of data,it is found that the current product has three different levels of user experience issues such as detours,communication modification of the position of getting on the train,and modification of the position of the issuing position.In order to improve the user experience of the network pickup,you can use the information in the more order process to comprehensively recommend the pick-up points,such as the positioning position of the driver at the time of picking up the order,the end of the order,and the user's historical taxi pehavior.The optimization of the strategy recommended by the on-street vehicle can reduce the cost of communication between the driver and the passenger;for the driver,the optimization of the on-site recommendation can increase the cost-effectiveness of the entire trip and obtain better profits;from the perspective of the enterprise,the precision of the on-site recommendation can be Improve the efficiency of pick-up,while increasing the user's stickiness.This thesis has analyzed the travel trajectory data,user location information and order information of drivers and passengers,and implemented a pick-up point recommendation system based on Spark.The system includes the basic aboard-point digging,log analysis and integration module,order extraction module,feature extraction module,sample label module,model training and offline evaluation module.The pick-up point recommendation system adopts machine learning method to solve the sorting the aboard position recommendation.Based on the basic aboard-point digging,the actual pick-up position of passengers is excavated for each order;the log analysis and integration module is to parse and integrate the relevant data of the order;the order data extraction module is based on the output of the log analysis and integration module.The random sampling of orders is the base for constructing training data;the feature extraction module extracts the corresponding feature vector for each candidate pick-up point;For different modeling solution,the sample label module can flexibly mark labels;the model training module is based on Spark MLlib and LightGBM frameworks,and the training problem is abstracted as "Binary Classification" or "Rank"" The offline effect evaluation module evaluates the performance of the model on the test data,and the business evaluation standard is offline fixed-point rate.The "post-order" pick-up point recommendation system has improved understanding of the user's scene,it has improved the user experience of aboard-point recommendation and enhanced the interpretability of the recommendation results.The thesis compares the machine learning model with the existing baseline model by dividing the traffic into two branches to do AB Test experiments.According to the order statistic,the model fixed-point rate has improved nearly 2 percentage point compared to baseline.