Process-Recommendation Technology Based On Multi-task Learning Research And Implementation

Capturing the development trend of user interest is very important for recommendation systems and search systems.User interest refers to the behavior trajectory left on the Internet.Using behavior trajectories to model the recommendation system can predict the next behavior,which is called serialization recommendation systems.This recommendation system needs to calculate the user’s historical behavior sequence before recommending the next behavior.Therefore,whenever a user has a new behavior,it needs to be added to the previous behavior sequence and recalculated.With the advent of the big data era,the online performance of the single-point estimation serialized recommendation system will be overloaded.Moreover,with the rapid development of the mobile Internet,user interests have become more and more diversified,and how to recommend future information of interest to users through their historical behavior trajectory has become an urgent problem to be solved.The process recommendation system based on multi-task learning proposed in this paper focuses on the four aspects of user intention understanding,user interest diversity,serialized recommendation algorithm online computing speed,and multi-task learning modeling.The main research contents are as follows:(1)Propose an intent recognition algorithm based on knowledge graph enhancement—Knowledge-Graph Based Intent Recognition algorithm,which is used as the recall stage of the recommendation system.First,it analyzes the modeling method of user intent recognition algorithm and the problem that the current evaluation indicators cannot effectively measure the diversity of results.In response to this problem,this paper uses the Deep Structured Semantic Model as the basic intent recognition model,using BERT(Bidirectional Encoder Representation from Transformers)model fusion knowledge map to capture the user’s behavior trajectory trend,the fusion of knowledge map can better describe user intentions,improve the accuracy of the estimation results and increase the diversity of the estimation results to a certain extent.In addition,the BLD algorithm is proposed to optimize the diversity of the estimation results.This scheme greatly improves the diversity of the results under the premise that the estimation accuracy is basically unchanged.Finally,the effectiveness and feasibility of the algorithm are verified through simulation experiments.The experimental results show that the Knowledge-Graph Based Dynamic Intent Recognition algorithm can effectively capture the interest trends of users,and it can also enhance the diversity of the estimated results.The subsequent stages provide the basis.(2)Propose a serialized recommendation algorithm based on sliding reasoning—Group-Wise Serialized Recommendation algorithm,which is used as the ranking stage of the recommendation system.First,it analyzes the noise problem in the interaction sequence of processing long user behavior trajectories.For this problem,the algorithm uses LSTM algorithm to solve the long-term dependence of sequence high-order information,sliding CNN algorithm to model the local dependence of the sequence,and Attention algorithm to Capture the information most relevant to the current intent.In addition,a sliding reasoning algorithm is proposed to recommend a set of future interest information sets for users based on historical behavior trajectories.The algorithm can effectively describe the user’s behavior trajectory information,and improve the online calculation speed through the sliding reasoning algorithm.Finally,the effectiveness and feasibility of the algorithm are verified through simulation experiments.The experimental results show that the Group-Wise Serialized Recommendation algorithm can effectively improve the accuracy of the estimation.(3)Propose a rerank algorithm based on multi-task learning algorithm,which is used as the rerank stage of the recommendation system.First,it analyzes the problems of the recall phase and the sorting phase.Because it is based on a data-driven algorithm,there will be a problem that the estimated result cannot be explained.For this problem,the algorithm uses the IOP(Inner Order Prediction)auxiliary task to analyze the output of the sorting phase.The results are reordered,and the algorithm can effectively improve the accuracy of the final estimation.(4)The prototype of the recommendation system proposed in this thesis realized on the basis of the e-commerce data set.The effectiveness of the system is verified,and the accuracy is improved.