Research On Implicit Representation And Multi-task Recommendation System Based On Knowledge Graph

More and more platforms provide services for us through the Internet,almost all over the food,clothing,housing,and transportation.It is becoming more and more difficult to publish and iterate recommendation systems with the exponential growth of users and data.The cold start of new users and projects greatly affects the normal service of the recommendation system.As users’ real intentions become more and more diversified,it is often necessary to improve users’ stickiness through multiobjective comprehensive improvement.Because of the above problems,this paper takes the recommendation system as the main research object,aiming at comprehensively improving accuracy and efficiency.Therefore,a series of research is required on the important modules of the recommendation system,such as recall,fine arrangement,and post-ranking optimization.The following schemes based on the above were proposed.(1)The implicit representation algorithm strategy based on pre-training in the recall stage was proposed to solve the problems of difficulty and high delay in the recall stage of the recommendation system while improving the accuracy;At the same time,it could not only be used in the recall stage sorting but also the post-sorting module could be used as a feature.The knowledge representation of user side information and user behavior was carried out in view of the complex structure,a large amount of data processing and high online requirements of the recall module of the recommendation system;At the same time,the model embedded a shared framework in the project as a sub-module of KACL(KNOWLEDGE AWARE COLLABORATIVE LEARNING,a collaborative learning algorithm based on fine granularity features of knowledge map),which could also be used alone.Pre-training implicit representation framework supported online and offline training.Experiments show that many indexes are higher than the benchmark model and can work in a small sample size environment.(2)An end-to-end sorting model based on the knowledge subgraph was proposed.Knowledge subgraph was introduced to improve the representation ability of unstructured data based on implicit representation framework,which further helped the precision of the sorting model to improve.KACL,a learning framework for fine arrangement,was proposed for the first time,which combines users’ historical comments with knowledge maps.First,a named entity recognition system based on open NER dataset learning was used to identify named entities corresponding to candidate features in unstructured reviews.We then used the Entity Linking System to map the entities identified in the first step to DBpedia(Data Base Pedia).Then we constructed a subgraph that dependent on the extracted entities and related parties and embedded the subgraph into the low latitude dense vector space as a pre-training item through the knowledge subgraph embedding model TransR.It was combined with PES(Pre-Train Embedding Service)pre-training framework for joint modeling.Structured and unstructured features were embedded into the classification model according to users’ comment behavior.Experimental results show that fine-grained feature analysis based on a knowledge graph is helpful to improve the precision of the KACL fine layout model.(3)Proposed a multi-task fine arrangement model based on knowledge distillation enhancement and gave a specific modeling scheme of e-commerce scene to improve the multi-task modeling ability in the recommendation system;At the same time,considering that multi-task learning would lead to a larger model;Therefore,a scheme based on knowledge distillation was proposed to solve the problem of large parameters.In this paper,a shared underlying model enhanced by knowledge distillation was proposed for multi-task learning of the fine layout model.Firstly,the gating network of multi-expert models was used as the shared bottom layer for learning task-specific representation.Each expert model was distilled by the model.Then,we used the neural network with a cooperative crossover as the task-specific prediction network for predicting a single target.Finally,the objective function was designed to optimize the total sales according to the final task output.Experiments show that each task has been significantly improved.(4)Proposed a personalized reordering model based on a knowledge map.The goal of the fine ranking model of recommendation system usually did not explicitly consider the mutual influence between candidates and the difference of user preferences or intentions.The performance of the model is further improved by introducing implicit representation based on knowledge subgraph and embedded pretraining to learn personalized coding functions for different users.The ranking is the core task in the recommendation system,which aims to provide users with an orderly candidate list.Datasets usually need to be marked to optimize global performance.The ranking model generated by learning the labeled data sets is also called Fine Ranking Model.However,this method is not optimal because the scoring function is applied to each candidate individually and does not explicitly consider the interaction between candidates and the differences of user preferences or intentions.Therefore,a personalized reordering model was proposed according to the results of the fine sorting model.This model can be used as a followup module of any fine ranking model to directly use the existing ranking feature vectors.By using the converter structure,the information of all candidates in the fine arrangement results is effectively encoded,and then the whole recommendation list was directly optimized.Specifically,the converter applies a self-attention mechanism to directly model the global relationship between any pair of candidates in the whole list.Therefore,it is proved that the performance of the model can be further improved by introducing implicit representation based on knowledge subgraph and embedded pre-training to learn personalized coding functions for different users.The experimental results have verified that the reordering model has an obvious improvement effect.