Research On Key Technologies Of Consumption Intention Identification And Prediction Based On Event Logic Graph

With the development of Internet e-commerce,more and more goods can be purchased online.For e-commerce websites,analyzing the user's consumption intentions can push the corresponding products to the users.Therefore,the research on consumption intentions has important economic value.Specifically,the research content of this article includes the following three parts:1.Multi-Domain Consumption intention identification based on transfer learning.Consumption intention identification is a domain-related task,and it is costly to label data in each domain.Transfer learning is often used to solve this kind of problem.Deep neural networks are often used to learn transferable features,so that the model can be transferred from the source domain to the target domain.Conventional domain adaptation models transfer knowledge from one source domain to one target domain.It is unidirectional and single-source.In Chapter 1,we proposed a two-way multi-source transfer learning model,which significantly exceeded the baseline model in the consumption intent recognition task.2.Construction of event consumption graph.Many users' consumption intentions are triggered by events,such as "running" needs "running shoes" and so on.There are inherent patterns in the evolutionary relationship between events.We use the event logic graph to model this relationship.There is a common-sense relationship between events and consumption intentions,and we use consumption graphs to model this relationship.The event logic graph and the consumption graph are collectively called the event consumption graph.In Chapter 2,We explored the automatic construction method of event consumption graph.Based on pre-training model,we extracted the events and identified the temporal relationship between events.The correspondence between events and consumption intentions is rarely explicitly expressed in the text,so it is difficult to obtain it by direct extraction.We crawled the product reviews as a weak supervision signal,manually marked some data as a strong supervision signal,and proposed to use the bipartite graph embedding algorithm to effectively use the above two kinds of information.The experimental results of this method on the task of constructing the consumption graph exceed the baseline model.3.Consumption intention identification and prediction based on event consumption graph.There is a common-sense evolutionary relationship between events,so the user's consumption intention is not static.For example,a user's consumption intention changes from "marriage proposal" to "infant milk powder" when the event evolves from "marriage" to "give birth to a baby".Consumption intentions can be divided into short-term consumption intentions and long-term consumption intentions.The research on the two is called consumption intention identification and consumption intention prediction,respectively.In Chapter 3,we give the first Chinese data set for this task and propose a homo-hetero genous relationship attention model for identifying and predicting consumer intentions based on event consumption graph.This model can explicitly model the homogenous and heterogeneous edges in the event consumption graph,and the experimental results on the consumption intention identification and prediction tasks exceeded the baseline model.The model can also be extended to heterogeneous graphs with more node types.