Research And Application Of User Behavior Based On Causal Mechanism

With the popularity of online social networks,online videos,online games,and other online applications,online activities have become an important part of people's lives.In the context of marketing,online marketing becomes the main form in turn.In terms of online marketing,how to mine useful information from massive online user behaviors,guide the design of marketing strategies,reduce marketing costs,and increase marketing conversion rates are critical issues.Association-based analyses have achieved many theoretical and application results but are insufficient in guiding intervention marketing,which is a bottleneck for online marketing that needs to be broken through.Different from association-based analysis,causality analysis strictly distinguishes causality-effect asymmetry and can discover the causes behind the users' behaviors to guide decision-making.As a result,this thesis aims to conduct systematic research on several critical issues in online marketing from the perspective of causal mechanism.The massive,high-dimensional,and high-noise online user behaviors have brought great challenges to marketing research based on causal mechanisms,which are mainly in three aspects.(1)In terms of causal representation,existing user behaviors are mainly user clicks and trajectories without rich semantics.How to extract causal variables from these low-semantic raw data is the basis for subsequent causal inference.(2)In terms of causal structure,the causal variables of user behaviors show the characteristics of high-dimensionality and sparseness.How to discover the causal structure among the high-dimensional and sparse user behaviors is challenging.(3)In terms of marketing evaluations,the marketing process is not grouped randomly.How to deal with confounding factors on the observed data to achieve accurate marketing evaluations is critical.Focusing on the three challenges,the thesis is organized as follows:1.In terms of causal representation learning,two algorithms are proposed for the unstructured text data and user trajectory data respectively.(1)For the text data,in order to mine the meaningful user tags from the high-noise unstructured text data,a key-phrase mining method based causal representation learning method is proposed.Firstly,to discovery meaningful key phrases from the noisy text data,an improved frequent pattern mining based candidate key phrase discovery algorithm is proposed.Secondly,in view of the business needs of the label system for the "key phrase-tag" hierarchical structure,a label system updating method is proposed,based on the membership criteria and the representative key phrase screening criteria.(2)For the trajectory data,a causal variable representation learning algorithm based on user trajectory behavior is proposed,which addresses the low semantics of user trajectory and the lack of demographic data.Firstly,a sequential state model is proposed by integrating spatiotemporal trajectory with users' state and POIs information to establish a semantic bridge between the low-level features and user tags.Secondly,a representation learning method based on long and short memory neural network is proposed to address the noise and missing data of the user sequence state.Finally,a lightweight fine-tuning algorithm based on multi-task dependency is proposed by making full use of the correlation between multiple prediction tasks to predict the demographic attributes of users.This work provides a complete data foundation for subsequent causal structure learning.2.In terms of causal structure learning,a hybrid causal structure discovery algorithm is proposed to discover the causal relationship between high-dimensional and sparse user behaviors.Firstly,an iterative causal discovery framework by alternately employing the causal detection and causal direction inference steps is proposed which greatly improves the reliability of discovered causal network in the high-dimensional data.Secondly,a causal direction inference method by combining V-structure and additive noise model is proposed,which alleviates the Markov equivalence problem in sparse data.Finally,theoretical correctness of the algorithm is provided.The effectiveness of the proposed method has been verified on multiple simulated data and real application scenarios.This work takes the lead in exploring the causal structure of user tags and provides an important basis for advertising and other marketing strategies.3.In terms of causal effect evaluation,a fine-grained confounder balancing-based marketing effect evaluation method is proposed to eliminate the influence of confounding factors.Firstly,a user-based fine-grained confounder balancing objective function is designed by combing deep learning modeling with causal inference,to address the marketing scenario with large control samples and rich descriptions of the users.Secondly,the K-nearest neighbor criterion is introduced to find the most suitable control group users for each marketing group user,which avoids interference errors brought by large samples.Finally,through the deep hybrid network,time series and non-time series features are effectively integrated to achieve efficient sample weight learning,and the causal effect evaluation is solved based on a learned weight function.The effectiveness of this work is verified in a membership card marketing scenario of an enterprise.To sum up,this thesis conducts systematic research on several critical issues in online marketing from the perspective of causal mechanism and deals with the shortcomings of traditional association-based methods,which are significant in theory and application.