Research On Personalized Marketing Algorithm And Application Based On User Preference

In recent years,with the rapid development of mobile internet technology,human society has entered the big data era with an information explosion.In the era of big data,the problem of information overload caused by massive information makes users unable to efficiently process information.In addition,the problem of market inequality caused by the information gap also becomes more and more prominent.In such a social background,the marketing industry is facing an unprecedented challenge.To overcome this challenge,personalized marketing combined with big data and artificial intelligence technologies has become a new development trend in the marketing industry.Personalized marketing is a data-driven marketing schema.It adopts big data and artificial intelligence technologies to deeply analyze user data,and then models user preference.Finally,it can provide users with appropriate products or services.However,in the actual scenario,modeling user preference typically encounters the following key challenges:1)User attribute diversity.Users will show completely different characteristic attributes in different scenarios,which will have a huge impact on user preference modeling.2)User behavior diversity.User behaviors are diverse,and there are potential correlations among different user behaviors,which will affect each other and even form user behavior patterns.To address the above limitations,this dissertation systematically carries out a series of exploratory research on user portrait modeling and user behavior analysis.In general,the major contributions of this dissertation can be summarized as follows:Firstly,this dissertation explores the methods of user portrait modeling.Specifically,we focus on the financial products marketing task and aim to utilize the user portrait modeling method to identify potential clients with a strong intention to purchase financial products.However,there are some critical problems encountered in real practice that make this task challenging:1)User feature redundancy.In financial scenarios,there are a large number of user features irrelevant to the potential client identification task.These irrelevant user features will not only reduce the training speed of the model but also increase the storage cost.Even worse,they can have a detrimental impact on accurately modeling user portraits.2)Sample selection bias.The data distribution bias between the training samples and the actual samples will affect the generalization effect of the model.3)Data sparsity.The sparse training data is not conducive to the training of model parameters.To this end,this dissertation proposes the Multitask Feature Extraction Model(MFEM).Inspired by the filter and embedded feature selection methods,a two-stage feature selection algorithm is proposed,which can efficiently and accurately select user features that are highly correlated with the task objective from an incredibly huge number of user feature fields.In addition,MFEM also designs some related subtasks and adopts the framework of multi-task learning to alleviate the problem of sample selection bias,thereby improving the generalization of the target task.Finally,MFEM adopts the mechanism of hard parameter sharing in multi-task learning,which solves the problem of parameter learning in data-sparse scenarios.Secondly,this dissertation studies the user behavior analysis methods.In terms of two different types of user behaviors in the bundle recommendation scenario,namely user-bundle and user-item interactions,it can be considered that these two behaviors reflect the distinctive preferences of users.In addition,since bundles are made up of items,users’ preferences for bundles and items will affect each other.Therefore,how to model the potential relationship between these two behaviors is the key to the bundle recommendation task.Moreover,it is extremely difficult to learn robust and discriminative representations due to the sparse interactive behavior data.To solve these problems,this dissertation proposes the framework of Dual-view Tripartite Graph Contrastive Learning(TGCL).Based on the two types of user behaviors and the bundle-item affiliations,TGCL constructs two tripartite graphs from the perspective of bundles and items,respectively.Then,TGCL adopts an efficient graph convolution method to update the representation vectors,and an even-numbered layer combination operation is proposed to capture cooperative information from homogeneous nodes.Finally,TGCL also designs a dual-view contrastive learning module to construct the node self-discriminating task from two different views of user preference and bundle structure.It greatly improves the robustness and discrimination of the learned representation vectors.Finally,this dissertation conducts extensive experiments to verify the cffectiveness of MFEM and TGCL.For MFEM,all the experiments are conducted on a real-world dataset from a famous fintech bank,and the results show that the MFEM model can effectively improve the accuracy of potential client identification.For TGCL,this dissertation conducts experiments on four bundle recommendation datasets belonging to different application scenarios.The experimental results show that the performance of TGCL greatly improves compared to the existing state-of-the-art baseline models.