基于点击流网络的再次购买意愿预测模型

Studies on Forecasting Customer's Repurchase Based on Click-Stream Networks

预测客户重复购买意愿是电商平台制定营销策略的关键,是促进客户重复购买、提升企业利润的重要手段。研究表明,消费者点击行为能够刻画其决策过程,对点击行为进行建模有助于提升再次购买意愿预测的准确度。基于此,本文提出基于点击流网络的再次购买意愿预测模型。该模型以消费者购买决策模型(EBM)等理论为依据,借助复杂网络刻画消费者的点击行为,并提取反映商品热度和消费者行为的可解释性特征;然后采用经典机器学习模型预测消费者是否会在7天内再次购买同一商品,并借助Shapley值解释特征在预测模型中的作用,从而使预测方法具备可解释性并启发实践活动。实验结果表明,该预测模型具备较好的准确性、可解释性和稳健性,能够用于电商平台预测消费者再次购买意愿。

As the e-commerce market matures,competition among e-commerce platforms has become increasingly intense.The difficulty of attracting new customers is far greater than maintaining existing ones,making customer repurchases a crucial means for e-commerce platforms to increase profits.Predicting customers repurchase tendencies/frequencies is key to formulating marketing strategies for these platforms,attracting widespread attention in fields such as marketing,operations research,statistics,and computer science.Predicting customer repurchase tendencies also helps marketers understand the main factors affecting consumer loyalty,thereby better serving platform customer relationship management.Existing research often relies on theories of consumer behavior,proposing hypotheses and using methods like surveys and structural equation modeling to confirm factors influencing consumer repurchases.Some studies adopt data-driven approaches,using models like random forests to predict consumers repurchase intentions.As e-commerce accumulates more data,data-driven research methods are gaining importance.However,these methods are limited to modeling frequency domain indicators and struggle to depict consumers online browsing trajectories.Consumers online shopping behaviors not only record their product-seeking process but also reflect their shopping intentions,which can,to some extent,indicate their repurchase intentions.Common approaches transform online shopping behaviors into frequency domain indicators like click counts for modeling,which fails to effectively depict the popularity of clicked products on e-commerce platforms and also obscures the interaction between consumers and products.Complex network analysis methods offer new insights into mining online consumer behaviors and have been applied to some extent.Studies show that the number of links to a product associats with its demand,and the centralization of similar product networks impacts the demand for focal products.Therefore,using complex networks to depict consumers online clicking behaviors and extracting relevant features can significantly improve the accuracy of repurchases prediction.Beyond accuracy,marketing is more concerned with model interpretability.An interpretable prediction model can help us grasp the factors affecting consumer repurchase intentions,thereby avoiding risks due to unmet marketing expectations.This study proposes an interpretable consumer repurchase prediction model based on clickstream networks.The model,grounded in consumer behavior theory,employs complex network methods to measure users browsing activities and extracts features that characterize product popularity and consumer behavior,ensuring a degree of interpretability of the extracted features.It then uses classic machine learning models to predict whether a consumer will repurchase the same product within 7 days.Through a series of comparative experiments,the study demonstrates that the three sets of features extracted based on consumer behavior theory—product click features,consumer click features,and interaction click features—all enhance the accuracy of repurchase predictions.Moreover,the removal of any one category of features from the feature set constructed from the clickstream network significantly decreases prediction accuracy compared to the model with complete features,further confirming the necessity of including clickstream features in the prediction model.In terms of the model s interpretability,the features extracted on the basis of consumer behavior theory in this study have inherent interpretability.This is further confirmed by post-hoc analysis using Shapley values,which also validate the importance of the extracted features.Finally,robustness analysis,including Lasso feature selection and adjusting the proportion of training samples,also proves that the method proposed in this study has a stable effect.Therefore,the interpretable consumer repurchase prediction model based on clickstream networks proposed in this study shows relatively good performance in terms of prediction accuracy,interpretability,and robustness.This research interprets the role of clickstream networks in predicting repurchase intentions from a big data-driven perspective.Compared with classic theory-driven studies,this research may not reveal the causal relationship between clicks and repurchase intentions,but by modeling repurchase intentions,it can provide references and insights for business operations management.We believe that in the process of making recommendations,businesses should,on the one hand,recommend products with a high likelihood of repurchase to consumers;on the other hand,they should reduce recommendations of products with particularly low purchase intentions to consumers.To enhance consumer purchase intentions,it is necessary to combine theory-driven approaches for argumentation,which is where data-driven methods fall short.In terms of research methodology,although the features extracted in this paper based on theories such as consumer behavior have a certain degree of accuracy,robustness,and interpretability,they are still limited compared to the automatic feature extraction of deep learning methods.Regarding the research data,the method in this paper only uses one month s data,and both the indicators and data have certain limitations.However,as a data-driven research method,it holds practical significance.In future research,we will explore the use of more advanced methods for modeling,such as deep graph neural networks,and further propose more management-relevant research questions based on business practice,develop more data,and test these in the application process within businesses.