In this paper, the analysis and prediction method of e-commerce user behavior based on machine learning algorithm is deeply discussed. In the part of user behavior analysis, this study counts the key indicators such as user access frequency, access duration and browsing path, divides users into different groups through clustering algorithm K-means, and describes the behavior characteristics of each group in detail. Furthermore, the FP-Growth algorithm is used to mine the association rules between products, revealing the hidden preferences and deep needs of users in the shopping process, which provides data support for optimizing the product recommendation system. In the aspect of user behavior prediction, this paper constructs a prediction model based on integrated learning (Stacking strategy), which integrates many machine learning algorithms such as decision tree (DT), random forest (RF), gradient boosting decision tree (GBDT) and neural network. Through cross-validation and grid search, the model parameters are optimized, and the accurate prediction of users’ future behavior is realized. The experimental results show that the ensemble learning method has obvious advantages in accuracy, precision, recall and F1 score, reaching an accuracy of 0.90, showing a strong predictive ability. This study not only enriches the theoretical system of e-commerce user behavior analysis and prediction, but also provides practical solutions for e-commerce platform to realize personalized recommendation, optimize inventory management and logistics distribution.