Sequence-to-pattern analysis for predicting buying decisions on imbalanced clickstream data

The e-commerce industry has transformed online shopping. To better understand customer behavior and buying decisions, anonymized clickstream data from the SIGIR eCOM 2021 Data Challenge was analyzed. The analysis iterated between data preprocessing and exploratory data analysis (EDA) before predictive modeling. Constraint-based Seq2Pat, which included the Dichotomic Pattern Mining (DPM) technique, was employed to identify common browsing patterns among customers. Using the ratio method in the SequenceMatcher class of difflib, the obtained patterns were mapped with the patterns from the preprocessed clickstream dataset, and the sequences with the highest similarity score were identified. Preprocessed imbalanced clickstream data with various ratios of buyer and non-buyer groups, namely 4:96, 3:97, 2:98, and 1:99, were prepared by adjusting the thresholds for the similarity score, and their prediction performance was observed. Logistic Regression (LR) achieved high prediction performance across imbalanced clickstream datasets of different ratios, with a ratio of 4:96 performing exceptionally well, with 90.95% average recall and 95.26% average F1-score. Businesses can use this knowledge as a guide to preprocess clickstream data and automate it. Moreover, LR can be used to predict potential buyers. Finally, the results from EDA and prediction support effective marketing campaigns, accurately target potential buyers, and increase the conversion rate.