Constructing Purchase-Related Candidate Sets: Time-Ordered Prediction and Product Co-Occurrence

Abstract

Clickstream logs capture observable product interactions rather than consumers’ latent consideration sets. Accordingly, previously interacted-with products are treated as a behavior-based candidate pool rather than as direct evidence of consideration. This study examines whether recent interactions can help rank products within this pool by their probability of purchase on the following day and whether a smaller candidate set can be effectively constructed. The analysis uses the 2019 JDATA competition dataset from JD company, comprising 37.2 million actions between 1 February and 15 April 2018. Without loss of generality, the largest anonymized category in the pre-April period (Cate 7) is selected for the primary analysis. The sample contains 1.52 million actions from 109,352 users involving 45,284 products. Using five-day histories, a Bayesian logistic model is estimated on target days 6–9 April, an F1-maximizing threshold is calibrated on 10 April, and performance is tested on 11–15 April. Across 1,873,710 test user–product–days, the model achieves an ROC-AUC of 0.7582 and a PR-AUC of 0.00559, compared with 0.7550 and 0.00473 for a views-recency benchmark. At the calibrated threshold, precision is 0.0124 and recall is 0.0735. Only 19.15% of test purchase pairs fall within the prior-interaction universe; among the 1,826 eligible user-days, Hit@5 reaches 0.776, compared with 0.741 for page views alone. The threshold retains 11,982 candidate records, from which a co-occurrence graph with 1,980 nodes and 7,148 edges is constructed. The results support the use of recent interactions as a narrow purchase-related filter, but not as a direct measure of cognitive consideration or product competition.