Repeat-Explore-Aware Grocery Next-Basket Recommendation with Time-Decayed LightGCN

Grocery shopping is highly repetitive, yet customers also introduce new products into their buying routine over time, creating a repeat-explore challenge for next-basket recommendation. This study examines this trade-off in offline grocery recommendation using the Dunnhumby Complete Journey and Ta-Feng datasets. The task is formulated at the household-product level and evaluated using overall, repeat-specific, and explore-specific Top-K metrics. We compare global and personal frequency baselines with GraphSAGE, GCN, LightGCN, and weighted and time-decayed graph variants, and propose two hybrid models: TL-HFE, which combines an exploitation-oriented LightGCN ranking over the learnable catalog with a history-filtered exploration head, and TL-PPR, which combines a non-parametric personal-popularity repeat head with a LightGCN-based exploration head and household-specific quota interleaving. On Dunnhumby, Personal Top-Frequency achieves Overall, Repeat, and Explore Recall@20 values of 0.20, 0.40, and 0.00, while TL-PPR achieves 0.15, 0.29, and 0.02. On Ta-Feng, TL-PPR achieves 0.19, 0.67, and 0.07, compared with 0.18, 0.80, and 0.03 for Personal Top-Frequency. Paired bootstrap tests confirm that TL-PPR significantly improves exploratory recommendation over Personal Top-Frequency, although repeat recovery remains stronger for the baseline. Overall, the findings show that grocery NBR should be evaluated through separate repeat and explore perspectives rather than aggregate accuracy alone, especially when product discovery is a practical objective.