A Self-Supervised Graph Transformer Framework for Explainable Sequential Recommendation on Amazon Beauty Reviews

Sequential recommenders often obtain accurate next-item predictions by exploiting short-range transitions, yet these shortcuts can obscure longer-term preferences and weaken explanation fidelity. This risk is pronounced in catalogs containing multiple product identifiers with identical titles, where repeat-aware evaluation can overstate generalization. This study proposes a Calibrated Candidate-Transition Self-Supervised Graph Transformer, termed SSGT-CTF, which integrates a Transformer sequence encoder, review semantics, user-item graph propagation, item-transition priors, self-supervised objectives, validation-based calibration, and attention-transition rationales. Experiments used 252,056 raw Amazon Beauty reviews; after exact deduplication and iterative 5-core filtering, 26,606 interactions from 2,256 users and 1,220 items were divided by a chronological leave-one-out protocol. Across three random seeds, SSGT-CTF produced mean NDCG@10 values of 0.6300, 0.6921, and 0.5759 under repeat-aware, no-repeat, and no-repeat-title evaluation, respectively, and it exceeded the neural baselines in every protocol. The Markov baseline remained stronger in the first two protocols, whereas SSGT-CTF surpassed it in the no-repeat-title protocol; the paired seed-42 gain was 0.0129 with a Wilcoxon p-value of 0.0022, and deletion analysis across 120 cases showed larger score reductions after removing top rationales than after random deletion for one to three removals. The findings indicate that candidate transitions are indispensable but must be evaluated under title-aware controls, and future work should validate the model on newer catalogs, correct implementation-sensitive padding conventions, and test cross-domain robustness.

Keywords: explainable recommendation, graph transformer, self-supervised learning, sequential recommendation, transition modeling