Process-Context-Routed Recommendation: A Process-Mining-Guided Mixture-of-Experts Framework

Recommender systems increasingly exploit sequential interactions, but higher-level process context is rarely used to organize model specialization. This study proposes a process-context-routed recommendation framework in which e-commerce interaction prefixes are represented through activity frequencies, directly-follows transitions, and prefix-level descriptors, then assigned to process-derived contexts that route queries to context-specific ranking experts. Using the RetailRocket dataset, a stability-screened MiniBatch K-Means procedure over K = 3–20 and ten initializations selects 11 contexts. The proposed model is compared with popularity, ItemKNN, first-order Markov, BPR matrix factorization, temporal-context and process-state rankers, direct process-derived feature injection, and a global sequence-aware LambdaMART ranker. On the final chronological test partition, NDCG@10 reaches 0.402141 with the routed mixture-of-experts model and 0.399758 with the global sequential model. This is a relative gain of 0.596%; the bootstrap 95% confidence interval for the difference is [0.001215, 0.003544]. The gain remains positive across all five LightGBM seeds. Random routing and category-only routing also improve performance, showing that generic specialization accounts for some of the gain. The process-derived assignment provides a further, modest benefit on average. Whether this benefit extends beyond the dataset studied here remains to be tested through external replication and online evaluation.