Enhancing Dropshipping Performance Through Hybrid Recommendation Engines

Abstract

Small dropshipping stores often present the same popular products to all customers rather than providing personalised recommendations. Personalisation, however, represents one of the few aspects of the shopping experience that such stores can directly control. This study presents a lightweight hybrid recommendation engine that combines content-based filtering, item-based collaborative filtering, and trend signals under the budget and infrastructure constraints of small businesses. The system was evaluated using the public Amazon Beauty dataset, comprising 1.35 million ratings from 883,753 users across 23,838 products, and compared with tuned ItemKNN, PureSVD, and Bayesian Personalized Ranking (BPR) baselines using a temporal train–test split and full-catalogue ranking. Introducing item-based collaborative filtering increased NDCG@10 from 0.0029 to 0.0418, while the best-performing hybrid configuration, with weights of 10% content, 70% collaborative, and 20% trend, achieved an NDCG@10 of 0.0425 ± 0.0013 and significantly outperformed the individual components. Although statistically reliable, the resulting improvement was small in practical terms, corresponding to approximately one additional successful recommendation per 862 recommendation lists. The relative advantage increased to 52% when only one-quarter of the training data was available but reversed for users with longer interaction histories. Content-based recommendations derived from product attributes provided the weakest signal, challenging the assumption that attribute similarity is particularly effective for beauty products. This finding was consistent across the metadata configurations evaluated, including the product-type field, which was the only fully reliable product attribute available in the dataset. The collaborative neighbourhood model achieved approximately twice the ranking performance of the latent-factor baselines while requiring approximately one-fiftieth of their training time. On commodity hardware, the deployed system generated recommendations with a median latency of 34 ms and required approximately 25 s for an offline rebuild. Six propositions with explicit boundary conditions are formulated to characterize the observed relationships between data sparsity, evidence distribution, and recommendation accuracy. The central proposition is that accuracy under sparse conditions depends more strongly on the alignment between model parameters and the locations of available evidence than on model sophistication alone. As the evaluation is limited to offline experiments, future work should assess whether these improvements translate into changes in customer engagement and conversion through live A/B testing.