Learning-Driven E-Commerce Decision Systems: Sales Forecasting, Dynamic Pricing, and Personalised Recommendation

Modern e-commerce operates under rapid market shifts, large assortments, and heterogeneous (often multimodal) data, which strain conventional analytics. This Ph.D. thesis develops a coherent methodological pathway that spans three interdependent components of decision making: sales forecasting, dynamic pricing, and recommendation, together with scalable implementations suited to digital retail. First, a Voting Network–based Candidate Product Classification (VN-CPC) framework is introduced for sales forecasting and product selection. By combining tree-based and neural predictors through an optimised voting scheme, VN-CPC improves prediction accuracy across diverse Amazon categories and reduces error relative to single models and standard ensembles; in our evaluations, VN-CPC yields materially lower RMSE, and its three-stage clustering isolates high-potential products to guide inventory and promotion planning. Second, to translate forecasts into price action, the thesis presents ItaNet, a deep reinforcement framework for dynamic pricing with multimodal state inputs. Across extensive simulations and comparative baselines, ItaNet raises average profit margins by more than 16.7% while limiting volume impact to 0.3–2.0%. The accompanying data set over 34,000 promotions spanning 27 categories with text, image, and numeric features supports a reproducible study of pricing policies at scale. Third, it addresses non-stationarity in user behaviour for recommendations. A nonstationary transformer for sequential recommendation improves deep learning benchmarks (about 8% reduction in LogLoss and up to 2% higher F1) and also strengthens reinforcementlearning policies by increasing cumulative reward under identical serving constraints. Finally, a unified multimodal sequential recommender (UMRec) based on a Mamba backbone and gated fusion integrates IDs, images, and text in a single architecture. On MovieLens-1M, MIND, and PixelRec-200K, UMRec surpasses strong baselines, improving Recall@10 by up to 12.7% and NDCG@10 by up to 23.3%. It also reduces memory usage by roughly 50% versus BERT-base text models and cuts training time by about ×5 relative to ViT-based image pipelines. Together, these contributions advance the modelling of demand, the translation of forecasts into prices, and the personalisation of exposure. The resulting frameworks provide practical accuracy gains and resource efficiency, and point toward integrated systems that jointly optimise forecasting, pricing, and recommendation in digital commerce.