What question did this study set out to answer?

This investigation aims to improve purchasing intent prediction in e-commerce by addressing class imbalance and leveraging a hybrid deep learning architecture.

June 5, 2026Open Access

ContextAware Hybrid Deep Learning for Intent Prediction in Imbalanced Ecommerce Datasets

KDKaran DarjiGujarat Technological University JPJal PatelGovernment Dental College and Hospital KBKaran P. BhattGujarat Technological University

Key Points

This investigation aims to improve purchasing intent prediction in e-commerce by addressing class imbalance and leveraging a hybrid deep learning architecture.
Developed a Context-Aware Three-Tower Hybrid Deep Learning Architecture utilizing collaborative, sequential, and contextual domains.
Replaced traditional Binary Cross-Entropy loss with Focal Loss to mitigate majority-class bias.
Analyzed 2.7 million rows of behavioral data using various embedding and pooling techniques.
Achieved a Validation AUC-ROC of 88.7% with the Hybrid model.
The fully integrated Hybrid Engine showed a prediction confidence of 36.83%, outperforming Collaborative (29.31%) and Content-based (31.58%) models.
Demonstrated the effectiveness of combining sequential and cyclical contextual data for improved prediction accuracy.

Abstract

Modern e-commerce ecosystems generate massive volumes of high-velocity behavioral data, yet extracting actionable purchasing intent remains a significant challenge due to extreme class imbalance and the non-linear nature of human decision-making. In a typical retail environment, over 96% of digital footprints represent non-conversion events (passive browsing), creating approximately a 24:1 ratio that frequently leads to majority-class bias in traditional Machine Learning models (Fatta et al., 2017, p. 164). This investigation presents a mathematically grounded, Context-Aware Three-Tower Hybrid Deep Learning Architecture designed to unify disparate behavioral modalities into a single predictive framework. Unlike static recommendation systems, the proposed architecture treats intent as a dynamic function of three independent latent spaces. Tower 1 (Collaborative Domain) utilizes high-dimensional Embedding Layers to map visitor and item identifiers into a shared latent space, capturing long-term preference vectors. Tower 2 (Sequential Domain) processes the user's recent browsing trajectory using Global Average Pooling to synthesize short-term micro-intent. Tower 3 (Contextual Domain) integrates the environmental state by mapping continuous temporal data into cyclical Sine and Cosine embeddings, enabling the network to comprehend the periodic nature of shopping behavior. To overcome majority-class dominance, standard Binary Cross-Entropy (BCE) loss is replaced with a specialized Focal Loss function, which down-weights the contribution of easily classified negative examples (window shoppers) and focuses optimization on genuine purchasing signals. The dense fusion layers are further stabilized using L2 Weight Decay with a penalty factor of 0.001 to prevent overfitting on the 2.7 million-row Retailrocket dataset. Experimental results demonstrate that the Hybrid model achieves a Validation AUC-ROC of 88.7%. An Ablation Study further confirms the synergistic advantage of the architecture: the fully integrated Hybrid Engine achieved a prediction confidence of 36.83%, representing a substantial absolute gain over isolated Collaborative (29.31%) and Content-based (31.58%) baselines. These findings indicate that fusing sequential trajectories with cyclical temporal context, optimized via focal-weighted gradients, provides a strong blueprint for real-time prescriptive analytics in e-commerce.

Ask AI

Helpful

Bookmark

View Full Paper