Next basket prediction system

Abstract

The next‐basket prediction (NBP) problem aims to anticipate the items a user will include in their upcoming purchase, a task of growing importance in retail environments such as supermarkets and online grocery platforms. By using a customer’s past shopping behaviour, supermarkets can personalise product suggestions, streamline the purchasing process, and improve user satisfaction. However, this task presents significant challenges: user preferences are highly dynamic, purchases are often irregular, and meaningful patterns can be obscured by sparse data. Furthermore, unlike traditional recommendation tasks which focus on suggesting a single item, NBP requires predicting a coherent set of items—a full basket—demanding models that understand both relational and sequential dependencies. Previous research has tackled NBP using collaborative filtering, frequency‐based methods, and deep learning models. While collaborative filtering captures user similarities, it often struggles with data sparsity. Frequency‐based techniques are simple to implement and interpret, but fail to adapt to changing user intent. Deep learning and reinforcement learning models offer greater flexibility but can be computationally demanding and difficult to tune. To address these limitations, this project proposes three approaches. A Graph Neural Network (GNN) is employed to capture relational patterns between users and products by modelling their interactions on a bipartite graph structure. A Proximal Policy Optimisation (PPO) agent, based on reinforcement learning, is designed to model sequential dependencies and user decision‐making over time. Finally, a lightweight meta‐learner dynamically combines the outputs of both models using a voting scheme that considers confidence scores, reward feedback, and predictive accuracy. This stacking‐based ensemble uses the GNN’s precision and the PPO’s exploratory capabilities to generate balanced and personalised basket predictions. Evaluation on the Instacart dataset reveals that the proposed systems can surpass other models and established baselines. The GNN achieved high binary hit rate (BHR@10: 0.8510), while the PPO excelled in item‐level recall (Recall@20: 0.5153). The meta‐learner delivered the strongest overall performance (BHR@20: 0.9150, Recall@20: 0.4348), validating the effectiveness of model fusion. These results suggest that combining structural and temporal learning through meta‐learning offers a robust and scalable solution to the next‐basket prediction problem.