Measuring and optimizing beyond-accuracy objectives in next basket recommendation

Recommender systems are multi-objective systems that connect users with information aligned with their preferences. Besides accuracy, beyond-accuracy objectives such as item fairness and diversity are essential for a sustainable system and users' long-term engagement. This thesis studies the evaluation and optimization of beyond-accuracy objectives for next basket recommendation (NBR), where users exhibit repetitive and exploratory purchase behaviors over sequences of item sets. First, we measure the item fairness of representative NBR methods using seven fair ranking metrics covering equal opportunity and statistical parity. We then test their sensitivity to basket size, position weighting models, and user repeat behavior, and identify the most robust and applicable metrics. Second, we propose a two-step repetition–exploration (TREx) framework to verify a short-cut strategy for better accuracy and beyond-accuracy performance. We also reveal a loophole in the overall evaluation protocol that masks the poor accuracy of exploratory recommendation, and refine it by evaluating repetitive and exploratory recommendations separately. Third, we propose two model-agnostic re-ranking algorithms that jointly mitigate repeat bias and optimize item fairness or diversity, and extend them to unified and combined NBR methods. Finally, we investigate the semantic-ID-based generative paradigm for diverse and exploratory basket recommendation. We propose a mask-prompted generative basket recommendation architecture, where a bidirectional encoder encodes the basket sequence and a lightweight decoder autoregressively generates the semantic ID of the masked item. We further propose a balanced RQ-KMeans tokenizer that increases codeword assignment entropy to improve downstream recommendation diversity.