Capacity-aware fair POI recommendation combining transformer neural networks and resource allocation policy

Abstract

Point of Interest (POI) recommendations have primarily focused on maximizing user satisfaction, while neglecting the needs of POIs and their operators. One such need is recommendation exposure, which can lead to envy among the POIs. Some POIs may be under-recommended, while others may be over-recommended, resulting in dissatisfaction for both staff and users due to long queues or overcrowding. Existing work has not addressed the trade-off between satisfying user preferences and being fair to POIs, which typically aim to operate at capacity. Therefore, we introduce the POI fair allocation problem to model this issue, taking into account both user satisfaction and POI exposure fairness. To address this problem, we propose a fair POI allocation technique that balances user satisfaction and POI capacity-based exposure simultaneously. Our proposed model utilizes existing (transformer neural networks and attention LSTM model) personalized POI recommendation models that capture users’ spatio-temporal influences and interests in POI visits. We then propose POI capacity-based allocation using the over-demand cut policy and under-demand add policy, which ensures POI exposure ratio and envy-freeness up to certain thresholds. We evaluate the performance of our proposed model on five datasets containing real-life POI visits. Experimental evaluations show that our proposed model outperforms baselines in terms of user and POI-based evaluation metrics. To ensure reproducibility, we have publicly shared our source code at Codeocean.

Topics & Concepts