Cold-start Management in POI Recommendation via Reinforcement Learning and Spatial Proximity Exploration

Abstract

In this abstract we summarise the research presented in a paper appearing in the proceedings of the 31st ACM Conference on User Modeling, Adaptation and Personalization (UMAP 2023) [1]. A major goal of any destination management organisation (DMO) is the exposure of an unbiased selection of the destination Points of Interests (POIs) to their visitors. However, with their online portals, destinations can hardly achieve that goal as they usually ofer more visibility to popular POIs, compared to less popular ones. Recommender Systems (RSs) can help to generate personalised and unbiased POI recommendations, however, since they are usually trained on users' feedback (e.g., check-ins or reviews), they end up recommending well know items, to users that have a consolidated history of system interaction. Hence, RSs fail to suggest new POIs to new users, and to solve the combined new user and new item problem.

-

POIs [ 3 ], a choice prediction Markov Model [ 4 ], Q B A S E [ 2 ], and L G L M F [ 5 ].

To test the ability of these RSs to solve the new user and new item problems, we have collected, from tourism “experts” of two target destinations (Rome and Florence), a reliable ground truth composed by novel POIs, i.e., not already present in the model training set. This data consists of a selection of new POIs that can be considered as relevant for new tourists to a destination. We have generated next POI recommendations for a uniform spatial distribution of tourists and analysed cross-sectional dimensions, such as, concentration, popularity, area and catalogue covered by the recommended POIs.

We have observed that Q E X P outperforms the considered baselines in recommending relevant new items to new users. Moreover, Q E X P produces a larger variety of POI recommendations, and it significantly reduces the recommendations of popular, possibly known items, while promoting the discovery of the destination full area. We argue that Q E X P can support the construction of more efective POI RSs and that can be employed in “critical” destinations that are facing the issues brought by unregulated tourism policies, such as those connected with overtourism. Q E X P , in combination with a well-designed HCI strategy, can help these destinations in mitigating overtourism by helping to promote uncovered areas of the destination while considering the general needs and wants of tourists.

References

[1] D. Massimo , F. Ricci , Combining reinforcement learning and spatial proximity exploration for new user and new poi recommendations , in: Proceedings of the 31st ACM Conference on User Modeling, Adaptation and Personalization , UMAP '23 , ACM , 2023 , pp. 422 - 431 . doi:1 0 . 1 1 4 5 / 3 5 6 5 4 7 2 . 3 5 9 2 9 6 6 .

[2] D. Massimo , F. Ricci , Popularity, novelty and relevance in point of interest recommendation: an experimental analysis , J. Inf. Technol. Tour . 23 ( 2021 ) 473 - 508 . doi:1 0 . 1 0 0 7 / s 4 0 5 5 8 - 0 2 1 - 0 0 2 1 4 - 5 .

[3] P. Sánchez , A. Bellogín , Applying reranking strategies to route recommendation using sequence-aware evaluation, User Modeling and User-Adapted Interaction 30 ( 2020 ) 659 - 725 . URL: https://doi.org/10.1007/s11257-020 -09258-4. doi:1 0 . 1 0 0 7 / s 1 1 2 5 7 - 0 2 0 - 0 9 2 5 8 - 4 .

[4] W. Zheng , X. Huang , Y. Li , Understanding the tourist mobility using gps: Where is the next place? , Tourism Management 59 ( 2017 ) 267 - 280 . URL: https://www.sciencedirect. com/science/article/pii/S0261517716301509. doi:h t t p s : / / d o i . o r g / 1 0 . 1 0 1 6 / j . t o u r m a n . 2 0 1 6 . 0 8 . 0 0 9 .

[5] H. A. Rahmani , M. Aliannejadi , S. Ahmadian , M. Baratchi , M. Afsharchi , F. Crestani , Lglmf: Local geographical based logistic matrix factorization model for poi recommendation , in: F. L. Wang , H. Xie , W. Lam , A. Sun , L.-W. Ku, T. Hao , W. Chen , T.-L. Wong , X. Tao (Eds.), Information Retrieval Technology , Springer International Publishing, Cham, 2020 , pp. 66 - 78 .