Abstract
To encourage a more sustainable tourism, such as, visiting less popular and less crowded points of interest (POIs), recommendations must influence tourists preferences, and eventually change their usual behaviour. This can be achieved by proposing less crowded POIs that are either: (i) unknown but valuable, i.e., matching users’ preferences, or (ii) already known, but describing them so that they appear more salient and valuable. In small urban environments, such as small tourist cities — the focus of our research — often the recommendations do not introduce novel POIs, as tourists already know the available city attractions. Therefore, the primary way to influence tourists is to increase the salience of recommended POIs, which in our application scenario are also less crowded. While various techniques can affect recommendation salience (e.g., explanations), this paper does not investigate how to increase salience. Rather, we focus on the question: how can a given level of salience affect tourists’ behaviour and the quality of their experiences? The traditional train-test split testing protocol is inadequate for answering this question, as it fails to capture the causal relationship between recommendations and behaviour change. Meanwhile, online A/B testing and user studies may not always be feasible or appropriate.
This paper proposes an offline data-driven protocol to evaluate how recommendation salience impacts on tourists’ choices and the subsequent quality of the experiences tourists make. Specifically, we: (i) model and fit the choice generating process, (ii) model how recommendations alter this process, (iii) identify a relevant experience quality metric, and (iv) simulate the effect of recommendation salience on this metric. To demonstrate the usefulness of our protocol, we conduct two case studies in two small Italian cities, Verona and Brixen. In each city, we simulate how the salience of non-personalised and sustainable recommendations, i.e., designed to promote less crowded tourist attractions, may affect the actual quality of the visit experiences of the tourists. This simulation offers valuable insights for urban designers aiming to implement recommender systems on-site.