In recent years a renewed focus has been posed on the promotion of Artificial Intelligence (AI) methods to support the definition of new means to understand, study, preserve, valorise, and enjoy the Cultural Heritage (CH). In particular, the fruition of CH by the greater public can benefit from the availability of intelligent recommender systems and travel planners. On the basis of appropriate data and meta-data associated with cultural attractions and curated by experts, such systems help tourists to plan personalised cultural trips and enjoy tailored experiences. Moreover, territorial administrations, policy makers, and other stakeholders have the opportunity to promote territory, possibly focusing on less-known material or immaterial heritage. In this paper we present CHIP (Cultural Heritage Itinerary Planner), a recommender system and travel planner for cultural tourism. We present an overview of the system architecture and then focus on the solutions implemented for the Travel Recommender System (TRS) module, which provides tourists with relevant recommendations of points of interest, and for the Travel Planner (TP) module, which builds personalised itineraries based on the users' preferences and taking into account travel constraints. CHIP is currently being tested with data about the tourist attractions of the Umbria region in Italy, a region with a strong cultural vocation. Finally, we highlight future developments and goals of the project.
In the last decades Artificial Intelligence research, coupled with an ever-growing availability of both
structured and unstructured data, led to vast developments across several domains, including Cultural
Heritage (CH). The CH field, characterised by an inherent multi-disciplinary approach involving the
cooperation of humanities and science experts, is experiencing an increasing interest in the application of
cutting-edge technologies at all levels, and is a key target for future investments and research, as clearly
demonstrated by initiatives such as ECCCH [
In this paper we present CHIP (Cultural Heritage Itinerary Planner), a new framework that combines a recommender system and a travel planner. It is a powerful tool for travellers, able to provide usertailored Point of Interest (POI) recommendations in the context of a touristic territory (e.g., a city or a region), and to plan one or more itineraries built on the recommendation-phase results, while satisfying the travel constraints expressed by the user. CHIP is useful for tourists who want to optimize their time and to visit multiple attractions eficiently. It is being developed in the scope of RASTA 1, a broader MUR Italian project aimed at the promotion of the territory and cultural heritage through the use of advanced technologies to develop personalised user experiences.
A Recommender System (RS) is a computer-based intelligent systems that exploit Big Data and Artificial
Intelligence (AI) techniques to determine the appreciation of a user for an item and provide it with a
list of additional items they may be interested in [
Content-based approaches are particularly efective in cultural tourism recommendations [
For further details on the subject, we refer the reader to a recent broad survey about recommender systems in the tourism sector [8].
The problem of computing itineraries given a set of destinations to cover and a starting point has been widely studied in the literature. Prominent examples are the Travelling Salesman Problem (TSP) and the Orienteering Problem (OP) [9]; variants have been proposed to account for a vast set of constraints, including capacity, time, resources, travel means, etc. Itinerary planning problems are usually NP-hard, 1Realtá Aumentata e Storytelling Automatizzato, Augmented Reality and Automated Storytelling - Italian MUR PON project. and therefore heuristics and meta-heuristics have been widely considered to speed-up the computation at the expense of optimality. The specific problem of planning cultural itineraries[ 10] has been dubbed Tourist Trip Design Problem (TTDP) in [11]; the goal is to compute itineraries that respect a set of constraints, while maximizing one or more objectives, such as profit for the user with respect to his/her preferences, time spent on the road, or number of visited attractions. The constraints may include budget, travel times, opening times, transport means, and POIs scores, increasing the complexity of the problem. Several approaches for the TTDP have been proposed, mainly based on Operational Research. The basic version of the problem is modeled as an OP [11]. Variants are proposed to account for additional constraints, e.g., the Team Orienteering Problem (TOP) for multi-day itineraries and the Orienteering Problem with Time Windows (OPTW) for POIs opening times. Other approaches for the TTDP are based on a related problem, the Vehicle Routing Problem (VRP), whose goal is to define strategies to serve customers eficiently with a fleet of vehicles; under this approach, each day of the trip is modeled as a diferent vehicle, and customers are modeled as POIs [ 12]. For further references, we refer the reader to a comprehensive survey on TTDP [13].
CHIP is based on a micro-services architecture, to grant scalability and modularity. Each backend module communicates with the others through gRPC, whereas the frontend modules, i.e. the GUIs for tourists and administrators, communicate with the backend modules through REST APIs. An overview of the architecture is shown in Figure 1.
A prototype version of CHIP is available at the URL https://mozart.diei.unipg.it/rasta/. It is based on data provided by Umbriatourism, the tourism ofice of the Umbria regional administration, which collected a total of 624 POIs, each associated with its geographic coordinates and a textual description in English language. We describe the system modules in the following paragraphs.
The POI-TOI scorer backend module represents the core of the CHIP framework; it is responsible for
assigning each POI a relevance score for each TOI. Informally speaking, the essence of the problem is
to “estimate” the “conceptual distance” between TOIs and POIs, and then report those POIs that are
“closer” to the TOIs. The question behind this problem is how to use the descriptions, and perhaps
additional features of POIs and TOIs, to deduce the relevance of the diferent POIs with respect to the
various TOIs. In the approach of [
In this paper a machine-learning based approach is adopted to build latent-space representations of
POIs and TOIs starting from textual content, e.g., POIs textual descriptions and sets of words semantically
related to TOIs. POIs descriptions are usually written in some natural language by travel experts; it
is therefore necessary to perform simple Natural Language Processing tasks, such as tokenisation
and stopwords removal, to retain semantically-rich words only that describe the POI context. Word
embeddings can then be computed and subsequently aggregated to provide a single description-level
embedding for each POI; common aggregation strategies include component-wise averaging or cosine
centroid. Likewise, an aggregated embedding is computed for each TOI. TOIs can either be associated
with a textual description, like POIs, or linked to semantically-related keywords, e.g., in an ontology.
Nevertheless, the same embedding strategies can be applied. Once the embeddings are obtained, the
proximity between a POI and each TOI can be determined. The cosine similarity between the POI
representation and each TOI is computed. Consequently, if we consider a number of distinct TOIs,
for some positive integer , each POI is represented by a vector ∈ R, where ∈ [
To evaluate the POI-TOI scorer performance, its output has been compared to a ground truth provided by 12 experts in the cultural domain in the Umbria region. These experts have been asked to manually annotate a sample of about 12% of the 624 POIs with a relevance score for each TOI; a minimum of three annotators was employed for each POI, and the scores were averaged TOI-wise. Then, for each TOI, the rankings induced by the aggregate manual evaluation have been compared to those of the POI-TOI scorer output. In our experiment, we did not include the Food/Wine TOI, as it was dificult for the domain experts to clearly classify POIs that were particularly relevant for this category. From the resulting data, we observed medium to high correlation levels for the Religion, Landscape, and Craftsmanship TOIs. On the other hand, the results for Sports and Culture suggest the importance of further refinements of our techniques for these TOIs (see Table 1).
The User-Profile Matchmaker backend module ranks the POIs for a user according to the user preferences and the POI-TOI relevance scores computed by the POI-TOI Scorer module (see Section 2.1). In a Travel Recommender System one of the main challenges is to tailor the POI recommendations to the user preferences, which may be highly specific in terms of preferred activities/events or cultural attractions; thus, after modeling user profiles and POIs, adequate techniques are needed to accurately match them and provide custom recommendations.
When new users access CHIP, they are asked to express their preferences for each TOI, on a scale from 1 to 10; the user is then associated with a preference vector ∈ R6, where each component specifies the appreciation towards TOI , in the same order as the POI-TOI proximity vector . To assess the correlation between each POI and the user’s preferences, a similarity measure is computed between the preference vector and the POI relevance vector ; the similarity values are then used to induce a POI ranking, in which the top-placed POIs are considered more adherent to the user’s preferences; as in the case of the POI-TOI Scorer, cosine similarity was selected as similarity measure. Currently, CHIP does not provide explanations about recommendations to users; common explanation strategies on POIs rely on reviews text [15], which for less popular destinations may not be available in suficient quantity or even at all, as in the case of the Umbria region. The numerical vector representation of POIs and user preferences according to the same TOIs, however, allows the categorisation of POIs according to its most representative TOIs. The main categories for recommended POIs could be presented to users and correlated to their preferences to explain the ranking choices. Furthermore, showing salient passages of POIs textual description which closely match the preferred topics could enhance the user trust in the system. Finally, to evaluate the recommendations quality, a user study is currently being designed.
The Travel Planner (TP) backend module computes personalised itineraries for users on the basis of their travel preferences, constraints, and the categorisation of POIs in the territory. The problem has been modeled as a Prize Collecting Vehicle Routing Problem (PCVRP) [16], a variant of the conventional Vehicle Routing Problem in which the number of vehicles available is not suficient to serve all the destinations/customers. A selection criterion is employed to include a subset of the destinations only to both respect resource constraints and optimise one or more objectives, e.g., total profit collected, total travel time, total travelled distance. Formally, in a PCVRP, a complete weighted graph = (, ) is built from the set of destinations plus a starting node usually named depot; each edge ∈ is weighted according to a metric, e.g., Euclidean Distance, road distance, or travel time. A fleet of vehicles is available to serve the destinations; each destination has an associated profit and, under some formulations, an optional resource demand . The goal is to maximise the total profit ∑︀=1 ∑︀=1 , where = 1 if destination is visited by vehicle , subject to constraints, such as demand, time, and utilisation (all available vehicles must reach at least one destination). Furthermore, tours must not overlap, i.e., two vehicles must not visit the same destinations and all the vehicles must return to the starting point (the depot) at the end of the day. Such formulation is akin to that of the Tourist Trip Design Problem, in which single day or multiple, non-overlapping days itineraries must be planned to maximise tourists’ satisfaction and respect trip constraints. A mapping from PCVRP to TTDP is naturally induced, by modelling each day of the trip as a diferent vehicle, POIs as destinations/customers, and the user-POI similarity scores as prizes for visiting POIs; round-trips from the depot mimic tourists’ behaviour, departing from and returning to a fixed point, e.g., a hotel or country-house. This proves true especially in a small regional touristic context such as Umbria, where destinations are available at a fairly short distance. The CHIP travel planner employs an approach based on PCVRP, and handles the following constraints: maximum duration of a daily tour, maximum visiting time for a POI (i.e. filtering out POIS which require longer visits), number of days, lists of POIs that must be included in or excluded from the tour, number of proposed alternative tours, and transport mean, i.e., car, bicycle, or foot; the average visiting time required for each POI is factored in the itinerary computation. An adequate response time by the travel planner is required to improve the user experience, and therefore strategies are adopted to speed-up the calculation. On the one hand, a guided local search meta-heuristic allows escaping local minima while preserving eficiency[ 17]; on the other hand, POIs with similarity scores under a certain threshold = 0.5, are excluded from the itinerary computation. The TP module of CHIP has been implemented with the OR-Tools library by Google [18]; to model prizes, a penalties mechanism is employed, by assigning higher penalties for excluding POIs with a high user similarity score. The travel distances between POIs and from the trip starting point, needed to build the complete travel graph, are provided by OpenRouteService [19] for each type of transport mean.
The POI-TOI Manager handles the communication between other backend modules and the POI-TOI database, where persistent data on POIs, TOIs, computed itineraries and users are stored; it centralises the database access logic, ensuring the correctness and execution of read and write operations.
The frontend functionalities provide separate interfaces for end-users (i.e., tourists) and administratorsstakeholders. For the end-users the system ofers the following functionalities: • Expressing their preferences on the TOIs: with this information a user profile is created, which is then used by the User-Profile Matchmaker (see Section 2.2) to generate a ranking of the POIs that are most relevant and similar to the tourist needs; • Expressing the user travel constraints (see Section 2.3); this information is used by the Travel
Planner to compute travel itineraries to present to the user; • Expressing a feedback on whole itineraries and single POIs (registered tourists only). This feedback is stored and will be used to dynamically adapt the user profile over time (the user profiling functionality is currently under development).
In Figure 2 a screenshot of the user interface is shown. After collecting preferences and constraints, the user is provided with a set of alternative itineraries; for each itinerary, the detailed travel plan for each day is visualized, both as an ordered list of POIs and as an itinerary on a map.
Administrators, instead, can manage POIs and TOIs within the system. They can gather knowledge about tourism dynamics to develop further insights and to modify information related to individual POIs, such as location, description, or visiting time. In addition, they will also be able to manage the diferent TOIs, ensuring that the information is always up-to-date and relevant. Changes made to POIs and TOIs trigger the POI-TOI Scorer module (see Section 2.1), which automatically recomputes scores between POIs and TOIs, thus ensuring an accurate and consistent evaluation.
We presented CHIP, a Recommender System and Travel Planner for Cultural Tourism. The core modules of the system have been described, along with the user and administrator frontend functionalities. CHIP is currently tailored to provide travellers with robust, powerful, and flexible tools to plan personalised cultural itineraries in the Umbria region, based on their preferences. However, the architecture and techniques adopted by the system are general enough to replicate the service for other geographic regions, if suficient data is available.
In the near feature we plan to extend the system with additional functionalities. In particular, the current user-profiling strategy, based on an initial user preferences collection phase, will be augmented with an explicit and implicit feedback mechanism on POIs and itineraries, to further improve user models. The current approach relies on computing accurate and robust latent representations of POIs and TOIs, a result achieved for a subset of the considered TOIs (see subsection 2.1); additional representation strategies are currently being tested for the POI-TOI Scorer module, e.g., employing advanced language models such as TourBERT, to better estimate the POI-TOI proximity. The combination of continuous user profile refinements through feedback mechanisms and accurate embedding mechanisms will further improve the quality of the recommendations provided to users; preliminary qualitative tests show that the planned itineraries already reflect the users’ preferences. Further tests will be conducted, both before and after public availability of the system, to better assess the quality of the recommendations and improve them.
Regarding the chosen TOIs, administrators already have the possibility of changing existing TOIs or introducing new ones. The proposed approach can easily scale to a higher number of TOIs, provided that accurate and suficiently long descriptions are available for TOIs and POIs, to compute embeddings; interrelated topics will likely be described by semantically related context words, leading to similar latent space representations and, subsequently, to numerically close POI-TOI scores. For such reasons, it can be expected that, using cosine similarity, the POIs would still be correctly matched to users, under the condition that the preferences expressed by the user are coherent with the semantics of the topics. However, it must be noted that requiring the user to rate a too high number of diferent TOIs may be detrimental to the user experience and induce confusion; perhaps more interesting is the possibility of dynamically changing TOIs with respect to the time of the year or promotional strategies planned by stakeholders.
Concerning the travel planner, the system will be extended to plan non round-trip itineraries.
Finally, in the broader scope of the RASTA project, new cultural heritage fruition strategies are currently under development to integrate VR/AR techniques, that will allow designing virtual tours to visit POIs and enable travellers to enjoy immersive experience during their trips. Finally, suitable Retrieval-Augmented Generation (RAG) techniques will pave the way for new storytelling experiences in which a POI narrates its history interactively.