**1. Introduction**

During the last decades, the abundance of online data resources has encouraged the rapid spread of information, but it is also responsible for *information overload* (https://www. interaction-design.org/literature/article/information-overload-why-it-matters-and-howto-combat-it, accessed on 10 September 2022). A Recommendation System (RS) is an advanced search tool that alleviates this overload by suggesting content that is likely to meet the preferences and needs of potential users [1,2]. RSs have gained popularity in different fields through the creation of mostly mobile applications for delivering personalized information services to the end user. In this context, research efforts have been made in the Cultural Heritage (CH) domain to identify different profiles of cultural visitors, classify them into distinct types, and exploit such classifications in order to provide personalized suggestions of potential cultural POIs (points of interest) through RSs [3–6].

In an era where the typical cultural visitor holds smartphones and uses digital technologies to facilitate their trips, they expect to receive personalized suggestions when and where they should need them. In this ubiquitous computing environment, a mobile recommendation system (MRS) can act as a *mediator* between their visiting preferences and the available cultural content, with the objective of providing useful recommendations of potential POIs [7,8]. But in order to exploit such knowledge about visitors, relevant information must be provided to the recommender. Thus, when beginning to develop an MRS, the main question would be, "*What information is required and how to elicit it?*"

A critical fact to take into account when considering what information is required as input on behalf of the cultural visitor (often termed in the bibliography as *user feedback*) is that, especially at an early phase of a visit, they may not be consciously aware of their desires and thus not be in a position to state them explicitly [9,10]. An MRS intends to make the visitor more conscious of their desires during a visit by *classifying* them

**Citation:** Konstantakis, M.; Christodoulou, Y.; Aliprantis, J.; Caridakis, G. ACUX Recommender: A Mobile Recommendation System for Multi-Profile Cultural Visitors Based on Visiting Preferences Classification. *Big Data Cogn. Comput.* **2022**, *6*, 144. https://doi.org/ 10.3390/bdcc6040144

Academic Editor: Denis Helic

Received: 30 September 2022 Accepted: 25 November 2022 Published: 28 November 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

according to different *visiting profiles* using a variety of criteria. For example, Walsh [11] and Özel [12] classify cultural visitors based on criteria such as personal motivation, travel behavior characteristics, or demographics. McKercher [13] classifies visitors based on *cultural centrality* (high/low), i.e., the importance of cultural motives when choosing a destination and the *depth of user experience* (deep/shallow) intended when visiting cultural content. Missaoui [3] uses a combination of contextual information (such as location and time) with content from the visitors' social media interactions in order to provide personalized suggestions. Nevertheless, there is a common agreement that the effectiveness and reliability of the classification of cultural visitors in an MRS should be based on their *visiting preferences* as the primary classification criterion [14–18].

To answer the second part of the question, a variety of user feedback elicitation techniques are used to obtain the desired information from the visitor [19]. These techniques may be *explicit* (i.e., requiring some action) or *implicit*. *Explicit* techniques can be further distinguished into direct (e.g., collecting information through questionnaires, ratings, or free-text comments) or indirect, i.e., engaging the visitor in activities that do not appear directly relevant to profiling (e.g., gamification). In *implicit* techniques, on the other hand, visiting preferences are automatically deduced primarily by monitoring the visitor's online activity (e.g., "checking-in" places, social network activity, or browsing history). According to Antoniou [19] and Kanoje [2], in reality, some *combination* of both techniques is highly recommended since, in this way, both the (more static) characteristics and the (more dynamic) behavioral information of the cultural visitor are retrieved and combined in a way that can eventually lead to recommendations that are closer to the visitors' *current* desires and needs. However, such an approach requires a relatively high level of visitor engagement in order to be efficient, and thus actual effectiveness cannot be guaranteed [20,21].

Drawing on the above, in this paper, we propose ACUX Recommender (ACUX-R), an MRS, for personalized recommendation of cultural POIs to visitors based on their visiting preferences. The classification of visitors implemented in ACUX-R presents the following features:


The rest of the paper is structured as follows. Section 2 reviews related work. Section 3 describes the proposed system. Section 4 presents the evaluation of the system. Finally, conclusions and future research points are drawn in Section 5.

#### **2. Related Work**

Various RSs have been developed for the CH domain with the objective of assisting cultural visitors in planning their trips. As mentioned above, a critical factor for effective recommendation is to elicit the correct information about the user. In this context, a variety of approaches for collecting user information have been developed and proposed in the relevant literature. These include *content*, *collaborative*, *knowledge*, *demographic*, and *hybrid* approaches [23]. Meanwhile, Burke [24] argues that content-based and knowledge-based recommendation approaches are more frequently applied in the CH domain.

Indeed, ample RSs have been developed for the CH domain, applying content-based and/or knowledge-based approaches for collecting user information. Neidhardt [21,25] presents PixMeAway, a content-based RS that provides personalized recommendations of POIs to visitors. PixMeAway combines profiles from Golberg's [26] and Gibson's [18] visitor typologies in order to present a new typology, referred to as the *seven-factor model*. First, the visitor is prompted to choose among a set of pictures of POIs that they consider appealing when thinking of vacation. Next, the pictures are mapped to the aforementioned model, and a score is calculated for each factor according to the visitor's selections in order to determine their profile. Finally, a set of POIs is recommended to the visitor based on the deduced profile.

Grün [15] introduces Go2Vienna, a knowledge-based RS that provides recommendations of POIs within the city of Vienna. Go2Vienna also classifies visitors according to the *seven-factor model*. First, the cDOTT ontology (core Domain Ontology of Travel and Tourism) is employed for measuring the similarity between visiting preferences. Then, using the Pearson correlation coefficient, the similarity between the profiles of the sevenfactor model and the visiting preferences is calculated in order to determine the visitor profile and recommend an initial set of POIs. Furthermore, if the visitor is not satisfied with the recommendations, they can rate the suggested POIs by stating positive/negative feedback, which is used to refine their profiles and deliver an updated set of POIs.

PicTouRe [27] is a newer content-based version of PixmeAway which also adopts the *seven-factor model* for classifying cultural visitors. PicTouRe allows visitors to upload three to seven pictures of their choice and sort them in order of preference. Then, the system determines the visitors' profile by mapping the uploaded pictures with the seven-factor model, where each factor receives a score according to the picture's ordering. Furthermore, PicTouRe allows visitors to refine their profile using sliders that increase/decrease the percentage of each of the seven factors.

Pythia [28], City Trip Planner [29], and MyMytilene [30] follow a knowledge-based approach to collect user information, combining contextual information with visiting preferences as classification criteria.

TRIPMENTOR [31,32] is a bilingual (Greek/English) content-based MRS for Android and iOS devices, suggesting personalized routes for cultural visitors in Athens based on their visiting preferences. TRIPMENTOR is enriched with small gamification mechanisms that aim to enhance user engagement through social interaction and dynamically update the list of recommended POIs.

Regarding the collaborative approach, Herzog [33] proposes TourRec, a collaborative MRS for Android devices that recommends personalized routes to individual visitors or groups. First, TourRec determines the popularity of POIs by measuring the number of visits per POI and by matching geo-tagged photos (obtained from Flickr) with the POI's coordinates. Then, the visitor's profile is determined by combining the POI popularity, visiting preferences, and travel constraints (i.e., time limitations or the need to start/end at specific POIs). Finally, the system recommends routes of POIs that match the deduced profile. Figueredo [34] presents Find Natal, a collaborative MRS for both Android and iOS devices that recommend POIs to cultural visitors using social media photos and previous users' ratings and comments as user input information.

Moreover, various hybrid approaches have been proposed. Missaoui [3] presents LOOKER, a hybrid MRS for Android devices that delivers personalized POI recommendations to visitors, using a content-based filtering module that filters content (i.e., reviews in social posts) that the visitor has generated on social media. Then, using language models, the filtered content is converted into visiting preferences and is combined with contextual information to determine the visitor's profile. Based on the deduced profile, personalized recommendations of POIs are shown on a map or in a list, along with reviews of previous visitors. Logesh [35] introduces PCAHTRS, a personalized context-aware hybrid RS that uses contextual information, previous user reviews, and POI similarity in order to recommend POIs to cultural visitors. Finally, Meehan [36] presents VISIT, a hybrid RS that uses a

combination of collaborative, content-based, and demographic approaches for classifying visitors in order to recommend POIs.

The literature review showed that, unlike ACUX-R, the great majority of RSs developed for the CH domain do not classify their users into distinct visitor profiles. Rather, the user provides the required information (usually visiting preferences, demographics, or/and contextual information), and the RSs directly suggest POIs based on that information. On the other hand, RSs that do perform user classification as an intermediate step for providing recommendations most of them classify visitors into multiple profiles (multi-label classification) and also allow them to manually fine-tune their assigned profile (as is the case with ACUX-R).

#### **3. ACUX Recommender**

#### *3.1. ACUX-R Architecture*

ACUX-R has been developed following a typical three-tier architecture, using Google's Android Studio and Flutter Software Development Kit (SDK) (see Figure 1):

	- - Content data, i.e., information about the available POIs (such as name, description, location, GPS data, or images).
	- - User data, i.e., information regarding the user's visiting preferences and assigned profile, together with other personal information (e.g., account details).
	- - Classification data: i.e., the knowledge required for classifying (i) the visitors and (ii) the POIs available, according to visiting preferences.

**Figure 1.** The ACUX-R 3-tier architecture.
