Season-aware attraction recommendation method with dual-trust enhancement

Abstract

Attraction recommendation plays an essential role in tourism. For example, it can relieve information overload for tourists and increase sales for tourism operators. When making travel decisions, tourists depend heavily on the personal preferences and suggestions from people they trust. However, most existing attraction recommendation methods focus on the tourist preferences for topics of attractions, yet overlook the seasonality in topic preferences. Additionally, extant studies are generally based on a single type of trust, which may represent trust relations inaccurately. In order to overcome these issues, we propose a novel season-aware attraction recommendation method based on the seasonal topic preferences and dual-trust relations. Firstly, we capture tourists’ seasonal topic preferences by analyzing their travel histories along two dimensions: time and attraction. Secondly, we develop a dual-trust relationship (DTR) model based on familiarity-based trust and similarity-based trust, in contrast to existing studies that purely focus on a single type of trust. Thirdly, we propose a novel season-aware attraction recommendation method named SAR-DTR. In a specific season, it predicts ratings based on both topic preferences in the given season and suggestions from tourists they trust. To demonstrate the superiority of the proposed method to other approaches, an empirical study with real-world data was conducted. The experimental results regarding both prediction and recommendation performance are reported.

Keywords

Attraction recommendation seasonal topic preference similarity-based trust familiarity-based trust

1 Introduction

Tourism industry has witnessed an astonishing growth in recent years. According to the report by the World Travel & Tourism Council (WTTC) in 2016 1 , the direct and indirect contribution of travel & tourism accounted for 10.2% of the world’s gross domestic product (GDP) and approximately 10% of all jobs. Apparently, the tourism industry has emerged as a key driver of social and economic development. However, the boom challenges tourism operators with fierce competition and tourists with information overload. To survive in this market, tourism operators utilize travel recommender systems to attract tourists. For tourists, they frequently employ travel recommender systems to relieve information overload as recommender systems can provide tailored information based on their preferences [1]. Given the critical role of recommender systems, the development of effective travel recommendation methods is highly desirable.

Tourism activity is closely related to tourist preferences [2]. In recent years, inferring tourist preferences for topics of attractions from user-generated content has been an emerging issue. Nevertheless, the seasonality in topic preferences, a distinctive feature of tourism, is overlooked by the majority of studies. Tourism is actually a seasonal phenomenon [3]. Tourists’ topic preferences are characterized by seasonality, which indicate the dynamic changes of topic preferences with seasons. It’s mainly attributed to the distinctive characters of attractions and different needs of tourists with respect to seasonal context. Although topic preference has drawn much attention in travel recommendation, research on seasonal topic preference is scarce.

Because tourism activity is difficult to evaluate before consumption, tourists heavily depend on the information search when making travel plans. The proliferation of Internet facilitates their access to rich travel information. However, online information can be false, which may lead to misguided travel plans. Therefore, trust has been adopted by recommendation methods to enhance the reliability of results. Existing trust-based travel recommendation methods are generally based on the trust explicitly issued by tourists or the similarity-based trust [4 –6]. Actually, trust inferred from the familiarity between users commonly called familiarity-based trust [7], can potentially be exploited to advance the performance. It reflects the fact that the tourists take acquaintance as a major information source [8]. Since both similarity-based trust and familiarity-based trust exert considerable influence on travel decision-making process, it would be better to consider these two complementary types of trust when making travel recommendations.

To overcome these limitations, we infer the tourists’ seasonal topic preferences, develop a dual-trust relationship (DTR) model, and propose a novel season-aware attraction recommendation method. We initially introduce the construction of DTR model based on familiarity-based trust and similarity-based trust. Subsequently, we propose a season-aware attraction recommendation method named as SAR-DTR, in a specific season, which generates recommendations based on seasonal topic preferences and suggestions from trusted tourists. Finally, using the large scale real-world data, we conduct an empirical study to demonstrate the performance improvement achieved by the proposed method.

The remainder of this paper is organized as follows. In Section 2, the literature related to our study is reviewed. Section 3 presents the inference of seasonal topic preferences and the DTR model. In Section 4, the proposed attraction recommendation method is introduced in detail. The empirical study and results are discussed in Section 5, and we conclude the paper with limitations and future studies in Section 6.

2 Literature review

Tourist preferences are extremely important to making satisfactory travel recommendations. User preferences can be inferred explicitly and implicitly [9]. The explicit ways infer user preferences from interactions and feedbacks [10] and the implicit ways capture user preferences by analyzing user’s behaviors [11]. In tourism, tourists are usually required to provide preferences explicitly. For example, Garcia et al. [12] generated travel recommendations based on the general likes provided by tourists. In [13], Vansteenwegen et al. proposed an expert system, which requires tourists to enter the personal interest and plans routes for tourists. Inferring preferences from user-contributed data, one of typical explicit ways, has drawn much attention. Besides, because topic is more general and more flexible in describing characters of attractions and tourist preferences, it has been adopted by a growing body of research. For example, in [14], Jiang et al. extracted topics of user preferences by analyzing the textual description of photos and recommended points of interest according to the similar users. Xu et al. [15] used community-contributed geotagged photos to obtain the topic distribution of user travel histories and then recommended locations to tourists.

The literature review suggests that existing studies focus on the topic preferences but overlook the seasonality. To overcome this limitation, we inferred tourists’ seasonal topic preferences from their travel histories. Given that attractions are the fundamental components of destinations to attract tourists [16], we focus on the attraction recommendation in this study.

To reduce risks from trustless users, the trust-based recommendation methods have gained extensive attention. There are many types of trust with respect to its provenance [17]. The similarity between users contributes to the generation of trust, known as similarity-based trust. The rationales behind the similarity-based trust can be explained by social influence theory which indicates that users share similar preferences with people they are connected with [18]. In paper [19], Ziegler et al. proved that there is a strong correlation between trust and interest similarity. Golbeck [20] demonstrated the correlation between trust and profile similarity. Fernandez-Gago et al. [21] presented a trust model based on the context similarity network. In addition, familiarity is frequently adopted to measure the trust levels as well. For instance, in [22], Zhang et al. proposed an improved familiarity-based trust model based on the familiarity measurement. Another related work is [23], which studied the familiarity-based trust in the mobile social networks based on the social interactions. The trust-based travel recommendation methods are mostly based on the trust issued by tourists. According to the trust expressed by users, Avesian et al. [4] proposed a trust-enhanced recommender system, which recommends skiing routes based on the user experiences and trust score. In paper [5], García-Crespo et al. evaluated services according to the semantic similarity between the preferences of users and those of his trusted friends. Hinze et al. [6] proposed the concept of geographic trust based on the geographic distance and presented a trust-based recommendation service.

One limitation of methods above is that they are purely based on one type of trust, leading to inaccurate description of trust relations between users. Actually, tourists prefer to seek suggestions from people they are familiar with [8]. Thus, it would be reasonable to consider the familiarity-based trust when recommending attractions. To our best knowledge, travel recommendation methods have not paid attention to the trust model that considers both the similarity-based trust and familiarity-based trust. Therefore, in this study, we construct a dual-trust model based on the above two types of trust and further develop a personalized season-aware attraction recommendation method with dual-trust enhancement.

3 Dual-trust relationship model

3.1 Familiarity-based trust model

Familiarity-based trust refers to the trust based on the personal familiarity [7]. Studies indicate that social interactions usually contribute to the familiarity between users [24]. The virtual interactions, such as following others, visiting homepages, commenting on the content, leaving message, can lead to higher familiarity between them. The familiarity in this paper depends on three primary types of virtual social interactions, namely “following others”, “visiting homepage” and “commenting on the content”, with the objective of making the proposed trust model work with majority of social medias and simultaneously reveal the familiarity between users precisely. Suppose two tourists u₁ and u₂ in the social media SM, each of three kinds of virtual interactions and its value is described as follows.

Following others: This functionality makes it easy for tourists to get the latest information from people they are interested in. It leads to the repeated exposure and thus generates familiarity [24]. Denote the “following others” as $s_{u_{1}, u_{2}}^{f}$ . In this study, $s_{u_{1}, u_{2}}^{f}$ is defined as Equation (1): $s_{u_{1}, u_{2}}^{f} = {\begin{matrix} 1, & u_{1} followed u_{2} \\ 0, & otherwise \end{matrix}$ (1)

Visiting homepage: Visiting homepage offers opportunity for tourists to gather information from others’ homepages to support their travel decisions. Let $s_{u_{1}, u_{2}}^{v}$ denotes the “visiting homepage” and it is defined as Equation (2): $s_{u_{1}, u_{2}}^{v} = {\begin{matrix} 1, & u_{1} visited homepage of u_{2} \\ 0, & otherwise \end{matrix}$ (2)

Commenting on the content: Tourists can comment on the contents that are published by others to share opinions or seek advice. Denote the “commenting on the content” as $s_{u_{1}, u_{2}}^{c}$ . $s_{u_{1}, u_{2}}^{c}$ is defined as Equation (3): $s_{u_{1}, u_{2}}^{c} = {\begin{matrix} 1, & \begin{matrix} u_{1} commented on the content \\ published by u_{2} \end{matrix} \\ 0, & otherwise \end{matrix}$ (3)

Based on the description of social interactions, the degree of familiarity from u₁ to u₂, denoted by f (u₁, u₂) can be calculated by aggregating three types of interactions. The f (u₁, u₂) is formulated as Equation (4): $f (u_{1}, u_{2}) = s_{u_{1}, u_{2}}^{f} + s_{u_{1}, u_{2}}^{v} + s_{u_{1}, u_{2}}^{c}$ (4)

The more types of social interactions between tourists indicate the higher level of familiarity between tourists.

Inspired by the positive correlation between familiarity and trust [25], we define the familiarity-based trust from u₁ to u₂, FT (u₁, u₂), as Equation (5): $FT (u_{1}, u_{2}) = \frac{f (u_{1}, u_{2})}{\max (f (u_{1}, u))}$ (5) where u is a tourist in the social media SM. As indicated in Equations (4) and (5), the familiarity-based trust is positively associated with social interactions. Without loss of generality, the more familiar tourists are, the higher trust levels they have. Hence, FT (u₁, u₂) implies how much u₁ considers u₂ trustworthy based on the previous social interactions.

3.2 Similarity-based trust model

Similarity-based trust refers to the trust that is generated according to the degree of similarity between users, which is generally the similarity in preferences [19, 20]. Because the tourist’s topic preferences differ with respect to seasonal context, we define the similarity-based trust as the trust based on the similarity in tourists’ seasonal topic preferences. Rather than using the questionnaire survey, we capture seasonal topic preferences by analyzing travel histories that are published on social medias by tourists themselves.

Denote the season set as S = {spring, summer, autumn, winter} and the topic space as TS = {t₁, t₂, …, t_m}, where t₁, t₂, …, t_m represent different topics. For a specific tourist u in social media SM, denote his travel histories as H_u = {h₁, h₂, …, h_w}. A visit record in H_u is denoted by h = (p, a), which refers that tourist u visited attraction a in month p. To infer the seasonal topic preferences of u, each visit record is analyzed along travel time (i.e., month) dimension and attraction dimension. Two dimensions are mapped into seasons and topic space, respectively. To map the travel time into seasons, a year is separated into four seasons by the meteorological division method. Consequently, the method groups March, April and May into spring, June, July and August into summer, September, October and November into autumn, December, January and February into winter. To map attractions into topic space, the topics of attractions that are defined with a unified classification standard are obtained. Then, for a specific tourist u, his seasonal topic preferences are inferred as follows.

Firstly, for each visit record h = (p, a) in travel histories H_u, the month p can be mapped into season s and the attraction a can be mapped into topic space, as a result of which, attraction a is represented as a m-dimensional topic vector, denoted by V_a. The elements in V_a are 0/1. That is, if attraction a is featured with topic t, t ∈ TS, the corresponding element in V_a is 1, otherwise 0. Now, a visit record is denoted as h = (s, V_a). Secondly, according to the season s in season set S, sum up the topic vectors of attractions that u visited in season s. We will get a 4 × m matrix, where the elements are denoted as $p_{u}^{s, t}$ . Finally, the seasonal topic preferences matrix can be constructed by normalizing the elements $p_{u}^{s, t}$ into [0, 1], as shown as Equation (6): $p_{u}^{s, t} = \frac{p_{u}^{s, t}}{\sum_{i = 1}^{m} p_{u}^{s, i}}$ (6) where i denotes topic.

Denote the seasonal topic preferences matrix of tourist u as STP_u, which presents the preferences of tourist u for topics with respect to the seasonal context. ${STP}_{u}^{s}$ is a m-dimensional vector and denotes the topic preferences of tourist u in season s and the element $p_{u}^{s, t}$ represents the preference of u for topic t, t ∈ TS in season s.

Inspired by the positive correlation between similarity and trust [19, 20], the similarity-based trust is defined as the relevance of tourists’ seasonal topic preferences. Given the seasonal topic preferences matrices of two tourists u₁ and u₂, denoted as STP_{u
₁} and STP_{u
₂}, the similarity-based trust value, ST (u₁, u₂), is formulated as Equation (7): $ST (u_{1}, u_{2}) = \frac{\sum_{s = 1}^{p} \sum_{t = 1}^{m} p_{u_{1}}^{s, t} \times p_{u_{2}}^{s, t}}{\sqrt{\sum_{s = 1}^{p} \sum_{t = 1}^{m} (p_{u_{1}}^{s, t})^{2}} \sqrt{\sum_{s = 1}^{p} \sum_{t = 1}^{m} (p_{u_{2}}^{s, t})^{2}}}$ (7) where the $p_{u_{1}}^{s, t}$ and $p_{u_{2}}^{s, t}$ are the elements in STP_{u
₁} and STP_{u
₂}, respectively; ST (u₁, u₂) indicates how trustworthy u₁ considers the information provided by u₂ to be in terms of similarity in preferences. Note that ST (u₁, u₂) could be negative and the magnitude of negativeness indicates the untrustworthy degree.

3.3 DTR trust model

DTR aims at measuring the credibility of tourists in acting as the information source of attraction recommendation. Therefore, the DTR should firstly ensure the reliability of tourists, and then evaluate the relevance of information provided by them. Although the familiarity-based trust and similarity-based trust play significant role in seeking travel advice, they are different in reference significance. Specifically, the proposed similarity-based trust is derived from user-contributed data (i.e., the travel histories), the authenticity of which is the prerequisite for the evaluation of similarity-based trust. In reality, the large amount of false information online poses a great challenge for the proposed similarity-based trust. In comparison, the familiarity-based trust fundamentally depends on the social interactions that are initialized by tourists themselves, which makes familiarity-based trust independent of false information. To ensure the reliability of information from others and enhance the reliability of recommendation results, we take the familiarity-based trust as foundation of DTR. Furthermore, the acquaintances may differ in preferences and suggestions from acquaintances with different preferences are irrelevant. Thus, we further differentiate the familiarity-based trust according to the similarity-based trust.

Based on the analysis above, the DTR trust level from u₁ to u₂, denoted by T (u₁, u₂), is computed by incorporating the similarity-based trust into familiarity-based trust, as shown in Equation (8): $T (u_{1}, u_{2}) = {\begin{matrix} 0, & FT (u_{1}, u_{2}) = 0 \\ \frac{FT (u_{1}, u_{2}) + ST (u_{1}, u_{2})}{2}, & otherwise \end{matrix}$ (8)

Note that because trust relations between users are more stable, the DTR relationships in this paper are independent of seasonal context. According to the DTR values, a weighted trust network can be constructed.

4 Season-aware attraction recommendation method

Tourists usually make travel plans based on both personal preferences and evaluations (ratings) from trusted tourists to improve the tour quality and reduce potential risks [26]. In this section, we propose a novel season-aware attraction recommendation method, which predicts tourists’ ratings on attractions based on their topic preferences in the given season and ratings from trusted tourists. The proposed method is named as SAR-DTR. The general scheme of SAR-DTR is shown as Fig.1 and the notations used in the method are listed in Table 1.

Fig.1

General scheme of SAR-DTR.

Table 1

Notations used in recommendation equations

Notation	Description
${\hat{R}}_{s}^{p} (u, a)$	The predicted rating of tourist u on attraction a in season s based on seasonal topic preferences
sim_s (u₀, u)	The preference similarity of u₀ and u in season s
$U_{u}^{p}$	The neighbors of tourist u selected according to the similarity in preferences
${\bar{R}}_{u}$	The average attraction rating of u
$U_{u}^{t}$	The neighbors of tourist u selected according to the trust relations
N _k	The set of tourists in kth-level
t _u₀,u	Trust value from u₀ to u
V _a	Topic vector of attraction a
sim (a, b)	The topic similarity of attraction a and b
$\hat{R} (u, a)$	The predicted rating of tourist u on attraction a based on his ratings on similar attractions
${\hat{R}}_{t} (u, a)$	The predicted rating of tourist u on attraction a based on trust relations
${\hat{R}}_{s} (u, a)$	The predicted rating of tourist u on attraction a in season s based on preferences and trust relations
β	A parameter utilized to balance the preference-based rating and trust-based rating

4.1 Preference-based rating prediction

Studies indicate that users with similar preferences in the past are more likely to be interested in the same items in the future [27]. In order to predict preference-based ratings according to the tourists with similar preferences, it is important to accurately identify the similar tourists among all tourists in the travel recommender system. As aforementioned, the topic preferences are obviously featured with seasonality. Thus, the preference-based rating ${\hat{R}}_{s}^{p} (u_{0}, a_{0})$ is generated according to the similar tourists in the given season, which is different from existing methods that predict ratings based on the similarities without considering seasonality. Let s denote season and ${STP}_{u_{0}}^{s}$ and ${STP}_{u}^{s}$ represent the topic preferences of tourist u₀ and u in season s, which are m-dimensional vectors. The preference similarity of tourist u₀ and u in season s can be estimated by computing the relevance of their topic preferences vectors ${STP}_{u_{0}}^{s}$ and ${STP}_{u}^{s}$ , which is defined as Equation (9): ${sim}_{s} (u_{0}, u) = \frac{{STP}_{u_{0}}^{s} \times {STP}_{u}^{s}}{∥ {STP}_{u_{0}}^{s} ∥ \cdot ∥ {STP}_{u}^{s} ∥}$ (9)

The higher the sim_s (u₀, u) is, the more similar of u₀ and u are in preferences in season s. Note that because of the seasonality in topic preferences, the similarities between tourists change with seasonal context. According to the similarities between tourists, the top-n most similar tourists can be identified, denoted by $U_{u_{0}}^{p}$ . In this paper, we set n to 5.

Utilizing preference similarity between target tourist u₀ and other similar tourists in $U_{u_{0}}^{p}$ , we predict the preference-based rating of tourist u₀ on attraction a₀ in season s as Equations (10), (11) and (12): $\begin{matrix} {\hat{R}}_{s}^{p} (u_{0}, a_{0}) \\ = {\bar{R}}_{u_{0}} + \frac{\sum_{u \in U_{u_{0}}^{p}} {sim}_{s} (u_{0}, u) (R_{u, a_{0}} - {\bar{R}}_{u})}{\sum_{u \in U_{u_{0}}^{p}} {sim}_{s} (u_{0}, u)} \end{matrix}$ (10) ${\bar{R}}_{u_{0}} = \frac{\sum_{a \in A_{u_{0}}} R_{u_{0}, a}}{| A_{u_{0}} |}$ (11) ${\bar{R}}_{u} = \frac{\sum_{a \in A_{u}} R_{u, a}}{| A_{u} |}, \forall u \in U_{u_{0}}^{p}$ (12) where ${\bar{R}}_{u_{0}}$ and ${\bar{R}}_{u}$ indicate the average attraction rating of u₀ and u. ${\hat{R}}_{s}^{p} (u_{0}, a_{0})$ refers to the level to which tourist u₀ likes attraction a₀ based on how other similar tourists feel about a₀ weighted by the similarities in season s.

4.2 Trust-based rating prediction

According to the transitivity property of trust, the tourists who are indirectly connected with trust relations are also trustworthy, making more reliable information available for trust-based recommendation methods [28]. Thus, trust-based rating is generated according to multiple levels of tourists with the objective of eliminating the problems generated from information deficiency as well as increasing the credibility of recommendation results. We regard the tourists who are directly trusted by u₀ and are selected to make recommendations as first level, the selected tourists directly trusted by tourists in first level as second level, and so on. Denote the tourist in kth level as N_k. Note that k may be infinite and for the target tourist u₀, the reliability of tourists in N_k will decrease with the increment of k [29]. Thus, it would be better to set the upper limit of k. According to the “three degrees of influence rule” [30], we set the maximum of k to 3. Besides, as tourists differ in trustworthiness, they have different reference significance [31]. Instead of selecting tourists randomly, the tourists with higher trust values should be selected with higher priority. For a specific tourist u₀, we will identify the top-n most trustworthy tourists as his neighbors, denoted by $U_{u_{0}}^{t}$ . The n in this paper is set to 5.

When generating trust-based rating ${\hat{R}}_{t} (u_{0}, a_{0})$ , the tourists in $U_{u_{0}}^{t}$ will be firstly considered. For tourist u in $U_{u_{0}}^{t}$ , if u has scored a₀, denote the rating as R_u,a₀. R_u,a₀ will be retained to predict ${\hat{R}}_{t} (u_{0}, a_{0})$ and we will not further consider the tourists who are trusted by u, since u has given the rating explicitly. Tourism is a relatively infrequent activity, which leads higher sparsity of rating matrix. When the tourist did not score the target attraction, the average of his ratings on the attractions that has the same topics can be utilized as his rating on target attraction. That is, if u did not score a₀, we will adopt the average of his ratings on similar attractions as rating from u on a₀, denoted by $\hat{R} (u, a_{0})$ . Specifically, attractions with same topics are similar in characters to some extent. It is reasonable to assume that ratings on similar attractions reflect the evaluation on the target attraction. The similarity between attractions is computed based on their m-dimensional topic vectors, which are generated in section 3. The similarity between attraction a and attraction b is defined as Equation (13): $sim (a, b) = \frac{V_{a} \times V_{b}}{∥ V_{a} ∥ \cdot ∥ V_{b} ∥}$ (13) where V_a and V_b refer to the topic vectors of a and b. According to similarities of attractions, the top-n most similar attractions that u has scored can be identified, denoted as A_u. We adopt the average ratings from u on similar attractions as his rating on a₀, as shown in Equation (14): $\hat{R} (u, a_{0}) = \frac{\sum_{b \in A_{u}} R_{u, b}}{| A_{u} |}$ (14) where R_u,b denotes the rating on similar attraction.

Although $\hat{R} (u, a_{0})$ can reflect the evaluation of u on a₀ to some degree, it would be better to consider an adequate number of ratings to mitigate the impact of noisy information [31]. Besides, the homepages of tourists are usually open to others, which foster the flow of information in the entire network. Furthermore, tourists who are indirectly connected with u₀ also exert influence on him [30]. Therefore, if u does not score a₀ and k does not reach the maximum, after generating $\hat{R} (u, a_{0})$ , we will further consider the neighbors of u, namely $U_{u}^{t}$ . For the tourists in $U_{u}^{t}$ , repeat the above process. This process will continue until k reaches the maximum or all tourists in kth level scored a₀. Then, for a specific tourist u in N_k who did not scored a₀, his final predicted rating is computed as Equation (15): $R_{u, a_{0}} = \frac{\hat{R} (u, a_{0}) + \frac{\sum_{v \in U_{u}^{t}} t_{u, v} R_{v, a_{0}}}{\sum_{v \in U_{u}^{t}} t_{u, v}}}{2}$ (15) where $\hat{R} (u, a_{0})$ indicates rating based on his ratings on similar attractions; $\frac{\sum_{v \in U_{u}^{t}} t_{u, v} R_{v, a_{0}}}{\sum_{v \in U_{u}^{t}} t_{u, v}}$ is rating from his neighbors. We take the average of two parts as his final evaluation of a₀.

According to the ratings from direct neighbors of u₀ on a₀, we define the trust-based rating ${\hat{R}}_{t} (u_{0}, a_{0})$ as Equation (16): ${\hat{R}}_{t} (u_{0}, a_{0}) = \frac{\sum_{u \in U_{u_{0}}^{t}} t_{u_{0}, u} R_{u, a_{0}}}{\sum_{u \in U_{u_{0}}^{t}} t_{u_{0}, u}}$ (16) where t_u₀,u indicates the trust level from u₀ to u.

Since tourists heavily depend on the personal preferences and suggestions from trusted tourists to make travel decisions, we generate rating according to preference-based rating and trust-based rating with a parameter β, as shown in Equation (17): ${\hat{R}}_{s} (u_{0}, a_{0}) = β {\hat{R}}_{s}^{p} (u_{0}, a_{0}) + (1 - β) {\hat{R}}_{t} (u_{0}, a_{0})$ (17) where ${\hat{R}}_{s}^{p} (u_{0}, a_{0})$ denotes the rating based on preferences, ${\hat{R}}_{t} (u_{0}, a_{0})$ denotes the rating based on trust relations and β ∈ [0, 1]. In particular, if there is no similar tourist in season s to predict ${\hat{R}}_{s}^{p} (u_{0}, a_{0})$ , ${\hat{R}}_{s} (u_{0}, a_{0}) = {\hat{R}}_{t} (u_{0}, a_{0})$ . If ${\hat{R}}_{t} (u_{0}, a_{0})$ can not be generated because u₀ does not trust anybody, ${\hat{R}}_{s} (u_{0}, a_{0}) = {\hat{R}}_{s}^{p} (u_{0}, a_{0})$ . Note that although seasonal topic preferences are considered in both preference-based rating and trust-based rating, they play different roles. The seasonal topic preferences in the preference-based rating aim at evaluating the level to which tourists like attractions in terms of their own preferences. In the prediction of trust-based rating, seasonal topic preferences are utilized to evaluate the trustworthiness of information provided by other tourists.

5 Evaluation and analysis

5.1 Experiment setup

To evaluate our proposed recommendation method with real-world data, we crawled data from Mafengwo (www.mafengwo.cn), a famous online social network that is specialized in tourism in China. Because Mafengwo does not provide the topics of attractions, we collect the topics from a famous online tourism operator in China Ctrip (www.ctrip.com). We conducted the experiments with attractions in China and tourists who have visited at least one attraction and simultaneously followed more than one tourist in Mafengwo. With the objective to establish the trust network, we selected 5 tourists randomly who satisfy the condition above as starting nodes and further moved to the tourists who are followed by the selected tourists. This process continued until there are 6,676 tourists in the network. Subsequently, the tourism data and social data about the tourists in the dataset were collected from Mafengwo. Ultimately, the topics of attractions were collected from Ctrip according to the names of attractions. The statistics of the experimental dataset are shown in Table 2.

Table 2
Statistics of the experimental dataset

Tourists 6,676

Attractions 10,827

Ratings 174,112

Topics 147

Following others 394,461

Visiting homepages 3,582

Commenting on the content 55,505

Social interactions 453,548

Average ratings per tourists 26.08

Average social interactions per tourist 67.94

Average ratings per attraction 16.08

According to the proposed DTR model, the DTR values are calculated based on the social interactions and seasonal topic preferences. In the experiment, 437,966 familiarity-based trust relations are established and the average familiarity-based trust level is 0.636. Besides, 214,705 similarity-based trust relations are established between tourists who are connected with familiarity-based trust relations. Based on the familiarity-based trust and similarity-based trust, we finally establish 437,966 DTR trust relations and the average of DTR values is 0.361.

5.2 Recommendation strategies and performance metrics

In the experiment, we demonstrate the superiority of the proposed SAR-DTR method to the four benchmark methods. Firstly, we select the preference-based method that is based on seasonal topic preferences only (named as STP), to demonstrate the importance of the trust relations in attraction recommendation. STP is introduced in subsection 4.1. Secondly, the trust-based recommendation method that is purely based on dual-trust relations (named as DT) and detailed in subsection 4.2 is selected to confirm the necessity of seasonal topic preferences. Finally, two classic collaborative filtering methods, the user-based collaborative filtering method (UCF) [32] and item-based collaborative filtering method (ICF) [33], are selected to demonstrate the superiority of SAR-DTR to the traditional methods and the crucial roles of both seasonal topic preferences and trust relations.

To comprehensively evaluate the prediction performance of our recommendation method, several indicators are used, including Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), NRMSE, Coverage and F-Measure. Additionally, to evaluate the recommendation performance, the Recommendation Hit Rate (RHR) and Recommendation Loss Rate (RLR) are used.

The prediction accuracy denotes the closeness of predicted rating to the actual rating [29]. The MAE and RMSE, two widely used evaluation metrics, are adopted to measure the prediction accuracy [34]. The smaller the MAE and RMSE are, the more accurately the method predicts. MAE and RMSE are defined as Equations (18) and (19): $MAE = \frac{\sum_{u, i} | R_{u, i} - {\hat{R}}_{u, i} |}{N}$ (18) $RMSE = \sqrt{\frac{\sum_{u, i} (R_{u, i} - {\hat{R}}_{u, i})^{2}}{N}}$ (19) where R_u,i denotes the actual rating that tourist u gave to attraction i; ${\hat{R}}_{u, i}$ refers to the predicted rating that is generated by the recommendation method and N indicates the number of tested ratings.

Because of the insufficient information, some ratings can not be predicted. The metric Coverage is applied to measure the ability of recommendation methods in predicting ratings in the context of information insufficiency. The Coverage is defined as the percentage of pair of 〈tourist, attraction〉, for which the method can predict the ratings [35]. The larger the Coverage is, the more ratings can be predicted. It is represented as Equation (20): $Coverage = \frac{P}{T}$ (20) where P denotes the pairs of 〈tourist, attraction〉 that can be predicted and T indicates the total number of tested ratings.

After converting RMSE into metric NRMSE, which is in the range of [0,1], we apply F-Measure to combine metrics RMSE and Coverage [35], as Equations (21) and (22). $NRMSE = 1 - \frac{RMSE}{4}$ (21) $F - Measure = \frac{2 \times NRMSE \times Coverage}{NRMSE + Coverage}$ (22)

The higher F-Measure denotes the better performance of recommendation methods with respect to both prediction accuracy and coverage.

An attraction i is recommended to tourist u when the predicted rating ${\hat{R}}_{u, i} \geq δ$ , where δ is the predefined recommendation threshold. The recommendation performance of methods can be evaluated by the metrics RHR and RLR, which are separately defined as Equations (23) and (24) [36]: $RHR = \frac{| (u, i) ∣ {\hat{R}}_{u, i} \geq δ, R_{u, i} \geq δ |}{| (u, i) ∣ {\hat{R}}_{u, i} \geq δ |}$ (23) $RLR = \frac{| (u, i) ∣ {\hat{R}}_{u, i} < δ, R_{u, i} \geq δ |}{| (u, i) ∣ {\hat{R}}_{u, i} < δ |}$ (24) where R_u,i indicates the actual rating that tourist u gave to attraction i. The Recommendation Hit Rate evaluates the probability that the attractions recommended by recommendation method are also scored highly by tourists. The Recommendation Loss Rate computes the probability that the attractions evaluated highly by tourists are not recommended by the method.

5.3 Experimental results

In order to evaluate the performances of recommendation methods, the dataset is divided into two parts according to the time of travel: the travel activities occurred before 2015 as training dataset (163,734 ratings), and those occurred in 2015 as evaluation dataset (10,378 ratings). The training dataset is utilized to infer tourists’ seasonal topic preferences, construct the DTR trust network and predict ratings for the pair of 〈tourist, attraction〉 in evaluation dataset. The evaluation dataset is utilized to evaluate the prediction performance and recommendation performance. Note that the SAR-DTR method and STP method take into consideration of seasonal context, while the other three methods (DT, ICF and UCF) are indifferent to the season. We not only compare their overall performance, but also compare the seasonal performance of SAR-DTR with benchmark methods. The overall performance is computed by averaging the seasonal performances across the four seasons.

We firstly compare the performance of SAR-DTR with different β. From Equation (17), we can observe that the bigger the β is, the more important role the preference-based rating plays in rating prediction. We firstly examine the influence of β on prediction performance and recommendation performance of SAR-DTR. The β ∈ (0, 1) and the step is set to 0.1 in this experiment.

Table 3 shows that β has a trivial effect on the prediction performance of SAR-DTR. For example, with respect to the overall performance of SAR-DTR with different β on MAE and RMSE, the maximum differences are merely 0.0258 and 0.0365, respectively. In terms of Variance, we can observe that with the change of β, the variation of the index values are very small and the performance on Coverage does not change. The experimental results indicate that β has no significant influence on SAR-DTR. To make the proposed SAR-DTR integrally perform best, we determine β according to the overall performance of SAR-DTR on F-Measure, an aggregate indicator that synthetically considers the prediction accuracy and Coverage. As shown in Table 3, the SAR-DTR obtains the best overall prediction performance on F-Measure when β = 0.4.

Table 3
Prediction performance of SAR-DTR with different β

Season Metric β Variance

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Spring MAE 0.4516 0.4433 0.4380 0.4354 0.4354 0.4376 0.4417 0.4478 0.45 5 0.00005308

RMSE 0.5976 0.5869 0.5791 0.5742 0.5724 0.5736 0.5780 0.5853 0.5955 0.00008827

Coverage 0.9705 0.9705 0.9705 0.9705 0.9705 0.9705 0.9705 0.9705 0.9705 0

F-Measure 0.9066 0.9081 0.9092 0.9099 0.9102 0.9100 0.9094 0.9083 0.9069 0.00000179

Summer MAE 0.4618 0.4556 0.4520 0.4508 0.4516 0.4540 0.4581 0.4644 0.4724 0.00005164

RMSE 0.6031 0.5957 0.5914 0.5902 0.5920 0.5970 0.6049 0.6157 0.6293 0.00016984

Coverage 0.9725 0.9725 0.9725 0.9725 0.9725 0.9725 0.9725 0.9725 0.9725 0

F-Measure 0.9067 0.9077 0.9084 0.9085 0.9083 0.9076 0.9064 0.9049 0.9029 0.00000352

Autumn MAE 0.4547 0.4466 0.4405 0.4368 0.4365 0.4390 0.4448 0.4525 0.4620 0.00007873

RMSE 0.5950 0.5831 0.5749 0.5706 0.5704 0.5742 0.5818 0.5933 0.6083 0.00016811

Coverage 0.9772 0.9772 0.9772 0.9772 0.9772 0.9772 0.9772 0.9772 0.9772 0

F-Measure 0.9099 0.9116 0.9128 0.9134 0.9134 0.9129 0.9118 0.9101 0.9080 0.00000347

Winter MAE 0.4347 0.4278 0.4245 0.4244 0.4271 0.4320 0.4393 0.4490 0.4607 0.00015234

RMSE 0.5749 0.5657 0.5603 0.5587 0.5609 0.5670 0.5767 0.5899 0.6064 0.00025232

Coverage 0.9830 0.9830 0.9830 0.9830 0.9830 0.9830 0.9830 0.9830 0.9830 0

F-Measure 0.9153 0.9166 0.9174 0.9176 0.9173 0.9164 0.9150 0.9131 0.9108 0.00000518

Overall MAE 0.4507 0.4433 0.4388 0.4369 0.4377 0.4407 0.4460 0.4534 0.4627 0.00007430

RMSE 0.5927 0.5829 0.5764 0.5734 0.5739 0.5780 0.5854 0.5961 0.6099 0.00014783

Coverage 0.9758 0.9758 0.9758 0.9758 0.9758 0.9758 0.9758 0.9758 0.9758 0

F-Measure 0.9096 0.9110 0.9120 0.9124 0.9123 0.9117 0.9107 0.9091 0.9072 0.00000304

Season	Metric	β	Variance
Spring	MAE	0.4516	0.4433	0.4380	0.4354	0.4354	0.4376	0.4417	0.4478	0.45 5	0.00005308
	RMSE	0.5976	0.5869	0.5791	0.5742	0.5724	0.5736	0.5780	0.5853	0.5955	0.00008827
	Coverage	0.9705	0.9705	0.9705	0.9705	0.9705	0.9705	0.9705	0.9705	0.9705	0
	F-Measure	0.9066	0.9081	0.9092	0.9099	0.9102	0.9100	0.9094	0.9083	0.9069	0.00000179
Summer	MAE	0.4618	0.4556	0.4520	0.4508	0.4516	0.4540	0.4581	0.4644	0.4724	0.00005164
	RMSE	0.6031	0.5957	0.5914	0.5902	0.5920	0.5970	0.6049	0.6157	0.6293	0.00016984
	Coverage	0.9725	0.9725	0.9725	0.9725	0.9725	0.9725	0.9725	0.9725	0.9725	0
	F-Measure	0.9067	0.9077	0.9084	0.9085	0.9083	0.9076	0.9064	0.9049	0.9029	0.00000352
Autumn	MAE	0.4547	0.4466	0.4405	0.4368	0.4365	0.4390	0.4448	0.4525	0.4620	0.00007873
	RMSE	0.5950	0.5831	0.5749	0.5706	0.5704	0.5742	0.5818	0.5933	0.6083	0.00016811
	Coverage	0.9772	0.9772	0.9772	0.9772	0.9772	0.9772	0.9772	0.9772	0.9772	0
	F-Measure	0.9099	0.9116	0.9128	0.9134	0.9134	0.9129	0.9118	0.9101	0.9080	0.00000347
Winter	MAE	0.4347	0.4278	0.4245	0.4244	0.4271	0.4320	0.4393	0.4490	0.4607	0.00015234
	RMSE	0.5749	0.5657	0.5603	0.5587	0.5609	0.5670	0.5767	0.5899	0.6064	0.00025232
	Coverage	0.9830	0.9830	0.9830	0.9830	0.9830	0.9830	0.9830	0.9830	0.9830	0
	F-Measure	0.9153	0.9166	0.9174	0.9176	0.9173	0.9164	0.9150	0.9131	0.9108	0.00000518
Overall	MAE	0.4507	0.4433	0.4388	0.4369	0.4377	0.4407	0.4460	0.4534	0.4627	0.00007430
	RMSE	0.5927	0.5829	0.5764	0.5734	0.5739	0.5780	0.5854	0.5961	0.6099	0.00014783
	Coverage	0.9758	0.9758	0.9758	0.9758	0.9758	0.9758	0.9758	0.9758	0.9758	0
	F-Measure	0.9096	0.9110	0.9120	0.9124	0.9123	0.9117	0.9107	0.9091	0.9072	0.00000304

The recommender systems recommend attractions to tourists only when the predicted rating is equal to or greater than a predefined level named as recommendation threshold, which is denoted by δ. δ will influence the recommendation performance and the higher δ indicates the more strict recommendation condition. In this experiment, δ ∈ {3, 4, 5}. Figure 2 depicts the overall recommendation performance of SAR-DTR with different β under each δ. As we can see in Fig. 2, when β = 0.4, the SAR-DTR generally has the highest RHR and the lowest RLR under every δ.

Fig.2

The overall recommendation performance of SAR-DTR with different β.

According to the analysis above, the proposed SAR-DTR exhibits the best overall prediction performance and recommendation performance when β = 0.4. We will subsequently show the performance comparison between SAR-DTR method and other benchmark methods under β = 0.4. Table 4 compares the prediction performances of different recommendation methods. From Table 4, we can draw several conclusions as follows.

Table 4

Prediction performance comparison

Method	Season	MAE	RMSE	Coverage	F-Measure
SAR-DTR	Spring	0.4354	0.5742	0.9705	0.9099
	Summer	0.4508	0.5902	0.9725	0.9085
	Autumn	0.4368	0.5706	0.9772	0.9134
	Winter	0.4244	0.5587	0.9830	0.9176
	Overall	0.4369	0.5734	0.9758	0.9124
STP	Spring	0.5074	0.6377	0.6120	0.7083
	Summer	0.5431	0.6992	0.5971	0.6929
	Autumn	0.5240	0.6644	0.6852	0.7523
	Winter	0.5348	0.6802	0.7204	0.7713
	Overall	0.5273	0.6704	0.6537	0.7312
DT		0.4603	0.6043	0.9745	0.9074
ICF		0.6208	0.7815	0.8379	0.8209
UCF		0.6044	0.7654	0.8379	0.8230

Firstly, under the condition that both SAR-DTR and STP account for the seasonal context, SAR-DTR substantially outperforms STP in each season. In terms of overall performance, the MAE and RMSE of SAR-DTR are separately 0.0904 and 0.097 lower than those of STP. Meanwhile, SAR-DTR outperforms STP by 49.27% with respect to Coverage and its F-measure is 24.78% higher than STP as well. The performance advantages of SAR-DTR over STP verify that the consideration of trust relations in attraction recommendation can significantly improve the prediction accuracy and coverage.

Secondly, in each season, SAR-DTR precedes DT, in terms of MAE, RMSE and F-Measure, under the condition that both methods consider trust relations. As for the overall performance, the MAE and RMSE of SAR-DTR are separately 5.08% and 5.11% lower than those of DT. In the meantime, compared with DT, the SAR-DTR has 0.55% higher F-Measure. Although the Coverage of SAR-DTR in spring and summer are slightly lower than those of DT, its overall performance is superior to DT. The experimental results confirm that the consideration of seasonality can increase prediction accuracy without reducing coverage.

Thirdly, compared with the two traditional methods, the performance improvement of SAR-DTR is more significant. With regard to the overall performance comparison between SAR-DTR and ICF, the MAE and RMSE of SAR-DTR are respectively 29.62% and 26.63% lower, while its Coverage and F-Measure are separately 0.1379 and 0.0915 higher. In contrast with UCF, the SAR-DTR has 27.71% lower MAE and 25.08% lower RMSE, meanwhile, the performances of SAR-DTR on Coverage and F-Measure are separately 0.1379 and 0.0894 higher.

Table 5 compares the recommendation performance of proposed SAR-DTR method with those of other methods under each recommendation threshold δ, in terms of RHR and RLR. We can observe that compared with other four benchmark methods, the SAR-DTR has the highest RHR and lowest RLR. That is, the propose SAR-DTR can recommend more satisfactory attractions to tourists and lose less business opportunities. Concretely speaking, comparing SAR-DTR with STP, we can observe that under each δ, SAR-DTR significantly outperforms STP with respect to both RHR and RLR. Table 5 shows that SAR-DTR precedes STP in each season. Besides, averaged across the four seasons, the RHRs and RLRs of SAR-DTR also outperform those of STP under each δ. Even under the strict condition, namely δ = 5, the RHR of SAT-DTR reaches 0.8587, which indicates that when the δ is set to 5, on average, 85.87% attractions recommended by SAR-DTR are also scored highly by tourists. In comparison, the RHR of STP is 0.7442, 13.33% lower than that of SAR-DTR. Meanwhile, the RLR of SAR-DTR under δ = 5 is 0.0654 lower than that of STP. In other word, 100 attractions recommended by SAR-DTR can lose 6.54 less business opportunities than that recommended by STP. As for the performance comparison between SAR-DTR and DT, the overall performances of SAR-DTR are better than those of DT, despite that in spring and autumn, the RHR of SAR-DTR is lower than that of DT. Given δ = 4 as an example. SAR-DTR outperforms DT on RHR by 1.77%, meanwhile, the RLR of SAR-DTR is 5.42% lower than that of DT. Compared with ICF and UCF, the performance advancement of SAR-DTR on recommendation performance is more significant. In each season, SAR-DTR is obviously superior to ICF and UCF under each δ, in terms of RHR and RLR. Averaged RHRs and RLRs across four seasons, the overall performances of SAR-DTR substantially precede ICF and UCF as well. For example, under δ = 4, the SAR-DTR outperforms ICF by 14.19% and UCF by 10.34% with respect to RHR. In the meantime, the RLR of SAR-DTR is 22.96% lower than that of ICF and 20.79% lower than that of UCF.

Table 5

Recommendation performance comparison

Method	Season	δ = 3		δ = 4		δ = 5
		RHR	RLR	RHR	RLR	RHR	RLR
SAR-DTR	Spring	1	0	0.9917	0.7355	0.8462	0.3367
	Summer	1	0	0.9831	0.6441	0.8774	0.3324
	Autumn	1	0	0.9877	0.7398	0.8077	0.3304
	Winter	1	0	0.9904	0.7363	0.9036	0.3308
	Overall	1	0	0.9882	0.7139	0.8587	0.3326
STP	Spring	1	0.9783	0.9738	0.8163	0.7946	0.3933
	Summer	1	1	0.9717	0.7265	0.7044	0.4050
	Autumn	1	1	0.9779	0.7994	0.7258	0.3893
	Winter	1	0.9844	0.9645	0.8425	0.7521	0.4044
	Overall	1	0.9907	0.9720	0.7962	0.7442	0.3980
DT		0.9998	1	0.9710	0.7548	0.8571	0.3360
ICF		0.9999	0.9976	0.8654	0.9267	0.2875	0.3689
UCF		0.9999	0.9976	0.8956	0.9013	0.3552	0.3664

Because tourism activities are of low frequency, attraction recommendation may suffer severe cold-start problem. We specially compare the recommendation results of cold-start tourists. In this paper, the tourists with less than five visit records are identified as the cold-start tourists [31]. Note that the cold-start tourists here are not referred to the isolated tourists in trust network. Table 6 shows the overall performance comparison for cold-start tourists. We can observe that SAR-DTR performs well in dealing with cold-start problem. As shown in Table 6, although SAR-DTR has 1.04% lower Coverage than DT, it outperforms STP that ignores trust relations and DT that overlooks seasonal topic preferences, in terms of all the other metrics. Compared with ICF and UCF, the performance advantages of SAR-DTR on cold-start problem are more outstanding.

Table 6

Comparison results for cold-start tourists

	Prediction performance				Recommendation performance
	MAE	RMSE	Coverage	F-Measure	δ = 3		δ = 4		δ = 5
					RHR	RLR	RHR	RLR	RHR	RLR
SAR-DTR	0.4434	0.5771	0.8493	0.8521	1	0	0.9792	0.6476	0.8500	0.2124
STP	0.6620	0.8116	0.1687	0.2751	1	0	0.9550	0.6609	0.4792	0.3394
DT	0.4563	0.5929	0.8582	0.8550	1	0	0.9649	0.6788	0.7500	0.2269
ICF	0.8018	0.9917	0.4253	0.5433	1	1	0.7444	0.9521	0.1111	0.2540
UCF	0.8724	1.0677	0.4253	0.5383	1	1	0.7681	0.9362	0.1395	0.2585

6 Conclusion

With the objective to improve the performance of attraction recommendation method by advancing the inference of both tourist preferences and trust relations, we proposed a season-aware attraction recommendation method with dual-trust enhancement (SAR-DTR). It contributes to the literature on four aspects.

We proposed the inference method of seasonal topic preferences, which can capture tourists’ topic preferences in a specific season from their travel histories.

We constructed the dual-trust relationship (DTR) model that infers trust relations based on similarity-based trust and familiarity-based trust, in contrast to existing studies that merely focus on a single type of trust.

We proposed the season-aware attraction recommendation method (SAR-DTR) that recommends attractions according to the seasonal context, seasonal topic preferences and DTR trust relations.

The empirical study was conducted with real-world data to compare the performance of SAR-DTR with benchmark methods. The experimental results demonstrate that the SAR-DTR outperforms the other methods in terms of both prediction performance and recommendation performance and that the comprehensive consideration of seasonal topic preferences and trust relations contributes to the performance improvement of SAR-DTR. The results also demonstrate the performance advantages of SAR-DTR over the benchmark methods with respect to the cold-start tourists.

There are some directions for further research. First, the simulation study will be considered with the objective to gain deep insight into the formulated problem. Second, we assume that three types of social interactions are equally important in Equation (4). The computing methodology for familiarity can be further advanced by exploring their relative weights.

Footnotes

World Travel & Tourism Council.

Acknowledgments

This research is supported by the National Natural Science Foundation of China under Grant Nos. 71671038 and 41571133.

References

Yera

and Martínez

, Fuzzy tools in recommender systems: A survey, International Journal of Computational Intelligence Systems10(1) (2017), 776–803.

Huang

and Bian

, A Bayesian network and analytic hierarchy process based personalized recommendations for tourist attractions over the Internet, Expert Systems with Applications36(1) (2009), 933–943.

Ahas

, Aasa

, Mark

Ü.

, Pae

and Kull

, Seasonal tourism spaces in Estonia: Case study with mobile positioning data, Tourism management28(3) (2007), 898–910.

Avesani

, Massa

and Tiella

, A trust-enhanced recommender system application: Moleskiing, Proceedings of the 2005 ACMsymposium on Applied computing, 2005, pp. 1589–1593.

García-Crespo

, Chamizo

, Rivera

and Mencke

, Colomo-Palacios

and Gómez-Berbś

J.M.

, SPETA: Socialpervasive e-Tourism advisor, Telematics and Informatics26(3) (2009), 306–315.

Hinze

and Quan

, Trust-and location-based recommendations for tourism, On the Move to Meaningful Internet Systems: OTM2009, SpringerBerlin Heidelberg, 2009, pp. 414–422.

Qiao

, Li

, Su

and Cao

, A context-awareness dynamic friend recommendation approach for mobile social network users, International Journal of Advanced Intelligence3(2) (2011), 155–172.

Llodrá-Riera

and Martínez-Ruiz

M.P.

, Jiménez-Zarco

A.I.

and Izquierdo-Yusta

, A multidimensional analysis of the information sources construct and its relevance for destination image formation, Tourism management48(2015), 319–328.

Hanani

, Shapira

and Shoval

, Information filtering: Overview of issues, research and systems, User Modeling and User-Adapted Interaction11(3) (2001), 203–259.

10.

Alton-Scheidl

, Schmutzer

, Sint

P.P.

and Tschertru

, Votingand rating in Web4Groups, 1997.

11.

Adomavicius

and Tuzhilin

, Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions, IEEE Transactions on Knowledge and Data Engineering17(6) (2005), 734–749.

12.

Garcia

, Sebastia

and Onaindia

, On the design of individual and group recommender systems for tourism, Expert Systems with Applications38 (2011), 7683–7692.

13.

Vansteenwegen

, Souffriau

, Berghe

C.V.

and Oudheusden

D.V.

, The city trip planner: An expert system for tourists, Expert Systems with Aplications38(6) (2011), 6540–6546.

14.

Jiang

, Qian

, Shen

and Mei

, Travel recommendation via author topic model based collaborative filtering, International Conference on Multimedia Modeling, Springer International Publishing, 2015, pp. 392–402.

15.

Z.X.

, Chen

and Chen

, Topic based context-aware travel recommendation method exploiting geotagged photos, Neurocomputing155 (2015), 99–107.

16.

Richards

, Tourism attraction systems: Exploring cultural beavihor, Annals of Tourism Research29(4) (2002), 1048–1064.

17.

Lesani

and Montazeri

, Fuzzy trust aggregation and personalized trust inference in virtual social networks, Computational Intelligence25(2) (2009), 51–83.

18.

Sharma

and Bedi

, CCFRS – Community based Collaborative Filtering Recommender System, Journal of Intelligent & Fuzzy Systems32 (2017), 2987–2995.

19.

Ziegler

C.N.

and Golbeck

, Investigating interactions of trust and interest similarity, Decision Support Systems43(2007), 460–475.

20.

Golbeck

, Trust and nuanced profile similarity in online social networks, ACM Transactions on the Web3(4) (2009), 12:1–12:33.

21.

Fernandez-Gago

, Agudo

and Lopez

, Building trust from context similarity measures, Computer Standards & Interfaces36 (2014), 792–800.

22.

Zhang

, Ghorbani

A.A.

and Cohen

, A familiarity-based trust model for effective selection of sellers in multiagent ecommerce systems, International Journal of Information Security6(5) (2007), 333–344.

23.

Hao

, Min

, Lin

, Luo

and Yang

L.T.

, Mobi-Fuzzy Trust: An efficient fuzzy trust inference mechanism in mobile social networks, IEEE Transactions on Parallel and Distributed Systems25(11) (2014), 2944–2955.

24.

Smith

E.R.

, Miller

D.A.

, Maitner

A.T.

, Crump

S.A.

, Garcia-Marques

and Mackie

D.M.

, Familiarity can increase stereotyping, Journal of Experimental Social Psychology42(4) (2006), 471–478.

25.

Gefen

, Ecommerce: The role of familiarity and trust, Omega28(5) (2000), 725–737.

26.

María Munar

, Tourist-created content: Rethinking destination branding, International Journal of Culture, Tourism and Hospitality Research5(3) (2011), 291–305.

27.

De Meo

, Nocera

, Terracina

and Ursino

, Recommendation of similar users, resources and social networks in a Social Internetworking Scenario, Information Sciences181(7) (2011), 1285–1305.

28.

Kim

Y.A.

, An enhanced trust propagation approach with expertise and homophily-based trust networks, Knowledge-Based Systems82 (2015), 20–28.

29.

Tang

, Hu

and Liu

, Social recommendation: A review, Social Network Analysis and Mining3(4) (2013), 1113–1133.

30.

Christakis

N.A.

and Fowler

J.H.

, Connected: The surprisingpower of our social networks and how they shape our lives, Little, Brown, 2009.

31.

Deng

, Huang

and Xu

, Social network-based service recommendation with trust enhancement, Expert Systems with Applications41(18) (2014), 8075–8084.

32.

X.Y.

and Khoshgoftaar

T.M.

, A survey of collaborative filtering techniques, Advances in Artificial Intelligence, 2009, pp. 1–19.

33.

Nilashi

, Ibrahim

O.B.

and Ithnin

, Hybrid recommendation approaches for multi-criteria collaborative filtering, Expert Systems with Applications41(8) (2014), 3879–3900.

34.

Vashistha

, Khuranab

and Bedi

, A fuzzy hybrid recommender system, Journal of Intelligent & Fuzzy Systems32(2017), 3945–3960.

35.

Jamali

and Ester

, Trustwalker: A random walk model for combining trust-based and item-based recommendation, Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2009, pp. 397–406.

36.

Y.M.

, Wu

C.T.

and Lai

C.Y.

, A social recommender mechanism for e-commerce: Combining similarity, trust, and relationship, Decision Support Systems55(3) (2013), 740–752.

Tourists	6,676
Attractions	10,827
Ratings	174,112
Topics	147
Following others	394,461
Visiting homepages	3,582
Commenting on the content	55,505
Social interactions	453,548
Average ratings per tourists	26.08
Average social interactions per tourist	67.94
Average ratings per attraction	16.08