Social platform based interval valued intuitionistic fuzzy location recommendation system

Abstract

The improvements in mobile technologies led to the wide adaptation and triggered the demand for location based services. In this respect, examining user similarities enable the analysis of user interests in terms of the determination of purchasing preferences and actual needs. User similarities are generally extracted from consumer life style, demographical information or the reflections from previously sent messages. In spite of the fact that these factors may not directly influence the purchasing decision, uncertain or lack of information can be encountered while establishing recommendation systems. Thus, researchers try to search other indicators that can reflect customer characteristics from spatial data, digital contribution in social media and search history for preferable representation of the changes in purchasing tendency. In this study, social platform based interval valued intuitionistic fuzzy location recommendation system is proposed by considering three common social platforms: Trip Advisor, Zomato and Foursquare. To perform restaurant offers to appropriate social platform users, a sentiment analysis is conducted to selected restaurants and number of negative, positive and neutral comments are extracted. After that, restaurant and location information are examined by using user, restaurant and location clustering via fuzzy clustering. Finally, intuitionistic fuzzy similarity matrix based collaborative filtering is used for restaurant offers to similar users.

Keywords

Location based systems recommendation systems interval valued intuitionistic fuzzy sets

1 Introduction

Because of increasing demand on mobile technologies and applications, location based systems can present customers’ special interests in a particular location and time [33]. More generally, location based systems (LBS) ensure content providers to send convenient messages to application users considering their visiting experience in a specific location [16]. The aim of these services is tracking and navigating to transfer time based push messages.

As a result of increasing penetration rate of mobile advertising, targeted advertising, which allows sending personalized information (including personalized price offers) to customers, has drastically affected from LBSs with regard to companies’ marketing and communication strategies. As a consequence of this situation, LBSs must reflect consumer interests more sensitively for implementing targeted advertising systems.

During the recent years, location identification adapted recommender systems in targeted advertisement is conducted a substantial base which enhance companies’ transmission channels to customers with limited advertisement budgets [36]. Thus, LBSs are relied on time and preference based services for tracing and tracking the current position of application users. The main aim is the prediction of future visits by considering previous places [10]. On the other hand, users are commonly irritated from instant push up messages because of unconcerned context and irrelevant price offers [41]. Additionally, privacy concerns that users are not tend to share their current locations, are another aspect for not to be interested in these push up messages. These concerns cause the low penetration of LBSs [30]. Furthermore, user relevancy and customer needs sustainably vary when their current and future locations are analyzed. Therefore, updating contexts to the recent ones enables generally-accepted and commonly used recommender systems. For this reason, companies are tend to find out more effective methodologies that ensure both individualized concepts and commonly preferable alternatives. To cope with digressiveness problems, previously visiting daat is required. Hence, potential location identification based recommendation systems can be evaluated as the most crucial point to determine the user purchasing probability when considering future visits [33].

In addition to all these, social platforms are another part that firms are utilized for both social relation data and current positions of the users for better understanding of customer needs and expectations. With the development of social networking, it is extremely beneficial to recommend advertisements based on social information of users. Social platform based recommender systems utilize the social relation data, such as friendships and current locations of the users to improve recommendation results. The main contribution of using social platforms for LBSs is that personal offers for customers can be conducted considering their user data and comments and also, shared location information. These circumstances enhance recommender system sensitivity for personalized location offers [22].

Although there are crucial advantages of using social platforms location recommendation systems, vague information can be appeared due to the lack of user shares and profile specification. For that reason, fuzzy set theory provides coping with this imprecise information in an efficient manner with the help of linguistic variables. Human reasoning works better on the range of the linguistic values as it better captures the vagueness. Taking this as an advantage, the recommender system can use fuzzy information to interpret user locations, social platform data (such as current location, previously visited locations or comments), and similarity of two users more efficiently [23].

From this point of view, this study focuses on performing an intuitionistic fuzzy set theory based recommendation system by combining the data gathered from three commonly used social platforms: Zomato, Trip Advisor and Foursquare. To perform restaurant offers to appropriate social platform users, a sentiment analysis is conducted to selected restaurants and number of negative, positive and neutral comments are extracted. After that, restaurant and location information are practiced by adapting user, restaurant and location clustering via fuzzy clustering. Finally, intuitionistic fuzzy similarity matrix based collaborative filtering is performed for restaurant offers to similar users.

2 Literature review

This study relies on the combination of sentiment analysis and location based systems. Thus, a brief literature review is presented to reveal former studies and applications and highlight the points that have been ignored in the literature.

2.1 Sentiment analysis

Sentiment Analysis is the process of the classification of opinions about a particular topic from a written text which contains positive, negative or neutral comments by well-developed lexicons. The data is mostly taken from documents, sentence, entity or aspect. The wide application of sentiment analysis can be realized both in the studies from academicians and practitioners ([26]). For instance, companies are more prone to understand people’s perceptions and consumer reviews on products in order to enhance their services ([9]). Additionally, purchase decision on marketing ([3, 12]) and tourism management ([1]) are two of the most studied topics for sentiment analysis to explore different opinion groups.

From literature review, sentiment analysis is mainly adapted in terms of applying sentiments to the documents or focusing on seeking sentiments of words, subjective explanations, subjective clauses and topics stored in web sites or databases [29]. Besides that, sentiment analysis is studied on the purpose of identifying opinions for different business intelligence areas [18]. Former studies mostly cope with customer segmentation ([38, 40]), social text summarization ([43]), social and conventional media ([44]), social reviews ([14]) cinema audience evaluation via Internet Movie Database (IMDb) ([24]), e-learning ([4]) stock market predictions and financial evaluations ([8]), market evaluations considering social networks ([22]) and early identification of emerging political topics on Twitter ([35]) etc.

Sentiment analysis has generally suffered from three main issues: aspect detection, opinion word detection and sentiment orientation identification [45]. Aspect is the major indicator for the determination of the topic of an opinion [19]. Review ratings and overall ratings can be stated from numerical values. On the other hand, text comments should be analyzed for the individual assessment of the opinions which could be difficult to extract from the document directly. For example, a comment can include positive idea about services but it can also implies negative idea related to the menu. Thus, aspect level based sentiment analysis can be useful for understanding the customer satisfaction level according to different co-subjects [5].

In addition to the advantageous proceeds of sentiment analysis, differentiation of the comments by clustering methods decreases the need of dealing with huge amount of training data with a high level of accuracy ([25, 34]). Clustering methods appeared in the literature were generally practiced with hierarchical methods, compartmentalize methods, density based methods that search sets appeared in low densities and plotting raster based approaches that decrease the computation time ([2, 32]). Note that the papers which used clustering and sentiment analysis together generally concentrate on implementing different types of methods in the first stage of sentiment analysis. However, if clustering is applied after the sentiment analysis step, tweets could be evaluated with more precise classes [29].

In our study, first the corpus which is a large and structured set of texts that are prepared for further analysis, is gathered by text cleaning process. For assigning a polarity (positive or negative) to each word and query, both the synonyms and antonym should be processed with automatically or manually detected seed words, After that, transforming the unstructured data to term-document matrix (TDM) is conducted for displaying the corpus. The TDM is described with a matrix form in which the documents or sources can be presented with the rows and terms represented with the columns. After TDM is constituted, the proceeding stage is to determine novel and useful patterns from this matrix. For describing the themes of comments, classification and clustering methods are two of the most useful methods to clarify the number of positive, negative and neutral comments as adapted in this study [42].

As mentioned before, social media analysis has been studied mostly for the further analysis of sharing comments and giving information via Internet as a straightforward way of understanding customer needs. Additionally, most of the studies emphasis on gathering essential information from social media using data and text mining approaches. Although it is a well-adapted method among researches and great number of applications were appeared in the literature, the gap in analyzing the outcomes of these methods has not been filled [39]. Thus, this study proposes making convenient outcomes from social media reviews in order to understand the needs of customers and extract business value.

2.2 Location based recommendation systems

Location-based systems can be defined as the overall services that use the spatial location of the user to enhance a special marketing exhorter [28]. In other words, location-based systems gather real time location data of consumers, establish mobile connection to other mobile devices and push a relevant content when they appeared in appropriate field. This enables the improvements in location-aware functions as realized in shopping malls [7]. The main contribution of these systems is extracting knowledge from where the services are settled. The knowledge can be gathered from geographic locations of users who downloaded the application-related tool in their mobile phones. In addition to that, sensor data captures interactions between users and locations via Bluetooth technology. Thus, in order to increase the brand awareness, large-scale retail companies began to utilize LBSs for their retail activities [47]. Other specific applications can be summarized as location-based advertising and tourist route recommender systems [20].

From the reflections of literature review, LBSs have been widely adapted in indoor positioning applications such as location-based advertising and location-based mobile advertising, mobile shopping ([48]) recognition of previously visited places for making further predictions ([37]), and social media based recommender systems [17]. As seen from these studies, location-based applications facilitate the tracking and traceability of individuals’ physical moving in indoor fields. Thus, positioning and segmentation algorithms play critical roles in interpreting individuals’ personal positions and visiting tendencies to make suggestions for location-specified promotions.

Although the advantages of location-based systems are obvious, some drawbacks are discussed when practicing. The problems that users realized are mainly consist of privacy issues and concerns of the messages: customers do not tend to share their location data via Bluetooth and generally, do not want to get push up messages. Besides privacy issues, context of the promotion message should be related to the users interests [6]. For instance, location based recommendation system is generally requires the data gathering from diversified social platforms for better construction of relationship between customers, locations (shopping malls) and their relationships with each other [11]. According to the mentioned situations above, motivation of this study is location based recommendation system configuration by considering previously visited location data and users. Also the identification of alternative locations with respect to the similarities are assigned for providing wide range of location based alternatives. Thus, location clustering can be adapted for constructing the relativeness of the restaurants and alternatives by considering the similarities of location clusters and restaurant groups.

3 Interval valued intuitionistic fuzzy sets and interval valued intuitionistic fuzzy location based system methodology

In this part, a brief background of interval valued intuitionistic fuzzy set theory is given and proposed methodology is explained with theoretical details for better understanding.

3.1 Preliminaries for interval valued intuitionistic fuzzy set theory

Intuitionistic fuzzy sets (IFSs) are defined using membership and non-membership degrees in order to present the imprecise information. The interval-valued intuitionistic fuzzy set (IVIFS) is a modified form of IFSs that contains T[0,1] as a subinterval of interval [0,1] and assume that X be a set as a non-empty set. An interval valued intuitionistic fuzzy set in X is identified as $\tilde{S} = {〈 x, μ_{\bar{s}} (x), v_{\bar{s}} (x) | x \in X 〉}$ and $μ_{\bar{s}} : X \to T [0, 1]$ is the membership degree and $v_{\bar{s}} : X \to T [0, 1]$ is the non-membership degree for $0 ⩽ μ_{\bar{s}} (x) + v_{\bar{s}} (x) ⩽ 1$ . Note that, $μ_{\bar{s}} (x)$ and $v_{\bar{s}} (x)$ have closed intervals with lower and upper limits. $μ_{\bar{s}} (x)$ can be represented using $μ_{\bar{s}} (x)^{-}$ and $μ_{\bar{s}} (x)^{+}$ . Additionally $v_{\bar{s}} (x)$ can be stated with $v_{\bar{s}} (x)^{-}$ and $v_{\bar{s}} (x)^{+}$ considering $0 ⩽ μ_{\bar{s}} (x)^{+} + v_{\bar{s}} (x)^{+} ⩽ 1$ and $μ_{\bar{s}} (x)^{-} + v_{\bar{s}} (x)^{-} ⩽ 0$ . In terms of closed interval limits, algebraic operations of interval valued intuitionistic fuzzy sets (IVIFSs) are given for two interval valued intuitionistic fuzzy numbers (IVIFNs) considering ${\tilde{S}}_{1} = ([a_{1}, b_{1}], [c_{1}, d_{1}])$ and ${\tilde{S}}_{2} = ([a_{2}, b_{2}], [c_{2}, d_{2}])$ as below: $\begin{matrix} {\tilde{S}}_{1} + {\tilde{S}}_{2} & = & ([a_{1} + a_{2} - a_{1} a_{2}, b_{1} + b_{2} - b_{1} b_{2}], \\ [c_{1} c_{2}, d_{1} d_{2}]) \end{matrix}$ (1) $\begin{matrix} \tilde{δ S_{1}} & = & ([1 - (1 - a_{1})^{δ}, 1 - (1 - b_{1})^{δ}], [c_{1}^{δ}, d_{1}^{δ}]) \end{matrix}$ (2) where δ ⩾ 0. Score and accuracy functions are calculated for an IVIFN Š = ([a, b] , [c, d]) as follows: $s (Š) = \frac{1}{2} (a + b - c - d)$ (3) $h (Š) = \frac{1}{2} (a + b + c + d)$ (4) where s ( Š ) , h ( Š ) ∈ [0, 1]. If s ( Š ₁) < s ( Š ₂), then Š ₁ < Š ₂. If s ( Š ₁) = s ( Š ₂), accuracy function is considered. Note that if h ( Š ₁) < h ( Š ₂), in that case Š ₁ < Š ₂and additionally if s ( Š ₁) = s ( Š ₂), so that h ( Š ₁) < h ( Š ₂) and we can say that Š ₁ = Š ₂ [49]

To calculate the distance measure of Š ₁ and Š ₂, the following formula is adapted:

$\begin{matrix} d ( & Š_{1}, & Š_{2}) & = & \frac{1}{4} (| a_{1} - a_{2} | + | b_{1} - b_{2} | \\ + | c_{1} - c_{2} | + | d_{1} - d_{2} |) \end{matrix}$ (5)

For 0 ⩽ d ( Š ₁, Š ₂) ⩽1.

Assume that $X = ({\tilde{x}}_{ij})_{mxn}$ be a matrix and suppose that all ${\tilde{x}}_{ij}$ s are IVIFNs for i = 1, 2, . . . , m and j = 1, 2, . . . , n considering b_ij + d_ij ⩽ 1. This matrix is evaluated as interval valued intuitionistic fuzzy matrix (IVIFM). The distance measure between X₁ and X₂ is identified in the following way:

$\begin{matrix} d (X_{1}, X_{2}) & = & \frac{1}{4 mn} \sum_{i = 1}^{m} \sum_{j = 1}^{n} (| a_{ij}^{(1)} - a_{ij}^{(2)} | \\ + | b_{ij}^{(1)} - b_{ij}^{(2)} | + | c_{ij}^{(2)} - c_{ij}^{(2)} | \\ + | d_{ij}^{(1)} - d_{ij}^{(2)} |) \end{matrix}$ (6) such that 0 ⩽ d (X₁, X₂) ⩽1.

IVIFS based similarity measure for Š ₁, Š ₂ is given in the following: $sim (Š_{1}, Š_{2}) = \frac{d (Š_{1}, Š_{2}^{c})}{d (Š_{1}, Š_{2}) + d (Š_{1}, Š_{2}^{c})}$ (7) where $Š_{2}^{c} = ([c_{2}, d_{2}], [a_{2}, b_{2}])$ considering $0 ⩽ sim (\tilde{S_{1}}, \tilde{S_{2}}) ⩽ 1$ .Particularly, the similarity measure can be transformed to IVIFM form with ${\tilde{x}}_{ij}^{(1)}$ and ${\tilde{x}}_{ij}^{(2)}$ as seen in Equation (8): $\begin{matrix} sim (X_{1}, X_{2}) \\ = \frac{\sum_{i = 1}^{m} \sum_{j = 1}^{n} d ({\tilde{x}}_{ij}^{(1)}, ({\tilde{x}}_{ij}^{(2)^{c}}))}{\sum_{i = 1}^{m} \sum_{j = 1}^{n} d ({\tilde{x}}_{ij}^{(1)}, {\tilde{x}}_{ij}^{(2)}) + \sum_{i = 1}^{m} \sum_{j = 1}^{n} d ({\tilde{x}}_{ij}^{(1)}, ({\tilde{x}}_{ij}^{(2)^{c}}))} \end{matrix}$ (8)

Where, ${\tilde{x}}_{ij}^{(2)^{c}} = ([c_{ij}^{2}, d_{ij}^{2}], [a_{ij}^{2}, b_{ij}^{2}])$ using the constraint0 ⩽ sim (X₁, X₂) ⩽1 [46].

3.2 The proposed methodology

We propose a location based recommendation system based on the combination of sentiment analysis that determines the related location’s view according to the customers (opinion targeting) and clustering based on the similarity of users and locations. Figure 1 represents the structure of the proposed location based recommendation system which is comprised on four key components: Data collection and preprocessing, sentiment analysis, grouping users and locations and recommendation system structure. First of all, data is needed to be described in interval valued intuitionistic fuzzy number form in order to apply the recommended algorithm. This can be already achieved by considering the data gathered from diversified social media platforms and thus, data should be stated with intervals in order to reflect different aspects of ratings. Secondly, review texts and review titles should be analyzed as negative, positive and neutral comments and these comments should be named as “number of positive comments”, “number of negative comments” and “number of neutral comments” for each restaurant. Additionally, users and locations are grouped according to interval valued fuzzy c means clustering before the implementation of interval valued intuitionistic fuzzy recommendation system. The extracted groups are utilized for making proper location suggestions for different customer groups.

Fig. 1

The structure of the proposed location based for recommendation system restaurants.

3.2.1 Sentiment analysis

The sentiment analysis applied in this study is adapted from Turban et al. [13]’s study. The steps of the methodology can be summarized as follows:

Step 1-Establishing the corpus: Corpus is a large and structured set of texts that are prepared for further analysis. The aim of this phase is to gather all of the consumer reviews and comments from the related social platforms which are relevant to the problem. Quality and number of the documents considered in the corpus are two of the most substantial factor for the effectiveness of the separation. Note that, the corpus should be defined and collected using manual or automated techniques such as software programs that periodically collect data from diversified sources. The source of data may contain HTML files, emails, web blogs or textual documents. As these various types of documents are collected, they should be organized before the application of sentiment analysis. In this study, user reviews formed the corpus.

Step 2- Sentiment Analysis: As described from previous section, sentiment analysis includes the utilization of natural language processing, document analysis and computational linguistics to clarify and extract subjective information in source materials. The term opinion mining is also used instead of sensitivity analysis. The aim of sentiment analysis is to determine the position of a user considering a relevant topic or the determination of overall contextual polarity of a document. The attitude may be his or her judgment or the intended emotional communication.

Step 3- Setting up the Term Document Matrix: Term Document matrix (TDM) aims to transform the unstructured data into a structured representation of the corpus. The TDM is a matrix form in which the documents are represented by the rows and terms are stated by the columns. In Fig. 2, a sample TDM matrix is given.

Fig. 2

A sample TDM matrix.

The numerical values in the TDM matrix denote the relationship between the terms and the documents such as how frequently a term appears in a document with various measures. For converting unstructured data into a TDM matrix, the underlying assumption is that the meaning of the documents can be introduced within the matrix with the measures in the cells. On the other hand, all the terms are not equally important for representing a document due to the auxiliary verbs which should be excluded from the process. Note that the mentioned terms are called “stop terms” and they should be precisely determined before or during the TDM phase. Additionally, “include terms”, which explain the predefined terms that are chosen for establishing TDM, should be described in advance. Besides including terms, the “synonyms”, which the pairs of terms are refined at the same time, should be clarified. Finally, substantial phrases can be extracted after obtaining the entire class instead of separate words. Because the success of the sentiment analysis relies on the accuracy of TDM, the phase involves feedbacks that the analyzer can track the former tasks and renovates them for better outcomes.

Step 4- Specifying the themes: After building the TDM and dealing with the dimensionality problem, the next task is to extract novel and beneficial patterns from this matrix. In order to determine the themes of reviews, k-means clustering is utilized. From k means clustering adaptation, theme groups can be extracted for further recommendation system adaptation.

Step 5- Analysis of the results: The overall results of sentiment analysis are evaluated with defined themes to analyze the performance of each theme.

3.2.2 Fuzzy c means clustering

Among other clustering approaches, fuzzy c-means (FCM) is identified as the most commonly utilized clustering technique which is relied on the minimization of the deviation to cluster centers. Minimization of objective function can be denoted as a nonlinear optimization problem as follows: $J (Z, U, V) = \sum_{i = 1}^{c} \sum_{j = 1}^{N} (μ_{ij})^{m} {∥ z_{j} - v_{j} ∥}^{2}$ (9)

Note that Z shows the data set, U presents the fuzzy partition matrix, V denotes cluster centers’ vector. Additionally, N denotes the number of observations, c shows the number of proper clusters and μ is stated as the membership value. Additionally, m is fuzzifier that identifies the fuzziness degree. In the formula, z_j - v_i is called as the distance measure for observation j and cluster center i [31].

The phases of fuzzy c-means (FCM) algorithm are identified by Babuska [31]:

Establish U = [u_ij] matrix, U⁽⁰⁾

At t-step: Define the vectors of the centers V^(t) = [v_i] using U^(t) $v_{i} \frac{\sum_{i = 1}^{N} μ_{ij}^{m} \cdot z_{j}}{\sum_{i = 1}^{N} μ_{ij}^{m}}$

Update the values of U^(t), U^(t+1)

μ_{ij} = \frac{1}{\sum_{i = 1}^{c} {(\frac{∥ z_{j} - v_{i} ∥}{∥ z_{j} - v_{t} ∥})}^{\frac{2}{m - 1}}}

If ∥U^(t+1) - U^(t) ∥ < ɛ is the stopping criteria; then return to step 2.

3.2.3 Interval valued fuzzy c means clustering

The methodology of interval valued fuzzy c means is an extended form of fuzzy c means clustering. In this clustering paradigm, two values of the fuzzifier (m1 and m2) should be considered for upper and lower limits. Thus, two objective functions are needed for conducting the clustering process [15]. $\begin{matrix} J_{m_{1}} = \sum_{i = 1}^{c} \sum_{j = 1}^{n} (u_{ij})^{m_{1}} d_{ij}^{2} \\ J_{m_{2}} = \sum_{i = 1}^{c} \sum_{j = 1}^{n} (u_{ij})^{m_{2}} d_{ij}^{2} \end{matrix}$ (10) wheren denotes the number of observations and c is stated as the number of classes. The Euclidean distance is calculated as $d_{ij} = d (x_{j} - v_{i}) = ∥ x_{j} - v_{i} ∥ = {[\sum_{k = 1}^{h} (x_{jk} - v_{ij})^{2}]}^{1 / - 2}$ where m denotes the number of features and attributes while u_ij ∈ [0, 1] presents the membership degree of x_j according to the centroid v_i. After solving the minimization problem, upper and lower membership degrees are extracted as follows:

$\begin{matrix} {\bar{u}}_{ij} & = & {\begin{matrix} \frac{1}{\sum_{k = 1}^{c} {(\frac{d_{ij}}{d_{kj}})}^{\frac{2}{m_{2} - 1}}} if \frac{1}{\sum_{k = 1}^{c} (\frac{d_{ij}}{d_{kj}})} < \frac{1}{c} \\ \frac{1}{\sum_{k = 1}^{c} {(\frac{d_{ij}}{d_{kj}})}^{\frac{2}{m_{2} - 1}}} if \frac{1}{\sum_{k = 1}^{c} (\frac{d_{ij}}{d_{kj}})} ⩾ \frac{1}{c} \end{matrix} \\ \underline{u} ij & = & {\begin{matrix} \frac{1}{\sum_{k = 1}^{c} {(\frac{d_{ij}}{d_{kj}})}^{\frac{2}{m_{2} - 1}}} if \frac{1}{\sum_{k = 1}^{c} (\frac{d_{ij}}{d_{kj}})} ⩾ \frac{1}{c} \\ \frac{1}{\sum_{k = 1}^{c} {(\frac{d_{ij}}{d_{kj}})}^{\frac{2}{m_{2} - 1}}} if \frac{1}{\sum_{k = 1}^{c} (\frac{d_{ij}}{d_{kj}})} < \frac{1}{c} \end{matrix} \end{matrix}$ (11) where 1 ⩽ i ⩽ c and 1 ⩽ j ⩽ n. For the defuzzification, mean of upper and lower bounds are computed both for membership degrees and the centroids of the clusters [15].

3.2.4 Interval valued intuitionistic fuzzy recommendation system based on multi criteria

Definition 1. Multi Criteria Interval valued Intuitionistic Fuzzy Recommendation System (IVIFRS): Let P, S, D be users, locations and restaurants respectively. Each user P_i (∀ i ∈ {1, . . . , n}) and location S_j (∀ i ∈ {1, . . . , m}) are supposed to have the same features. Recommendation system matching for each user to a restaurant is relied on X and Y which include s interval valued intuitionistic linguistic labels. Hence, restaurants, D_j (∀ i ∈ {1, . . . , k}) will consist of s interval valued intuitionistic linguistic labels. The utility function, T is a mapping defined on (X, Y) written below [21]: $\begin{matrix} R : X \times Y \to D_{1} \times D_{2} \times . . . \times D_{k} \\ [μ_{1 X} (x)^{-}, μ_{1 X} (x)^{+}], [v_{1 X} (x)^{-}, v_{1 X} (x)^{+}] \\ 〈 [μ_{2 X} (x)^{-}, μ_{2 X} (x)^{+}], [v_{2 X} (x)^{-}, v_{2 X} (x)^{+}] 〉 \\ \dots \\ [μ_{sX} (x)^{-}, μ_{sX} (x)^{+}], [v_{sX} (x)^{-}, v_{sX} (x)^{+}] \\ [μ_{1 Y} (y)^{-}, μ_{1 Y} (y)^{+}], [v_{1 Y} (y)^{-}, v_{1 Y} (y)^{+}] \times \\ 〈 [μ_{2 Y} (y)^{-}, μ_{2 Y} (y)^{+}], [v_{2 Y} (y)^{-}, v_{2 Y} (y)^{+}] 〉 \\ \dots \\ [μ_{sY} (y)^{-}, μ_{sY} (y)^{+}], [v_{sY} (y)^{-}, v_{sY} (y)^{+}] \\ [μ_{1 D} (D_{1})^{-}, μ_{1 D} (D_{1})^{+}], [v_{1 D} (D_{1})^{-}, v_{1 D} (D_{1})^{+}] \times \\ 〈 [μ_{2 D} (D_{1})^{-}, μ_{2 D} (D_{1})^{+}], [v_{2 D} (D_{1})^{-}, v_{2 D} (D_{1})^{+}] 〉 \\ \dots \\ [μ_{sD} (D_{1})^{-}, μ_{sD} (D_{1})^{+}], [v_{sD} (D_{1})^{-}, v_{sD} (D_{1})^{+}] \times \dots \\ [μ_{1 D} (D_{k})^{-}, μ_{1 D} (D_{k})^{+}], [v_{1 D} (D_{k})^{-}, v_{1 D} (D_{k})^{+}] \\ \times 〈 [μ_{2 D} (D_{k})^{-}, μ_{2 D} (D_{k})^{+}], [v_{2 D} (D_{k})^{-}, v_{2 D} (D_{k})^{+}] 〉 \\ [μ_{sD} (D_{k})^{-}, μ_{sD} (D_{k})^{+}], [v_{sD} (D_{k})^{-}, v_{sD} (D_{k})^{+}] \end{matrix}$ (12)

Multi criteria IVIFRS includes the fundamental functions expressed in the following:

Prediction: Express the values of [μ_lD (D_i) ^-, μ_lD (D_i) ⁺] , [v_lD (D_l) ^-, v_lD (D_l) ⁺] , ∀l ∈ {1, . . . , s} , ∀ i ∈ {1, . . . , k} .

Recommendation: Find out i^* ∈ [1, s] to ensure

\begin{matrix} i^{*} & = & arg, \max_{i = \bar{1, s}} {\sum_{j = 1}^{k} w_{j} ([μ_{iD} (D_{j})^{-}, μ_{iD} (D_{j})^{+}], \\ [v_{iD} (D_{j})^{-}, v_{iD} (D_{j})^{+}])} \end{matrix}

where w_j ∈ [0, 1] is the weight of D_j considering

\sum_{j = 1}^{k} w_{j} = 1

[21].

Considering Equation (8), assume that X₁ and X₂ be two IVIFMs for multi criteria based IVIFRSs. The interval valued intuitionistic fuzzy similarity degree (IVIFSD) is stated in the following way:

$\begin{matrix} SIM (X_{1}, X_{2}) & = & α \sum_{i = 1}^{s} w_{& 1 i} {\tilde{S}}_{1 i} + β \sum_{i = 1}^{s} w_{2 i} {\tilde{S}}_{2 i} \\ + δ \sum_{h = 3}^{t} \sum_{i = 1}^{s} w_{hi} {\tilde{S}}_{hi} \end{matrix}$ (13) where $\tilde{S}$ is the IVIFSM between X₁ and X₂. W = (w_ij) (∀ i ∈ {1, . . . , t} , ∀ j ∈ {1, . . . , s}) is the weight matrix of IVIFSM for X₁ and X₂ that enables $\sum_{i = 1}^{s} w_{1 i} = 1, \sum_{i = 1}^{s} w_{2 i} = 1, \sum_{i = 1}^{s} w_{hi} = 1 \forall h \in {3, . . ., t}$ and α + β + δ = 1. These formulas can be adapted for user based, item based and rating based similarity degrees in recommendation systems when β = δ = 0, δ = α = 0, and α = β = 0, respectively [21]. The similarity matrix elements are calculated as follows: ${\tilde{S}}_{1 i} = 1 - \frac{1 - exp (- 1 / - 2 (| \sqrt{μ_{iX}^{(1)} (x)} - \sqrt{μ_{iX}^{(2)} (x)} | + | \sqrt{v_{iX}^{(1)} (x)} - \sqrt{v_{iX}^{(2)} (x)} |))}{1 - exp (- 1)} ⩽$ for ∀i ∈ {1, . . , s} ${\tilde{S}}_{2 i} = 1 - \frac{1 - exp (- 1 / - 2 (| \sqrt{μ_{iY}^{(1)} (y)} - \sqrt{μ_{iY}^{(2)} (y)} | + | \sqrt{v_{iY}^{(1)} (y)} - \sqrt{v_{iY}^{(2)} (y)} |))}{1 - exp (- 1)}$ for ∀i ∈ {1, . . , s} ${\tilde{S}}_{hi} = 1 - \frac{1 - exp (- 1 / - 2 (| \sqrt{μ_{iD}^{(1)} (D_{h - 2})} - \sqrt{μ_{iD}^{(2)} (D_{h - 2})} | + | \sqrt{v_{iD}^{(1)} (D_{h - 2})} - \sqrt{v_{iD}^{(2)} (D_{h - 2})} |))}{1 - exp (- 1)}$ (14) for ∀i ∈ {1, . . , s} and ∀h ∈ {3, . . , s} [21].

The formulas to predict the values of linguistic labels of user P_{u(∀u∈{1,...,n})} to location S_{j(∀j∈{1,...,m})} according to the restaurants D₁, D₂, . . . D_k are addressed as given in [21]: $\begin{matrix} [μ_{\dot{ID}}^{P_{u}} (D_{j})^{-}, μ_{\dot{ID}}^{P_{u}} (D_{j})^{+}] \\ = \frac{\sum_{v = 1}^{n} SIM (P_{u}, P_{v}) \times [μ_{\dot{ID}}^{P_{v}} (D_{j})^{-}, μ_{\dot{ID}}^{P_{v}} (D_{j})^{+}]}{\sum_{v = 1}^{n} SIM (P_{u}, P_{v})} \end{matrix}$ (15) where ∀i ∈ {1, . . . , s} , ∀ j ∈ {1, . . . , k} , ∀ u ∈ {1, . . . , n}. $\begin{matrix} [v_{\dot{ID}}^{P_{u}} (D_{j})^{-}, v_{\dot{ID}}^{P_{u}} (D_{j})^{+}] \\ = \frac{\sum_{v = 1}^{n} SIM (P_{u}, P_{v}) \times [v_{\dot{ID}}^{P_{v}} (D_{j})^{-}, v_{\dot{ID}}^{P_{v}} (D_{j})^{+}]}{\sum_{v = 1}^{n} SIM (P_{u}, P_{v})} \end{matrix}$ (16) where ∀i ∈ {1, . . . , s} , ∀ j ∈ {1, . . . , k} , ∀ u ∈ {1, . . . , n}([21, 23]).

Note that composition operations are required for the aggregation of multiple criteria. The composition operation is adapted as follows: $\begin{matrix} (\begin{matrix} a_{11}^{(1)} & \dots & a_{1 s}^{(1)} \\ ⋮ & ⋱ & ⋮ \\ a_{t 1}^{(1)} & \dots & a_{ts}^{(1)} \end{matrix}) ■ (\begin{matrix} a_{11}^{(2)} & \dots & a_{1 s}^{(2)} \\ ⋮ & ⋱ & ⋮ \\ a_{t 1}^{(2)} & \dots & a_{ts}^{(2)} \end{matrix}) = \\ (\begin{matrix} a_{11}^{(12)} & \dots & a_{1 s}^{(12)} \\ ⋮ & ⋱ & ⋮ \\ a_{t 1}^{(12)} & \dots & a_{ts}^{(12)} \end{matrix}) that a_{1 i}^{(12)} = \\ (\min (μ_{iX}^{(1)} (x), μ_{iX}^{(2)} (x)), \max (v_{iX}^{(1)} (x), v_{iX}^{(2)} (x))) \\ for \forall i \in {1, . . ., s} \end{matrix}$ (17)

4 Application

Since the increasing potential of online marketing and sales operations, social networks have become an essential tool for the extension of communication with the customers. Thus, diversified types of online congregations, firms, blogs and forums focused on the development of their social network attraction. Therefore, data gathered from social networks should be evaluated for constructing location based systems. In addition to that, the need of competitive analysis of the shopping malls should be extracted from the differentiation as positive and negative. Contribution of the feelings to recommendation system will be beneficial especially in offering specialized alternatives (here restaurants) [36].

4.1 Data collection and preprocessing

The data was gathered from various social media platforms (Trip Advisor, Zomato and Foursquare) and includes 251 restaurants, 201 users and 136 places. Data contains both user information and location information. The reasons for using different social media platforms are (1) Reflecting different users’ reviews and ratings (2) Combining diversified data sources for better information extraction, (3) Provide missing information from other social media platforms. Trip Advisor and Foursquare are selected because they are well known and widely used across the world. They also provide enormous amount of user content which enhance the results of the proposed methodology. In addition to social media data, Google search engine is also utilized for reflecting number of Google reviews, average hours spent in the shopping mall and rating scores. Location data contains only the restaurants in Istanbul and numerical variables are standardized for better representation in clustering process. The list of the variables that are used in the analysis can be found in Table 1.

Table 1
Variables used for location based recommendation system provided by users

Variables for user clustering Source Variables for location clustering (for shopping malls) Source Variables for restaurant clustering Source

User nickname (or ID) Tripadvisor Location name (or ID) Tripadvisor, Zomato, Foursquare Restaurant name (or ID) Tripadvisor

User membership period (time) Tripadvisor Ranking Google search Tripadvisor, Zomato, Foursquare Ranking (things to do list in Trip Advisor) Tripadvisor Zomato, Foursquare

# of previously visited places Tripadvisor Average point Tripadvisor, Zomato, Foursquare Average price Tripadvisor Foursquare

# of contributions Tripadvisor # of positive reviews Tripadvisor Recommended time Google search

# of added photos Tripadvisor # of negative reviews Tripadvisor Average score Tripadvisor

Level of participation Tripadvisor # of votes Zomato, Foursquare Location (shopping mall) Tripadvisor

# of reviews Tripadvisor Recommended time Google search Speed Tripadvisor

Average score of previously visited places Tripadvisor Score distribution Tripadvisor, Zomato, Foursquare Service point Tripadvisor

# of followers Tripadvisor # of total reviews Google search, Tripadvisor, Zomato, Foursquare Flavor Tripadvisor

# of helpful votes Tripadvisor # of reviews from foreigners Tripadvisor Value Tripadvisor

# of reviews from locals Tripadvisor Location Tripadvisor

# of positive reviews

# of negative reviews

Variables for user clustering	Source	Variables for location clustering (for shopping malls)	Source	Variables for restaurant clustering	Source
User nickname (or ID)	Tripadvisor	Location name (or ID)	Tripadvisor, Zomato, Foursquare	Restaurant name (or ID)	Tripadvisor
User membership period (time)	Tripadvisor	Ranking	Google search Tripadvisor, Zomato, Foursquare	Ranking (things to do list in Trip Advisor)	Tripadvisor Zomato, Foursquare
# of previously visited places	Tripadvisor	Average point	Tripadvisor, Zomato, Foursquare	Average price	Tripadvisor Foursquare
# of contributions	Tripadvisor	# of positive reviews	Tripadvisor	Recommended time	Google search
# of added photos	Tripadvisor	# of negative reviews	Tripadvisor	Average score	Tripadvisor
Level of participation	Tripadvisor	# of votes	Zomato, Foursquare	Location (shopping mall)	Tripadvisor
# of reviews	Tripadvisor	Recommended time	Google search	Speed	Tripadvisor
Average score of previously visited places	Tripadvisor	Score distribution	Tripadvisor, Zomato, Foursquare	Service point	Tripadvisor
# of followers	Tripadvisor	# of total reviews	Google search, Tripadvisor, Zomato, Foursquare	Flavor	Tripadvisor
# of helpful votes	Tripadvisor	# of reviews from foreigners	Tripadvisor	Value	Tripadvisor
# of reviews from locals	Tripadvisor	Location	Tripadvisor
# of positive reviews
# of negative reviews

Before implementing sentiment analysis, the following steps are conducted: (1) Identify the sentiment of user reviews as positive and negative. (2) Remove reviews which are not meaningful and have missing character(s) (3) Remove stopwords and punctuations in the corpus (4) Converting word into lower cases. After number of positive and negative reviews are gathered, locations and users are grouped for constituting homogenous groups. These groups can be beneficial for offering a great variety of locations to diversified users.

Since location data is gathered from various sources, ranking, average point, average price and score distribution are stated as intervals. A sample dataset is given in Table 2. Note that user and restaurant information can be represented in the similar manner as Table 2.

Table 2

A sample dataset for location clustering

Location	Ranking	Average point	# of positive reviews	# of negative reviews	# of votes
Viaport AVM	4265	[2.2, 3] [3, 3.1]	[16,][80, 92]	[21, 82][82, 90]	[40, 221][221, 262]
ÜmraniyeMeydan AVM	3567	[2.9, 3][3,.2]	[2, 43]	[12, 80]	[20, 193]
Istinye Park AVM	340	[3.9, 4.2][4.2, 4.4]	[98, 167][167, 450]	[23, 45][45, 80]	[196, 343][343, 838]
Recommended time (hour)	Score distribution (1:bad and 5:very good)	# of total reviews	# of reviews from foreigners	# of reviews from locals
[0, 8, 2]	[1, 4][4, 5]	43	25	18
[0, 8, 2]	[1, 3][3, 4]	16	2	13
[1.5, 3]	[4, 5]	714	320	308

Finally, data for interval valued intuitionistic fuzzy recommendation system is prepared. The data contains four dimensions: user group (m), restaurant group (r) location group (n) and ratings (k). In Table 3, a sample of data is presented for better understanding before adapting recommendation system methodology.

Table 3

A sample of data for location based recommendation system

User group	Location group	Overall rating	Restaurant group	Average Speed	Average Service Point	Average Flavor	Average Value
P ₁	S ₁	[2.74, 2.87][2.87, 3.09]	D ₁	[2.12, 2.15],[2.15, 2.67]	[2.45, 2,86][2.86, 2.93]	[2.45, 2.55],[2.55, 2.63]	[2.03, 2.09][2.09, 2.76]
P ₁	S ₂	[1.98, 2.65][2.65, 2.98]	D ₂	[2.76, 2.89][2.89, 2.98]	[1.87, 2,21][2.21, 2.34]	[1.43, 1.56][1.56, 1.89]	[1.97, 2.75][2.75, 3.17]
P ₁	S ₃	[1.76, 2.98],[2.98,3.09]	D ₃	[2.34, 2.40][2.40, 2.56]	[2.45, 2.65][2.65, 2.87]	[2.01, 2.13][2.13, 2.45]	[1.56, 2.34][2.34, 3.29]
P ₁	S ₄	[2.31, 2.91][2.91, 3.16]	D ₄	[2.67, 2.87][2.87, 2.96]	[2.17, 2.32][2.32, 2.45]	[2.59, 2.87][2.87, 2.97]	[2.04, 2.61][2.61, 3.56]

Location data obtains shopping malls in Istanbul, recommended time to spend, location ranking, number of comments, score distribution (Excellent, Very good, Moderate, Bad and Terrible) and foreign and local comments from the mentioned social platforms in order to reflect different aspects of the shopping malls. Additionally, restaurant data includes restaurant ranking, average price, recommended time as seen in Google search, number of comments in Zomato and Trip Advisor, average score on Zomato and Foursquare, number of votes, number of negative and positive comments, average point of places nearby the restaurants are also implicated in restaurant grouping. User data is gathered from Trip Advisor including membership time period, participation level, number of contributions, average score of previously visited places, number of helpful votes, number of added photos and number of followers. Note that some of the users do not have enough necessary information that the relevant columns are marked as “Null”.

4.2 Proposed approach

Step 1. Sentiment analysis: In the first stage, sentiment analysis is adapted for determining the classes of negative, positive and neutral comments for locations from Trip Advisor. After text cleaning process in which misspelled words are eliminated, seed words are selected for noticing positive comments as (’perfect’, ‘great’, ‘good”, ‘very good’, ‘fair’, awesome’,’fantastic’), negative comments as (‘puff’, ‘off’, ‘bad’, ‘poor’,’terrible’) and neutral (‘don’t know’, ‘no comment’) as appeared with Turkish words.

The next step of the sentiment analysis is the gathering of the term-document matrix that determines the words as “frequency numbers” for handling the significance and importance levels of the terms appeared in the related comment texts. The literature presents various functions to be used for constituting the TDM. In this study, TF-IDF (term frequency-inverse term frequency) function is used. Note that TF-IDF is a combination of term frequency and inverse document frequency that produces a composite weight for each term in related document. By utilizing TDM, the unstructured comment data is transformed to structured form. Finally, clustering the comments based on TDM data is adapted via Rapidminer 5.3 considering the preprocessing of the data. For instance, comments are separated to words, transformed to lower case and filtered from irrelevant prepositions and conjunctions. After the preprocessing, k-means clustering is adapted by determining the most appropriate number of clusters (k). Note that different values between 2 and 8 is applied and according to the average distances from cluster centers, best clustering is performed when k = 3. The number of extracted comments for some of the locations (shopping malls) are presented in Table 4.

Table 4
Number of positive, negative and neutral comments from Trip Advisor

Location name Akasya Kanyon City’s İs-tinye< Buyaka Zorlu Aqua Florya Asto-ria İstanbul Marina Özdi-lek Nauti-lus Palla-dium Can-park Capitol

Positive 23987 25403 426 48800 48391 21097 22208 18413 23664 22982 51519 41234 45742 14974

Negative 111 1234 3456 53 6738 10566 5757 1799 7477 3753 9099 420 2905 5573

Neutral 456 976 881 672 106 232 101 267 130 433 749 690 558 593

Location name	Akasya	Kanyon	City’s	İs-tinye<	Buyaka	Zorlu	Aqua Florya	Asto-ria	İstanbul Marina	Özdi-lek	Nauti-lus	Palla-dium	Can-park	Capitol
Positive	23987	25403	426	48800	48391	21097	22208	18413	23664	22982	51519	41234	45742	14974
Negative	111	1234	3456	53	6738	10566	5757	1799	7477	3753	9099	420	2905	5573
Neutral	456	976	881	672	106	232	101	267	130	433	749	690	558	593

Step 2- User, location and restaurant clustering: As mentioned before, location similarity and grouping is necessary as the initial step for location prediction of future visiting tendencies of the customers. Thus, in the second stage, users, restaurants and locations are grouped using interval valued fuzzy c means clustering. In order to determine the optimum number of groups, Xie-Beni index values for different values of c parameter are determined and four clusters are gathered from whole dataset of users and locations. Second, cluster centers are randomly assigned and after that, distance between cluster centers to the given data is measured. Finally, locations are formed as groups similarly using R package as seen from Figs. 4 and 5. In this step, user clustering is explained in details to provide better understanding of the clustering process.

Fig. 3

Trip Advisor screen shots for some of the properties.

Fig. 4

Cluster plot for locations.

Fig. 5

Cluster plot for users.

After processing twenty iterative steps of fuzzy clustering to user data, the algorithm is stopped with an acceptable error. As seen from Table 5, 5 of the observations are assigned in Cluster 1 (‘passive social media users’-P1), Cluster 2 (‘prestigious place followers using active social media’-P2), and Cluster 4 (‘frequently visiting users using social media actively’-P4) and 3 observations are appeared in Cluster 3 (‘causal people using active social media’-P3).

Table 5

Centroid and spread points of clusters

	User group 1	User group 2	User group 3	User group 4
Centroid CL1	(0.005,0.039,0.114)	(0.009,0.051,0.098)	(0.013,0.131,0.183)	(0.027,0.231,0.536)
Centroid CL2	(0.087,0.163,0.504)	(0.233,0.418,0.848)	(0.014,0.125,0.314)	(0.021,0.219,0.431)
Centroid CL3	(0.028,0.145,0.205)	(0.017,0.056,0.338)	(0.031,0.162,0.447)	(0.214,0.356,0.765)
Centroid CL4	(0.123,0.263,0.672)	(0.103,0.206,0.441)	(0.214,0.525,0.704)	(0.422,0.459,0.661)

Additionally, Table 6 indicates the centroid points from observed data and score function is calculated for making comparisons properly. Therefore, initially prototypes are presented as interval valued fuzzy numbers. The last columns of Table 5 denote the sum of squares (SSQ) between the object $x_{j} ([μ_{j}^{-}, μ_{j}^{+}], [v_{j}^{-}, v_{j}^{+}])$ and centroids $α_{i} = ([β_{i}^{-}, β_{i}^{+}], [θ_{i}^{-}, θ_{i}^{+}])$ and also, $SSQ (x_{j}, α_{i}) = ({∥ μ_{j}^{-} - β_{i}^{-} ∥}^{2} + {∥ μ_{j}^{+} - β_{i}^{+} ∥}^{2} + {∥ v_{j}^{-} - θ_{i}^{-} ∥}^{2}$ $+ {∥ v_{j}^{+}, θ_{i}^{+} ∥}^{2})$ Consider that lower SSQ values between objects imply properly assigned object to related centroids. For instance, objects including lowest SSQ values are assigned with higher membership degrees excluding User ID 13.

Table 6

Membership degrees of each cluster for each object for user clustering

User ID	Cluster 1	Cluster 2	Cluster 3	Cluster 4	Total SSQ	User ID	Cluster 1	Cluster 2	Cluster 3	Cluster 4	Total SSQ
1	0.986	0.006	0.005	0.003	0.0484	10	0.010	0.803	0.046	0.141	0.2902
2	0.905	0.003	0.060	0.032	0.0210	11	0.100	0.515	0.045	0.34	0.6214
3	0.001	0	0.319	0.68	0.1209	12	0.900	0.014	0.084	0.002	0.1320
4	0.035	0.042	0.823	0.100	0.1674	13	0.001	0.001	0.996	0.002	0.1338
5	0.869	0.011	0.018	0.102	0.1543	14	0	1	0	0	0.1145
6	0.107	0.190	0.033	0.67	0.4365	15	0.109	0.002	0.145	0.744	0.1054
7	0.101	0.004	0.793	0.102	0.2219	16	0.002	0.205	0.072	0.721	0.2870
8	0	0	0	1	0.0802	17	0	1	0	0	0.1245
9	0.801	0.003	0.008	0.188	0.0834	18	0.009	0.915	0.042	0.034	0.098

In addition to user grouping, location groups are denoted as “popular place-S1”, “rarely visited place-S2”, “middle income oriented place-S3” and “rich customer oriented place-S4” for indicating the location property of the cluster. At last, restaurants are grouped as “cheap restaurant-D1”, “prestigious restaurant- D2”, “beef and steak restaurant-D3” and ”café-restaurant- D4”.

Step 3-Interval valued Intuitionistic Fuzzy recommendation system: After gathering clusters, multi criteria interval valued intuitionistic fuzzy recommendation system (IVIFRS) is applied to user, restaurant and location groups considering transformed IVIFN based data of restaurants. Recommendation system matching for each user group to a specific restaurant is applied by the determination of similarity measures and predicting the values of linguistic labels of user P_u to location S_j and naturally, restaurants. Results are listed as seen from Table 7. An illustrative example is given to explain the adaptation of IVIFRS to clustering section in the following:

Table 7

Training dataset for the values being predicted according to the proposed method

P	D₁	D₂	D₃	D₄
P ₁	[0.4, 0.1][0.5, 0.2]	[0, 2, 0.1][0.6, 0.7]	[0, 6, 0.1][0, 3, 0.5]	[0.2, 0.4][0.3, 0.4]
P ₂	[0.3, 0.4][0.5, 0.5]	[0, 2, 0.3][0.4, 0.6]	[0.4, 0.5][0.6, 0.5]	[0.3, 0.4][0.5, 0.5]
P ₃	[0.2, 0.3][0.3, 0.4]	[0, 6, 0.1][0, 3, 0.5]	[0, 2, 0.4][0.5, 0.4]	[0.3, 0.2][0.6, 0.8]
P ₄	*	*	*	*

According to the similarity calculation, the values given in Table 7 are utilized to predict P₄. The parameters are applied as α = 0and δ = 0.5, and also w_1i = w_2i = w_3i = 0.2 respectively. Similarity between other users are calculated as SIM (P₁ - P₄)=0.76, SIM (P₂ - P₄)=0.71 and SIM (P₃ - P₄)=0.43. The predictive values of IVIFRS based methodology are given as $\begin{matrix} Restaruant (P_{4}) = \\ 〈 D_{1} = {, [0.23, 0.13]} \\ D_{2} = {[0.45, 0.45], [0.51, 0.62]} \\ D_{3} = {[0.34, 0.45], [0.21, 0.12]} \\ D_{4} = {[0.37, 0.43], [0.31, 0.52]} 〉 \end{matrix}$

Finally, defuzzification is adapted as D₁ = 0.54, D₂ = 0.23, D₃ = 0.45 and D₄ = 0.29.

From Table 8, restaurants are recommended from restaurant group lists to users and location groups. As seen from the results, users who visit popular place-(S1) generally prefer most crowded shopping malls. In the similar manner, users who generally prefer to go middle income oriented place are tend to eat local food based restaurants. Causal people using active social media are exactly opposite of the people who visit popular places and locations are mostly focused on middle income level because of their tendency to go favored restaurants which are “esoteric” that address “high income” group.

Table 8

Restaurant recommendation results to each user group with each location group

User-location group	Recommended restaurant group	Location group	User-location group	Recommended restaurant name	Location name
P1-S1	Burger King	Capitol	P1-S3	HD Iskender	Nautilus
P2-S1	SaitEfendi	Özdilek	P2-S3	Bursa Kebapevi	Canpark
P3-S1	Günaydın Steakhouse	İstinye Park	P3-S3	Plus Kitchen	Zorlu
P4-S1	Bodrum Mantı	Buyaka	P4-S3	GünaydınKöfteDöner	Palladium
P1-S2	Big Chefs	Astoria	P1-S4	Tike	İstanbul Marina
P2-S2	Midpoint	Aqua Florya	P2-S4	Kirpi Café	Buyaka
P3-S2	Happy Moon’s	Akasya	P3-S4	BalıkEv	City’s
P4-S2	Tazele	Kanyon	P4-S4	Pilezza	Özdilek

5 Comparison of the proposed methodology

The impact of sentiment analysis and clustering assisted interval valued FcM algorithm on fuzzy data can be realized by comparing the proposed methodology with several methodologies in the same context. The selected studies for the comparison process are Son and Thong [21]’s study which is related to medical diagnosis and Nilashi et al. [27]’s that proposes a recommender system for tourism industry. For better understanding of the comparison phase, these two studies are briefly selected for the reasons given as follows:

(1)Son and Thong [21]’s study presents intuitionistic fuzzy set theory adaptation to recommender systems for measuring the contribution of using interval valued intuitionistic fuzzy set theory rather than using intuitionistic fuzzy set theory in recommender systems and also, evaluating sentimentanalysis phase contribution to existing recommender systems.

(2) Nilashi et al. [27]’s study is selected for the similar study topic and comparison of social platform based recommender systems to demonstrate the contribution of fuzzy set theory to recommender systems.

The comparison is conducted based on Mean Absolute Error (MAE) and the computational time. The same methodologies are adapted to the given dataset and executed on Intel^® Celeron ^® CPU 847@ 1.10 Ghz 4 GB RAM. The results are considered using the intuitionistic defuzzification method presented in Albeanu and Vladicescu [15] for 10 more iterations. Results are presented in Table 9.

Table 9
The comparative results of Son and Thong [21] and Nilashi et al. [27]’s studies with respect to proposed methodology

# of Iterations MAE of proposed methodology MAE for Son and Thong [21]’s study MAE for Nilashi et al. [27]’s study Computational time of proposed methodology Computational time for Son and Thong [21]’s study Computational time for Nilashi et al. [27]’s study

10 0.329549 0.332072 0.34902 9.531862 10.4887 10.72052

20 0.313976 0.348645 0.303591 9.830731 9.303313 10.58915

30 0.306936 0.315018 0.343408 9.331028 11.89122 10.4728

40 0.347523 0.313377 0.324817 8.611162 11.83422 10.62169

50 0.325362 0.329224 0.338069 8.444247 9.427412 10.19122

60 0.311693 0.311109 0.341132 8.124302 9.167207 10.21279

70 0.306621 0.342644 0.332759 8.74614 9.620453 10.60639

80 0.309343 0.30305 0.349606 8.404803 9.759067 10.66614

90 0.323932 0.336973 0.347513 8.433736 9.254415 10.26861

100 0.316776 0.332972 0.334589 8.547861 10.52727 10.14095

# of Iterations	MAE of proposed methodology	MAE for Son and Thong [21]’s study	MAE for Nilashi et al. [27]’s study	Computational time of proposed methodology	Computational time for Son and Thong [21]’s study	Computational time for Nilashi et al. [27]’s study
10	0.329549	0.332072	0.34902	9.531862	10.4887	10.72052
20	0.313976	0.348645	0.303591	9.830731	9.303313	10.58915
30	0.306936	0.315018	0.343408	9.331028	11.89122	10.4728
40	0.347523	0.313377	0.324817	8.611162	11.83422	10.62169
50	0.325362	0.329224	0.338069	8.444247	9.427412	10.19122
60	0.311693	0.311109	0.341132	8.124302	9.167207	10.21279
70	0.306621	0.342644	0.332759	8.74614	9.620453	10.60639
80	0.309343	0.30305	0.349606	8.404803	9.759067	10.66614
90	0.323932	0.336973	0.347513	8.433736	9.254415	10.26861
100	0.316776	0.332972	0.334589	8.547861	10.52727	10.14095

According to the final results, proposed method is clearly outperformed considering with both MAE results and computational time especially when number of iterations are increasing. Some of the concluding remarks from the comparison phase are given below:

(1) Multi criteria interval valued intuitionistic fuzzy recommendation system (IVIFRS) is capable to perform the recommendation of the special offers using both for interval valued intuitionistic fuzzy data and semi- interval valued intuitionistic fuzzy data. Interval valued intuitionistic fuzzy data provides better presentation of uncertain data (recommended time, score distribution, average price etc.). This situation provides the late adaptation of defuzzification process which minimizes the information loss while performing clustering process.(2) Interval valued intuitionistic fuzzy recommendation system is integrated with clustering process which provides the broad access of the similar alternatives and increases the acceptance of the special offers (3) Using both semi-interval valued intuitionistic fuzzy sets and interval valued intuitionistic fuzzy sets enables the flexibility and usefulness of proposed approach as indicated in Son and Thong [21]’s study.

(4) Nilashi et al. [27]’s study applied dimension reduction (PCA) to the relevant data in a similar topic. Our dataset also includes considerable number of features. On the other hand, with using fuzzy adaptation to clustering techniques, dimension reduction phase is eliminated with using less computational time to more complicated dataset (interval valued intuitionistic fuzzy data presentation) as seen in Table 9.

(5) Although Son and Thong [21]’s fuzzy adaptation to recommender system is similar to our approach, interval valued intuitionistic fuzzy application to recommender system outperformed better when considering MAE and computational time results except one type of iteration in Table 9. This indicates the usefulness of sentiment analysis application outcomes (extracting positive and negative comments) to fuzzy clustering process and fuzzy recommender systems.

As seen from the overall results, interval valued intuitionistic fuzzy recommender system is adaptable to social platform data with many contributions: better uncertain data presentation, better computational time and MAE results. If the dataset size increases to large volumes, the amount of information to be processed will exceed the computing capacity of a single machine. In such cases, distributed computing that divides the original dataset into smaller subsets that are obtained with multiple nodes of a cluster, commonly used. Thus, each node proceeds only the divided data stored in it. The outputs exist from all the nodes and then these outputs are aggregated to obtain the final result [50].

6 Conclusions

In this paper, we focused on implementing interval valued intuitionistic fuzzy recommendation system using sentiment analysis to reveal number of positive, negative and neutral comments made by users and additionally, user, location and restaurant clustering are conducted to provide more options to recommendation system users. In order to implement user, location and restaurant clustering properly, data is manually gathered from social media platforms before restaurant recommendation system is established. For this reason, data is extracted from Trip Advisor, Zomato and Foursquare for adapting the clustering process that will ensure broad range of alternatives in mentioned three dimensions. Note that, users, restaurant and location clustering are implemented using interval valued fuzzy c means clustering separately and user data considers the previously visited location scores which enable the representation of historic base of the users. Then, IVIFRS methodology is conducted for restaurant recommendation utilizing the formed user, restaurant and location groups.

The reasons for conducting interval valued intuitionistic fuzzy numbers based decision making process are: (1) interval valued fuzzy sets provide more degrees of freedom in decision making process when users are not exactly express the membership and non-membership degrees. Thus, modelling the uncertainty will be easier compared with Type 1 fuzzy numbers and triangular fuzzy numbers, (2) Compared with other fuzzy sets, IVIFSs are able to present uncertainty of inaccurate information by closed interval based memberships and non-memberships. This ensures detailed representation of intuitionistic fuzzy sets while evaluating vague information [21].

Results demonstrate that using such this social platform data considering various locations with interval valued fuzzy numbers has the potential to represent customers’ life style and interests by not using crisp numbers for getting insight about each individual user and location properties before implementing personalized recommender systems. In addition to that, fuzzy recommendation system is applicable and effective especially not having crisp values in the related database. The novelty of the paper is not only the effective usage of social platform data via sentiment analysis and fuzzy clustering. In addition to these properties, the formulas to predict the recommendation possibility based on similarity measure using interval valued intuitionistic form of fuzzy recommender system and late adaptation of defuzzification method are other aspects of the effective usage of recommendation systems. Besides that, the paper fulfills the gap indicated in Son and Thong [21]’s study that the small and medium sized real dataset implementation of intuitionistic recommendation system can be beneficial and also, hybrid algorithm based on fuzzy clustering adaptation will enrich recommendation system performance in advance.

To sum up, the proposed recommender system has some advantages when compared with other fuzzy recommender based systems: (1) Better presentation of data that was gathered from diversified sources such as Zomato, Trip Advisor and Google search by using interval valued intuitionistic fuzzy dataset. (2) Contribution of text data (comments) to recommendation process by using sentiment analysis. (3) Fuzzy c means clustering is utilized for gathering broad range of restaurant offers to specified customer groups (segments). (4) Contribution to imprecise social platform data to location based recommender systems is obvious as seen in Table 9.

For further studies, applications of fuzzy recommendation system with massive datasets should be examined in order to check the efficiency of the proposed approach. Secondly, other applications of fuzzy clustering frameworks such as entropy based clustering c-medoids ensure the implementation of the study and generalization of fuzzy recommendation systems. Finally, other theoretical analysis of IVIFRS operations can be conducted using t – norm. These future works will enrich the dissemination of fuzzy recommendation systems both for practical and academic perspectives.

References

Alaei

, Becken

and Stanic

, Sentiment Analysis in Tourism: Capitalizing on Big Data, Journal of Travel Research (2017), doi: 10.1177/0047287517747753

Bagheri

, Saraee

, Jong

and Care

, More about customers: Unsupervised domain-independent aspect detection for sentiment analysis of customer reviews, Knowledge-Based Systems 52 (2013), 201–213.

Bali

, Agarwal

, Poddar

, Harsole

and Zaman

N.M.

, Consumer’s. Sentiment Analysis of Popular Phone Brands and Operating System Preference, International Journal of Computer Applications 155(4) (2016), 975–8887.

Ortigosa

, Martín

J.M.

and Carro

R.M.

, Sentiment analysis in Facebook and its application to e-learning, Computers in Human Behavior 31 (2014), 527–541.

Giachanou

and Crestani

, Like it or not: A survey of Twitter sentiment analysis methods, ACM Computing Surveys 49 (2016), 28:21–28:41.

Pingley

, Yu

, Zhang

, Fu

and Zhao

, A context-aware scheme for privacy-preserving location-based services, Computer Networks 56 (2012), 2551–2568. B. Oztaysi, C. öner, D.H. Beyhan Market Analysis Using Computational Intelligence: An Application for GSM Operators Based onTwitter Comments, Intelligent Techniques in Engineering Management (2015), 483–501.

Anagnostopoulos

, Hadjiefthymiades

and Kolomvatsos

, Time-optimized user grouping in Location Based Services, Computer Networks 81 (2015), 220–244.

Kearney

and Liu

, Textual sentiment in finance: A survey of methods and models, International Review of Financial Analysis 33 (2014), 171–185.

Lipizzi

, Dessavre

D.G.

, Iandolib

and Marquez

J.E.R.

, Social Media Conversation Monitoring: Visualize Information Contents of Twitter Messages Using Conversational Metrics, Procedia Computer Science 80 (2016), 2216–2220.

10.

C.H.

, Kao

S.C.

, Wu

C.C.

and Huang

, Location-aware service applied to mobile short message advertising: Design, development, and evaluation, Information Processing & Management 51(5) (2015), 625–642.

11.

Gavalas

, Konstantopoulos

, Mastakas

and Pantziou

, Mobile recommender systems in tourism, Journal of Network and Computer Applications 39 (2014), 319–333.

12.

Costa

, Ferreira

, Brito

, Bittencourt

, Holanda

, Machado

and Marinho

, A framework for building web mining applications in the world of blogs: A case study in product sentiment analysis, Expert Systems with Applications 39 (2012), 4813–4834.

13.

Turban

, Sharda

and Delen

, Decision Support and Business Intelligence Systems, Ninth Edition (2011) Upper Saddle River, New Jersey: Prentice hall.

14.

Xianghua

, Guo

, Yanyan

and Zhiqiang

, Multi-aspect sentiment analysis for Chinese online social reviews based on topic modeling and HowNet lexicon, Knowledge-Based Systems 37 (2013), 186–195.

15.

Albeanu

and Vladicescu

F.L.P.

, Intuitionistic fuzzy numbers in software reliability modeling, J Sustainable Energy 1(1) (2010), 30–34.

16.

Tussyadiah

I.P.

, A concept of location-based social network marketing, Journal of Travel & Tourism Marketing 29(3) (2012), 205–220.

17.

and Li

, A Location Recommender Based on a Hidden Markov Model: Mobile Social Networks, Journal of Organizational Computing and Electronic Commerce 24(2-3) (2014), 257–270.

18.

Ortigosa-Hernandez

, Rodr?guez

J.D.

, Alzate

, Lucania

, Inza

and Lozano

J.A.

, Approaching Sentiment Analysis by using semi-supervised learning of multi-dimensional classifiers, Neurocomputing 92 (2012), 98–115.

19.

Bhattacharjee

and Petzold

, What drives consumer choices? Mining aspects and opinions on large scale review data using distributed representation of words, 2016 IEEE 16th International Conference on Data Mining Workshops (2016), 908–915.

20.

Cheverst

, Davies

, Mitchell

, Friday

and Efstratiou

, (2000) Developing a context-aware electronic tourist guide: Some issues and experiences, CHI ’00 Proceedings of the SIGCHI conference on Human Factors in Computing Systems: 17-24 (2000), The Hague, The Netherlands—April 01 - 06, 2000.

21.

Son

L.H.

and Thong

N.T.

, Intuitionistic fuzzy recommender systems: An effective tool for medical diagnosis, Knowledge-Based Systems 74 (2015), 133–150.

22.

Qiu

, Rui

and Whinston

, Social network-embedded prediction markets: The effects of information acquisition and communication on predictions, Decision Support Systems 55 (2013), 978–987.

23.

Cheng

L.C.

and Wang

H.-A.

, A fuzzy recommender system based on the integration of subjective preferences and objective information, Applied Soft Computing 18 (2014), 290–301.

24.

García-Cumbreras

M.Á.

, Montejo-Ráez

and Díaz-Galiano

M.C.

, Pessimists and optimists: Improving collaborative filtering through sentiment analysis, Expert Systems with Applications 40 (2013), 6758–6765.

25.

Hadano

, Shimada

and Endo

, Aspect Identification of Sentiment Sentences Using A Clustering Algorithm, Procedia - Social and Behavioral Sciences 27 (2011), 22–31.

26.

Zhang

, Zhang

and Vo

D.T.

, Neural Networks for Open Domain Targeted Sentiment, Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (2015), 612–621, Lisbon, Portugal, 17-21 September 2015.

27.

Nilashi

, Bagherifard

, Rahmani

and Rafe

, A recommender system for tourism industry using cluster ensemble and prediction machine learning techniques, Computers & Industrial Engineering 109 (2017), 357–368.

28.

Mobile Marketing Association, Mobile. Location Based Services Marketing Whitepaper, Tech. Rep, Mobile Marketing Association, (2011).

29.

and Wu

D.D.

, Using text mining and sentiment analysis for online forums hotspot detection and forecast, Decision Support Systems 48 (2010), 354–368.

30.

Abbas

, Michael

and Michael

M.G.

, The regulatory considerations and ethical dilemmas of location-based services (LBS): A literature review, Information Technology & People 27(1) (2014), 2–20.

31.

Babuska

, Fuzzy and Neural Control Disc Course Lecture Notes, Retrieved October 5, 2012 from http://www.dcsc.tudelft.nl/disc_fnc/transp/fncontrol.pdf (2009).

32.

Schumaker

R.P.

, Zhang

, Huang

C.-N.

and Chen

, Evaluating sentiment in financial news articles, Decision Support Systems 53(3) (2012), 458–464.

33.

Fan

, Lau

R.Y.K.

and Zhao

J.L.

, Demystifying Big Data Analytics for Business Intelligence Through the Lens of Marketing Mix, Big Data Research 2(1) (2015), 28–32.

34.

Kakarlapudi

and Garla

, PASW Chakraborty, Clustering and Sentiment Analysis Using Tweets from Twitter, Practical Text Mining and Statistical Analysis for Non-structured Text Data Applications (2012), 557–583.

35.

Rill

, Reinel

, Scheidt

and Zicari

R.V.

, PoliTwi: Early Detection of Emerging Political Topics on Twitter and the Impact on Concept-Level Sentiment Analysis, Knowledge-Based Systems (2014), DOI: 10.16/j.knosys.2014.05.008

36.

öner

S.C

and öztay?i

, An Interval Valued Hesitant Fuzzy Clustering Approach for Location Clustering and Customer Segmentation, Advances in Fuzzy Logic and Technology (2017), 56–70.

37.

T.H.N.

, Ryu

K.H.

and Park

, A method for predicting future location of mobile user for location-based services system, Computers & Industrial Engineering 57 (2009), 91–105.

38.

Kamakura

W.A.

and Gessner

, Consumer sentiment and buying intentions revisited: A comparison of predictive usefulness, Journal of Economic Psychology 7(2) (1986), 197–220.

39.

, Zha

and Li

, Social media competitive analysis and text mining: A case study in the pizza industry, International Journal of Information Management 33 (2013), 464–472.

40.

Huth

W.L.

, Eppright

D.R.

and Taube

P.M.

, The indexes of consumer sentiment and confidence: Leading or misleading guides to future buyer behavior, Journal of Business Research 29(3) (1994), 199–206.

41.

Shin

and Tsui-Chuan Lin

, Who avoids location-based advertising and why? Investigating the relationship between user perceptions and advertising avoidance, Computers in Human Behavior 63 (2016), 444–452.

42.

Bai

, Predicting consumer sentiments from online text, Decision Support Systems 50 (2011), 732–742.

43.

Hijikata

, Ohno

, Kusumura

and Nishida

, Social summarization of text feedback for online auctions and interactive presentation of the summary, Knowledge-Based Systems 20(6), 527–541.

44.

, Duan

and Cao

, The impact of social and conventional media on firm equity value: A sentiment analysis approach,(55), Decision Support Systems (2013), 919–926.

45.

Chang

Y-C.

, Ku

C-H.

and Chen

C-H.

, Social media analytics: Extracting and visualizing Hilton Hotel ratings and reviews from Trip Advisor, International Journal of Information Management (2018), https://dx-doi-org.web.bisu.edu.cn/10.1016/j.ijinfomgt.2017.11.001

46.

Yue

, Deriving decision maker’s weights based on distance measure for interval-valued intuitionistic fuzzy group decision making, Expert Systems with Applications 38 (2011), 11665–11670.

47.

Zou

and Huang

K.W.

, Leveraging location-based services for couponing and infomediation, Decision Support Systems 78 (2015), 93–103.

48.

Yang

W.S.

, Cheng

H.C.

and Dia

J.B.

, A location-aware recommender system for mobile shopping environments, Expert Systems with Applications 34 (2008), 437–445.

49.

Garg

, A new generalized improved score function of interval-valued intuitionistic fuzzy sets and applications in expert systems, Applied Soft Computing 38 (2016), 988–999.

50.

Elkano

, Galar

, Sanz

and Bustince

, CHI-BD: A fuzzy rule-based classification system for Big Data classification problems, Fuzzy Sets and Systems 348 (2018), 75–101.