Developing an MQ-LSTM-Based Cultural Tourism Accelerator with Database Security

Abstract

Cultural tourism (CT), which enhances the economic development of a region, aids a country in reinforcing its identities, enhancing cross-cultural understanding, and preserving the heritage culture of an area. Designing a proper tourism model assists tourists in understanding the point of interest without the help of a local guide. However, owing to the need for the analysis of different factors, designing such a model is a complex process. Therefore, this article proposes a CT model for peak visitor time in Riyadh, a city in Saudi Arabia. The main objective of the framework is to improve the cultural tourism of Riyadh by considering various factors to help in improving CT based on recommendation system (RS). Primarily, the map data and cultural event dataset were processed for location, such as grouping with Kriging interpolation-based Chameleon (KIC), tree forming, and feature extraction. After that, the event dataset’s attributes were processed with word embedding. Meanwhile, the social network sites (SNS) data like reviews and news were extracted with an external application programming interface (API). The review data were processed with keyword extraction and word embedding, whereas the news data were processed with score value estimation. Lastly, the data were fused, corresponding to a historical site, and given to the Multi-Quadratic-Long Short-Term Memory (MQ-LSTM) recommendation system (RS); also, the recommended result with the map was stored in a database. Lastly, the database security was maintained with locality sensitive hashing (LSH). From the experimental evaluation with multiple databases including the Riyadh Restaurants 20K dataset, the proposed recommendation model achieved a recommendation rate (RR) of 97.22%, precision of 97.7%, recall of 98.27%, and mean absolute error (MAE) of 0.0521. This result states that the proposed RS provides higher RR and reduced error compared to existing related RSs. Thus, by attaining higher performance values, the proposed model is experimentally verified.

1. Introduction

Tourism is valued in most countries, owing to its contribution to economic, social, and cultural development. Previously, by sharing travel information with clients and suppliers, travel agencies aid in sustainable tourism development, which makes positive contributions to ensuring socio-cultural and economic sustainability [1]. However, tourism has entered a novel era of smart tourism with the rapid development of modern information as well as communication technology and the continuous refinement of the tourism market demand [2]. But, reading through all such information for planning their trip or leisure time is challenging for travelers, owing to the massive volume of information on tourism or leisure spots and activities on the internet. In this case, a travel recommender could assist by supporting the decision-making process in areas, namely, the selection of attractions, mode of transport, choice of destinations, finalization of routes, identification of appropriate accommodations and restaurants, et cetera [3]. Moreover, the big data era has triggered a complete change in CT development. Informationization has resulted in a fast spread of public opinion, which is directly associated with tourism destination survival [4]. Hence, researchers developed a smart CT information system comprising tourism historical destinations that combines the graphical interface system (GIS) and the historical information content of historical sites [5].

Moreover, in smart CT, the contextual suggestion has emerged as a modified RS. For providing the tourists with a list of suggestions grounded in contexts such as location, time of day, or else the day of the week (weekdays or weekends), the contextual RS is integrated with larger databases [6]. Grounded in the point of interest ratings and recommendations, various contextual tourism RSs were developed since they are valuable to tourists and enable them to explore new places for visiting [7]. Moreover, for the tourism recommendations, several machine learning (ML) and deep learning (DL) approaches like collaborative filtering and convolutional neural network (CNN) [8], were developed. But, the recommendation could not be reliable only with the contextual information. Therefore, for the modeling of cultural heritage sites, a GIS model, namely, ArcGIS 10.5, was utilized during CT [9]. This GIS interpretation mainly aims at enabling visitors to understand the place, sensitizing them to the surroundings, and making them aware of the significance of the nature and monuments of the visiting point of interest [10].

For predicting persuasive point of interest (POI) recommendations, existing research works [11] performed Activity and Behavior induced Personalized Recommender System (ABiPRS) as a hybrid approach. However, this research did not explore well on noise management in the group recommender systems. Also, the model of [12] used various collaborative filtering approaches based on similarity models to perform an efficient recommendation to reduce the search cost of the customers. Nevertheless, there were several malicious as well as non-malicious noises present in the rating data that distorted the recommendation quality. [13] used multi-attention-based group recommendation models using neural networks to attain accurate recommendations. Based on the preference data, a vector representation of features was extracted for recommendation. However, deep semantic features to differentiate the preferences of tourists were not determined. Existing works in [14] used ML approaches to aid in developing effective hybrid recommendation systems. However, the use of ML techniques affects the performance of RS while using large amounts of data.

Thus, for sustainable CT, a database model should be developed to provide information on the cultural sources of a region centered on spatial and contextual information. However, to get hold of tourists, any country should consider various factors like peak visitors’ time of the prevailing heritage place and its significance to providing convenient CT. Hence, for enhancing tourism, this article proposes an MQ-LSTM-based tourism model for Riyadh, the capital city of Saudi Arabia. Since the events have a strong as well as wide cultural effect on the local society, the cultural events of Riyadh are considered in this research [15], which could attract tourists.

Problem Statement

Even though some research like [16,17,18,19], are available for the tourism recommendation model, they exhibit several limitations, which are given as

• During the busy hours in historical sites, the recommendation model for analyzing the culture of the place by recommending the most nearby cultural events is scarcely developed.
• In conventional works, when the recommendation is created only grounded on the reviews, the historical importance was neglected, which affects the recommendation of CT.
• When a tourism-centric database is created, there is no guarantee that the database is used by legal users only. This would cause a security threat to the historical sites, which are the cultural symbol of a country.
• Most of the existing research works did not concentrate on analyzing the historical sites or tourist places of Riyadh city based on the restaurant details that tourists accommodate. This greatly impacts the RS.

By analyzing the aforementioned issues, the proposed framework’s objective is to develop a smart CT model in a secure database, and its contributions are

• To develop KIC and Cramer’s’ V Correlated-Minimal Spanning Tree (CVC-MST) schemes for analyzing the possible nearest cultural events to the historical sites and suggested to tourism planners.
• To develop an MQ-LSTM recommendation model considering the location, time, and review, along with historical significance.
• To establish an LSH hashing for providing the security of the tourism recommendation database using the user attributes.

The remaining part is organized as follows: the related works are described in Section 2; the proposed research work is explicated in Section 3; the outcomes are elucidated in Section 4; and, finally, the paper is concluded in Section 5.

2. Related Works

Author [20] established an RS for rural tourism routes grounded on the multiple data sources’ fusion model. Centered on the idea of tensor orthogonal decomposition, the RS was developed. The experimentation revealed that the developed model attained the lowest Error. Nevertheless, the algorithm was developed considering the ratings only, which could not be reliable since some popular places receive lower ratings. [21] presented a mobile RS for multi-profile cultural visitors grounded on visiting preferences classification named ACUX Recommender (ACUX-R). To assign profiles for cultural visitors, ACUX topology was leveraged. A study and questionnaire-based evaluation proved that the ACUX-R model satisfied the cultural visitors. But, the model was applicable only to local people, which limited the model to in situ visitors only. [22] employed ArcGIS 10.5 as an analysis tool for the temporal and spatial distribution of A-level tourist attractions in Chengyu. By utilizing the geographic detector model, an analysis was performed by the model on the factors affecting the distribution of tourist attractions. As per the outcomes, the impact factor of A-level tourist attraction was owing to the traffic, topography, river system, and socio-economic development model. Yet, the model neglected the cultural events, which might be an A-list tourist attraction. [23] recommended a DL-centric spring festival holiday tourism data mining model. Primarily, the tourism information and key factors, which affected the choice of tourist attractions, were excavated. After that, grounded in the analyzed key factors, the recommendation was given. The experimental analysis revealed a better convergence rate of the scenic spots after optimization. The scheme also had a shortcoming, namely, poor recommendation quality. [24] established an ML technique to predict over-tourism in Spain. The ML models utilized were support vector machine (SVM), linear regression, and naïve Bayes. The descriptive analysis displayed that the indicator and objective thresholds acted as the indication of overcrowding situations. Nevertheless, the exercise performed stated that the utilized competitiveness index was not a determinant of over-tourism. [25] illustrated a heuristic fuzzy approach for the assessment and management of tourism sustainability. For handling the involved variables’ inherent uncertainty as well as vagueness, fuzzy was wielded in this framework. Afterward, for the assessment of the relationship betwixt the obtained ranked outcomes and the obtained life score quality, a correlation analysis was introduced. The algorithm was applied for the Italy region and found to ensure a high level of versatility. However, with fuzzy logic, inappropriate decision-making rules would deteriorate the assessment results. [26] recommended a model for the sentiment analysis of Chinese tourists toward Malaysia’s cultural heritage grounded on online travel reviews. The sentiment analysis leveraged vocabulary filtering, semantic clustering technology, and BERT, along with co-occurrence analysis. Lastly, the BERT-based emotions predicted with BERT obtained enhanced sentence predictive performance. But, the model was developed by being grounded in online content only, which might be false reviews. [27] employed a personalized day tour design for urban tourists considering carbon dioxide emissions. For designing a personalized day tour route in Chengdu, China, the model developed an evolutionary system grounded on reinforcement learning. The outcomes indicated that the developed algorithm was superior to the selected baseline techniques. Nevertheless, for the day tour design, a limited number of factors were considered. [28] presented a multi-agent-based scheme for the recommendation of cultural tours. The multi-agent model enclosed beliefs, desires, and intentions for the plan recommendation with the estimated user behavior and item similarity. The preliminary experimental results on the system validated the developed approach. However, in the multi-agent model, if the evaluation failed on beliefs, the action was not applied to the world. [29] suggested a multiple itinerary tourist’s recommendation engine named MULTITOUR. The recommendation was grounded on the tourists’ interests and different constraints. The outcomes exhibited that regarding the accuracy, the MULTITOUR algorithm outperformed the baseline techniques. But, with the geotagged photos, long-time tourism could not be effectively recommended. [16] designed a tourism RS based on user preferences to provide personalized services. This work used user reviews on tourism social network data to extract user preferences for better RS. The experimental analysis of Trip Advisor platform data demonstrated better f-measure in comparison with previous models. However, some of the required factors for RS, such as geographical information systems, various conditions of visitors, and their environmental conditions were neglected in this work. This affects the effectiveness of the RS. [17] developed an RS to rank tourist attractions via online reviews grounded on aspect-level SA as well as multi-criteria decision making with intuitionistic and hesitant fuzzy information. For ranking tourism recommendations, the preference information of the potential tourist and experienced tourist were combined for determining the weight criteria. Experimental outcomes illustrated the better performance of this model compared to that of some other prevailing methodologies. On the other hand, in the HF-decision matrix construction, the non-membership degree information is not considered, even though more decision-making information was comprised for completely reflecting the experienced tourists’ hesitation, which might result in loss of some decision-making information. [18] introduced a technique for classifying cultural tourism attractions grounded on tourists’ preferences detected by their citywide travel trajectories. For detecting the four categories of cultural tourism, cluster analysis was performed. The study contributes an innovative methodology for differentiating cultural tourism attractions, which aids in targeting potential tourists. Nevertheless, social media platforms could only cover a part of the population (mainly the younger generation), causing a biased user sample. [19] projected community tourism in emerging cities by applying the gamification mechanism for enhancing the standard of products and services in tourism for offline as well as online operations. As per the outcomes, higher sensitivity and accuracy were achieved for the recommended location for the tourists, entrepreneurs, and tourism development agencies while utilizing the applications. However, the main drawback in this work was that the RS takes more time to provide the output as a result of the large number of required parameters. [30] projected a personalized list of tourist attractions for every single tourist grounded on the similarity of users’ desires and interests, reputation, trust, relationships, and social communities. The developed technique’s superiority was shown by the outcomes over other common approaches. Yet, the use of user social relations along with their opinion in tourism recommender systems have been less considered.

3. Proposed Methodology for the Cultural Tourism Accelerator in Riyadh

In 2000, Riyadh was chosen as the cultural capital of the Arab world by the United Nations Educational, Scientific and Cultural Organization (UNESCO) [31]. The city encloses various cultural centers. Therefore, a tourism model based on the MQ-LSTM is developed to enhance tourism in Riyadh to spread its culture. The proposed framework utilized festival events data about Riyadh, the geographical location of historical sites, reviews about the historical sites from social media platforms, and reviews about the restaurants in Riyadh City. These data are analyzed in the proposed framework to give better RS for tourists. The festival events data about Riyadh City provides insightful information about the right time to visit the place. This makes the tourism and the traditional value of the city effective with a greater number of visitors. Moreover, the consideration of geographical location and the social media review data provides a brief view of others’ opinions and perspectives about the particular location while visiting. This helps the upcoming visitor to visit the best historical sites and best-reviewed sites to explore. In addition, the usage of restaurant reviews helps the tourists to accommodate while visiting. The reviews about the restaurants near the historical sites make them plan an effective accommodation. The consideration of all these data makes the tourism RS more effective by helping the visitors by providing all the insightful details about the city. Primarily, the spectral grouping of the festival events was performed in the proposed work based on the GIS map. Then, based on the location and historical places, tree construction was carried out to effectively map the location with the historical site. In the meantime, missing values present in the dataset were removed for effective recommendation. Then, the categorical data present in the dataset were converted into the numerical format by using a pre-trained dictionary named Bidirectional Encoder Representations from Transformers (BERT). Following this, the features from the dataset as well constructed tree were extracted and inputted into the RS. Along with this input, the social network review data and restaurant reviews were taken, and keywords related to the historical sites were extracted using the Rapid Automatic Keyword Extraction (RAKE) technique. Then, BERT was imputed to transform the categorical data followed by score value calculation based on term frequency-inverse document frequency (TF-IDF) contents. These were further inputted to RS to provide effective tourism recommendations by considering all these factors. Figure 1 elucidates the proposed work’s block diagram.

3.1. Input Data

In the proposed system, the GIS map $\left(M\right)$; festival and events dataset $\left(F\right)$; and social networking site (SNS) data, namely, Google reviews $\left(R\right)$ and news $\left(N\right)$ data were taken as input data $\left(I\right)$ and expressed as:

$$I=〈F,M,R,N〉$$

(1)

Primarily, the locations of the mapped historical sites and the cultural events along with the festivals near the historical sites data were processed. This process was for suggesting nearby cultural events and festivals, and also for mitigating the waiting time of tourists who wish to know about the culture of Riyadh. To identify the nearest points, the locations of historical sites and cultural events were combined. By utilizing the KIC clustering technique, the nearest nodes (i.e., nearest events of historical sites) were grouped as a single cluster. Moreover, the Chameleon clustering was considered, owing to its efficiency in constructing graph-based clustering. However, the Chameleon clustering struggles in forming a single group with a few data values. Therefore, to overcome this problem, the Kringing interpolation (KI) technique is wielded in Chameleon clustering. The overview of KIC is depicted in Figure 2.

Graph construction: Initially, in KIC clustering, the sparse graph is constructed for the location data in $〈F,M〉$ grounded on the K-nearest neighbors (KNN), and the location data is signified as:

$$L_{hs}=\left\{L_1,L_2,\dots ,L_z\right\}$$

(2)

$$L_{ce}=\left\{L_1,L_2,\dots ,L_x\right\}$$

(3)

where in $L_{hs},L_{ce}$ symbolizes the historical site and cultural event location set, and $L_z,L_x$ represents the locations of the $z^{th}$ historical site and $x^{th}$ cultural event. In the KNN-based graph generation, a node $\left(L_{hs}\right)$ is connected with its $\kappa$ closest neighbor $\left(L_{ce}\right)$ by an edge that is given weight equal to the inversion of their distance. Hence, the graph constructed is defined as $G$. After that, to form spectral clusters, the partitioning and merging of the graph process are carried out. We present a pseudocode of proposed KIC grouping as Algorithm 1.

Partitioning: From the generated graph $G$, the graph separation is performed with a graph partitioning algorithm, such that each cluster $Q_c$ is divided into sub-clusters $Q_u,Q_v$. This splitting is executed for minimizing the weight of the edges by cutting the edges. Therefore, the initial sub-clusters $Q_u,Q_v$ are obtained with the graph partitioning algorithm and the edge cut is signified as $\Im \left(Q_u,Q_v\right)$.

Merging: The relative inter-connectivity $\left(\alpha \right)$ and relative closeness $\left(\beta \right)$ are determined after the sub-clusters are obtained. The KIC clustering merges the pair of sub-clusters for which $\alpha$ and $\beta$ are high. Thus, to merge and form a cluster even with a small number of data, the KI-based relative inter-connectivity is estimated as:

$$\alpha \left(Q_u,Q_v\right)=\frac{2.\chi \left(Q_u,Q_v\right)\left|\Im \left(Q_u,Q_v\right)\right|}{\left|{\Im }^{Q_u}\right|+\left|{\Im }^{Q_v}\right|}$$

(4)

where $\chi \left(.\right)$ symbolizes the distance function, and ${\Im }^{Q_u}$,${\Im }^{Q_v}$ represents the edge cut of sub-clusters. After that, the relative closeness betwixt the sub-clusters is given as:

$$\beta \left(Q_u,Q_v\right)=\frac{\left(\left|Q_u+Q_v\right|\right)W_{\Im \left(Q_u,Q_v\right)}}{\left|Q_u\right|W_{{\Im }^{Q_u}}+\left|Q_v\right|W_{{\Im }^{Q_v}}}$$

(5)

where $W_{{\Im }^{Q_u}},W_{{\Im }^{Q_v}}$ indicates the edges’ average weights of the edges that belong to the min-cut bisector of sub-clusters $Q_u,Q_v$, respectively, and $W_{\Im \left(Q_u,Q_v\right)}$ indicates the edges’ average weight that connects the vertices in $Q_u$ to the vertices in $Q_v$. Then, after the $\alpha \left(Q_u,Q_v\right)$ and $\beta \left(Q_u,Q_v\right)$ are determined, the sub-clusters with $\max \left(\alpha ,\beta \right)$ are merged to form spatial clusters, which are denoted as:

$$Q_c=\left\{Q_1,Q_2,....,Q_{\delta }\right\}$$

(6)

where $Q_{\delta }$ signifies the ${\delta }^{th}$ cluster with the respective historical site and its nearest possible cultural events.

Algorithm 1 Pseudocode of proposed KIC grouping

Input: Locations $\left\{L_{hs},L_{ce}\right\}$ Output: grouped locations Begin    Initialize locations, distance function $\left(\chi \right)$    For input locations do       Construct graph $G$ with KNN       Perform partitioning of $G$ to $\Im \left(Q_u,Q_v\right)$       //Merging of sub-cluster pairs       Estimate KI relative interconnectivity closeness $\alpha \left(Q_u,Q_v\right)$       Estimate relative closeness $\frac{\left(\left|Q_u+Q_v\right|\right)W_{\Im \left(Q_u,Q_v\right)}}{\left|Q_u\right|W_{{\Im }^{Q_u}}+\left|Q_v\right|W_{{\Im }^{Q_v}}}$          If $\left(\alpha ,\beta ==high\right)$ then             Merge sub-clusters          Else             Compare other sub-clusters          End If       End For    Return cluster $Q_c$ End

3.2. Tree Construction

After the locations are spatially clustered, by utilizing the proposed Cramer’s’ V Correlated-Minimal Spanning Tree (CVC-MST), the historical site with its nearest possible events is formed as a tree. This process is to arrange the historical site location and its nearest events in order. For tree construction of the historical site based on the nearest events, the minimal spanning tree is selected, owing to its efficiency in path forming from the clustered graphs. However, the tree construction takes more time to create a child by looping repeatedly. Therefore, to solve this problem, the Cramer’s V Correlation (CVC) in a minimal spanning tree is introduced.

Sorting: From the graph $Q_c$, the edges are sorted in increasing order such that the parent node is formed as the historical site $\left(H\in Q_c\right)$; afterward, the child nodes $\left\{c_v\in Q_c\right\},v=1,2,\dots ,p$ are kept added to the tree grounded on the CVC correlation betwixt the parent and child nodes.

Correlation: The CVC correlation $\left(C\left(H,c_v\right)\right)$ between $H$ and $c_v$ is equated to form the child without looping as:

$$C\left(H,c_v\right)=\sqrt{\frac{{\chi }^2\left(H,c_v\right)}{p\times \omega }}$$

(7)

where ${\chi }^2\left(H,c_v\right)$ implies the chi-squared test result value, and $p,\omega$ signifies the total number of child nodes to be formed and the degree of freedom, correspondingly.

Construction: With the determined correlation value, the child node having a higher correlation is directly added to the parent node. Hence, by following this procedure, a reliable tree structure is formed. The developed tree is symbolized as $T$. After the tree $T$ is constructed, the features of nodes, such as edge weight, number of vertices, end-node, and starting nodes are extracted. The extracted features $\left(e\right)$ are given as:

$$E_e=\left\{e_1,e_2,e_3,e_4\right\}$$

(8)

where $E_e$ illustrates the feature set.

3.3. Event Data Processing

After the location of historical sites and cultural events features are obtained, the attributes of the cultural events are processed and the attribute set is represented as $A_n$, which encloses the name, month, and date of the events. After $A_n$ is extracted, the data are pre-processed with missing value removal. This process is performed to mitigate the misclassification. Then, the pre-processed results $\left(P_m\right)$ is given as

$$P_m=\left\{P_1,P_2,....,P_X\right\}$$

(9)

where $P_X$ epitomizes the preprocessed cultural event data. As the dataset $\left\{P_1,P_2,....,P_X\right\}$ contains the string values, such as month and festival name, it is then vectorized with the Bi-directional Encoder Representation from Transformers (BERT).

Tokenization: The BERT’s input layer is the pre-processed cultural event data $P_m$. In the input layer of BERT, the tokens of each string $\left\{P_1,P_2,....,P_X\right\}$ are determined and represented within different classes (i.e.,$\left[cls\right]$), and each $\left[cls\right]$ is separated by the separation function $\left[sep\right]$. The tokens of text are exemplified as:

$$t_m=\left[t_1,t_2,\dots ,t_X\right]$$

(10)

where $t_X$ is the $X^{th}$ tokenized word.

Embedding layers: After that, the tokenized input $\left(t_m\right)$ are given to the embedding layer, which performs token embedding, segment embedding, and position embedding. Embedding is the representation of words in vector form.

BERT transformer layers: The token, segment, and position embedding are summed up and given to the BERT transformer encoder layer, which contains 12 transformers with 12 attention mechanisms and millions of parameters. The encoder encodes the string values, whereas the decoder determines the significant keywords and gives contextual embedding as:

$$c_m=\left\{c_1,c_2,.....,c_X\right\}$$

(11)

Output layers: After the strings are processed by the transformer layers, it is then given to output layers to obtain the embedded output. In the output layer of BERT, strong links between the sentences are determined, which aids in word embedding. The output layer contains a simple classifier model with a fully connected layer and sigmoid activation. The loss $\left(\mathcal{l}\right)$ in the classifier output is computed as:

$$\mathcal{l}=\frac{1}{2}\left(\gamma -{\displaystyle \sum _{m=1}^X{\hat{c}}_m}\right)$$

(12)

where $\lambda$ illustrates the target word embedding score, and $\hat{c}$ is the output word embedding, which is computed utilizing:

$${\hat{c}}_m={\displaystyle \sum _{\wp =1}^{hid}\frac{1}{1+\exp \left(-H_{\wp }\right)}}$$

(13)

where $hid$ signifies the total number of hidden neurons in the fully connected layer and $H_{\wp }$ is the output of the ${\wp }^{th}$ hidden layer. Therefore, by utilizing BERT, word embedding is obtained in which the strings are converted to the corresponding vector format $\left({\hat{c}}_m\right)$.

3.4. Processing with SNS Data

For a better understanding of the historical significance and public opinion of the historical sites, the SNS Google reviews as well as news $\left(R,N\right)$ are extracted with the Google External API and processed.

Processes with the Review Data

While recommending a tourist place, the opinion of previous visitors takes a vital role as the visitors give their opinion on the previously visited places. This will be helpful while planning for a travel to historical sites. Here, the review content set $\left(R_q\right)$ is extracted, which is expressed as:

$$R_q=\left\{R_1,R_2,\dots ,R_{\epsilon }\right\}$$

(14)

where $R_{\epsilon }$ illustrates the ${\epsilon }^{th}$ extracted review.

(i) Keyword Extraction

Only the keywords are extracted to recommend the opinion on the visited cultural sites since the reviews have large text contents. The keywords are extracted with the unsupervised Rapid Automatic Keyword Extraction (RAKE) algorithm. For determining the most relevant words or phrases in a review text, the RAKE uses a list of stop words and phrase delimiters.

(a) Locating keywords

In RAKE, $R_q$ is split into arrays of words by the word delimiters. After that, the candidate keywords are located by removing the stop words like articles and prepositions. The words appearing between two stop-list words and punctuation marks are marked as candidate keywords. Thus, the identified keywords set $\left(k_w\right)$ are represented as:

$$k_w=\left[k_1,k_2,\dots ,k_y\right]$$

(15)

Here, $k_y\in R_q$ specifies the $q^{th}$ located candidate keyword in review $R_q$.

(b) Building score-weight matrix

After locating the keywords, the score value of the located keywords is determined, and a score-weight matrix is constructed by calculating the word degree $\left(D_w\right)$, word frequency $\left(f_w\right)$, and ratio of the degree to frequency $\left(d{f}_w\right)$. Word frequency $\left(f_w\right)$ specifies the repeated occurrence of a term in the review content, and $D_w$ implies the degree of co-occurrence of the words in the review content. Then, the ratio of degree to frequency is given as:

$$d{f}_w=\frac{D_w}{f_w}$$

(16)

where $d{f}_w$ implies the words that predominately occur in longer candidate keywords. Therefore, with these score values, the score-weighted-matrix $\left(mx\right)$ is formed.

(c) Extracted keywords

From the matrix $\left(mx\right)$, RAKE looks for pairs of keywords that adjoin one another. After that, a new keyword is formed as the combination of such keywords with its interior stop words. By summing the member keyword scores present in the matrix $\left(mx\right)$, the score of the new combined candidate keyword is estimated. The scores of the new keywords are given as $\left(s_1,s_2,..,s_{\varphi }\right)$.

After the candidate keywords are scored, the top $\aleph$ scored words are selected as the keyword $\left(T_{\hslash }\right)$ of the review content and extracted. This is represented as:

$$T_{\hslash }=\left[T_1,T_2,\dots ,T_{\mathsf{\Omega }}\right]$$

(17)

where $T_{\mathsf{\Omega }}$ illustrates the ${\mathsf{\Omega }}^{th}$ extracted keyword. Then, for computational convenience, the extracted text is converted to embedding with BERT and is signified as ${\hat{T}}_{\hslash }$.

3.5. Processing of News Data

After the review content is processed, the Google news data $\left(N\right)$, which contains the news data about the historical sites, is processed to know its historical significance. This process is executed to avoid suggesting the tourist spot only grounded on the reviews as the reviews cannot be always trusted. The extracted news set is expressed as:

$$N_{ƛ}=\left\{N_1,N_2,\dots ,N_{\mathsf{\Gamma }}\right\}$$

(18)

where ${\mathsf{\Gamma }}^{th}$ news data is notated as $N_{\mathsf{\Gamma }}$.

(a) Determining score value

After the news data are extracted, the news data’s frequency is determined, and the score value for the news feed corresponding to the historical site is estimated. By doing this, if the historic site appears frequently in the news but has no satisfactory reviews, it can be recommended as a better tourist spot. The score value is determined with the TF-IDF.

Step-1: Term frequency estimation

The TF-IDF determines the weight value for the historical site-centric words and determines the frequency of such words in the extracted news data. The term frequency of each word is determined utilizing:

$$\zeta =\frac{N\left(nw\right)}{\underset{nw\in N_{ƛ}}{\max }\left(N\left(nw\right)\right)}$$

(19)

where, $N$ illustrates the frequency of historical site terms $\left(nw\right)$ in news.

Step-2: Determine inverse document frequency

After the term frequency is determined, the inverse document frequency (IDF) is estimated for estimating the $nw$ in the overall news content. The IDF $\left(\psi \right)$ is expressed as:

$$\psi =\ln \left(\frac{\mathsf{\Gamma }}{U\left(nw\right)}\right)$$

(20)

where $\mathsf{\Gamma }$ specifies the total number of news, and $U\left(.\right)$ implies the total number of news contains the term $\left(nw\right)$.

Step-3: Calculate the product

After that, the TF-IDF score $\left(S_{hs}\right)$ is estimated centered on the product of term frequency and IDF as:

$$S_{hs}=\zeta \times \psi$$

(21)

Hence, the determined score values of the historical sites are represented as:

$$S_{hs}=\left\{S_1,S_2,..,S_{\eta }\right\}$$

(22)

where $S_{\eta }$ signifies the score value estimated for the ${\eta }^{th}$ historical site. From the score values estimated, the higher score values are regarded as the often-repeated term, and the historical site that has a higher score can be recommended to tourist planners for CT.

3.6. Data Fusion

After all the data are processed, the data are fused for training and testing the recommendation model. Thus, the fused data are given as:

$$Y=\left[L,K\right]$$

(23)

$$\left.\begin{array}{l}L=E_e\oplus {\hat{c}}_m\\ K={\hat{T}}_{\hslash }\oplus S_{hs}\end{array}\right\}$$

(24)

where $Y$ signifies the fused data, which comprises the details of features of location, time and name of the festival, and score value, as well as keywords of the respective Riyadh historical sites.

3.7. Recommendation

Lastly, the fused data $Y$ is given to the MQ-LSTM to give recommendations grounded on the user’s location and preference. Long short-term memory (LSTM) is selected, owing to its effectiveness in providing context recommendations. However, LSTM has the disadvantage of having a vanishing gradient problem without the proper activation of neurons. Thus, to solve this problem, the multi-quadratic (MQ) activation is applied in the LSTM model. Figure 3 elucidates the MQ-LSTM’s architecture.

Input: The MQ-LSTM receives $Y$ along with spontaneously generated two vectors within the MQ-LSTM (i.e., hidden state $\left(sh\right)$ and the cell state $\left(cs\right)$), which is taken as the input at the time instant $\left(\tau \right)$. Given the three input vectors $\left(Y,sh,cs\right)$, the MQ-LSTM regulates these vectors through the gates like the input gate $\left({\xi }_j\right)$, forgot gate $\left({\omega }_j\right)$, and output gate $\left({\upsilon }_j\right)$. The input gate, an output gate, and a forget gate control the state of memory cells. The MQ-LSTM process is mathematically represented as:

$${\xi }_{\tau }=\rho \left({\varpi }_j\left[s{h}_{\tau -1},Y_{\tau }\right]+B_j\right)$$

(25)

$${\omega }_{\tau }=\rho \left({\varpi }_{\omega }\left[s{h}_{\tau -1},Y\right]+B_{\omega }\right)$$

(26)

$${\upsilon }_{\tau }=\rho \left({\varpi }_{\upsilon }\left[s{h}_{\tau -1},Y\right]+B_{\upsilon }\right)$$

(27)

$$c{s}_{\tau }={\omega }_{\tau }\cdot c{s}_{\tau -1}+{\xi }_{\tau }.\overline{c{s}_{\tau }}$$

(28)

$$\overline{c{s}_{\tau }}=\tanh \left({\varpi }_{cs}.\left[s{h}_{\tau -1},Y\right]+B_{cs}\right)$$

(29)

where ${\varpi }_{\varpi },{\varpi }_{\upsilon },{\varpi }_j$ illustrate the weight values of the input, forgot, and output gates, respectively; $s{h}_{\tau -1}$ specifies the previous hidden state instant; $B$ indicates the bias value; and $\rho$ signifies the MQ activation, which is equated as:

$$\rho =\frac{\exp \left(VY+W\right)}{1+\exp \left(VY+W\right)}$$

(30)

where $W,V$ are random variables, which are changed to form multi-quadratic activation functions. ${\xi }_j,{\omega }_j,{\upsilon }_j$ are chained together to form an MQ-LSTM cell, where each cell in MQ-LSTM acts as the memory module. The information from the previous memory module $\left(s{h}_{\tau -1}\right)$ is passed to the next cell to predict the final contextual output.

Output: Lastly, the MQ-LSTM output is determined from the hidden state $\left(s{h}_{\tau }\right)$ as:

$$s{h}_{\tau }={\upsilon }_{\tau }\times \tanh \left(c{s}_{\tau }\right)$$

(31)

where $s{h}_{\tau }$ epitomizes the final output, which is the hidden state of the final cell. This final value represents the final recommended contextual information of the CT spot. By doing this contextual information, the tourism planners can be well aware of the tourist spot and its nearby possible cultural events. Therefore, during peak visitor time, a tourist does not have to wait for a long time to visit a historical site. The Proposed pseudocode of MQ-LSTM is given as Algorithm 2.

Algorithm 2 Pseudocode of proposed MQ-LSTM

Input: Fused data $Y=\left[L,K\right]$ Output: contextual recommendation Begin    Initialize states $\left(sh,cs\right)$, gates $\left(\xi ,\omega ,\upsilon \right)$, time instance $\left(\tau \right)$    Set initial $sh=0,cs=0$    For time instant $\left(time\le \tau \right)$do       Determine cell states, input gates and forget gates $\overline{c{s}_{\tau }},{\xi }_{\tau },{\omega }_{\tau }$        Perform MQ activation $\rho$        Update cell state with $c{s}_{\tau }={\omega }_{\tau }\cdot c{s}_{\tau -1}+{\xi }_{\tau }.\overline{c{s}_{\tau }}$        Estimate output gate $\left({\upsilon }_{\tau }\right)$ value    End For    Return output $s{h}_{\tau }$ End

For the use of CT planners, this tourism recommendation along with the spatial map developed with ArcGIS is stored in a database.

3.8. Database Security

Since the recommendation model comprises more information about cultural event details, the database cannot be accessed by legal users only. There are a lot of chances for the information to be corrupted by malicious users (hackers) who might change the data stored in the database. Hence, to avoid this problem, the user is authenticated for utilizing the data stored in the database. A legitimate user $\left({\lambda }_{user}\right)$ registers their details, such as name $\left(\Re \right)$, current location $\left(\mu \right)$, date $\left(\mathsf{\Phi }\right)$, preference $\left(P\right)$, and type of trip package $\left(C\right)$. This is mathematically represented as:

$${\lambda }_{user}\stackrel{reg}{{\displaystyle \to }}〈\Re ,\mu ,\mathsf{\Phi },P,C〉$$

(32)

During the registration process, a hash code is formed grounded on the registered time $\left(rt\right)$ and name of the user $\left(\Re \right)$. With the LSH, the hashcode is generated. The LSH algorithm is designed to generate hash digests for the message $\left(rt,\Re \right)$, such that low distances betwixt the digests indicate that the corresponding messages are likely to be similar. Primarily, the messages $\left(rt,\Re \right)$ are processed by utilizing a sliding window of length $\nu$ and populates a hash bucket $\left(hb\right)$. After that, the messages $\left(rt,\Re \right)$ are passed through the hash function of SHA-256 to obtain the hash value. This hash value is stored in the bucket $\left(hb\right)$. After that, by estimating the quartile points from the bucket $\left(hb\right)$, the bucket $\left(hb\right)$ is sorted. Then, the hashcode digest header is computed. The first three bytes of the SHA-256 hash are reserved for the digest header. The header of the digest contains the following parts. The first byte $\left(Z_1\right)$ contains the checksum of the $\left(rt,\Re \right)$ string byte with the modulo function. The second byte $\left(Z_2\right)$ is computed from the logarithm of the string $\left(rt,\Re \right)$ length $\left(st{r}_{len}\right)$, which can be represented as:

$$Z_2=\log \left(st{r}_{len}\right)$$

(33)

The third byte $\left(Z_3\right)$ is the result obtained grounded on the second-byte ratio $\left(K_2\right)$ and first-byte ratio $\left(K_1\right)$ as:

$$Z_3=K_1<<<4|K_2$$

(34)

where

$$K_1=\left(Z_1\ast 100/Z_3\right)\mathrm{mod}16$$

(35)

$$K_2=\left(Z_2\ast 100/Z_3\right)\mathrm{mod}16$$

(36)

After the header is obtained, the digest body is constructed from the bucket array. In this, by reversing the order in the bucket, the bucket is read. The last element of the string is read first, and the first element is read last. Lastly, the reversed elements are converted to hex form and are appended into the SHA-generated hashcode. Therefore, the generated hashcode at the user and the administer side is illustrated as $\left(i_{hash}\right)$.

Hashcode matching: The matching process takes place during the login after the hashcode $\left(i_{hash}\right)$ is generated on the administrator and user side. During the login phase, the hashcode $\left(i_{hash}\right)$ generated during the registration time is utilized. If the $i_{hash}$ provided by the user side and the administrator side during the login phase are the same, then the user is considered as a legitimate user and is eligible to access the data stored in the database. However, when a malicious user tries to obtain the data using $\Re$ of a legitimate user and at a different time, a more similar hash digest will be produced but not the same hash value. Therefore, such requests will be declined in the proposed model.

Hence, it is analyzed that with the proposed CT accelerator model, the improvement in tourism can be achieved by the increased security and attractiveness of the recommendation model.

4. Results and Discussion

In this section, the detailed exploration of the final outcome of the proposed tourism RS is explained. The proposed methodology is employed in the working platform of PYTHON utilizing the data collected from user reviews and ratings about particular places of Riyadh on social media platforms, GIS map, and the Riyadh Restaurants 20K dataset.

4.1. Dataset Description

In this section, the dataset used for the proposed RS and the corresponding examples of data processing steps are detailed for better analysis. The dataset used for the proposed work was the Riyadh Restaurants 20K dataset. It contains restaurant details about all the restaurants in Riyadh with their latitude and longitude locations, prices, likes, photos, tips, ratings, and addresses to provide better recommendations. Also, the data from user reviews and ratings on the popular places in Riyadh were collected from social media sites by using web scraping. Web scraping automatically extracts data from social media platforms using a web scraper, which is a software tool to extract data. After collecting the required data, preprocessing was carried out to remove the missing columns. From the total number of data collected, 80% of the data was used for training, and the remaining 20% of the data was used for testing purposes to show the effectiveness of the proposed framework.

4.2. Performance Analysis

In this phase, the performance assessment of the proposed recommendation model was compared with the existing algorithms and related works to show the efficacy of the proposed framework. The assessment for proposed spectral clustering using KIC was compared based on grouping time, and the proposed MQ-LSTM was compared with prevailing classifiers to make predictions. This was analyzed by means of metrics, such as RR, precision, recall, sensitivity, training time, FDR, and NPV. The performance analysis of each phase is discussed further.

Performance Analysis of Spectral Grouping

The proposed KIC’s performance was analogized with other conventional methodologies like Chameleon clustering, Robust Clustering using Links (ROCK), Balanced Iterative Reducing and Clustering using Hierarchies (BIRCH), and Clustering Using Representatives (CURE).

Table 1. Performance validation of the proposed KIC.

Techniques	Grouping Time (ms)
Proposed KIC	36,017
Chameleon	43,129
BIRCH	53,012
CURE	65,894
ROCK	78,653

illustrates the time consumed for grouping by the proposed KIC and the existing models. For proving the spectral grouping’s efficacy, the time taken for grouping must be low. The proposed mechanism consumed 36,017 ms to group the locations of the sites. However, the existing models required higher grouping time than the time consumed by the proposed model. The inclusion of Kriging interpolation in the existing Chameleon technique improved the grouping process to a greater extent.

4.3. Performance Analysis of the Recommendation System

This phase analysed the proposed MQ-LSTM’s efficacy by comparing its outcomes with the existing techniques like LSTM, recurrent neural network (RNN), deep neural network (DNN), and artificial neural network (ANN).

The recommendation output of the proposed MQ-LSTM centered on a long tail plot is illustrated in Figure 4. This plot explores the popular trends in user–item interaction data grounded on the IDF score values of the historical sites. The head refers to the portion of historical sites that often have a high volume of recommendations. Even though the long tail contained the majority of items, recommendations for those items only account for a small portion of all recommendations. Therefore, the proposed model exhibited accurate recommendations along with user satisfaction and convenience.

Figure 5 elucidates the performance of the proposed MQ-LSTM grounded on the recommendation rate (RR), precision, F-measure, specificity, and recall. The red lines indicate the difference between the high impact and low impact while the blue line indicates the number from highly recommended sites to low. The proposed model attained an RR of 97.22%, whereas the existing models attained a lower RR of 96.32% for LSTM, 94.3% for RNN, 91.25% for DNN, and 90.6% for ANN. Likewise, the proposed model obtained higher precision, F-measure, specificity, and recall values of 97.7%, 97.98%, 94.94%, and 98.27%, respectively. Thus, the proposed MQ-LSTM had better performance, owing to the effective usage of location information, user reviews, and IDF scores; therefore, it enhanced the historical site recommendations in Riyadh city.

Table 2. Performance analysis of the proposed MQ-LSTM.

Techniques	MR (%)	NPV (%)	MCC (%)
Proposed MQ-LSTM	2.78	96.15	93.53
LSTM	3.68	94.95	92.52
RNN	5.7	93.14	91.43
DNN	8.75	91.39	90.31
ANN	9.4	89.35	88.17

demonstrates the proposed MQ-LSTM’s performance grounded on the misclassification rate (MR), negative predictive value (NPV), and Mathews correlation coefficient (MCC). The NPV and MCC of the proposed method exhibited increases of 1.2% and 1.01%, respectively, when analogized with the existing LSTM. Likewise, the NPV and MCC values of other existing methods also provided lower performance. In analyzing Table 1, the proposed method achieved a lower MR of 2.78%. Overall, the usage of multi-quadratic activation in the LSTM recommender showed superior recommendation output.

Figure 6 depicts the false negative rate (FNR), mean absolute error (MAE), and false discovery rate (FDR) of the proposed and existing recommendation models. These measures indicated the negative recommendations made by the model and should be lower to prove the better performance of the model. Accordingly, the proposed MQ-LSTM achieved the lowest values of 0.023 of FDR, 0.0521 of MAE, and 0.0173 of FNR. Hence, it was revealed that the MQ activation was more reliable and superior to the existing recommendation approaches.

Table 3. Performance comparison regarding training time.

Techniques	Training Time (ms)
Proposed MQ-LSTM	45,005
LSTM	53,487
RNN	67,543
DNN	78,643
ANN	90,049

displays the time taken for training by the proposed MQ-LSTM and existing methods. It exhibited a decrease in training time of 8482 ms in comparison with the existing LSTM, 22,538 ms in comparison with RNN, 33,638 ms in comparison with DNN, and 45,044 ms in comparison with ANN. This shows that the recommendations made by the proposed technique were more accurate than the conventional frameworks.

Figure 7 exhibits the proposed MQ-LSTM and the existing model’s ROC curve. An effectual technique of assessing the recommendation model’s quality or performance is named ROC. The dotted lines indicate the chance level. By plotting true positive rate (TPR) against false positive rate (FPR), it was constructed. It was clear from the ROC curve that superior performance was exhibited by the proposed model in recommending the historical sites in Riyadh to the existing recommendation models.

Comparative Measurement with the Literature Papers

Here, the proposed methodology’s performance was contrasted with conventional recommendation approaches discussed in the literature survey.

Figure 8 depicts the performance of the proposed MQ-LSTM-based RS with existing systems developed by [23,24,26]. From the analysis, it is clear that by obtaining a high RR of 97.22%, the proposed model exhibited superior performance. The recommendations by the prevailing approaches were comparatively lower than the proposed model. Therefore, the comparison revealed that considering the IDF score values, Google reviews and spectral grouping improved the proposed RS for the tourism system for Riyadh’s historical sites.

Discussions: Here, the experimental evaluation of the proposed work by utilizing the aforesaid datasets is discussed. The proposed system’s performance was analyzed concerning a long tail plot, as shown in Figure 4. It explores the popular trends in user-item interaction data grounded on the IDF score values of the historical sites. Then, the rate of recommendation, precision, recall, f-measure, specificity, and recall are discussed in Figure 5. This mentions how effectively the proposed framework recommends the place for tourists based on the comparison of attained true positive, true negative, false positive, and false negative values. The higher the recommendation rate, the better the outcomes. With this perspective, the proposed recommender system attained the RR of 97.22%, whereas compared with existing techniques as recommendation rate, they attained lower values. Likewise, the proposed model obtained higher precision, F-measure, specificity, and recall values of 97.7%, 97.98%, 94.94%, and 98.27%, respectively. Thus, the proposed MQ-LSTM had better performance owing to the effective usage of location information, user reviews, and IDF scores; therefore, it enhanced the historical site recommendations in Riyadh City. Based on error metrics, the proposed MQ-LSTM achieved the lowest values of 0.023 (FDR), 0.0521 (MAE), and 0.0173 (FNR), as shown in Figure 6. Hence, it was revealed that the MQ activation was more reliable and superior to the existing recommendation approaches. Regarding the time taken for recommendation in Table 2, the proposed MQ-LSTM took only 8482 ms to provide a recommendation. On the other hand, by using existing techniques for recommendation, they gave higher training time. Similarly, from the related recommendation system comparison, the RR of the proposed MQ-LSTM-based RS was higher than the prevailing technique. From the overall discussion and analysis, it was concluded that with the utilization of various data sources as well as effective grouping of data and feature extraction, the proposed model improved the CT of Riyadh when compared to existing systems.

5. Conclusions

A CT accelerator based on MQ-LSTM with database security is proposed in this paper. The proposed work underwent various steps like spectral grouping, attribute extraction, tree construction, word embedding, keyword extraction, score value estimation, and finally a recommendation grounded on the collected details. After that, for evaluating the proposed mechanism’s performance, the performance, as well as the comparative analysis, were executed. The developed framework could handle various uncertainties and render more promising results. From the analysis, the proposed recommendation model achieved an RR of 97.22%, precision of 97.7%, recall of 98.27%, and MAE of 0.0521. The error rate attained by the proposed recommendation system was very low (2.78%). In addition, the FDR and FNR rates were also very low, at about 0.023 and 0.0173. Also, to cluster the location data, the proposed KIC took 36,017 ms, which was very low. From these results, it is evident that the spectral grouping of historical sites based on the location and the consideration of the hybrid data, such as restaurant details, and historical sites of Riyadh City with event details makes the tourism recommendation more effective with improved accuracy. Along with this, the consideration of reviews about the historical sites and restaurants scrawled from the social media sites makes the recommendation more effective. This helps to improve tourism in Riyadh City, which benefits the country. Also, this effective RS helps the visitors to make effective plans to visit the city on time of events to explore more about the city. Overall, the proposed system performed well in contrast to the comparable systems concerning all metrics. However, the usage of BERT embedding in this framework makes the system more expensive and it requires more computation. This affects the overall computation complexity of the proposed RS. In the future, this work will be improved by considering factors, namely, (i) the accessibility and utilizing automatic computation of itineraries as well as context-aware data for developing an advanced cultural tourism RS, and (ii) the usage of the explicit semantic analysis (ESA) mechanism for the traveler-centric recommender application, as well as augmenting the data in every single point of interest. Also, integrating more than one technique is essential for assessing the POI popularity in social media networks utilizing the SA system for effective RS. Thus, the work will be extended in the future by using more advanced and less complex embeddings, collecting details regarding all the tourist spots in Riyadh, and providing a tourism database model for providing secure and reliable RS.