Recommendation System for Sustainable Day and Night-Time Cultural Tourism Using the Mean Signed Error-Centric Recurrent Neural Network for Riyadh Historical Sites

Abstract

To accommodate user-specific requirements and preferences, a travel Recommendation System (RS) gives a customized place of interest. The prevalent research did not provide solutions to some essential situations for cultural tourism, including relevant time, environmental conditions, and stay places. Thus, the existing RS models led to unreliable cultural tourism recommendations by neglecting essential factors like personalized itineraries, environmental conditions of the cultural sites, sentiment analysis of the hotel reviews, and sustainable cultural heritage planning. To overcome the above factors, a day- and night-time cultural tourism RS utilizing the Mean Signed Error-centric Recurrent Neural Network (MSE-RNN) is proposed in this paper. The proposed work develops an efficient RS by considering historical data, Geographic Information System (GIS) map location, hotel (stay place) reviews, and environmental data to access day and night cultural tourism. First, from the Geographic Information System (GIS) map and hotel data, the historical and hotel geolocations are extracted. Currently, these locations are fed to Similarity-centric Hamilton Distance-K-Means (SHD-KM) for grouping the nearest locations. Next, hotels are ranked utilizing the Tent Mapping-centric Black Widow Optimization (TM-BWO) approach centered on the locations. In addition, using Bidirectional Encoder Representations from Transformers (BERT), the essential keywords from the historical geo-locations are embedded. In the meantime, the sites’ reviews and timings are extracted from Google. The extracted reviews go through pre-processing, and the keywords from the pre-processed data are extracted. For the extracted keywords, polarity is calculated centered on the Valence-Aware Dictionary for Sentiment Reasoning (VADER). Concurrently, utilizing the Reference-centric Pearson Correlation Coefficient (R-PCC), the timings of the sites are segregated. Lastly, for providing a recommendation of tourist sites, the embedded words, ranked hotels, and segregated timings, along with the pre-processed environment and season data, are fed to the MSE-RNN classifier. At last, the experimental evaluation verified that other recommendation systems were surpassed by the proposed approach.

1. Introduction

Since the tourism sector generates job openings and business opportunities, it is an essential component of the social and economic activity of a plethora of countries [1]. It is also regarded as a tool for fighting poverty with initiatives like the World Tourism Organization’s Sustainable Tourism-Eliminating Poverty (ST-EP) program. In recent days, for the tourism industry, cultural tourism has steadily become a significant source of income [2]. Accounting for an estimated 40% of global tourism by the United Nations World Tourism Organization (UNWTO), cultural tourism and heritage tourism are considered main elements of international tourism consumption (UN Tourism, 2016) [3]. A lot of tourism development has been grounded in the consumption of cultural sites and attractions, namely museums, art galleries, and heritage sites, mainly in the most important cities globally [4].

One of Saudi Arabia’s most important heritage sites is Riyadh. Many archeological sites of historical significance surround Riyadh. To promote tourism, the Saudi government hosts the Riyadh season with a diversity of events. The palace’s attractiveness and Riyadh Provincial’s monuments should be exposed to the public [5]. With the assistance of others’ blogs, reviews, ratings, photographs, and other public views, people can estimate the destined places regarding food, events, stays, or journeys. However, manual analysis of available choices has become more complicated and time-consuming with an increase in the number of digital sources [6]. In the tourism sector, for both managers and tourists, there are various GIS tools utilized for route planning, economic analysis, and so on. However, the most recent advancements in data analysis that are recommended in a difficult way with poor visualizations could not benefit stakeholders [7]. Making a tourism plan for Riyadh by solely depending on any single piece of data, either Google reviews, site location, or the hotel’s location, is not sufficient. Also, the environmental conditions that include the seasons, namely summer, rainy, or snowy situations, are predefined for relishing the tourism experience. Moreover, the timing of the sites, whenever the site is opened or closed or looks attractive, is essential to know.

Many Recommendation Systems (RSs) have been built to minimize the amount of effort that an individual wants to put into recognizing the correct travel plan. The information filtering model type that anticipates user preferences and gives a list of appropriate possibilities is named RS [8]. The overload of information issue is minimized by RS, which can offer tourists more opportunities for examining the nearest destination [9]. Recent studies have shown that the tourism industry often utilizes personalized Collaborative Filtering (CF) RS approaches, which make use of the information supplied by users about assessments or ratings. This might engender a sparsity problem when the user assessments are poor [10]. Contrarily, traditional CF techniques centered on nearest neighbor algorithms exhibit important issues with scalability and efficiency [11]. Further, the RS is developed based on the user’s preferred time, especially at night, by processing the dynamic time along with the tourism data. However, the existing works did not estimate both the day-time and night-time permits to the spot so that the user could visit the places at their convenience. Thus, using the MSE-RNN system and R-PCC, a sustainable day- and night-time cultural tourism recommendation for Riyadh historical sites is proposed in this work.

1.1. Research Questions

Some of the research questions that are considered to be solved by the proposed system are listed below.

(i)

How do you recognize the suitable time (day-time or night-time) for visiting the tourist sites in Riyadh?

(ii)

How do you localize historical sites and the hotels surrounding them during tourism?

(iii)

How do you determine the positive and negative reviews about the opening and closing times of the tourism sites in Riyadh?

(iv)

How do you identify the environmental conditions around the tourist sites and plan your visit accordingly?

1.2. Problem Definition

Existing approaches centered on RS have a few limitations, which are given below.

Suitable time for visiting the sites was neglected in existing works. This lacks information for the tourists about the quality and facilities of the tourist spots during the day and night.

Suggesting nearby tourist spots is mainly focused on; however, the place of stay in case of night-time sightseeing is ignored in existing research.

The possible environmental conditions of historical site tourism, namely season, pollution, etc., were ignored in existing research. This resulted in inefficient RS and a poor tourism experience.

The analysis of superior choices for a stay place is a difficult task as it helps the tourists to prefer a good place for the stay. Prevailing approaches failed to focus on it.

Thus, to alleviate the mentioned problems, this paper proposes MSE-RNN-centric tourism RS for Riyadh historical sites. Their major contributions are explained below:

The R-PCC approach is proposed to separate the time for visiting the sites.

For recommending the place of stay, information regarding hotels and their geolocations was considered.

A new environment condition score and season dataset are utilized.

The TM-BWO algorithm is established for recommending the best hotels.

The work’s remaining parts are arranged as follows: the related works concerning the proposed approach are analyzed in Section 2; the proposed tourism recommendation methodology is described in Section 3; the proposed approach’s results and discussion centered on performance metrics are displayed in Section 4; and, at last, the work is wrapped up with future work in Section 5.

2. Literature Survey

A tourism Recommendation System (RS) was established by extracting users’ reviews on tourism social networks [12]. Using preprocessing, semantic clustering, and sentiment analysis, we were able to determine the tourists’ preferences. According to the evaluation outcomes, the model enhanced the F-measure (FM) criterion. The model ignored contextual information like the opinions of individuals and stay recommendations, even though essential information about the weather, location, and time was utilized.

Ref. [13] propounded a social-hybrid RS in the social commerce circumstance that renders tourist attractions. Trust factors in recommendation resources, namely outlier detection in user ratings, were encompassed, and social relationships amongst users were deployed. The approach’s superiority over other general approaches was exhibited by the experimental outcomes. However, a very limited number of registered comments along with the users were processed in this system.

A hybrid RS was built in [14] for tourism centered on Big Data and Artificial Intelligence (AI). For a particular visit duration, the recommended trip planner designed a complete program comprising heterogeneous tourism resources. By recommending the most appropriate items, the design enhanced the visitor experience. Nevertheless, the cultural sites’ environmental conditions were ignored.

A personalized travel RS with a recency effect was introduced in [15] to understand the current travel interests of Twitter users along with their friends and followers. For acquiring travel recommendations, travel tweets were utilized. The prevailing personalized recommendation systems were surpassed by this system. But, as this model was reiterated and the preferences were updated for a user who utilizes Twitter often, it was computationally expensive and needed data storage.

Ref. [16] established a hotel RS utilizing hotel reviews’ sentiment analysis. It rendered a fresh, rich, and disparate dataset of online hotel reviews crawled from Tripadvisor.com. The pre-trained word embeddings formed by the BERT were fed to a Random Forest classifier. Utilizing fuzzy logic along with cosine similarity, the reviews were categorized into disparate groups. When analogized to top-notch systems, the outcomes were hopeful and much superior. However, the training process was slower.

A management toolkit for sustainable cultural heritage planning along with the management of overtourism in art cities was proposed in [17]. It was built within the structure of the Atlas Project in 2019. The baseline systems were surpassed by this model. Nevertheless, the current plans regarding crisis management, environmental concerns, cultural development plans, and change impacts were neglected.

Ref. [18] explored the two single tensor systems, comprising cultural groups, multi-criteria ratings, items, and users for the tourism recommendation. It utilized a trip advisor dataset, encompassing 13,000 users from 120 countries. As per the comparative analysis, the system’s performance was enhanced by considering Western and Eastern cultures. The evaluated performance showed the better performance of the developed model regarding accuracy and mean square error (MSE). However, other systems with cultural diversities exhibited superior performances.

The authors in [19] modeled and employed a personalized tourism RS centered on data mining along with the CF algorithm. It collaboratively examined the accumulated data from expenditure, global, and personal travel, along with other information. To update the present knowledge base in the filtering procedure, the maximally processed contextual data were extracted. High accuracy along with a high data handling rate were displayed by the data validation. Nevertheless, it did not concern web-centric location tracking approaches.

Ref. [20] developed a personalized RS in a smart product service model centered on an unsupervised learning (UL) system for examining user-provided data. It established an RS-integrating DL for offering customers personalized solutions. The suggested system’s benefits were proved by the outcomes. To save on labeling data costs, this system utilized UL; however, in gauging the system’s performance, it was weak as there was no ground truth.

A non-dominated sorting heuristic technique was suggested by [21] for city tourism recommendations. Transport mode options along with the spatial-temporal structures were considered. As per the outcomes, when compared with prior approaches, this technique was more advanced as well as generated sensible and customized itineraries. However, the issue of designing personalized itineraries in a multi-day tour situation along with hotel selection is not concentrated.

Ref. [22] investigated a knowledge-based automated system for recommending tourism. Primarily, the input data were collected from the Yelp dataset. Then, the data undergo preprocessing and vectorization to examine the words in the reviews. Further, the star ratings of the restaurants were predicted using different classifiers, namely decision tree, random forest, logistic regression, and support vector machines. Subsequently, the topics of the reviews were clustered using the K-means algorithm. Further, the salience and valence of the topic were derived, and then the restaurant list was recommended for the tourists with improved performance based on the acquired knowledge. However, the restaurant localization was not detected for the recommendation, which degraded the system’s efficiency.

Ref. [23] suggests big social data for planning cultural tourism. Initially, the user-generated content was acquired to examine insights about the tourism experience. From the content, the tourist perception and main characteristics of the site were detected with enhanced performance. However, the accommodations or hotels for the tourists were not focused, which affected the tourism experience.

Ref. [24] developed a community tourism RS in an adaptive manner. Firstly, the data related to tourism were collected by the tourist staff and entrepreneurs. Then, the standardized information provided by the gamification algorithm was demonstrated and represented using the pin reward icons on Google Maps. The results showed that the system recommended tourism with higher accuracy and sensitivity. Moreover, the data preprocessing was not carried out, resulting in irrelevant recommendations.

Ref. [25] presented a framework for assisting the tourist journey by using ChatBot. Primarily, a chatbot engine was designed based on the seq2seq model, which included a Recurrent Neural Network (RNN) and Gated Recurrent Unit (GRU) to provide tourism information. Further, the events regarding tourism were ingested automatically using the Enterprise service bus by acquiring the website data. The experimental results exhibited improved accuracy and loss function. However, the utilized network faced an overfitting issue, which affected the model’s learning and tourist recommendations.

Ref. [26] recommends cultural tourism by using the Optimized Weighted Association Rule (OWAR) algorithm. Initially, the data were obtained from the location-based social network and the network authorization. Further, the user data and tourism information were clustered using Density-Based Spatial Clustering of Applications with Nodes (DBSCAN). Then, the dynamic weights and time were added. Then, the OWAR algorithm was used to produce the tourism recommendation regarding the user’s planned time. The performance of the developed approach was enhanced with higher accuracy and F-measure. However, the utilized DBSCAN algorithm did not effectively cluster the data with huge density variations.

3. Proposed Tourism Recommendation System Framework

This paper proposes a better tourism Recommendation System (RS) utilizing MSE-RNN. In preparing the recommendation output utilizing the MSE-RNN classifier, environmental data, data from Google, and hotel data are utilized. The proposed model is shown in Figure 1.

The process involved in the proposed approach is described as follows:

3.1. Processing of Location-Based Data

At first, by utilizing the GIS map of Riyadh city, the location details of the various tourism sites $\left({\lambda }_{site}\right)$ in Riyadh are acquired. Meanwhile, the location of the hotels $\left({\lambda }_{hotel}\right)$ in Riyadh is acquired from the hotel dataset. The ${\lambda }_{site}$ and ${\lambda }_{hotel}$ are considered the input data for localizing the desired tourism sites and the appropriate hotel location. It is given as follows:

$${\lambda }_{site}=\left\{l_1,l_2,l_3,\dots ,l_n\right\}$$

(1)

$${\lambda }_{hotel}=\left\{L_1,L_2,L_3,\dots ,L_q\right\}$$

(2)

Here, the $n^{th}$ site’s location is depicted as $l_n$ and the $q^{th}$ hotel’s location is signified as $L_q$.

3.1.1. Localization

Here, utilizing the SHD-KM approach, the acquired locations ${\lambda }_{site}$ and ${\lambda }_{hotel}$ are grouped centered on the minimal distance between the sites and the hotels. Grounded on the given K-number of clusters, the k-means algorithm performs unsupervised grouping of the data. However, for large-dimensional data, the conventional Euclidean distance is not appropriate. Hence, the proposed study introduces similarity-centric Hamming distance estimation.

Primarily, the number of clusters $\left(K\right)$ is defined, and the location points $\left({\lambda }_{site}\right)$ are chosen to signify the cluster centroids. Next, the nodes (hotels) close to one another are assigned to the cluster centroid ${\lambda }_{site}$. Utilizing similarity-centric Hamming distance, the distance is calculated as follows:

$$S\left({\lambda }_{site},{\lambda }_{hotel}\right)=\left|1-\frac{2}{n}\eta \left({\lambda }_{site},{\lambda }_{hotel}\right)\right|$$

(3)

Here, the binary similarity measure is depicted as $S$, and the hamming distance is specified as $\eta$. The data are classified into $K$ clusters by considering the minimal distance between the centroid and the hotels. Until all the nodes are clustered, the above procedure is repeated. Subsequently, the final clusters $\left(C_N\right)$ are indicated as follows:

$$C_N=\left\{C_1,C_2,C_3,\dots ,C_K\right\}, N=1,2,\dots ,K$$

(4)

Here, the $K^{th}$ cluster is depicted as $C_K$. The SHD-KM approach’s procedure is explained in Algorithm 1.

Algorithm 1: SHD-KM Technique.

Input: Locations $\left({\lambda }_{site}, {\lambda }_{hotel}\right)$ Output: Clusters $\left(C_N\right)$ Begin    Initialize Cluster Centres $\left({\lambda }_{site}\right)$    For each cluster centres do     Compute distance between each hotel using the equation, $$S\left({\lambda }_{site},{\lambda }_{hotel}\right)=\left|1-\frac{2}{n}\eta \left({\lambda }_{site},{\lambda }_{hotel}\right)\right|$$     Assign Hotels to the nearest Site    End for    Compute new centroid    Return final cluster $C_N$ End

3.1.2. Hotel Ranking

Here, utilizing the TM-BWO algorithm, the hotels in the clusters $\left(C_N\right)$ are ranked centered on the minimal distance and the maximal ratings of the hotels. This phase chooses BWO since it emits the least fit data during the cannibalism. However, the algorithm converges slowly due to the random initialization of BW spiders. Hence, to solve this issue, the tent map is established in the BWO for initialization.

For finding the optimal solution to the optimization issue, the algorithm replicates the macroscopic along with the microscopic principles governing spider populations. In this, the hotels in the clusters $\left(L_n\in C_N\right)$ except the centroids are regarded as the population members, and it is signified as follows:

$$L_n=\left\{L_1,L_2,L_3,\dots ,L_K\right\}, n=1,2,\dots K$$

(5)

The TM-BWO algorithm works in four phases, namely population initialization, procreation, cannibalism, and mutation, as given below.

Population Initialization: Centered on the tent map, the algorithm population $L_n$ is initialized as follows:

$$L_n+1=\left\{\frac{\frac{L_n}{\wp }, 0<L_n\le \wp }{ \frac{\left(1-L_n\right)}{\left(1-\wp \right)}, \wp <L_n\le 1 } \right.$$

(6)

Here, the mapping constant is depicted as $\wp$. At present, each member’s fitness is estimated. In this context, fitness $\left(fit\right)$ is centered on minimal distance $\left(S_{\eta }\right)$ and increased ratings $\left(\gamma \right)$. It is computed as follows:

$$fit\left(L_n\right)=\langle min\left(S_{\eta }\right),max\left(\gamma \right)\rangle$$

(7)

Procreate: In this, the members’ unique mating produces a fresh generation. Centered on their procreation rate, a set of spiders’ is chosen, as mother and father spiders are randomly designated from the population to mate. The offspring production is given as follows:

$$\left\{\begin{array}{c}{Y}_1=\alpha L_1+\left(1-\alpha \right)L_2\\ Y_2=\alpha L_2+\left(1-\alpha \right)L_1\end{array}\right.$$

(8)

Here, the mother and father spiders are signified as $L_1$ and $L_2$, correspondingly, the offspring is signified as $Y_1$ and $Y_2$, and an array of random numbers is indicated as $\alpha$. Currently, the offspring and the mothers are classified based on their fitness.

Cannibalism: This phase removes the weak members and preserves the best members in the population, centered on the cannibalism rating. Cannibalism encompasses three phases, which are:

Sexual Cannibalism: During or after mating, the mother spider eats her husband. The existing mothers are included in the subsequent generation.

Sibling cannibalism: Due to the shortage of food sources, the spiders with the best fitness eat their siblings.

Cannibalism between offspring and mother: Powerful offspring may even eat their mother. Otherwise, the solution will take the place of its mother and enter the subsequent generation if parents generate a solution with a high fitness value.

Mutation: In this case, the mutation rate determines whether the members will mutate. All chosen members randomly exchange two elements in the array. Finally, the members with the best fitness are updated.

Until the termination criteria are satisfied, the process is repeated. At last, the best solution is ranked, and $L_{rank}$ symbolizes the ranked hotels. TM-BWO’s procedure is explained in Algorithm 2.

Algorithm 2: TM-BWO Technique.

Input: Clustered Hotels $\left(L_n\right)$ Output: Ranked Hotels $\left(L_{rank}\right)$ Begin    Initialize the population using tent map as, $$L_n+1=\left\{\begin{array}{c}\frac{L_n}{\wp }, \\ \frac{\left(1-L_n\right) }{\left(1-\wp \right)},\end{array}\right. \begin{array}{c}0<L_n\le \wp \\ \wp <L_n\le 1\end{array}$$   Initialize the procreation, cannibalism and mutation rate   Evaluate the fitness as, $$fit\left(L_n\right)=\langle min\left(S_{\eta }\right),max\left(\gamma \right)\rangle$$   While $\left(i<i_{max}\right)$    Select the parents randomly    $\mathbf{For} i=1 to n parents$ do     Select $L_1$ and $L_2$     Generate Offspring using the notation, $$\left\{\begin{array}{c}{Y}_1=\alpha L_1+\left(1-\alpha \right)L_2\\ Y_2=\alpha L_2+\left(1-\alpha \right)L_1\end{array}\right.$$     Destroy fathers     Eliminate weak members based on cannibalism rate    End For   Calculate the members to mutate    For $i=1 to n$ do     Mutate Randomly by exchanging the elements     Generate the new solution    End For    Save the best solution    End While    Return ${\mathrm{L}}_{\mathrm{rank}}$ End

3.1.3. Keyword Embedding

Meanwhile, from the GIS map location, the keywords $\left(K\right)$ that determine the sites’ location details $\left({\lambda }_{site}\right)$ are extracted and embedded utilizing the BERT algorithm. A pre-trained approach rendered with an encoder component that accepts the input text data and gives the recommendation utilizing the decoder is termed BERT. It consists of an output layer, a BERT transformer layer, and an embedding layer.

Primarily, the tokens for each keyword $\left(K\right)$ are determined and indicated within diverse classes, and the separation function separates all classes. The token $\left(T\right)$ is signified as follows:

$$T=\left[t_1,t_2,\dots ,t_c\right]$$

(9)

Here, the $c^{th}$ tokenized word is depicted as $t_c$. Currently, these tokens are provided to the embedding layer.

Embedding layer: The representation of words in vector form is termed embedding. This layer executes the processes, which are given below.

Token embedding: In this, the words are transformed into a fixed-dimensional vector with a class and separation function added to the beginning as well as the end of the sentences.

Segment embedding: For classifying the diverse inputs with binary coding, segment embedding for a word is performed.

Position embedding: For differentiating between a word’s contextual meaning and the meaning of the sentence according to how the word is positioned, position embedding is deployed.

BERT transformer layer: The token, segment, and position embeddings are added up and fed to the BERT transformer encoder layer, which contains 12 transformers with 12 attention mechanisms and millions of parameters. The words are encoded by the encoder, the important keywords are determined, and contextual embeddings are provided by the decoder.

Output layers: The encoded output is provided to the output layer, which comprises a simple classifier system with a completely linked layer along with an activation function. The loss $\left(Loss\right)$ in the classifier output is computed as follows:

$$Loss=\frac{1}{2}\left(Tar-\sum K_{emb}\right)$$

(10)

Here, the target word embedding score is signified as $Tar$, and the output word embedding is indicated as $K_{emb}$. Hence, word embedding $K_{emb}$ is acquired using BERT.

3.2. Processing of Google Data

This phase extracts the reviews along with the closing and opening times of the sites in Riyadh from Google. The extracted reviews $\left(\Re \right)$ and timings $\left(\Im \right)$ are given as follows:

$$\Re =\left\{{\Re }_1,{\Re }_2,{\Re }_3,\dots ,{\Re }_n\right\}$$

(11)

$$\Im =\left\{{\Im }_1,{\Im }_2,{\Im }_3,\dots ,{\Im }_n\right\}$$

(12)

Here, the $n^{th}$ site’s review is depicted as ${\Re }_n$, and the $n^{th}$ site’s timing is symbolized as ${\Im }_n$. At present, the extracted data go through the process given below.

3.2.1. Keyword Extraction

For the purpose of providing adequate information concerning the tourist spots in Riyadh, the keywords present in the reviews $\left(\Re \right)$ are extracted using the YAKE approach. Statistical features that are extracted from individual documents are utilized by YAKE for finding the text’s most relevant keywords. It passes through five stages, which are given below.

Pre-processing of text: Tokenization is carried out in this phase, where the review data $\left(\Re \right)$ is partitioned into many meaningful subunits called individual terms $\left(\mu \right)$. Mostly, it relies on the spacing between the words in the text. This transformation causes enhanced classification outcomes.

Feature Extraction: Here, a list of individual terms is utilized as input, and it is transformed into a set of statistical attributes. Next, a set of five features captured each term’s qualities. Casing, Word Position, Word Frequency, Word Relatedness to Context, and Word DifSentence are the five developed features.

Casing $\left({\eta }_{cas}\right)$: Here, the word beginning with a capital letter, or an acronym, is given attention. Otherwise, the word’s aspect is reflected, which is computed as follows:

$${\eta }_{cas}=\frac{max\left(Upp\left(\mu \right)\right), \left(Acr\left(\mu \right)\right)}{log {\mu }_{freq}}$$

(13)

Here, the number of times the term begins with uppercase is depicted as $Upp\left(\mu \right)$, the number of times the candidate word is tagged as an acronym is signified as $cr\left(\mu \right)$, and the frequency of the term $\mu$ is indicated as ${\mu }_{freq}$.

Word Position $\left({\eta }_{pos}\right)$: Word position values the word that appears at the document’s beginning more highly, centered on the theory that appropriate keywords tend to focus more on the document’s beginning frequently. This is given as follows:

$${\eta }_{pos}=log\left(lo{g}_2\left(2+med\left({\mu }_{sen}\right)\right)\right)$$

(14)

Here, the median function is depicted as $med$, and the set of sentences where the word $\mu$ appears is symbolized as ${\mu }_{sen}$.

Word Frequency $\left({\eta }_{freq}\right)$: How often the word appears in the text is exhibited by this feature. It indicates that the more often a word occurs, the more important it is. The following equation prevents the bias at high frequencies:

$${\eta }_{freq}=\frac{{\mu }_{freq}}{mean\left(\mu \right)+1\times i}$$

(15)

Here, the standard deviation is indicated as $i$.

Word Relatedness to Context $\left({\eta }_{rel}\right)$: It determines the number of additional words that appear to the left of the candidate word.

Word DifSentence $\left({\eta }_{dif}\right)$: The number of times a candidate term appears in several sentences is counted in this feature. This process is computed as follows:

$${\eta }_{dif}=\frac{{\mu }_{freq}}{num\left(sen\right)}$$

(16)

Here, the total number of sentences is depicted as $num\left(sen\right)$. Every feature is merged as a single term ${\eta }_{all}$, which is taken for candidate score computation.

Computing term score: The combined term is assigned a weight $\overline{\omega }\left({\eta }_{all}\right)$ for computing the score. This weight feeds the process of generating keywords, which is described as follows:

$$\overline{\omega }\left({\eta }_{all}\right)=\frac{{\eta }_{rel}\times {\eta }_{pos}}{{\eta }_{cas}+\frac{{\eta }_{freq}}{{\eta }_{rel}}+\frac{{\eta }_{dif}}{{\eta }_{rel}}}$$

(17)

Candidate keyword list generation: The weight $\overline{\omega }\left({\eta }_{all}\right)$ is given to this phase. This is considered a sliding window, which generates several sequences of keywords. Then, every candidate keyword will be assigned a final word $\aleph$, which is computed as follows:

$$\aleph =\frac{{\prod }_{\mu }\overline{\omega }\left({\eta }_{all}\right)}{{\mu }_{freq}\times (1+{\sum }_{\mu }\overline{\omega }\left({\eta }_{all}\right))}$$

(18)

Deduplication and Ranking: At last, the similar terms are removed and the appropriate keywords are ranked as 1, 2,…, such that the lower the term score, the keyword will be more significant.

Hence, utilizing the YAKE, the keywords ${\aleph }_{ex}$ are extracted.

3.2.2. Polarity Estimation

Here, utilizing the VADER approach, the polarity of the keywords ${\aleph }_{ex}$ is determined. In this, the polarity score corresponds to the positive, negative, and neutral terms. A variety of sentiment lexicons that are frequently categorized as either positive or negative based on their semantic orientation are used by VADER. It not only informs us of the positivity along with the negativity scores but also of the polarity score of every term. The polarity $\left(P\right)$ estimation is computed as follows:

$$P=\left\{\begin{array}{c}\aleph { }_{ex}^{pos}\to score=1, if \zeta >0.001\\ \aleph { }_{ex}^{neu}\to score=0, if \zeta >-0.001, \zeta <0.001\\ \aleph { }_{ex}^{neg}\to score=-1, if \zeta <-0.001\end{array}\right.$$

(19)

Here, the positive, negative, and neutral words are depicted as $\aleph { }_{ex}^{pos}$, $\aleph { }_{ex}^{neu}$, and $\aleph { }_{ex}^{neg}$, correspondingly, and the compound value is signified as ζ. Hence, VADER estimates the polarity $\left(P\right)$ of the terms.

3.2.3. Timing Segregation

Meanwhile, utilizing the R-PCC, the timings $\Im$ extracted from Google are separated. PCC was chosen due to its effectiveness in gauging the association between the data points. However, since it calculates the correlation between each point, the time taken to find the correlation between the two points is longer. Hence, to avoid this issue, the reference value of 24 h is provided to the PCC. The R-PCC segregation $({\Im }_{seg})$ is computed as follows:

$${\Im }_{seg}=\frac{\sum \left({\tau }_{24}^x-\overline{\Im }\right)\left({\tau }_{24}^y-\overline{\Im }\right)}{\sqrt{\sum {\left({\tau }_{24}^x-\overline{\Im }\right)}^2}{({\tau }_{24}^y-\overline{\Im })}^2}$$

(20)

Here, the reference values of 24 h of x and y samples are depicted as ${\tau }_{24}^x$ and ${\tau }_{24}^y$, correspondingly, and the mean value of the timings is signified as $\overline{\Im }$. Hence, R-PCC compares the timings and finds whether the site is open 24 h a day or only during the day.

3.3. Environmental Data

The proposed study for recommending tourist spots in Riyadh considered the environment condition score, which includes season data, pollution details, crime ratings, and quality of life, in addition to the data gathered above. The gathered environmental data $\left(E\right)$ are given as follows:

$$E=\left\{e_1,e_2,\dots ,e_u\right\}$$

(21)

Here, the $n^{th}$ number of data points gathered is depicted as $e_u$. Here, the environment condition score is evaluated, in which the better weather conditions, reduced pollution, and crime rating, along with improved life quality, are declared the highly scored environment conditions and vice versa. At present, the environmental data are carried out for pre-processing.

Pre-Processing

Here, the data E go through preprocessing in two stages, such as numeralization and missing value removal.

Numeralization: To improve the recommendation system’s results, the strings in the data E are converted to numbers.

Missing value removal: The missing values present in the data are eliminated after performing numeralization, as the missing value can degrade the recommendation system’s performance. Further, the incomplete and missing data values are imputed by computing the mean of the available data to attain more relevant recommendations.

At last, the data are pre-processed and is symbolized as E_pre.

3.4. Data Fusion

Subsequent to gathering and processing several sets of data, each set is combined for the final recommendation. The pre-processed environmental data, embedded keywords, estimated polarity scores, ranked hotels, and segregated timings are fused and provided to the input of the recommendation system so that it gives perfect recommendations for the historical sites in Riyadh based on several factors. The fused data are given as follows:

$$\delta =\left\{L_{rank},K_{emb},P,{\Im }_{seg},E_{pre}\right\}$$

(22)

Here, the fused data, which is taken as input for the recommendation phase, is depicted as $\delta$.

3.5. Recommendation

In order to provide recommendations regarding the sites, the fused data are fed to the MSE-RNN RS. The sequence of the several data points can be processed with a Recurrent Neural Network (RNN). However, an exploding gradient occurs, which updates the weight values of higher values. Hence, to solve this issue, the Mean Signed Error is modified in the existing RNN. Figure 2 displays the proposed MSE-RNN architecture.

MSE-RNN comprises three layers. There are additional layers of units between the input and output layers, which are called hidden layers. At first, the input data are given to the input layer, which is then transferred to the hidden layer. In the hidden layers, the edges that are linked repeatedly yield input from two parts. Each hidden layer has its own set of weights and biases. The hidden layer’s output $({\Phi }_{hid})$ is formulated as follows:

$${\Phi }_{hid}^i=\Theta \left({\omega }_{in}\delta +{\omega }_{hid}{\Phi }_{i-1}+\nu \right)$$

(23)

Here, the weights of the input and hidden layers are indicated as ${\omega }_{in}$ and ${\omega }_{hid}$, respectively. The output of the prior hidden layer, which is stored in memory, is depicted as ${\Phi }_{i-1}$, the bias vector is given as $\nu$, and the sigmoid activation functions are indicated as $\Theta$, which is formulated as follows:

$$\Theta =\frac{1}{1+e^{-\delta }}$$

(24)

Next, for the recommendation, the output layer is responsible. The following expression determines the output layer $\left({\Phi }_{out}\right)$ as follows:

$${\Phi }_{out}=\Theta \left[{\omega }_{out}{\Phi }_{hid}+\nu \right]$$

(25)

Here, the weight of the output layer is depicted as ${\omega }_{out}$. Hence, the classifier recommends the output. Next, the MSE is computed to verify the accurate recommendation output. It is formulated as follows:

$$MSE=\frac{1}{\chi } {\displaystyle \sum }_{i=0}^n\overline{\Phi }-{\Phi }_{out}$$

(26)

Here, the total number of hidden neurons is depicted as χ, and the ground truth is symbolized as $\overline{\Phi }$. Hence, the recommendations for the historical sites in Riyadh are provided by MSE-RNN. The MSE-RNN RS suggests historical sites, nearby hotels, and respective timings when the keywords regarding the location are provided by the users. The succeeding result section evaluates the proposed system’s efficiency.

4. Results and Discussion

Here, regarding the recommendation process together with user satisfaction and convenience, the robustness of the proposed tourism Recommendation System (RS) built by the Python 3 platform is analyzed.

In practice, the processed information of historical site location, hotel ranking, polarity estimated Google review, segregated timings, and preprocessed environmental data are fused and fed as the input of the proposed MSE-RNN for training. As the MSE is computed at the output layer, an accurate recommendation is produced by the proposed network. The resultant recommendation recommends historical sites, the best hotels nearby the sites, and the timings suitable for visiting the sites in the digital message form to the user based on the user’s requested location. The datasets used for evaluating the proposed system are further described below.

4.1. Dataset Description

This paper utilizes five different datasets to analyze the performance of the proposed system. The historical sites of Riyadh are obtained from the “Smartscraper database” [27]. This dataset includes 94 historical places located in Riyadh, along with the contact number, email address, and geocoded address. Then, the hotel locations around the historical sites are recognized by using the “Datantify Global Company Database” [28]. This database consists of details about 6356 hotels around Riyadh, including contact information and map locations. Further, the hotel reviews are acquired from the “Saudi Arabia Booking.com” dataset [29]. This dataset comprises 1025 hotel details, such as the hotel name, city, rent details, and star ratings. Moreover, the environmental conditions of Riyadh were collected using the “Saudi Arabia Hourly Climate Integrated Surface Data” [30]. This dataset provides the previous five years’ hourly data about the wind, sky condition, visibility, air quality, dew, and sea level pressure of the Arabian country. Meanwhile, the pollution level in Riyadh was extracted from the “Annual Average of Air Pollutant Concentration by Monitoring Station” data portal [31]. This data portal consists of pollution measures for 1492 stations in Riyadh. From each of these datasets, 80% of the data is used for training, and 20% of the data is used for testing the proposed model.

4.2. Performance Analysis of MSE-RNN-Based Recommendations

This section investigates the proposed MSE-RNN’s efficacy by contrasting its results with the existing approaches, namely Recurrent Neural Network (RNN), Convolution Neural Network (CNN), Deep Belief Network (DBN), and Artificial Neural Network (ANN).

The most precise recommendations offered by the model utilizing the evaluation metrics, namely precision, recall, F-measure, and accuracy, are determined in Figure 3. Here, the proposed approach attained 96.16% accuracy and 98.32% FM, which are higher compared to the existing approaches. As the exploding gradient issues could be fixed by the MSE approach by maintaining its gradients at a certain level, the proposed method attained higher accuracy and FM over the existing approaches. The outcomes exhibit that the proposed model results in more precise recommendations and could give users satisfaction along with convenience by exploiting essential features like the environment, user preferences, weather, time, and location.

Figure 4 measures the quality of the proposed model based on specificity, Negative Predictive Value (NPV), and Mathews Correlation Coefficient (MCC). In the tourism recommendation field, the proposed approach rendered a specificity, NPV, and MCC of 98.97%, 96.97%, and 96.3%, respectively. The processing of GIS maps, hotel data, Google reviews, timings, and environmental data prior to the data training enhances the proposed recommendation outcome. This displays that the proposed model attains the most considerable outcomes compared to the existing approaches by solving the exploding gradient issue utilizing the MSE error derivative.

The False Positive Rate (FPR), False Discovery Rate (FDR), and False Negative Rate (FNR) of the proposed and existing classification approaches are measured in Figure 5. These measures signify the negative recommendations made by the system and should be lower to prove the system’s superior performance. In this regard, the proposed MSE-RNN achieved the least FPR (0.0103), FDR (0.0085), and FNR (0.025). As the MSE is calculated at the output layer of MSE-RNN, the positive and negative classes of recommendation are exactly identified. This shows that by rescaling the error derivative, the MSE approach minimized the likelihood of exploding gradient issues.

The MSE of proposed and existing approaches is exhibited in

Table 1. Performance comparison regarding MSE.

Methods	MSE
Proposed MSE-RNN	0.034
RNN	0.068
CNN	0.102
DBN	0.134
ANN	0.192

. A measure of error between paired observations, which should be lower for the model’s superior performance, is named MSE. Table 1 shows that the proposed approach’s MSE-RNN is 0.034 and RNN is 0.068; MES-RNN outperforms RNN and other methods. This displays that, when analogized to the existing classification approaches, the recommendations made by the proposed system are more precise.

The Critical Success Index (CSI) measures for the proposed and existing approaches are examined in Figure 6. The occurrence of correctly recommended items over the total number of recommendations is depicted as CSI. A higher CSI of 96.85% is achieved by the proposed network, exhibiting the RS’s superior performance. Figure 6 displays that the proposed MSE-RNN’s CSI outperforms RNN and other methods. As the input data are processed step by step to obtain every piece of information, along with day and night timing segregation, the proposed methods achieved improved performance regarding CSI. Thus, the rescaling of weights for solving the gradient exploding issue has a better influence on enhancing the model’s CSI.

4.3. Performance Analysis of R-PCC-Based Segregation

In this section, for verifying the proposed system’s segregation performance, the proposed R-PCC approach’s process time is compared against the existing PCC, Spearman Correlation Coefficient (SCC), Bayesian Approach (BA), and Kendall’s Correlation Coefficient (KCC) techniques.

The time taken by the proposed approach along with the existing approaches for the segregation process is examined in

Table 2. Segregation performance analysis.

Methods	Process Time (ms)
Proposed R-PCC	837
PCC	1176
SCC	1257
BA	1320
KCC	1582

. The model’s superior performance is illustrated by attaining more subtle types of discrimination within a minimum time. Regarding this, the proposed approach takes a process time of 339 ms less than the existing PCC technique. This displays that the segregation time is considerably decreased by a reduction in correlation time by setting a reference value, and thus achieves adequate performance in recommending tourism.

4.4. Performance Analysis of TM-BWO-Based Ranking

To prove the proposed approach’s efficiency in attaining the defined objective function, this section undergoes convergence analysis for the proposed TM-BWO and existing BWO, Grey Wolf Optimization (GWO), Particle Swarm Optimization (PSO), and Salp Swarm Optimization (SSO) techniques.

Figure 7 signifies the fitness between the proposed and existing approaches. The observation reveals that as the number of iterations escalates, there is an increase in fitness. The fitness achieved by the proposed approach is higher by 4.54 than BWO, 6.84 than GWO, 8.63 than PSO, and 10.59 than SSO when the number of iterations is 50. Thus, when analogized with the existing approaches, the initialization of BW spiders utilizing the TM approach results in quick convergence of the proposed system in ranking hotels as per the criteria defined in the objective function.

4.5. Performance Analysis of SHD-KM-Based Localization

This subsection contrasts the proposed SHD-KM’s performance over the existing KM, K-Medoid, Clustering Large Applications (CLARA), and Balanced Iterative Reducing and Clustering using Hierarchies (BIRCH) approaches regarding localization time.

The localization times of the proposed and existing clustering approaches are analogized in

Table 3. Localization performance comparison.

Methods	Localization Time (ms)
Proposed SHD-KM	9654
KM	10,491
K-Medoid	10,829
CLARA	11,934
BIRCH	12,382

. The analysis exhibits that the localization by the proposed model is much faster by taking 867 ms less compared to the existing KM approach. This depicts that a standard way of handling large data is the SHD approach, which executed well in minimizing localization time. Thus, the study infers that for localization, the proposed approach shows maximum performance and becomes much more effective.

4.6. Comparative Analysis

This section performs a comparative analysis between the proposed tourism RS and the existing models mentioned in Section 2.

In Figure 8, the performance of the proposed MSE-RNN-centric RS is contrasted with existing models built by [12,13,16] in Section 2. Figure 8 illustrates that the existing techniques of [12] attained a precision of 92.55%, [13] attained a precision of 93.50%, and [16] attained a precision of 86% for tourism RS. However, the proposed network obtained 98.50%, which is higher than the existing methods. Thus, the analysis of Figure 8 reveals that the proposed system shows optimal performance by enhancing the precision rate by 7.51% compared to the existing approaches. The multiple data, namely location, hotel reviews, hotel location, timings, and environmental data regarding the tourist spot, are deeply examined and processed by the proposed model prior to the recommendation. Therefore, the performance of the proposed technique is better than the prevailing techniques, where those essential data are not focused together for RS. Thus, as per the analysis, the adaptability of the proposed RS into the tourism model can be enhanced by incorporating hotel-centric user needs along with the affordability of online Google reviews.

4.7. Evaluation Summary

Here, the significance of the results attained by the proposed RS is validated by comparing it with some prevailing research works.

Table 4. Performance comparison regarding accuracy.

References	Technique Used	Accuracy (%)
Proposed	MSE-RNN	96.16
[15]	TTC	75.23
[16]	Ensemble classifier	92.36
[19]	CF algorithm	91.50
[22]	RF	87.59
[25]	DL-chatbot	94.67

represents the significance of the proposed approach for tourism recommendations in terms of accuracy. As the data learning and vanishing gradient problems of the classifiers were not focused on the Travel Tweet Classifier (TTC) of [15] and Ensemble classifier in [16], the inaccurate recommendation with 75.23% accuracy and 92.36% accuracy was attained. Further, the Collaborative Filtering (CF) algorithm-based RS by [19] attained an accuracy of 91.50%. However, the nearby stay places and their locations were not analyzed to develop an efficient RS. Moreover, the tourism information was utilized from website data in [25], yet the environmental conditions at the tourist sites were not determined. Also, the overfitting issue in the Deep Learning (DL) network was not resolved. Therefore, the accuracy reached 94.67%, which is lower than the proposed system’s accuracy. As the aforementioned issues in the existing works, including the environmental data, hotel location, reviews, historical site location, and timings, are concentrated in this work for tourism recommendation, a higher recommendation accuracy of 96.16% is achieved by the proposed framework. Thus, the proposed cultural tourism RS is specified as a more significant one than the existing works.

5. Conclusions

This paper developed an efficient Recommendation System (RS) for sustainable day- and night-time cultural tourism using MSE-RNN for Riyadh historical sites as part of the government’s tourism policy. The Google data and GIS data are examined in the proposed system by including four significant phases, namely localization, ranking hotels, timing segregation, and recommendation. The sites of tourist spots and the hotels were located within 9654 ms using the proposed SHD-KM. Further, the introduced TM-BWO algorithm ranked hotels based on the locations with better fitness. Also, the timings were acquired from Google, and then the day and night timings were segregated using the R-PCC method within 837 ms. By analogizing the outcomes of the proposed and existing systems, the proposed study’s performance has been recorded. In the study, the ranking of the hotels, which was evaluated using the proposed TM-BWO based on locations, improved the RS. Moreover, the preprocessing of environmental data, analysis of timings, and time segregation using the presented R-PCC enhanced the RS by fulfilling the research questions. To verify the efficiency of the proposed work, the MSE-RNN performance for RS is compared with the existing systems. As the gradient vanishing problem of the RNN is solved by the MSE computation, the proposed model achieved 98.16% accuracy, which is 1.39% better than the existing RNN, 2.88% than CNN, 5.28% than DBN, and 7.95% than ANN. Also, the day and night timing analysis made the proposed system more effective for the better tourism experience of users. The development of this effective tourism RS helps the tourists reach the tourism sites of Riyadh at the most appropriate time, check in to the best hotels, and enjoy the tourism experience. Hence, the overall estimation exhibits that, regarding all metrics, the proposed model has surpassed comparable models and provides superior tourism RS.

6. Limitations and Future Work Recommendations

Even though the recommendation of historical sites centered on the time of availability of the sites was concentrated mostly in this study, the availability of routes between historical sites and hotels was not focused. As multiple data points were considered for developing the tourism recommendation, the proposed model also consumed more time during the research. Further, the presented study did not analyze the imperfect or missing information in the hotel review and location data. Also, the performance assessment for the developed RS solely depends on the internal measures using the PYTHON platform. Real-time evaluation was not attempted in this work due to the higher resource requirements, expensiveness, and computational complexity. Hence, in the future, the proposed system will be further extended by focusing on parameter tuning and route availability detection to minimize the training time and enhance the recommendation system, respectively. Additionally, the missing or imperfect hotel reviews and location data will be imputed in the future to improve the recommendation accuracy. In order to further prove the proposed RS, the external assessment will be attempted in the future by applying the developed approach in a real-time environment.