Abstract
In the postmodern era of tourism, tourists’ behavior has undergone a substantial change and the demand of customized experience dominates the tourism market. The traditional single-objective travel route recommendation method fails to meet the multiobjective needs of users. To handle this problem, a multiobjective hybrid tabu search algorithm for urban travel route recommendations is proposed in this paper. First, the rating and level of attractions, as well as the corresponding information of hotels and restaurants within a certain radius, are considered. Then, based on this information, a crowdsensing scoring method is established. Second, the hybrid particle swarm genetic optimization algorithm is exploited to generate a single-object route, and the fast nondominated Pareto sorting algorithm is exploited to find the optimized solution. Then, the hybrid tabu algorithm is used to optimize the personalized route chosen according to multiple objects set by users. This algorithm combines the global search ability of the genetic algorithm and the neighborhood search ability of the tabu algorithm to prevent convergence from falling into a local optimum. Finally, the experiments are conducted on the real-world data collected from the Dianping and Ctrip web sites. The comparison with baseline algorithms indicates that the algorithm proposed in this paper provides accurate and reasonable route recommendations for users.
1. Introduction
The rapid development of mobile internet technology greatly promotes the maturity of various social media platforms. These platforms provide users with space to share their travel behavior through travel logs and check-in data, which is a good approach to solve some challenging problems, such as map matching [1], travel route recommendation [2], data annotation [3, 4], and trajectory analysis [5]. However, the data shared by network users are heterogeneous, and it is difficult to quantify the structure and semantics of data. Meanwhile, relevant research studies are incomplete, which makes the needs of personalized tourism difficult to meet. For example, some papers consider the types of attractions, the cost of tourism, and the distance between attractions, but some other factors reflecting the preference of tourists are ignored, such as the departure time, departure place, and travel time. These factors are essential for recommendation travel routes.
Due to the complexity of the route recommendation problem, a variety of travel recommendation algorithms have been proposed. Wen et al. [6] proposed an efficient keyword-aware travel route recommendation framework that exploits time objects for route recommendation. Liu et al. [7] proposed an efficient route recommendation algorithm to search for recommended popular routes with minimal travel cost. Liao and Zheng [8] proposed a hybrid heuristic algorithm based on random simulation for designing a personalized one-day trip in a time-object-dependent random environment. Jia et al. [9] proposed a travel route recommendation method that takes transportation modes into account and recommended the shortest route with mixed transportation modes as the object. Hsueh et al. [10] designed a personalized travel recommendation framework based on user collaborative filtering and time reference to explore user preferences. This framework can effectively plan personalized travel routes by exploiting the time users arrive at tourist attractions. Debnath et al. [11] proposed that a travel route was composed of a series of interest point positions and that the corresponding time information was composed of time perception and preference perception.
The related route recommendation algorithms can solve the single-object problem but cannot meet the multiobjective needs of users. As for travel route recommendations, users generally set multiple objects, such as the number of attractions, the duration of travel, and travel cost. As for multiobjective optimization problems, there may be competition among the optimization of objective functions. In this case, the objective functions cannot be optimized at the same time, so a compromise between multiple objective functions is required.
To realize personalized tourist route recommendation and meet the multiobjective needs of users for tourist routes, a hybrid tabu search algorithm combining users’ historical preferences with high recommendation accuracy is designed in this paper, which considers time, cost, distance, and other factors for route recommendation, such as departure time, travel duration time, and attraction tickets. The hybrid particle swarm genetic optimization algorithm is exploited to generate a single-object route, and the fast nondominated Pareto sorting algorithm is employed to find the optimal solution. Then, the proposed hybrid tabu algorithm is used to optimize the personalized route chosen based on user preferences and user-provided constraints. The main contributions of this paper are as follows:(i)First, this paper presents a method for obtaining the crowdsensing score. The method mainly calculates the location and social score of attractions based on the real data of attractions, hotels, and restaurants. Then, it generates the crowdsensing score by combining the location and social score of attractions. By combining the crowdsensing score with the time score of the attraction, the comprehensive crowdsensing score of the attraction is obtained.(ii)Second, based on the comprehensive crowdsensing score of the route, the hybrid particle swarm genetic optimization algorithm is exploited to generate a single-object route, and the fast nondominated Pareto sorting algorithm is employed to find the optimal solution. Then, the proposed hybrid tabu algorithm is used to optimize the personalized route chosen based on user preferences and user-provided constraints. The hybrid tabu search algorithm combines the global optimization ability of the genetic algorithm and the neighborhood search ability of the tabu algorithm to prevent the route from falling into a local optimal solution.(iii)Third, a large number of case experiments and comparative experiments are conducted on real datasets. The experimental results indicate that the proposed algorithm is more effective than the baseline algorithm.
The rest of this paper is organized as follows: The second section describes the work related to tourism route recommendations. The third section introduces the concept, research methods, and scoring method of some related definitions. Also, this section introduces the proposed hybrid tabu search algorithm. The fourth section analyzes the experiment and the experimental results. The conclusion of this paper is presented in the fifth section.
2. Related Work
2.1. Route Recommendation Algorithm
At present, a large number of achievements have been achieved in the field of route recommendations. Qi et al. [12] proposed two simple and general methods to represent the time-solution quality relationship of route recommendation algorithms: function fitting and artificial neural network training. These methods are suitable for benchmarking, performance comparison, and algorithm behavior analysis. Yao et al. [13] proposed a multiobjective hyperheuristic intelligent city route recommendation scheme based on reinforcement learning and designed a low-level heuristic algorithm to improve searching speed. Yang et al. [14] proposed a multifactor recommendation algorithm, which creates a set of candidate points of interest based on the locality of user activity. To improve the quality of recommendation, the relationship between friends, the category of interest points, and some popularity factors are considered for each candidate point of interest. Rashid and Mosteiro [15] proposed a new solution to the genetic algorithm. Specifically, the mutation technique is preserved, and the local search heuristic algorithm with strong climbing ability is used. Besides, the greedy algorithm is exposited to generate new offspring. Zheng et al. [16] proposed a four-step heuristic algorithm, which consists of the genetic algorithm and the differential evolution algorithm. The former generates the route, and the latter optimizes the time. Finally, the best route is generated.
However, there are few research studies on travel route recommendations. The main weaknesses of the current research are that the travel factors considered are incomplete or the scoring method is unreasonable. In this case, the research on travel route recommendations mainly considers single-objective travel route optimization, which cannot meet the needs of users with multiple objectives. To handle this problem, this paper considers many factors of route recommendations, and the hybrid tabu search algorithm is exploited to generate personalized travel routes to meet user preferences.
2.2. Factors in Route Recommendation
There are many factors that need to be considered for the recommendation of tourist routes, but the current research only focuses on the factors such as time, cost, and distance. Guofeng et al. [17] designed a comprehensive scoring method, and three factors were introduced into the scoring method including the rating of attractions, the rating of time to reach attractions, and the rating of time to open attractions. Uwaisy et al. [18] combined the results of tabu search with the concept of multiattribute utility theory to determine the optimal travel route based on popularity, cost, number of tourist attractions, and other indicators. Zhang et al. [19] took into account several factors in the route recommendation, such as distance between attractions, initial travel location, initial departure time, travel duration, total cost, rating, and popularity of attractions. Meanwhile, they used the comprehensive attraction index to rank tourist routes. Liu et al. [7] designed a popular traverse map that considers travel cost for route recommendations and uses the principle of minimum descriptive length to model travel costs. Based on this, popular travel routes with minimal travel costs are recommended. Gunawan and Lau [20] proposed a mathematical model for the tourist trip design problem. The proposed model extends the team-orienteering problem under time window constraints by incorporating more factors, such as different total travel time budgets and different start and end nodes for routes. Also, user travel history and text descriptions are taken as keywords. Hsueh et al. [10] proposed a time-constrained personalized travel recommendation framework that uses geographic characteristics and social relationships to recommend personalized travel that meets user preferences. Dugani et al. [21] presented an adaptive and sequential recommendation for travel routes. To bridge the gap between user preferences and travel routes, the recommendation method considers factors such as different locations, cost of each subject, access time, and seasonal distribution. Luan et al. [22] proposed a personalized travel recommendation method based on the maximum marginal correlation, and the method focused on the travel correlation and diversity in travel recommendations.
In addition to the abovementioned factors considered for the recommendation of tourist routes, other factors such as location of attractions, business hours of attractions, and surrounding supporting facilities can have an impact on the user’s decision. This paper adopts the idea of crowdsensing to score attractions in terms of the level, rating, number of evaluators, number of hotels and restaurants within a specific radius, opening time, travel time, ticket prices, and so on, so as to improve the rationality and accuracy of attraction quality quantification.
2.3. Multiobjective Optimization Modeling
The methods to solve the multiobjective optimization problem (MOOP) can generally be divided into three categories: the methods based on the aggregate objective function (AOF), the methods based on the constrained objective function (COF), and the methods based on the Pareto-compliant ranking (PCR). AOF-based methods need to first construct a single aggregate objective function that combines all the original objectives in the MOOP and then optimize the single aggregate objective function (SOOP) [11, 14] by solving the single-objective optimization problem. COF-based methods need to consider the maximum posterior probability because only one object is optimized in the COF method and all other objects are treated as additional constraints [23, 24]. The former two methods are both a priori methods, which only obtain specific compromise solution according to the weight vector. However, the weights of different objects are determined by users’ preferences before searching, and there is a certain deviation in the whole optimization process. In this case, the solutions of the a priori methods are often not optimized. PCR is also known as the a posteriori method, which first obtains a set of Pareto optimal solutions and then chooses the most appropriate one from these solutions according to users’ preferences. A posteriori methods can be divided into two categories: exact methods and heuristic methods. Some exact methods exploit the Dijkstra algorithm [25, 26] to obtain Pareto optimal solutions, but the computational complexity of these methods is very high when they are applied to large-scale networks [27]. The heuristic wave diffusion algorithm is used to solve the Pareto route.
Many methods are proposed to solve the multiobjective optimization problem. The specific algorithm to use depends on the characteristics of the multiobjective problem. This paper exploits heuristics to recommend travel routes and hybrid heuristics to solve optimal travel routes. The hybrid particle swarm genetic optimization algorithm is used to generate a single-object route first. Then, according to the user’s preference, the hybrid tabu search algorithm is used to optimize the object route.
3. Recommended Methods for Hybrid Tabu Travel Routes
This paper aims to dynamically determine the tourist routes according to users’ preferences. Meanwhile, according to the constraint conditions provided by users, the hybrid tabu search algorithm is used to recommend top-k travel routes, which can satisfy users’ personalized needs and maximize their satisfaction.
3.1. Symbols and Descriptions
In essence, the problem solved in this paper is to obtain the optimal solution of the multiobjective route. The relevant symbols for describing the problem are shown in Table 1.
Definition 1 (Attraction information). For each attraction a A, the information of the attraction can be represented as a vector , where is the recommended stay time in the attraction, is the ticket price, and is the level ranging from 1 to 5.
Definition 2 (Travel route). The set of travel routes is represented as . The number of attractions is i, i.e., . This set of travel routes consists of one or more attractions.
Definition 3 (Travel time). The travel time is the total time spent from the start position to the end position :The time from to includes the travel time from to the first attraction . The travel duration is the travel route time between two attractions, and the stay duration is the total visiting time of all attractions. is the stay time from the last attraction to the destination.
Definition 4 (Travel cost). The travel cost refers to the sum of all travel expenses during travel:
3.2. Recommended Framework
This paper proposes a new travel route recommendation framework, as shown in Figure 1. The framework is composed of three parts: crowdsensing score, multiple constraints, and hybrid tabu search algorithm. The function of the crowdsensing score includes preprocessing input interest points, obtaining the social score and location score of attractions by using the information of hotels and restaurants around attractions, calculating the dynamic time score according to the time between attractions, and calculating the similarity between the attraction tag and the user tag. These scores are taken as the input of the proposed travel route recommendation framework. The multiple constraints represent the user’s personalized constraints, including time constraint, cost constraint, and site constraint. The function of the hybrid tabu search algorithm includes dividing the neighborhood route, generating a single-objective route, finding the Pareto optimal set, and optimizing the route according to the user’s personalized preferences. By integrating the crowdsensing score and multiple constraints into the algorithm, the optimal travel route recommended to the user can meet the user’s personalized needs.
3.3. Crowdsensing Score Method
3.3.1. User Interest Score
Defining as the total set of tags of attractions, the user’s interest tag is and the tag of attractions is . Based on this, the user’s interest score can be represented as :
3.3.2. Crowdsensing Social Score
For attractions, the comments of netizens can reflect the reputation and features of attractions to a certain extent. The number of evaluations of the attraction is recorded as , and the user rates the site as . Besides, the level of the attraction is recorded as . Then, the crowdsensing social score of the attraction is expressed as follows:
The normalized value of iswhere and are the largest and the smallest crowdsensing social scores of the attraction, respectively.
As for hotels, the number of comments on the hotel is recorded as . The user rates the hotel as , and the level of the hotel is recorded as . Then, the crowdsensing social score of the hotel is expressed as follows:
As for restaurants, the number of comments on the restaurant is recorded as . The user rates the taste of the restaurant as , the environment as , and the service as . represents the total level of the rating, and is the level of the restaurant.
3.3.3. Crowdsensing Location Score
Assuming that the number of attractions within the radius of the attraction is , the social perception score of the attraction is standardized as . Then, the location score of the restaurant within the radius of the attraction can be represented as
The normalized value of is
The score of hotel location is
The score of restaurant location is
3.3.4. Comprehensive Location Score
According to the comprehensive score of the locations of restaurants, hotels, and attractions within the radius of the attraction, the crowdsensing comprehensive location score of the attractions is obtained, which is recorded aswhere are the corresponding weights.
3.3.5. Crowdsensing Score Fusion
Since the tag-of-interest matching score contains the user’s personal preference, the comprehensive score of crowdsensing is obtained from most of the user’s comments and geographic location analysis. Combining these two scores, the personalized comprehensive score based on crowdsensing, i.e., , can be represented aswhere is the balance factor used to adjust the weight of the interest tag matching score and the crowdsensing comprehensive score.
The normalized value of is
3.3.6. Time Score
Usually, a shorter distance between two points takes less stay time, and a more attractive attraction contributes to a higher score of the route. Based on this fact, the normalization of the stay time between two points iswhere is the stay time between two points.
3.3.7. Score Fusion
The dynamic fusion of the crowdsensing score and the time score is used to adjust the weight of the crowdsensing score and the time score to balance the total score of attractions.
3.4. Multiple Constraints
In the travel route recommendation, users often put forward some specific requirements, including the upper limit of the whole travel time, the upper limit of budget, the upper limit of the number of attractions, and indispensable attractions. These constraints reflect the personalized needs of users. Therefore, this paper proposes a multiconstraint model.
3.4.1. Time Constraints
The stay time at attractions should be more than half of the recommended time. Also, the time to reach attractions should be later than the opening time of the attractions. Besides, the time to reach the last attraction should be earlier than its closing time, and the total travel time should be less than the maximum time given by users. These constraints can be represented as
3.4.2. Cost Constraints
The sum of attraction tickets should not exceed the upper limit of user budget:
3.5. Design of the Hybrid Tabu Search Algorithm Based on Multiple Constraints and Multiple Objects
According to the definition given in the previous section and the establishment of the rating method, the details of the tabu search algorithm are introduced.
3.5.1. Multiobjective Hybrid Tabu Search Algorithm
In the multiobjective hybrid tabu algorithm, the variable neighborhood search algorithm is first exploited to generate a basic route that satisfies the user’s objectives and constraints. Then, taking the basic route as the initial population, the hybrid particle swarm genetic optimization algorithm is used to optimize the basic route. Also, the fast nondominated sorting algorithm is used to find the optimized route. Finally, the hybrid tabu algorithm is exploited to obtain the personalized and optimized route according to the user’s preferences. The specific steps are shown in Algorithm 1.
The conventional genetic algorithm cannot be directly applied to solve the top-k personalized travel route recommendation problem with multiple objectives and constraints. Considering that the genetic algorithm has better global search ability and the tabu search algorithm has better neighborhood search ability, the idea of the hybrid tabu search algorithm is first exploited to solve this problem and obtain multiple local optimal solutions in the process of iteration. By comparing the local optimal solutions, a global optimal solution is obtained. This idea can contribute to an excellent solution because it expands the search range. The specific steps of Algorithm 1 are as follows: Step 1. The variable neighborhood search algorithm is used to generate the basic route that satisfies the multiple objectives and constraints of users. Taking the basic route as the initial population, the hybrid particle swarm genetic algorithm is used to optimize the single-object route, and the fast nondominated sorting algorithm is used to obtain the Pareto solution of the route as the initial solution (see steps 1–2 for details). Step 2. We judge whether the iterative condition of the algorithm is satisfied. If not, the optimized results are output; otherwise, the crossover and mutation of the parent generation are performed to obtain the offspring (see steps 3–5 for details). Step 3. The neighborhood function is exploited to generate neighborhood solutions, and the candidate solution with the best fitness value is selected from these solutions and compared with the current optimal generation. If the candidate route is better, the current optimal generation is replaced with the candidate route; otherwise, the iteration continues (see steps 6–13 for details). Step 4. The best candidate route for the current iteration is compared with the best initial route. If the best current route is better, the initial route is replaced with the current route; otherwise, it is added to the tabu table (see steps 14–18 for details). Step 5. We calculate the route fitness and judge whether the termination condition of the algorithm is satisfied. If yes, the run is terminated and the result is output; otherwise, we go to Step 2 (see steps 19–26 for details).
|
4. Experiment
4.1. Experimental Settings
4.1.1. Dataset and Environment
The dataset obtained by using crawler software grabbing from Ctrip and Dianping has high authenticity. Specifically, the real data on 6114 attractions, 8171 restaurants, and 2528 hotels in Beijing were collected before August 2017. After filtering, 200 attractions were obtained and used as test datasets for this experiment.
The C-shaped 64-bit processing algorithm was exploited. Meanwhile, the experiment was conducted on a computer equipped with a 1.80 GHz Intel(R) Core(TM) i5-3337U CPU and 4.00 GB main memory.
The experiment was carried out with the real data from 200 classical attractions in Beijing. Under the constraints specified by users, the hybrid tabu search algorithm was used for travel route recommendations. The baseline algorithm is a greedy algorithm [17] and ant colony algorithm [23]. The evaluation metrics include running time, comprehensive attraction score, number of attractions, and interest richness.
4.1.2. Evaluation Metrics
(1) Interest Richness. The quality of a travel route is affected by many factors, such as the route time, the number of attractions, the degree of matching with users’ interests, and the type of points of interest. This paper takes interest richness as the metric to evaluate the recommended route, which is defined as follows:where represents the total time of the route, represents the number of attractions included in the route, and is the proportion of tags of all attractions in the recommended route occupying the user-specified tags of interest. A small number of attractions included in the route can also affect the user’s experience, so the absolute value is used in the above formula to limit interest richness. is the best number of attractions in the recommended route. If , interest richness increases with the number of attractions; otherwise, interest richness decreases with the number of attractions.
(2) Route Score. This paper takes the route comprehensive score as one of the metrics to measure the quality of the recommended travel routes. A higher route score means that the recommended route can better meet the users’ actual preferences.
(3) Time Performance. On the same data scale and hardware and software platforms, the time needed by the route recommendation algorithm to complete is an important indicator of the algorithm performance. The running time of the experiment is in seconds, and it is increased by 2 seconds to facilitate image drawing.
(4) Number of Attractions. Under the same constraints, the number of recommended attractions that can meet the needs of users and improve their satisfaction is about 6 spots.
4.1.3. Baseline and Proposed Algorithms
The genetic algorithm and the ant colony algorithm are compared with the hybrid tabu search algorithm proposed in this paper:(i)GA [16]: The genetic algorithm is a heuristic algorithm without backtracking. It divides the whole solution process into several stages and obtains the global optimal solution by solving the local optimal solution in each stage.(ii)ATP [22]: The ant colony algorithm puts all ants in a starting position and then selects the next edge step by step according to the transition probability. If the ant has traveled all the points or is not allowed to visit any more, the ant has completed its journey.(iii)MOVNS [28]: In this paper, a novel multiobjective and multiconstraint tour route recommendation method is proposed. First, ArcMap was used to model the actual road network. Then, a new interest label matching method and a utility function scoring method were created based on crowdsensing, and a personalized multiconstraint interest model was constructed. A variable neighborhood search algorithm and a hybrid particle swarm genetic optimization algorithm were presented for recommending top-k routes.(iv)MOTHA: The hybrid tabu search algorithm proposed in this paper exploits the variable neighborhood search algorithm and the hybrid particle swarm optimization genetic algorithm to optimize the original route. The top-k optimal solutions are obtained by using Pareto quick sorting, and then, the hybrid tabu search algorithm is used to generate personalized and optimized routes according to the user’s preferences.
4.2. Experiment Results
4.2.1. Sensitive Analysis
This paper sets the radius of crowdsensing to 2. The restaurant taste, environment, and service rating level are all 10. We set , and the default values of the three weighting coefficients are set to , , and , and . The default value of the balance parameter is set to 0.5 in the crowdsensing user’s personalized comprehensive score. As for the multiobjective hybrid tabu search algorithm, the default values of and are set to 100 and 190, respectively.
(1) Neighbor Threshold. Taking interest richness as the metric, it can be seen from Table 2 that the change in falls in the range of 150–200. Meanwhile, the interest richness under the score object increases with until = 190. The number of attraction objects is the same as the score object. Thus, is set to 190.
(2) Threshold. Figures 2(a) and 2(b) present the results of optimizing the route with the highest score. Figures 2(c) and 2(d) present the results of optimizing the route with the largest number of attractions. It can be seen from Figure 2(a) that the running time increases with the number of iterations. Also, as shown in Figure 2(b), interest richness is stable after the reaches 100, so is set to 100 under the score object. Figures 2(c) and 2(d) show the results of 20,000 iterations. It can be observed that the slope shown in Figure 2(c) is larger than that in Figure 2(a). Besides, large step growth before 8,000 iterations is shown in Figure 2(d), and the trend becomes stable thereafter. Therefore, the scenic object is set to 8000.
(a)
(b)
(c)
(d)
4.2.2. Performance Evaluation
(1) Performance Contrast. The parameters of the MOHTA algorithm were optimized. The MOHTA, GA, MOVNS, and ATP algorithms were made to run 10 times, respectively. Then, the maximum number of attractions, the highest route score, the highest interest richness, and the running time generated by these four algorithms were compared, and the results are illustrated in Figure 3.
(a)
(b)
(c)
(d)
As shown in Figure 3(a), the average number of attractions obtained by the MOHTA algorithm is 12, which is significantly larger than that of the ATP and GAs, indicating that the MOHTA algorithm is superior to the ATP and GAs in the recommended number of routes under the same conditions. Meanwhile, it can be seen from Figure 3(b) that the average score of the route obtained by the MOHTA algorithm is significantly higher than that of the ATP, MOVNS and GAs. Besides, the result shown in Figure 3(c) indicates that the MOHTA algorithm achieves greater interest richness than the ATPMOVNS and GAs. Especially, the MOHTA algorithm achieves twice the interest richness as GA. Moreover, it can be seen from Figure 3(d) that the average running time of the ATP algorithm is the longest and that the average running time of the MOHTA algorithm is longer than that of the GA, but the results are at the same magnitude. From the above results, it is concluded that the MOHTA algorithm is superior to the ATP, MOVNS and GAs in terms of the number of attractions, route score, interest richness, and running time, verifying the effectiveness of the MOHTA algorithm.
(2) Effect of the Single-Object Route. In this section, the single score object generated by MOHTA is compared with that of GA, MOVNS, and ATP, as shown in Figure 4.
(a)
(b)
(c)
(d)
The results of the score object optimization are illustrated in Figures 4(a) and 4(b), and the results of the number of attraction object optimizations are shown in Figures 4(c) and 4(d). It can be seen from Figures 4(a) and 4(b) that the route score generated by the MOHTA algorithm is higher than that generated by the ATP, MOVNS, and GAs. The number of attractions generated by the MOHTA algorithm in the score object is 6, which is greater than the number of attractions generated by the ATP algorithm. From Figures 4(c) and 4(b), it can be seen that the scores of top 10 routes generated by the MOHTA algorithm are compared between those of the ATP algorithm and GA. The reason is that the more attractions involved in the route, the shorter the recommended travel time of a single attraction. Although the route score generated by the MOHTA algorithm is low, the number of attractions recommended by the MOHTA algorithm is 12, which is more than twice the number of attractions generated by the GA, which can meet users’ needs for multiple attractions.
(3) Pareto Optimal Solution versus Single-Objective Solution. First, the solutions obtained by the Pareto, ATP, MOVNS, and GAs are compared, and the results are shown in Figure 5.
(a)
(b)
It can be seen from Figures 5(a) and 5(b) that 6 of the route scores and the number of attractions for the Pareto top 10 optimal solutions are greater than those for the ATP, MOVNS, and GAs. The reason is that the Pareto solution can balance two objects, and the Pareto optimal solution generated by the MOHTA algorithm is better than that generated by the ATP, MOVNS, and GA.
Then, the multiobjective Pareto route generated by the MOHTA algorithm is compared with the single-objective route, and the result is shown in Figure 6.
(a)
(b)
(c)
(d)
As shown in Figure 6(a), SO denotes the score as the only optimization object and AO denotes only the number of attractions as the only optimization object. It can be seen that the number of attractions in the route under the score object is fewer than that generated by the Pareto route. Meanwhile, the result in Figure 6(b) indicates that the route score generated by the score object route is higher than that of the Pareto route. The reason is that the score object route pays more attention to the route score than the Pareto route, so the route score is higher, and the number of attractions generated is not as many as that generated by the Pareto route. Besides, in Figure 6(c), it can be seen that the number of attraction object routes is more than that generated by the Pareto route, but in Figure 6(d), the route score generated by the number of attraction object routes is fewer than that generated by the Pareto route. The reason is that compared with the Pareto route, the attraction object route pays more attention to the number of attractions, and the generated score is not as high as that generated by the Pareto route. As can be seen from the above figure, the single-object route only focuses on one object, while the Pareto route is a multiobjective compromise, which balances the number of attractions and route scores.
4.2.3. Route Simulation
The hybrid tabu search algorithm is exploited in this paper. Meanwhile, ArcGIS is used to simulate the routes generated by the ATP algorithm with the highest quality, those generated by the GA, those generated by the MOVNS algorithm, the optimal score route generated by the MOHTA algorithm, the routes with the maximum number of attractions, and the Pareto optimal routes with a compromise between the score and number of attractions. The results are detailed in Figure 7.
(a)
(b)
(c)
(d)
(e)
(f)
As can be seen from the route of the GA in Figure 7(a), the route generated by the GA does not duplicate the route, but the scenic spots of the route are scattered, especially the last summer palace deviates from other routes. The simulation results of Figures 7(b)–7(e) show that the generated route attractions are scattered and that the routes between attractions overlap, with a small number of attractions. As can be seen in Figure 7(f), the number of attractions along the route is large, the distribution of attractions is concentrated, and the order of attractions is reasonable. In general, the routes recommended by Pareto’s multiobjective routes meet user constraints well. The score and the number of attractions are also more in line with user requirements.
5. Conclusion
Based on the multidimensional preferences of tourists, an objective function based on crowdsensing is constructed in this paper, which focuses on the multidimensional preferences of tourists to improve travel routes. The hybrid tabu search algorithm is exploited to optimize multiobjective personalized travel routes. The work in this paper can be further improved in three ways in the future. First, the data on tourists’ preferences can be collected from more sources and more heterogeneous individuals. Second, the running time of the hybrid tabu search algorithm can be optimized. Third, experiments need to be conducted on the attractions of other cities to validate the effectiveness of the hybrid tabu search algorithm.
Data Availability
The datasets generated during the current study are available from the corresponding author on reasonable request or from the URL https://doi.org/10.21227/cnfs-6p81.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Acknowledgments
This research was partially supported by the Natural Science Foundation of China (No. 62272006, 41930644), the Key Research and Development Project of Wuhu (No. 2022yf55), the Collaborative Innovation Project of Anhui Province (GXXT-2022-093), the Key Program in the Youth Elite Support Plan in Universities of Anhui Province (gxyqZD2019010 and gxyqZD2020004), the Key Program of the National Key Research and Development Program of China (No. 2018YFB2101300), the Natural Science Foundation of Anhui Province (No. 2108085MF214), and the National Trusted Embedded Software Engineering Technology Research Center, East China Normal University.