A planning model for dedicated tourist bus routes based on an improved genetic-greedy algorithm and machine learning

Model for determining the combined weights

AHP method is a kind of systematic analysis method proposed by Saaty (1977) to carry out effective assessment research by quantifying the relative importance of different indicators, which is now widely used in the weighting analysis of multi-level indicators. The main steps of the AHP method include the establishment of a recursive structural model, the construction of a judgment matrix, the calculation of weights, and the consistency test.

The entropy weight method (EWM) is a well-established technique for objective weight allocation. It is based on the degree of variation for each indicator, calculating the entropy weight for each through information entropy analysis. This method adjusts the weights assigned to each indicator and enables more objective weight distribution.

According to the above two weighting methods to determine the subjective and objective weights, respectively through the entropy weighting method and the AHP method find the weights by applying the following formula (1) for the synthesis can be obtained by calculating the integrated weights $w$. The weight α and β values in the formula are obtained using the AHP method and the entropy weight method, respectively. We selected AHP for its ability to structure complex decision problems into hierarchical levels and to derive criterion weights through expert-based pairwise comparisons. AHP incorporates a consistency check (CR < 0.1), ensuring reliable expert judgment. Its interpretability and hierarchical logic have made it a widely accepted tool in tourism planning and transport decision-making (Munier & Hontoria, 2021). However, AHP is inherently subjective. To enhance objectivity and reduce expert bias, we introduce the EWM, which determines weights by measuring the dispersion of evaluation data. EWM emphasizes criteria with higher variability, allowing data-driven emphasis on impactful factors (Zhu, Tian & Yan, 2020). This is particularly useful in our case, where real traveler feedback (processed via sentiment analysis) is used to extract satisfaction indicators. Compared to other multi-criteria decision making (MCDM) methods such as Decision-Making Trial And Evaluation Laboratory (DEMATEL) (which emphasizes interdependence but lacks prioritization mechanisms) or CRiteria Importance Through Intercriteria Correlation (CRITIC) (which assumes linear correlations and ignores subjective knowledge), our Analytic Hierarchy Process—Entropy Weight Method (AHP-EWM) hybrid framework strikes a balance between domain expertise and data variability. We acknowledge that AHP can be affected by rank reversal or expert inconsistency, and that entropy-based weighting may overemphasize noise in small datasets (Talero-Sarmiento et al., 2024). To mitigate these, we (1) enforce consistency ratio thresholds (CR < 0.10) and (2) smooth raw indicator values before entropy computation. This ensures the robustness and validity of our combined weighting scheme.

(1) $$w={\displaystyle \frac{\sqrt{{\alpha }_j\ast {\beta }_j}}{{\sum }_{j=1}^n\sqrt{{\alpha }_j\ast {\beta }_j}}.}$$

Selection of attractions for combinatorial empowerment TOPSIS modeling

The TOPSIS model is a comprehensive evaluation method that ranks evaluation objects according to their proximity to an idealized target. The following is the process of integrating the utilization of TOPSIS and the new weight vector:

First, because the lower the fare, transportation cost, etc., the better, while the higher the comfort and accessibility, the better, it is necessary to first do the uniform normalization preprocessing of the data through Equation. The matrix has been normalized to obtain the standardized scoring matrix. Then calculate the maximum value of each evaluation indicator and the minimum value of each evaluation indicator.

Finally, the closeness of each evaluation object to the maximum value to the maximum and minimum values is calculated.

Selection of attractions for the combined empowerment RSR model

Rank-sum ratio (RSR (Wang et al., 2015) method) is a statistical analysis method that combines the advantages of classical parametric statistics and modern nonparametric statistics.

We first used the TOPSIS method comprehensive evaluation score as the algorithmic starting score WRSR in RSR evaluation. Subsequently, we calculated the cumulative frequency and converted it to the probability unit probit value, and calculated the routear regression equations with Probit as the independent variable and RSR value as the dependent variable according to Eq. (2):

(2) $$WRSR=a+b\ast probit.$$

Finally, based on the fitted value of WRSR calculated by the regression equation, the evaluation objects are ranked and graded, and their validity is verified by comparing the two sides.

Improved genetic-greedy algorithm for determining lines

Greedy algorithm (Alur, Henzinger & Kupferman, 2002) was proposed by Huffman in 1950, which refers to the process of finding the optimal solution to the problem based on some greedy criteria, starting from the initial state of the problem, directly seeking the optimal solution of each step. Genetic algorithm (Liu et al., 2022) is a stochastic global search optimization method, proposed by Professor Holland in 1969. It is a stochastic global search optimization method, which prevents the algorithm from falling into a local optimal solution through crossover, mutation, iteration, and so on.

Therefore, this article proposes to improve the greedy algorithm with comprehensive weights and genetic algorithm ideas to realize the bus line planning to achieve a significant increase in the optimal convergence speed and global optimal exploration. The diagram of the designed algorithm framework is shown in Fig. 1. The steps and ideas of the improved algorithm are as follows:

• •

  Step 1: Greedy selection

  Using the classic greedy algorithm, we select the attraction with the highest combined weight value of the indicators along the route (choosing a positive indicator; if the indicator is inverse, we select the one with the lowest weight). We repeat this process until all attractions have been visited.

• •

  Step 2: Genetic mutation

  Based on the local optimal path results of the weighted greedy algorithm in step 1, the obtained solution and the start and end points of the random two attractions are exchanged to obtain several offspring that constitute the initial population.

• •

  Step 3: Iterative optimization

  For the principles and indicators of bus line optimization, the following Eq. (3) defines the fitness function to evaluate the fitness of different individuals. (3) $$f_1\left(s\right)=\sum _{i=1}^{n-1}{W}_{i,i+1}+W_{n,1}/f_2\left(s\right)={\displaystyle \frac{1}{{\sum }_{i=1}^{n-1}{W}_{i,i+1}+W_{n,1}}.}$$ $f_1$ is the function defined when the route indicator is forward. $f_2$ is the function defined when the route indicator is reversed. $s$ is each body in the initial population, i.e., the bus route scheme. $W_{i,i+1}$ is the weight of attraction i to i + 1. $n$ is the number of attractions and $W_{n,1}$ is the weight of the end attraction to return to the starting point.

• •

  Step 4: Select highly adapted individuals from the initial population for crossover and mutation

• •

  Step 5: Select a new generation of populations based on fitness

• •

  Step 6: Repeat steps 4–5 until termination conditions are met.

When repeating steps 4–5 to get many sub-generation data, it should be given its termination conditions to get the convergence results and prevent the algorithm from running indefinitely. Hence, we take the average fitness improvement value of 100 times after iterations as the judgment data. To enhance the replicability of the proposed method, we summarize the execution flow of the improved GGA as follows in a simple manner: (1) the algorithm uses a greedy strategy to construct an initial route. At each step, it selects the attraction with the highest combined score based on the weighted indicators. This process continues until all attractions are visited, producing a locally optimal route. (2) this initial solution is refined using a genetic algorithm framework. New candidate solutions are generated by randomly exchanging two nodes in the route to form a population. Each individual (i.e., route) is evaluated using a fitness function that incorporates both travel distance and congestion. (3) Third, individuals with higher fitness are selected for crossover and mutation operations to generate the next generation of solutions. The population is updated iteratively, and fitness is re-evaluated after each generation. The algorithm terminates when the average fitness improvement over 100 generations falls below a predefined threshold. The best-performing route at termination is selected as the final solution. In the fitness function used during population evaluation, the parameters α and β represent the relative importance of two key route-level indicators: travel distance and congestion. These values are derived from the previously calculated combined weights in Table S5, following the removal of indicators unrelated to route structure (e.g., ticket price, accommodation). After normalization, the resulting weights for distance and congestion are 0.1852 and 0.1468, respectively, which serve as α and β in the fitness evaluation. This ensures consistency between the attraction evaluation phase and the route optimization process.

Data set cleaning, model selection comparison, and indicator identification

In this study, we selected the Meituan APP as the source of review data and used Python to crawl customer reviews for travel products like “7-day tour” and “Everest 3-day tour” from Meituan and Ctrip platforms. After deduplication, 2,932 comments were processed using the Jieba library for word segmentation. By filtering stop words and eliminating distractions, relevant information was extracted based on a lexicon. We then employed Word2Vec to encode the comment data, dividing it into a training set (70%) and a test set (30%) for double cross-validation. The training set was further split into 10 samples for repeated trials to minimize randomness and ensure model stability for indicator selection.

Commonly used models for sentiment classification in large datasets include AdaBoost (Walker, 2021), support vector machine (Yang et al., 2022), naive Bayes (Hassan, Rafi & Shaikh, 2012), and k-nearest neighbors (KNN) (Gejun, Changsheng & Feng, 2010) models. Using the above 10 training sets, the model is classified as positive, neutral, and negative to conduct model training and accuracy test, and the accuracy an $R^2$ calculation results are used as evaluation indexes for model comparison. Figure 2 show the comparison of line graphs vs. box-and-line plots of the test accuracy results for the four models on the training set, respectively:

The R² parameters of the four models are 0.000465, 0.000464, 0.000921, and 0.00102, respectively. As can be seen, by the results of the $R^2$ calculations and the line lengths exhibited in Fig. 2, the four models of AdaBoost and SVM show the best stability. The simple Bayesian model and the KNN model show a higher instability; compared to AdaBoost accuracy, SVM has a greater lead, so choose SVM for the indicators to select the judgment model. To improve the technical transparency and reproducibility of our classification pipeline, we provide implementation details as follows: (1) SVM implementation, we used an SVM with a linear kernel, chosen due to the high-dimensional (1,000) and sparse nature of the TF-IDF feature space. Hyperparameters were tuned via grid search. The final settings were: ‘C = 1’, ‘gamma = ‘scale’’, ‘degree = 3’, ‘coef0 = 0.0’, ‘shrinking = True’, ‘probability = False’, and ‘tol = 1e−3’; (2) AdaBoost implementation, short for Adaptive Boosting, is an ensemble learning technique that combines multiple weak classifiers to form a strong classifier. AdaBoost was implemented using 100 CART-based decision stumps (i.e., ‘max_depth = 1’, ‘criterion = ‘gini’’) as weak learners. The learning rate was set to 0.8 based on empirical tuning for a balance between bias and variance; (3) Feature extraction and preprocessing. We first used the Jieba tokenizer to segment Chinese reviews and removed stop words. We then applied Word2Vec for semantic vector encoding and further transformed the output using TF-IDF weighting, resulting in high-dimensional sparse vectors used for classifier training; (4) Sentiment labeling and classification targets, each review was anually labels as Positive (containing words like ‘Highly satisfied’, ‘Pleasant’, ‘Would recommend’), Negative (containing words like ‘Bad review’, ‘Disappointed’, ‘Not recommend’), and Neutral (containing no explicit positive or negative sentiment). These labels served as classification targets; (5) Evaluation metrics, we report precision, F1-score and recall for all models over 10 random train/test splits. We also report metric variances to validate model stability.

The trained SVM classifies 30% of the test set. We take the good reviews and the bad reviews to carry out the keyword statistics, and the statistics show that the top six most frequently mentioned reviews of the tourism products in Tibet are: tickets, accommodation, comfort, motivation, distance, and degree of congestion. Therefore, we determine these six items as the key attention indicators of Tibet tourism satisfaction.

Data sources and visitor characterization

Popular attractions in the Tibet Autonomous Region include many long-distance attractions outside Lhasa, so we first screened out 10 attractions as the evaluation range, based on which we further evaluated to verify the model effect. In this study, 150 questionnaires were randomly transmitted within 72 h to determine mainstream tourists’ travel time, travel range, and attraction selection. The results of the survey are shown in Figs. S1, S2 and S3. To ensure sample diversity and reduce bias, participants were recruited through two complementary channels: (1) Online distribution, by posting the questionnaire link on social media platforms and travel forums; (2) On-site surveys, by setting up temporary stations at major tourist sites in Tibet (e.g., Potala Palace, Jokhang Temple). All participants took part voluntarily. In constructing the evaluation model, we selectively incorporated seven questionnaire items directly into the modeling pipeline as input features related to satisfaction estimation. These included: travel duration, travel range, primary travel motivation, preferred travel style, transportation preference, travel budget, and accommodation preference. These variables served as inputs for sentiment-driven feature construction and informed the weighting process used in the TOPSIS-RSR framework.

To ensure data reliability, we applied a four-stage cleaning strategy to the collected questionnaire data: (1) Logic consistency checks were conducted to identify contradictory answers (e.g., selecting both ‘solo travel’ and ‘traveling with family’); such responses were either confirmed with participants or excluded. (2) Extreme value handling involved detecting implausible values (e.g., travel budgets of ‘1 yuan’ or ‘1 million yuan’) and excluding them if no justification was provided. (3) Missing data were addressed using multiple imputation to retain partially completed surveys while minimizing information loss. (4) Duplicate entries were filtered out by examining IP addresses and device identifiers to ensure the independence of responses.

The top 10 tourist attractions are The Potala Palace (A), Namtso Lake (F), The Jokhang Temple (B), Yamdrok Yumtso (O), Bartsuncuo (K), Migdui glacier (L), Mount Qomolangma (M), Mount Kailash (P), Norbu Lingka (D), and Yarlung Zangbo Daxiagu (G).

Therefore, these 10 attractions were selected as samples in the TOPSIS evaluation model. On this basis, we asked 80 tourists and 20 experts to evaluate attractions and quantify indicators (e.g., comfort and other non-quantitative indicators) for the above 10 attractions. The scores are shown in Table S1.

Calculation of weights

TOPSIS attraction evaluation

The data for each indicator has been processed, so the evaluation matrix can be calculated as shown in Table S6. Each attraction was rated as follows in Table S7.

By the final TOPSIS evaluation of the results, the attractions were scored in the following order: 1. Mount Qomolangma, 2. The Potala Palace, 3. Norbu Lingka, 4. Yarlung Zangbo Daxiagu, 5. Namtso Lake, 6. Bartsuncuo, 7. Yamdrok Yumtso, 8. The Jokhang Temple, 9. Migdui glacier, 10. Mount Kailash.

Additional evaluation metrics

We acknowledge that the use of R² as a performance metric is inappropriate for sentiment classification tasks, which involve discrete categorical outputs rather than continuous variables. Accordingly, we used evaluation metrics that are more suitable for classification: F1-score and recall. These are widely accepted in the machine learning community for assessing model effectiveness in multi-class sentiment categorization. Specifically, we conducted classification experiments using 10 distinct test subsets to ensure robustness. For each model, we computed the mean and variance of both F1-score and Recall across these test sets, as reported in Tables S8 and S9. This multi-run evaluation framework captures the variability in model performance and ensures that the reported results are not subject to randomness or overfitting on a single test split. The results indicate that the SVM consistently outperforms other classifiers, including decision trees, random forests, and AdaBoost, in terms of both average performance and stability (i.e., lower variance). This provides strong empirical justification for selecting SVM as the core sentiment classification model in our pipeline.

Hyperparameter sensitivity analysis

To assess the robustness of our hybrid weighting framework, we conducted a sensitivity analysis by varying the contribution ratios of subjective (AHP-based) and objective (entropy-based) weights. Specifically, we tested three representative configurations: 10/90, 50/50, and 90/10. The resulting composite weights for each attraction indicator and the corresponding rankings are summarized in Tables S10 and S11. Our analysis reveals that the overall attraction rankings are relatively stable across different weight combinations, particularly for the top-ranked sites. This consistency suggests that the model’s prioritization is not overly sensitive to changes in the relative emphasis of expert judgment versus data-driven dispersion. For example, attractions such as X and Y consistently appear within the top three across all configurations, indicating their robustness under both subjective and objective criteria. Minor variations are observed in the middle- and lower-ranked items, which is expected given the closer composite scores in those tiers. These results support the structural validity of our dual-weighting scheme and affirm that the ranking outputs are not artifacts of arbitrary parameter settings, but rather reflective of genuine multidimensional performance differences.

Evaluation of RSR attractions

The integrated score index and ranking calculated by the TOPSIS method are the basis of the RSR algorithm. The probit value is calculated the routear regression equation is obtained by the steps, and the WRSR fitted value is obtained in accordance with the equation. The results of the RSR method and the fitted linear regression equations are shown in Fig. 3 below.

The fitted routear regression equation is as follows:

$$WRSR=-0.00934+0.0205\ast Probit.$$

As seen from the Origin fitting, the $R^2=0.82095$ and the residual sum of squares is 0.00101, which indicates that the model is reasonable and reliable, with good simulation effect, and meets the needs of the evaluation model.

After building the TOPSIS-RSR model evaluation, the reasonable WRSR fitting value ranking of the attractions is obtained, based on which the attractions are graded as recommended attractions, not recommended attractions.

Dijkstra’s algorithm optimal route solution based on weight optimization

Attraction evaluation is different from route planning. Therefore, we further narrow down the focus of the six indicators, taking distance, comfort, and congestion as three key indicators for route planning, of which the comfort of the route is closely related to the travel mileage, transportation, congestion (Shao et al., 2022). Usually, the longer the journey time, the passenger’s fatigue may increase, resulting in a reduction in the level of comfort. Consequently, the impact of comfort can be distributed across the dimensions of distance, transportation, and congestion (Ha, Lee & Ko, 2020). Among these, distance and congestion emerge as the two pivotal indicators (Yin et al., 2020).

We divided the congestion as follows: 5 (High), 4 (Moderately High), 3 (Medium), 2 (Medium-Low), 1 (Low). Data on the distance and congestion between the eight recommended attractions has been collected. In this analysis, the weights assigned were 0.1852 for distance and 0.1468 for congestion. The path scoring table can be obtained and make the undirected graph shown in Fig. S4.

Since this undirected graph follows the 2.4 GGA step to find the optimal line:

A -> B -> C-> G -> D -> E -> H -> F

Namely: The Potala Palace-> Namtso Lake-> The Jokhang Temple-> Norbu Lingka-> Yamdrok Yumtso-> Bartsuncuo-> Yarlung Zangbo Daxiagu-> Mount Qomolangma

In this example study, we compare the performance of GGA and Classical Genetic Algorithm (CGA) in determining the optimal travel route. The experimental results are remarkable, as the comparison of the iteration diagrams (Fig. 4) shows that the GGA reaches the optimal solution after 533 iterations. In comparison, the CGA needs 9,671 iterations to reach the same route convergence result. Compared with the classical genetic algorithm, which required 1,872.2 iterations to converge to an optimal route solution, the proposed GGA model achieves convergence in only 103.1 iterations, resulting in a 94.45% reduction in convergence steps. This significant improvement implies not only faster computational efficiency but also the model’s ability to more rapidly identify high-quality solutions that align with human satisfaction scores. In practical terms, the optimized routes generated by the GGA yield higher alignment with tourist preferences (as validated via expert satisfaction scoring) and reduce total transit time by avoiding redundant loops or inefficient sequences. Therefore, the improvement translates to both enhanced satisfaction and more efficient route execution in real-world scenarios.

Modeling evaluation

Since the model is designed to create a dedicated tourist bus route based on passenger satisfaction, we analyze the results by comparing the scores of the routes generated by the model with existing popular routes in Tibet. Additionally, we conduct evaluation surveys with mainstream tourists to assess their satisfaction levels regarding the tourist route. This allows us to analyze the model’s results from both perspectives.

First, we look for 30 tourists, each with corresponding traveling days and vote for the following five mainstream routes and six routes obtained from the model. The routes and statistical results are as follows (D is the model route):

• •

  Method 1:

  A: Lhasa–Linzhi–Yarlung Zangbo Daxiagu–Namcha Barwa Peak–Lulang Forest–Bomi–Ranwu Lake in Yunnan–Laiku Glacier;

  B: Lhasa–Shigatse–Everest Base Camp–Jomolang Monastery–Saga–Tarkin–Gonpozi–Marbunyongtso–Layangtso–Pulan–Zada Tulin–Gugu Dynasty Ruins–Tarkin–Saga–Shigatse;

  C: Lhasa–Namtso–Sacred Gate of Heaven–Nagchu–For example–Sapo Mountain–For example–Nagchu;

  D: The Potala Palace–Namtso Lake–The Jokhang Temple–Norbu Lingka–Yamdrok Yumtso–Bartsuncuo–Yarlung Zangbo Daxiagu–Mount Qomolangma;

  E: Lhasa–Linzhi–Bomi–Ranwu Lake–Laiku Glacier–Lulang Forest–Yarlung Tsangpo River Grand Canyon–Namcha Barwa Peak–Sosong Village–Linzhi;

  F: Lhasa–Yamdrok Yumtso–Karola Glacier–Manla Reservoir–Gyangzigu Fortress–Shigatse–Zashilumbu Monastery–Tingri–Jomol Monastery–Everest Base Camp–Tingri–Saga–Tarkin–Gonchabozi– Marbunyongtso–Layangtso–Pran–Zadazhaputra–ruins of Gugu Dynasty–Tarkin–Saga–Rikaze.

• The statistics and comparison results are presented in Figs. 5, including the number of travel days for tourists other than the typical 3-day trips. The D route received the fewest votes overall, while other categories of tourists on the D route received the highest number of votes. The D route garnered 70 votes, accounting for 46.67% of the overall votes, significantly surpassing the other five routes.

• •

  Method 2:

  We conducted route evaluation for mainstream tourists (traveling time of 7 days). After transmitting 330 questionnaires and 20 expert consultations, the total number of people who were satisfied with the route was 331, with a satisfaction rate as high as 94.6%.

The results obtained from the above two surveys comprehensively verified that the route obtained from the model has a high level of tourist satisfaction, proving the scientificity and accuracy of the route evaluation method.