A scenic spot tourism demand prediction method and system based on dynamic weight fusion
By preprocessing and dynamically weighting structured data and tourist review texts, the problem of dynamic changes in tourists' subjective feelings and public opinion evaluations in scenic area tourism demand forecasting is solved, achieving high-precision future visitor flow forecasting and decision support.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INNER MONGOLIA UNIVERSITY
- Filing Date
- 2026-03-31
- Publication Date
- 2026-07-07
AI Technical Summary
Existing technologies cannot effectively consider the dynamic changes in tourists' subjective feelings and public opinion in predicting tourism demand in scenic areas, resulting in insufficient prediction accuracy and adaptability.
A dynamic weight fusion-based approach is adopted. By preprocessing structured data and tourist review texts, structured feature vectors and sentiment temporal feature vectors are generated. Then, an adaptive dynamic Bayesian smoothing algorithm and a dynamic weight generation network are used to dynamically adjust the importance of the sentiment dimension and generate future visitor flow predictions.
It achieves fine-grained analysis of multiple sentiment dimensions, improves the adaptability and accuracy of the prediction model, can dynamically adjust the importance of features in different contexts, and outputs clear decision support reports.
Smart Images
Figure CN122347441A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of intelligent prediction of tourism demand in scenic areas, and in particular to a method and system for predicting tourism demand in scenic areas based on dynamic weight fusion. Background Technology
[0002] With the deepening of smart tourism development and the popularization of big data technology, using data models to predict tourist demand in scenic areas has become a key technical means to improve management efficiency and optimize visitor experience. Accurate visitor flow forecasting can help scenic areas allocate resources in advance, provide early warnings and guidance, and formulate marketing strategies.
[0003] Currently, the technical solutions in this field can be divided into the following three categories based on the data dimensions and model methods they employ; 1. Prediction schemes based on traditional structured data: This scheme relies entirely on structured data and primarily uses time series analysis models (such as ARIMA, seasonal ARIMA) or traditional machine learning regression models (such as support vector machines, random forests). The fundamental drawback of this scheme is that it completely ignores the impact of tourists' subjective feelings and public opinion on tourism demand, and cannot cope with demand fluctuations caused by unstructured factors such as online word-of-mouth.
[0004] 2. A prediction scheme incorporating overall sentiment analysis: Building upon traditional structured data prediction schemes, this scheme introduces overall sentiment analysis. First, using a sentiment lexicon or basic sentiment analysis model, sentiment calculations are performed on the collected scenic area reviews to derive a single numerical value or classification label representing the overall evaluation tendency. Then, this overall sentiment score is used as a new feature and input into the prediction model along with the traditional structured data for training and prediction. While this scheme incorporates textual data, its analytical granularity is too coarse. Compressing tourists' complex evaluations of transportation, service, and scenery into a single score results in the loss of a significant amount of valuable information. It fails to distinguish between "positive scenery but negative service" and "positive service but negative transportation," leading to limited discriminative power of features extracted from the text.
[0005] 3. Predictive schemes employing multi-dimensional sentiment analysis but static fusion: These schemes typically use predefined sentiment dimensions (such as "food, accommodation, transportation, sightseeing, shopping, and entertainment") and obtain tourists' sentiment ratings across each dimension through relatively complex text analysis models. However, such schemes suffer from significant bottlenecks in the feature fusion stage. They usually involve simply concatenating the sentiment features obtained from all dimensions with traditional structured features or fusing them using fixed weighting coefficients. This static fusion mechanism assumes that the importance of the "transportation" dimension is constant on sunny and rainy days, and on weekdays and holidays. This clearly contradicts tourists' actual, context-dependent decision-making logic, limiting the upper limit of the model's predictive accuracy and scenario adaptability. Summary of the Invention
[0006] The purpose of this invention is to provide a method and system for predicting tourism demand in scenic areas based on dynamic weight fusion. This method enables fine-grained sentiment analysis of multiple sentiment dimensions (scenery, transportation, food, tickets, and services), avoiding the loss of a large amount of valuable information. Furthermore, it can dynamically and adaptively adjust the importance of multiple sentiment dimension features under different contexts, significantly improving the adaptability and prediction accuracy of the time series prediction model.
[0007] This invention is implemented by the following technical solution: a method for predicting tourism demand in scenic areas based on dynamic weight fusion, comprising: S1: Collection and preprocessing of structured data and tourist review texts of scenic spots to generate structured feature vectors and high-quality tourist review texts; S2: Transform the preprocessed tourist review texts into sentiment temporal feature vectors; S3: Dynamically weight fusion is performed on the preprocessed structured data and the sentiment time series feature vector to generate time series passenger flow prediction values and dynamic weight vectors. S4: Post-process and visualize the time series passenger flow forecasts and dynamic weight vectors to output a decision support report.
[0008] Furthermore, step S1 includes the following steps: S10: Obtain structured data and tourist review texts of the scenic area; S11: Fill in missing structured data with mean / median; perform sine-cosine encoding on structured data with periodic features; perform one-hot encoding on structured data with categorical features; standardize continuous structured data to form multi-dimensional structured feature vectors; S12: Remove irrelevant characters from tourist review text, standardize encoding format, convert emojis and filter noisy comments, and output standardized, high-quality tourist review text.
[0009] Furthermore, step S2 includes the following steps: S20: Using a large language model and human expert verification strategy, sentiment intensity labels are assigned to the preprocessed tourist review texts to generate a dataset of sentiment intensity labels; S21: Supervised fine-tuning of the pre-trained language model using a dataset labeled with sentiment intensity; S22: Obtain the collection of comment texts generated on the same day across multiple sentiment dimensions, input each comment in the collection into a pre-trained language model, and output the predicted score for each comment; S23: Determine the set of valid comments for the day based on the valid comment determination threshold and the predicted score of each comment; S24: Given the set of valid comments for the day, use the adaptive dynamic Bayesian smoothing algorithm to calculate the sentiment score for the sentiment dimension for the day.
[0010] Furthermore, the pre-trained language model includes a composite loss function, which is: ; in, This represents the mean squared error loss term. This represents the consistency loss term in the Pearson correlation coefficient; , The hyperparameter representing the balance between the contributions of the two loss terms; The calculation formula is ; In the formula, To train the number of batches of visitor review texts, Indicates the first The true emotional label of a tourist's review text This represents the predicted value of the pre-trained language model; The formula for calculating Lpcc based on the Pearson correlation coefficient is as follows: ; In the formula, This represents the mean of the true labels for tourist review texts within the training batch. This represents the mean of the predicted values of visitor comment texts within the training batch.
[0011] Furthermore, step S23 includes the following steps: S230: If the predicted score of each comment is greater than the threshold for determining a valid comment, then the comment is a valid comment in the sentiment dimension for that day; S231: If the predicted score of each comment is less than or equal to the threshold for determining a valid comment, then the comment is removed.
[0012] Furthermore, step S24 includes the following steps: S240: Set a historical sliding time observation window; S241: Calculate the average sentiment score of all valid historical comments within the historical sliding time observation window; The formula for calculating the average sentiment score of all historical valid comments in the sentiment dimension is:
[0013] in, The average sentiment score for all historical valid comments in the sentiment dimension; Indicates the time of the day. For the first The emotional score in the emotional dimension of the day. This represents the sum of the emotional scores of all valid comments on that day, representing the emotional dimension. For the first The number of valid comments in the emotional dimension; S242: Determine the adaptive smoothing factor; S243: Calculate daily sentiment score; The formula for calculating daily sentiment score is:
[0014] in, Daily sentiment score, It is an adaptive smoothing factor; S244: Generate sentiment time-series feature vectors based on daily sentiment scores and historical dynamics.
[0015] Furthermore, step S3 includes the following steps: S30: Input the preprocessed structured data into the feature interactive coding layer to map out a low-dimensional contextual latent vector; S31: Input the low-dimensional contextual latent vector into a set of parallel linear transformation layers to generate the unnormalized raw importance score of the sentiment dimension. S32: Substitute the unnormalized raw importance scores of the sentiment dimension into the Softmax function with a temperature parameter to calculate the final normalized weight vector of the sentiment dimension. S33: Perform sparsity regularization of information entropy on the final normalized weight vector of the sentiment dimension; S34: Assign each element of the final normalized weight vector of the sentiment dimension, which is regularized by the sparsity of information entropy, to all the features of the sentiment dimension in the sentiment temporal feature vector, and generate an extended weight matrix with the same dimension as the sentiment temporal feature vector. S35: Multiply the extended weight matrix with the sentiment time-series feature vector to obtain the weighted sentiment feature vector; S36: The weighted sentiment feature vector and the structured feature vector are fused to form a fused feature vector; S37: Input the fused feature vector into the time series prediction model and output a sequence of predicted passenger flow values for the next few days.
[0016] Furthermore, the formula for calculating the low-dimensional contextual latent vector is as follows: ; in, Represented as a low-dimensional contextual latent vector; and For the learnable weights and biases of the feature interactive coding layer, Represents a non-linear activation function; The formula for calculating the unnormalized raw importance score of the sentiment dimension: , in, This represents the unnormalized raw importance score of the sentiment dimension. Represents mutually parallel transformation matrices. The dimension of the contextual latent vector is represented as a low-dimensional vector. The formula for calculating the final normalized weight vector of the sentiment dimension is as follows:
[0017] in, This refers to temperature hyperparameters. This represents the final normalized weight vector for the sentiment dimension; This represents the potential strength of the impact of the i-th emotional dimension on passenger flow in the context of the day. The formula for calculating sparsity regularization of information entropy is as follows:
[0018] in, This represents the regularization weight score.
[0019] Furthermore, step S4 includes the following steps: S40: Post-process and visualize the predicted passenger flow values of the time series, and analyze and visualize the final normalized weight vector of the sentiment dimension with sparsity regularization of information entropy. S41: Generate a decision report based on the final normalized weight vector of the time series passenger flow forecast and the visualized sentiment dimension, which are encapsulated in post-processing and visualization.
[0020] On the other hand, the present invention provides a scenic area tourism demand prediction system based on dynamic weight fusion, including: a multi-source data acquisition and processing module, a multi-dimensional sentiment quantification module, a dynamic feature fusion and prediction module, and a prediction result and decision support output module. The multi-source data acquisition and preprocessing module is used to collect and preprocess structured data from the scenic area and tourist review texts; The multi-dimensional sentiment quantification module is used to receive tourist comment texts collected and preprocessed by the multi-source data acquisition and preprocessing module, and to perform multi-dimensional sentiment quantification on the tourist comment texts; The dynamic feature fusion and prediction module receives structured data collected and preprocessed by the multi-source data acquisition and preprocessing module and tourist review texts quantified by the multi-dimensional sentiment quantification module. It also performs dynamic weight fusion on the structured data and the multi-dimensional sentiment quantified tourist review texts to generate time series passenger flow prediction values and dynamic weight vectors. The prediction results and decision support output module receives the time series passenger flow prediction value and dynamic weight vector output by the dynamic feature fusion and prediction module, and outputs a decision support report based on the time series passenger flow prediction value and dynamic weight vector.
[0021] Advantages of this invention: 1. In this invention, the pre-trained language model predicts the multidimensional sentiment intensity of pre-processed tourist comment texts using a composite loss function composed of a weighted combination of a mean squared error loss term and a Pearson correlation coefficient consistency loss term. The mean squared error loss term minimizes the Euclidean distance between the predicted multidimensional sentiment intensity value and the actual labeled multidimensional sentiment intensity value, ensuring that the pre-trained language model can accurately regress the specific sentiment intensity value. The Pearson correlation coefficient consistency loss term maximizes the linear correlation between the predicted multidimensional sentiment intensity value and the actual labeled multidimensional sentiment intensity value, ensuring that the pre-trained language model captures the relative transformation trend, thereby ensuring that the prediction of the pre-trained language model is both accurate and stable.
[0022] 2. Based on the valid comments of the day, an adaptive dynamic Bayesian smoothing algorithm is used to calculate the adaptive smoothing factor. The daily sentiment score is calculated based on the adaptive smoothing factor, the total number of valid comments of the day, and the sum of sentiment scores. This algorithm dynamically calculates the smoothing factor based on the distribution characteristics of historical data, abandoning fixed hyperparameter settings. Based on the daily sentiment score, derived features such as lag time series, moving average, standard deviation of fluctuation, and difference trend are constructed. By integrating the real-time state and historical dynamics, a complete multi-dimensional sentiment time series feature vector is built for the scenic spot every day, which comprehensively represents the temporal change pattern of emotions. It realizes fine-grained sentiment analysis of multiple sentiment dimensions (scenery, transportation, food, tickets, and services), avoids the loss of a lot of valuable information, and truly considers the subjective feelings and public opinion evaluation of tourists' tourism needs.
[0023] 3. This invention designs a dynamic weight generation network. This network takes traditional structured feature vectors (including external objective conditions such as the weather, holidays, and seasons) as input to generate weight vectors for weighting different emotional dimension features, thereby modulating the contribution of multiple emotional dimensions. Then, the weighted emotional features are fused with the structured feature vectors and processed by a time-series prediction model to finally output the future passenger flow forecast. This dynamic weight fusion mechanism can adjust the dynamic and adaptive allocation of the importance of multiple emotional dimension features under different scenarios, significantly improving the adaptability and prediction accuracy of the time-series prediction model. The time-series prediction model not only outputs the future passenger flow forecast but also outputs a clear explanation of the weight distribution, enabling managers to understand the main driving factors behind the forecast (e.g., whether positive reviews of "transportation" or improvements in "service" have driven demand), thus possessing extremely high practical decision support value. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] Figure 1 This is a flowchart illustrating a method for predicting tourist demand in scenic areas based on dynamic weight fusion, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a scenic area tourism demand prediction system based on dynamic weight fusion provided in an embodiment of the present invention. Detailed Implementation
[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0027] Example 1 like Figure 1 As shown, this embodiment of the invention provides a method for predicting tourism demand in scenic areas based on dynamic weight fusion, including: S1: Collection and preprocessing of structured data and tourist review texts of scenic spots to generate structured feature vectors and high-quality tourist review texts; The structured data for the scenic area includes historical visitor flow data, real-time weather conditions, and calendar time information. Historical visitor flow data is acquired at an hourly frequency and aggregated into daily cumulative visitor flow. Weather data is obtained through access to meteorological data interfaces and includes indicators such as temperature, precipitation, and weather phenomena. Calendar event data integrates information on statutory holidays, weekend markers, seasons, and specific festivals. Tourist review texts include review texts and related metadata such as ratings and posting dates collected from mainstream tourism service platforms. Preprocessing of structured data includes: filling missing structured data with the mean / median; for example, if there is no meteorological data, it is filled with the mean / median of historical meteorological data; if passenger flow data is abnormal on a certain day, the abnormal passenger flow data is removed and filled with the mean / median of historical passenger flow data.
[0028] Sine-cosine encoding is used for structured data with periodic characteristics; for example, Monday to Sunday, January to December: December and January are adjacent, but the numbers are very different; Sunday and Monday are also connected; the function of sine-cosine encoding is to map the cyclical period to circular coordinates, so that December to January and Sunday to Monday are mathematically closest and perfectly preserve the cyclical relationship.
[0029] One-hot encoding is used for structured data with categorical features. For example, weather (sunny / cloudy / rainy) and holidays (weekday / weekend / public holiday) are not numbers and cannot be directly input into the model. One-hot encoding transforms sunny / cloudy / rainy and weekday / weekend / public holiday into a string of 0s and 1s, such as rainy day: 001, sunny day: 100. This allows large language models and pre-trained language models to recognize discrete categories without interference from size or order. Continuous structured data is then standardized.
[0030] Preprocessing of tourist comment texts includes removing irrelevant characters, standardizing encoding formats, converting emojis, and filtering out meaningless noisy comments according to preset rules (such as minimum effective length and proportion of non-Chinese content) to ensure the quality of the text input to the downstream sentiment analysis model and output clean, standardized, and high-value text content.
[0031] S2: Transform the preprocessed tourist review text into a sentiment temporal feature vector; Multi-dimensional sentiment analysis includes the scenery, transportation, food, tickets, and services of the scenic area. A large oracle model is used to assign a continuous sentiment intensity label to the pre-processed tourist review texts. The sentiment intensity labels are then verified and corrected by human experts, resulting in a large-scale, high-quality dataset of sentiment intensity labels. Supervised fine-tuning of a pre-trained language model is performed using a dataset of sentiment intensity labels. This involves receiving a comment text and directly outputting a multi-dimensional sentiment intensity vector, where each dimension's value (e.g., between -1 and 1) represents the positive degree of the corresponding sentiment dimension. The training objective of the pre-trained language model is to minimize the mean squared error between the predicted values and the true labels, and to maximize the correlation coefficient between them. To ensure the accuracy of the model's predictions in absolute values and the consistency in sentiment change trends, the training objective is designed to minimize a composite loss function. This function consists of a mean squared error loss term. Consistency loss term with Pearson correlation coefficient ( The weighted combination of these loss functions results in the following composite loss function: ,in, This represents the mean squared error loss term. This represents the consistency loss term in the Pearson correlation coefficient; , The hyperparameter representing the balance between the contributions of the two loss terms; The mean squared error loss term aims to minimize the Euclidean distance between the predicted sentiment score and the true labeled score, ensuring that the model can accurately regress the specific sentiment intensity value. The calculation formula is ; In the formula, The number of visitor review texts in the training batch; the number of samples in the training batch. Indicates the first The true emotional label of a tourist's review text This represents the predicted value of the pre-trained language model.
[0032] Lpcc is built upon the Pearson correlation coefficient, aiming to maximize the linear correlation between predicted and true sequences. This forces pre-trained language models to capture the relative trends in sentiment intensity rather than focusing solely on absolute values, thus overcoming the central tendency effect common in regression models. Its calculation formula is: ; In the formula, This represents the mean of the true labels for tourist review texts within the training batch. This represents the mean of the predicted values of tourist comment texts within the training batch; By optimizing the multidimensional sentiment intensity vector of tourist review text using a composite loss function, the pre-trained language model can output sentiment feature vectors that are both accurate and have good trend interpretation, thus ensuring that the predictions are both accurate and stable.
[0033] Obtain the collection of comment texts generated on a given day for a specific sentiment dimension, and mark them as... Each comment in the comment text set is input into a pre-trained language model, which outputs a predicted score for the multidimensional sentiment intensity vector of each comment. The predicted score of the multidimensional sentiment intensity vector output by the pre-trained language model ranges from -1 to 1. A predicted score close to 0 indicates that the multidimensional sentiment intensity of the comment tends to be unmentioned / neutral, a predicted score of 0-1 indicates that the multidimensional sentiment intensity of the comment tends to be positive, and a predicted score of -1-0 indicates that the multidimensional sentiment intensity of the comment tends to be negative.
[0034] A valid comment determination threshold of 0.1 is set within the system. Based on this threshold and the predicted score of each comment's multidimensional sentiment intensity vector, the set of valid comments for the day is determined. This set of valid comments is denoted as […]. If the predicted score of each comment is greater than the threshold for determining a valid comment, then the comment is considered a valid comment in terms of sentiment for that day; if the predicted score of each comment is less than or equal to the threshold for determining a valid comment, then the comment is removed.
[0035] Given the valid comment set for the day, the sentiment score for the sentiment dimension is calculated using an adaptive dynamic Bayesian smoothing algorithm. The adaptive dynamic Bayesian smoothing algorithm includes the following steps: Set a historical sliding time observation window ,For example, If the timeframe is 30 days, then the number of valid comments from the past 30 days is selected as the reference set for calculating the number of valid comments for the current day; the average sentiment score of all valid historical comments within the historical sliding time observation window is calculated as the prior mean for the current day, denoted as . This represents the "basic reputation level" of this emotional dimension over 30 days. The formula for calculating the average sentiment score of all historical valid comments in the sentiment dimension is:
[0036] in, The average sentiment score for all historical valid comments in the sentiment dimension; Indicates the time of the day. For the first The emotional score in the emotional dimension of the day. This represents the sum of the emotional scores of all valid comments on that day, representing the emotional dimension. For the first The number of valid comments in the emotional dimension.
[0037] To determine the adaptive smoothing factor and the smoothing strength, the distribution of the number of valid comments per day within the historical sliding time observation window is statistically analyzed, i.e., the set... The 25th percentile of the set is selected and rounded down to become the adaptive smoothing factor (i.e., virtual sample size) for the day.
[0038]
[0039] Select the 25th percentile as The value signifies that the system uses the "number of comments on historically quiet days" as a baseline weight for trusting historical priors. For a dimension where the number of valid comments is consistently high, The value automatically increases, enhancing stability; for dimensions where comments are scarce, The value decreases automatically to ensure that the true feedback of the day is not overly diluted.
[0040] The daily sentiment score is calculated using the following formula:
[0041] in, Daily sentiment score, For adaptive smoothing factor; when At that time, the daily sentiment score is mainly determined by the number of valid comments on that day; when At that time, the daily sentiment score smoothly transitioned to the historical average. This enables robust and adaptive construction of sentiment features.
[0042] Building upon this foundation, to capture the temporal dynamics, inertia, and trends of emotions, we constructed a rich set of derived features for each emotion dimension. These features not only include the raw score for the day but also its lagged values (such as the previous 1 day, 3 days, and 7 days), moving averages for multiple time windows, statistics representing volatility (such as standard deviation), and indicators reflecting changing trends (such as first-order difference). Each scenic area generates a corresponding multi-dimensional temporal feature vector for emotions each day, integrating immediate status and historical evolution information, denoted as... As the core input reflecting the dynamic evolution of tourist word-of-mouth, it generates an emotional time-series feature vector based on the daily emotional score of the emotional dimension.
[0043] S3: Dynamically weight fusion is performed on the preprocessed structured data and the sentiment time series feature vector to generate time series passenger flow prediction values and dynamic weight vectors.
[0044] The impact of emotional characteristics on tourism demand is not static, but dynamically adjusts with changes in external objective conditions (such as whether it is a holiday or the weather). Therefore, a dynamic weight generation network was designed. Preprocessed structured data is used to obtain structured feature vectors, which are then input into a feature interaction coding layer. This layer performs deep fusion of the structured feature vectors, mapping them to low-dimensional contextual latent vectors. The formula for calculating the low-dimensional contextual latent vectors is as follows: ; in, Represented as a low-dimensional contextual latent vector; and For the learnable weights and biases of the feature interactive coding layer, Represents a non-linear activation function; The low-dimensional contextual latent vector is input into a set of parallel linear transformation layers to generate the unnormalized raw importance score of the sentiment dimension; the formula for calculating the unnormalized raw importance score of the sentiment dimension is as follows: , in, This represents the unnormalized raw importance score of the sentiment dimension. Represents mutually parallel transformation matrices. The dimension of the contextual latent vector is represented as a low-dimensional vector. The unnormalized raw importance scores of the sentiment dimension are substituted into the Softmax function with a temperature parameter to calculate the final normalized weight vector of the sentiment dimension. The formula for calculating the final normalized weight vector of the sentiment dimension is as follows:
[0045] in, This refers to temperature hyperparameters. This represents the final normalized weight vector for the sentiment dimension; This represents the potential strength of the impact of the i-th emotional dimension on passenger flow in the context of the day. when As the distribution approaches 0, it tends to be "one-hot," and the model tends to focus only on the one or two most important dimensions (high determinism). As the value approaches infinity, the distribution tends to become uniform, and the model tends to focus on all dimensions on an average basis.
[0046] Through end-to-end gradient descent training, the network parameters ( , The system will automatically adjust, enabling the model to learn to adjust parameters such as "automatically increase the traffic dimension (corresponding to...)" in rainy or snowy weather. Complex nonlinear situational rules such as "weights of )".
[0047] The final normalized weight vector of the sentiment dimension is regularized for sparsity of information entropy. The formula for calculating sparsity regularization of information entropy is as follows:
[0048] in, This represents the regularization weight score. When the weight distribution tends towards a uniform distribution (i.e., ...), ... When the model is extremely hesitant, the entropy value is at its maximum. (As a penalty) it is also the largest; the entropy is the smallest when the weight distribution tends to be peaked.
[0049] By minimizing this regularization term, the model is forced to actively adjust its parameters during backpropagation, learning to output a more "distinctive" weight distribution. Each element in the final normalized weight vector of the sentiment dimension, which regularizes the sparsity of information entropy, is assigned to all the features of the sentiment dimension in the sentiment time-series feature vector, generating an extended weight matrix of the same dimension as the sentiment time-series feature vector. The extended weight matrix is multiplied by the sentiment time-series feature vector to obtain a weighted sentiment feature vector. The weighted sentiment feature vector is then fused with the structured feature vector to form a fused feature vector. The fused feature vector is input into the time-series prediction model to output a sequence of predicted passenger flow values for the next few days.
[0050] This invention preferably employs a time-series prediction model that excels at handling long-sequence dependencies and is highly efficient. The entire dynamic weight generation network and the time-series prediction model constitute a unified, end-to-end deep learning architecture, sharing gradients during training and collaboratively optimizing prediction accuracy and the rationality of weight allocation.
[0051] S4: Post-process and visualize the time series passenger flow forecasts and dynamic weight vectors to output a decision support report.
[0052] The core function of the module that generates decision reports based on the post-processed and visualized passenger flow forecasts and the final normalized weight vector of the sentiment dimension (which uses sparsity regularization of information entropy) is to post-process and visualize this data. For example, it compares and analyzes predicted passenger flow data with historical data from the same period and capacity thresholds, generating passenger flow trend charts and early warning alerts. More importantly, it analyzes and visualizes dynamic weight vectors, such as generating radar charts or time-series graphs showing the changes in the influence weights of various sentiment dimensions during holidays, visually demonstrating how the weights of the "service" and "transportation" dimensions increase as the holiday approaches.
[0053] Ultimately, the time-series forecasting model outputs not a set of isolated predicted figures, but a structured decision support report. This report includes both quantitative forecasts and qualitative attribution analysis, clearly identifying the key emotional dimensions influencing future visitor flow (e.g., "Visitor flow during the National Day holiday is expected to be significantly affected by negative evaluations of the transportation dimension"). This provides scenic spots with targeted data insights for operations, marketing, and service optimization, truly achieving a closed loop from data intelligence to management intelligence.
[0054] Example 2 like Figure 2 As shown, a scenic area tourism demand prediction system based on multi-dimensional emotion and dynamic weight fusion is characterized by comprising: a multi-source data acquisition and processing module 1, a multi-dimensional emotion quantification module 2, a dynamic feature fusion and prediction module 3, and a prediction result and decision support output module 4. Multi-source data acquisition and preprocessing module 1 is used to collect and preprocess structured data of the scenic area and tourist comment text; The multi-dimensional sentiment quantification module 2 is used to receive tourist comment texts collected and preprocessed by the multi-source data acquisition and preprocessing module 1, and to perform multi-dimensional sentiment quantification on the tourist comment texts; The dynamic feature fusion and prediction module 3 is used to receive the structured data collected and preprocessed by the multi-source data acquisition and preprocessing module 1 and the multi-dimensional sentiment quantification module 2 to receive the multi-dimensional sentiment quantification of tourist review texts. It also performs dynamic weight fusion on the structured data and the multi-dimensional sentiment quantification of tourist review texts to generate time series passenger flow prediction values and dynamic weight vectors. The prediction results and decision support output module 4 is used to receive the time series passenger flow prediction value and dynamic weight vector output by the dynamic feature fusion and prediction module 3, and output a decision support report based on the time series passenger flow prediction value and dynamic weight vector.
[0055] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting tourism demand in scenic areas based on dynamic weight fusion, characterized in that, include: S1: Collection and preprocessing of structured data and tourist review texts of scenic spots to generate structured feature vectors and high-quality tourist review texts; S2: Convert the preprocessed tourist comment text into an emotional temporal feature vector; S3: Perform dynamic weight fusion on the preprocessed structured data and emotional time-series feature vector to generate time-series passenger flow prediction values and dynamic weight vectors. S4: Post-process and visualize the predicted passenger flow values of the time series and the dynamic weight vector, and output a decision support report.
2. The method for predicting scenic area tourism demand based on dynamic weight fusion according to claim 1, characterized in that, Step S1 includes the following steps: S10: Obtain structured data and tourist review texts of the scenic area; S11: Fill in missing structured data with mean / median; perform sine-cosine encoding on structured data with periodic features; perform one-hot encoding on structured data with categorical features; standardize continuous structured data to form multi-dimensional structured feature vectors; S12: Remove irrelevant characters from tourist review text, standardize encoding format, convert emojis and filter noisy comments, and output standardized, high-quality tourist review text.
3. The method for predicting scenic area tourism demand based on dynamic weight fusion according to claim 1, characterized in that, Step S2 includes the following steps: S20: Using a large language model and a human expert verification strategy, assign sentiment intensity labels to the preprocessed tourist comment texts to generate a dataset of sentiment intensity labels; S21: Train the pre-trained language model using the dataset of the emotional intensity labels; S22: Obtain a set of comment texts generated on the same day across multiple sentiment dimensions, input each comment in the set of comment texts into a pre-trained language model, and output the predicted score for each comment; S23: Determine the set of valid comments for the day based on the valid comment determination threshold and the predicted score of each comment; S24: Under the set of valid comments for the day, use the adaptive dynamic Bayesian smoothing algorithm to calculate the sentiment score of the sentiment dimension for the day.
4. The method for predicting scenic area tourism demand based on dynamic weight fusion according to claim 3, characterized in that, The pre-trained language model includes a composite loss function, which is: ; in, This represents the mean squared error loss term. This represents the consistency loss term in the Pearson correlation coefficient; , The hyperparameter representing the balance between the contributions of the two loss terms; The calculation formula is: ; In the formula, To train the number of batches of visitor review texts, Indicates the first The true emotional label of a tourist's review text This represents the predicted value of the pre-trained language model; The formula for calculating Lpcc based on the Pearson correlation coefficient is as follows: ； In the formula, This represents the mean of the true labels for tourist review texts within the training batch. This represents the mean of the predicted values of visitor comment texts within the training batch.
5. The method for predicting scenic area tourism demand based on dynamic weight fusion according to claim 3, characterized in that, Step S23 includes the following steps: S230: If the predicted score of each comment is greater than the threshold for determining a valid comment, then the comment is a valid comment in terms of sentiment dimension for that day; S231: If the predicted score of each comment is less than or equal to the valid comment determination threshold, then the comment is removed.
6. The method for predicting scenic area tourism demand based on dynamic weight fusion according to claim 3, characterized in that, Step S24 includes the following steps: S240: Set a historical sliding time observation window; S241: Calculate the average sentiment score of all valid historical comments within the historical sliding time observation window; The formula for calculating the average sentiment score of all historical valid comments in the sentiment dimension is: in, The average sentiment score for all historical valid comments in the sentiment dimension; Indicates the time of the day. For the first Emotional score in the emotional dimension of the day. This represents the sum of the emotional scores of all valid comments on that day, representing the emotional dimension. For the first The number of valid comments in the emotional dimension; S242: Determine the adaptive smoothing factor; S243: Calculate daily sentiment score; The formula for calculating daily sentiment score is: in, The emotional score for the day is the emotional dimension. It is an adaptive smoothing factor; S244: Generate a sentiment time-series feature vector based on the daily sentiment score and historical dynamics.
7. The method for predicting scenic area tourism demand based on dynamic weight fusion according to claim 1, characterized in that, Step S3 includes the following steps: S30: Input the preprocessed structured data into the feature interactive coding layer to map out a low-dimensional contextual latent vector; S31: Input the low-dimensional contextual latent vector into a set of parallel linear transformation layers to generate the unnormalized raw importance score of the sentiment dimension. S32: Substitute the unnormalized original importance scores of the sentiment dimension into the Softmax function with a temperature parameter to calculate the final normalized weight vector of the sentiment dimension. S33: Perform sparsity regularization of the final normalized weight vector of the emotion dimension using information entropy; S34: Assign each element of the final normalized weight vector of the sentiment dimension, which is regularized by the sparsity of information entropy, to all the features of the sentiment dimension in the sentiment temporal feature vector, and generate an extended weight matrix with the same dimension as the sentiment temporal feature vector. S35: Multiply the extended weight matrix with the sentiment time-series feature vector to obtain a weighted sentiment feature vector; S36: The weighted sentiment feature vector is fused with the structured feature vector to form a fused feature vector; S37: Input the fused feature vector into the time series prediction model and output a sequence of predicted passenger flow values for the next few days.
8. The method for predicting scenic area tourism demand based on dynamic weight fusion according to claim 7, characterized in that, The formula for calculating the low-dimensional contextual latent vector is as follows: ； in, Represented as a low-dimensional contextual latent vector; and For the learnable weights and biases of the feature interactive coding layer, Represents a nonlinear activation function; The formula for calculating the unnormalized raw importance score of the sentiment dimension is as follows: ， in, This represents the unnormalized raw importance score of the sentiment dimension. Represents mutually parallel transformation matrices. The dimension of the contextual latent vector is represented as a low-dimensional vector. The formula for calculating the final normalized weight vector of the sentiment dimension is as follows: in, This refers to temperature hyperparameters. This represents the final normalized weight vector for the sentiment dimension; This represents the potential strength of the impact of the i-th emotional dimension on passenger flow in the context of the day. The formula for calculating the sparsity regularization of the information entropy is as follows: in, This represents the regularization weight score.
9. The method for predicting scenic area tourism demand based on dynamic weight fusion according to claim 1, characterized in that, Step S4 includes the following steps: S40: Post-process and visualize the predicted passenger flow values of the time series, and analyze and visualize the final normalized weight vector of the sentiment dimension after the sparsity regularization of the information entropy. S41: Generate a decision report based on the final normalized weight vector of the time series passenger flow forecast and the visualized sentiment dimension, which are encapsulated in post-processing and visualization.
10. A scenic area tourism demand forecasting system based on dynamic weight fusion, characterized in that, include: The module includes a multi-source data acquisition and processing module, a multi-dimensional sentiment quantification module, a dynamic feature fusion and prediction module, and a prediction result and decision support output module. The multi-source data acquisition and preprocessing module is used to collect and preprocess structured data of the scenic area and tourist comment text; The multi-dimensional sentiment quantification module is used to receive tourist comment texts collected and preprocessed by the multi-source data acquisition and preprocessing module, and to perform multi-dimensional sentiment quantification on the tourist comment texts; The dynamic feature fusion and prediction module is used to receive structured data collected and preprocessed by the multi-source data acquisition and preprocessing module and the tourist comment text quantified by the multi-dimensional sentiment quantification module. It performs dynamic weight fusion on the structured data and the multi-dimensional sentiment quantified tourist comment text to generate a time series passenger flow prediction value and a dynamic weight vector. The prediction result and decision support output module is used to receive the time series passenger flow prediction value and dynamic weight vector output by the dynamic feature fusion and prediction module, and output a decision support report based on the time series passenger flow prediction value and the dynamic weight vector.