Business vitality spatio-temporal prediction and dynamic evaluation method and system based on multi-modal feature fusion

Abstract

The present application relates to the technical field of spatial data analysis, and specifically provides a business vitality spatio-temporal prediction and dynamic evaluation method based on multi-modal feature fusion, comprising the following steps: collecting multi-modal heterogeneous data of geographic space, text semantics, real-time behavior and economic attributes, and obtaining a structured data set after cleaning and desensitization; extracting features from each single-modal data, fusing through cross-modal attention, and obtaining a business vitality core feature vector in combination with an emotion-vitality correlation model; inputting the same into a spatio-temporal sequence prediction model, and outputting a future preset time period business vitality prediction result; constructing a quantitative-qualitative dual-dimension evaluation system in combination with real-time updating data, and obtaining a dynamic evaluation result through real-time calculation with a stream processing framework; and displaying and outputting a structured evaluation report through a visual panel. The present application provides timely and reliable decision support for business district operation optimization, business investment decision-making, regional economic monitoring and other scenarios, and fully meets the precision and real-time needs of business vitality analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of spatial data analysis, in particular to a business vitality spatio-temporal prediction and dynamic evaluation method and system based on multi-modal feature fusion. BACKGROUND

[0002] In the current field of business vitality evaluation and prediction, with the acceleration of urbanization and the diversification of consumer behavior, traditional methods have been difficult to meet the needs of precision and real-time. In the prior art, the analysis of business vitality usually relies on single-dimensional data (such as turnover in economic attribute data, passenger flow statistical data), and the trend is predicted by a simple statistical model (such as ARIMA time series model, traditional regression model), or only based on geographic spatial data (such as business circle location, surrounding supporting density) for static evaluation. This kind of method has some shortcomings, such as:

[0003] The existing technology focuses on single modal data (such as only using economic data or geographic data), ignoring the influence of text semantic modal (such as user evaluation, social media mentioned business related emotional feedback) and real-time behavior modal (such as consumer real-time check-in, traffic flow associated with store visit trend) on business vitality. Single modal data cannot fully describe the complex influencing factors of business vitality, resulting in one-sided feature representation and limited accuracy of subsequent prediction and evaluation.

[0004] Consumer sentiment (such as passenger flow growth caused by positive evaluation, vitality decline caused by negative public opinion) is a key dynamic factor affecting business vitality, but the existing technology does not construct a quantitative correlation model between emotional features and business vitality indicators, and cannot convert emotional information in text semantics into feature weights that can be used for prediction, resulting in prediction results that are difficult to reflect emotion-driven business vitality fluctuations.

[0005] Existing spatio-temporal prediction methods mostly process time dimension (such as only predicting single-day passenger flow trend of a business circle) or space dimension (such as only analyzing static geographic correlation between business circles), without constructing a collaborative modeling framework that integrates "time trend-space interaction", which cannot capture dynamic spatial influence between business circles (such as passenger flow transfer between adjacent business circles, holiday cross-business circle consumption linkage), resulting in a large deviation between the predicted business vitality in the future preset period and the actual value.

[0006] Existing business vitality evaluation mostly uses quantitative indicators (such as passenger flow growth rate, turnover deviation rate) for static calculation, lacks descriptive evaluation of qualitative dimensions (such as business vitality level, risk type label), and relies on batch processing (such as daily or weekly summary calculation), which cannot realize real-time dynamic update through stream processing framework, making it difficult to meet the demand of "immediate adjustment strategy" in business circle operation.

[0007] To address these issues, this invention provides a method and system for spatiotemporal prediction and dynamic evaluation of business vitality based on multimodal feature fusion. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, this invention provides a method and system for spatiotemporal prediction and dynamic evaluation of business vitality based on multimodal feature fusion, in order to solve the problems in existing technologies.

[0009] One embodiment of the present invention provides a method for spatiotemporal prediction and dynamic evaluation of business vitality based on multimodal feature fusion, including:

[0010] Collect multimodal heterogeneous data including geospatial modality, text semantic modality, real-time behavioral modality, and economic attribute modality; perform data cleaning and desensitization processing on the multimodal heterogeneous data to obtain a structured multimodal dataset.

[0011] Based on the corresponding feature extraction algorithm, feature extraction is performed on each single modality data in the structured multimodal dataset to obtain each single modality feature vector. The feature vectors of each single modality are fused through a cross-modal attention mechanism, and the core feature vector of business vitality is calculated based on the pre-built emotion-vitality association model.

[0012] The core feature vector of business vitality is input into a pre-constructed spatiotemporal sequence prediction model, and the spatiotemporal sequence prediction model outputs the prediction result of business vitality in a future preset time period.

[0013] Obtain real-time updated data and business vitality prediction results, construct a quantitative and qualitative dual-dimensional evaluation index system based on the real-time updated data and business vitality prediction results, and perform real-time calculation on the evaluation index system through a stream processing framework to obtain dynamic evaluation results;

[0014] The dynamic evaluation results are displayed through a visual interactive panel, and a structured evaluation report containing the business vitality prediction results and dynamic evaluation results is output.

[0015] This application also relates to a spatiotemporal prediction and dynamic evaluation system for business vitality based on multimodal feature fusion, including:

[0016] The data processing module is used to collect multimodal heterogeneous data including geospatial modality, text semantic modality, real-time behavioral modality and economic attribute modality, and to perform data cleaning and desensitization processing on the multimodal heterogeneous data to obtain a structured multimodal dataset.

[0017] The association calculation module is used to extract features from each single modality data in the structured multimodal dataset based on the corresponding feature extraction algorithm, obtain each single modality feature vector, perform feature fusion on each single modality feature vector through a cross-modal attention mechanism, and calculate the core feature vector of business vitality based on the pre-built emotion-vitality association model.

[0018] The result prediction module is used to input the core feature vector of business vitality into a pre-constructed spatiotemporal sequence prediction model, and output the prediction result of business vitality within a future preset time period through the spatiotemporal sequence prediction model.

[0019] The dynamic evaluation module is used to acquire real-time updated data and business vitality prediction results. Based on the real-time updated data and business vitality prediction results, a quantitative and qualitative dual-dimensional evaluation index system is constructed. The evaluation index system is calculated in real time through a stream processing framework to obtain dynamic evaluation results.

[0020] The results output module is used to display the dynamic evaluation results through a visual interactive panel and output a structured evaluation report containing the business vitality prediction results and the dynamic evaluation results.

[0021] This application also relates to a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for spatiotemporal prediction and dynamic evaluation of business vitality based on multimodal feature fusion.

[0022] This application also relates to a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for spatiotemporal prediction and dynamic evaluation of business vitality based on multimodal feature fusion.

[0023] The above embodiments provide a method and system for spatiotemporal prediction and dynamic evaluation of business vitality based on multimodal feature fusion, which has the following beneficial effects:

[0024] This invention overcomes the limitations of single-modal data in traditional business vitality analysis. By collecting and structuring multi-dimensional data including geospatial, textual semantics, real-time behavior, and economic attributes, it achieves deep feature fusion through a cross-modal attention mechanism. Simultaneously, it introduces a sentiment-vitality association model to quantify the influence weight of sentiment on business vitality, effectively improving the comprehensiveness and accuracy of the core feature vectors representing business vitality. Furthermore, it utilizes a spatiotemporal sequence prediction model to achieve collaborative modeling of "time trend-spatial interaction," significantly improving the accuracy of future business vitality predictions for preset time periods and avoiding biases caused by traditional spatiotemporal fragmented modeling. Subsequently, it constructs a dual-dimensional evaluation index system of quantitative and qualitative metrics and uses a stream processing framework to achieve real-time calculation and dynamic updating of evaluation results, compensating for the shortcomings of traditional single quantitative evaluation and batch processing modes. Finally, it presents the results intuitively through a visual interactive panel and a structured evaluation report, providing timely and reliable decision support for scenarios such as business district operation optimization, business investment decisions, and regional economic monitoring, fully meeting the needs for precision and real-time analysis of business vitality. Attached Figure Description

[0025] Figure 1 A flowchart of a method for spatiotemporal prediction and dynamic evaluation of business vitality based on multimodal feature fusion provided in an embodiment of the present invention;

[0026] Figure 2 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions in the embodiments of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.

[0028] Reference Figure 1 One embodiment of the present invention provides a method for spatiotemporal prediction and dynamic evaluation of business vitality based on multimodal feature fusion, comprising the following steps:

[0029] S10. Collect multimodal heterogeneous data including geospatial modality, text semantic modality, real-time behavioral modality, and economic attribute modality. Perform data cleaning and desensitization processing on the multimodal heterogeneous data to obtain a structured multimodal dataset.

[0030] In this embodiment, the core of step S10 is to complete the basic data preparation for business vitality analysis. Specifically, it involves collecting multimodal heterogeneous data covering four modalities. Among them, the geospatial modality refers to spatial data such as the geographical location of the business district, the distribution of surrounding residential / transportation stations, and the location of competing commercial properties; the textual semantic modality refers to semantically meaningful text data such as user reviews, content mentioning the business in social media, and public opinion information; the real-time behavioral modality refers to dynamic behavioral data such as real-time check-in records of consumers, traffic flow associated with the commercial property, and real-time number of people visiting the store; and the economic attribute modality refers to economically relevant data such as the turnover of merchants in the business district, customer flow statistics, and rent levels. Next, data cleaning and desensitization are performed on the collected multimodal heterogeneous data. Data cleaning removes missing, outlier, and duplicate values ​​to ensure data quality. Desensitization ensures compliance in data use by removing identifiers (such as deleting user phone numbers, ID card numbers, and other private information). After cleaning and desensitization, the multimodal data needs to be structured and integrated. Specifically, this is achieved by unifying data formats (e.g., standardizing the coordinate format of geospatial data, the encoding format of text semantic data, and the timestamp format of real-time behavioral data to standardized formats such as CSV or JSON), standardizing field naming (e.g., adding core fields such as "unique business district ID," "data collection timestamp," and "modal type label" to all modal data), and establishing association indexes (e.g., associating multimodal data in the same time and space through "business district ID + timestamp"). The final result is a structured multimodal dataset with a unified structure, standardized fields, and direct usability for subsequent feature extraction, providing a reliable data foundation for the feature extraction steps of each single modality.

[0031] It should be noted that the data cleaning and desensitization processes involved in this step employ conventional data processing methods well-known to those skilled in the art. Data cleaning can be achieved through conventional methods such as missing value imputation, outlier removal, and duplicate data deduplication. Desensitization can be completed through conventional methods such as de-identification of user privacy information (such as mobile phone numbers and ID card numbers) and encryption of sensitive data. The specific operations are existing mature technologies and will not be elaborated here. Meanwhile, the acquisition of the geospatial modality, text semantic modality, real-time behavioral modality, and economic attribute modality data can be achieved by those skilled in the art through conventional channels such as corresponding public or authorized platforms and official statistical websites. For example, geospatial data can be obtained from professional map service platforms, economic attribute data can be obtained from regional commercial statistics websites, text semantic data can be obtained from public social media platforms or merchant review platforms, and real-time behavioral data can be obtained from commercial entity operation systems or traffic data service platforms. The specific acquisition paths are clear and obvious to those skilled in the art and do not require additional limitations.

[0032] S20. Based on the corresponding feature extraction algorithm, feature extraction is performed on each single modality data in the structured multimodal dataset to obtain each single modality feature vector. The feature vectors of each single modality are fused through a cross-modal attention mechanism, and the core feature vector of business vitality is calculated based on the pre-built emotion-vitality association model.

[0033] In this embodiment, the core of step S20 is to transform the structured multimodal data into core feature vectors that can be used to predict business activity. The specific process is as follows: First, based on the corresponding feature extraction algorithms adapted to each single modality, feature extraction is performed on the geospatial modality, text semantic modality, real-time behavioral modality, and economic attribute modality data in the structured multimodal dataset. Specifically, geospatial modality data can have spatial correlation features extracted using geographic information encoding algorithms (such as converting latitude and longitude coordinates into spatial feature vectors); text semantic modality data can have semantic and emotion-related features extracted using natural language processing algorithms (such as the BERT model); real-time behavioral modality data can have behavioral frequency and trend features extracted using time-series feature extraction algorithms (such as sliding window statistics); and economic attribute modality data... The dynamic data can be used to extract vitality scale and fluctuation characteristics through statistical feature extraction algorithms (such as mean and variance calculation), and finally obtain the single-modality feature vector corresponding to each modality. Then, a cross-modality attention mechanism is used to fuse all single-modality feature vectors. This mechanism can dynamically allocate weights according to the actual impact of different modalities on business vitality (such as the real-time behavior modality having a higher weight than other modalities during peak consumption seasons), and output a preliminary fused feature vector. Finally, the preliminary fused feature vector is input into a pre-constructed sentiment-vitality correlation model. This model combines the quantitative correlation between sentiment features and business vitality indicators to optimize and adjust the feature weights of each modality in the preliminary fused feature vector, and finally calculates the core feature vector of business vitality that can accurately represent the key influencing factors of business vitality.

[0034] This step addresses the problems of isolated traditional multimodal features and unquantified impact of emotion on vitality by employing targeted unimodal feature extraction, dynamic cross-modal fusion, and emotion association optimization. It provides high-quality input data for subsequent spatiotemporal sequence prediction models while ensuring that the feature vectors can fully reflect the complex influencing mechanisms of business vitality.

[0035] S30. Input the core feature vector of business vitality into the pre-constructed spatiotemporal sequence prediction model, and output the prediction result of business vitality in the future preset time period through the spatiotemporal sequence prediction model.

[0036] In this embodiment, the core of step S30 is to use the high-quality feature vector obtained in the previous step to predict the future of business vitality. The specific process is as follows: First, the core feature vector of business vitality output in step S20 is input into a pre-constructed spatiotemporal sequence prediction model. This model is specifically designed to integrate spatiotemporal dimension information to process the business vitality prediction task. It can simultaneously capture the temporal change pattern of business vitality (such as daily peak customer flow trend and weekly vitality fluctuation cycle) and spatial correlation characteristics (such as customer flow transfer between business districts and regional consumption linkage) contained in the feature vector, rather than processing the time or spatial dimensions in isolation. Then, the spatiotemporal sequence prediction model is used to process the input core feature vector. The model will extrapolate the business vitality within a preset future time period based on the spatiotemporal correlation logic formed by internal training. The "preset future time period" can be set according to the actual application scenario (such as 1 day, 3 days, 7 days, etc.). Finally, the model outputs the business vitality prediction result within the preset future time period. This result is usually presented in the form of quantitative indicators (such as predicted customer flow, predicted turnover, predicted consumption frequency, etc. in each time period), which directly reflects the changing trend of future business vitality.

[0037] This step, through a specialized spatiotemporal sequence prediction model, solves the problems of fragmented spatiotemporal modeling and difficulty in capturing dynamic correlations in traditional prediction methods. It effectively utilizes comprehensive information in the core feature vector of business vitality to improve the accuracy of future business vitality prediction, while providing key prediction data support for the dynamic evaluation in the subsequent step S40.

[0038] S40. Obtain real-time updated data and business vitality prediction results. Based on the real-time updated data and business vitality prediction results, construct a quantitative and qualitative dual-dimensional evaluation index system. Calculate the evaluation index system in real time using a stream processing framework to obtain dynamic evaluation results.

[0039] In this embodiment, step S40 focuses on dynamically assessing business vitality based on prediction results and real-time data. The specific process involves: first, acquiring two key data sets: real-time updated data (including real-time multimodal feature data and corresponding time-period actual business vitality indicators, reflecting the current true state of business vitality); and second, the business vitality prediction results output in step S30 (used for comparative analysis with real-time data). Then, based on these two sets of data, a quantitative and qualitative dual-dimensional assessment indicator system is constructed. The quantitative indicators are directly calculable numerical indicators (such as the degree of deviation between predicted and actual values, the rate of vitality growth, etc.), while the qualitative indicators are descriptive assessment conclusions (such as business vitality level, risk type labels, etc.). Next, this dual-dimensional assessment indicator system is input into a stream processing framework, which performs real-time calculations and correlation analysis on the indicators (such as dynamically updating quantitative indicator values ​​and matching qualitative labels). Finally, a dynamic assessment result is obtained, which is updated synchronously with real-time data updates, reflecting the real-time changes in business vitality and the accuracy of the prediction.

[0040] This step, by combining a dual-dimensional indicator system with a stream processing framework, solves the problem of traditional assessments relying solely on quantitative indicators and lacking real-time performance. It enables a comprehensive and dynamic assessment of business vitality, providing timely information for decisions such as adjustments to business district operations.

[0041] S50. Display the dynamic evaluation results through a visual interactive panel and output a structured evaluation report containing the business vitality prediction results and the dynamic evaluation results.

[0042] In this embodiment, the core of step S50 is to output the prediction and evaluation results of business vitality in an intuitive and standardized form for easy practical application. The specific process is as follows: First, the dynamic evaluation results obtained in step S40 are displayed through a visual interactive panel. This panel can use a chart-based display method (such as using a line graph to present the vitality trend, a heat map to show the spatial distribution, and a dashboard to display the real-time evaluation level) and supports interactive operations (such as filtering specific time periods and business districts, and clicking to view indicator details), making the dynamic evaluation results (including quantitative indicator values ​​and qualitative label changes) intuitive and easy to understand. At the same time, a structured evaluation report containing the business vitality prediction results of step S30 and the dynamic evaluation results of step S40 is output. This report needs to integrate information according to a preset framework (such as a summary of prediction results, real-time evaluation indicator analysis, trend judgment, risk warning, and optimization suggestions), and adopt a standardized format (such as PDF or structured tables) to ensure that the content is clear and the data source is traceable.

[0043] This step, through the combination of visualization and structured reporting, solves the problems of traditional results presentation formats being monotonous and information transmission being inefficient. It enables different users, such as technical personnel and operations managers, to quickly obtain key information and effectively support the implementation of business decisions.

[0044] In one embodiment, step S20, the construction steps of the pre-built emotion-vitality association model include:

[0045] S201. Obtain multimodal historical data including geospatial modality, text semantic modality, real-time behavioral modality and economic attribute modality, perform data cleaning and desensitization processing on the multimodal historical data to obtain a structured historical multimodal dataset;

[0046] S202. Extract text semantic modality historical data and economic attribute modality historical data within the corresponding time period from the structured historical multimodal dataset; extract sentiment features from the text semantic modality historical data to obtain a historical sentiment feature vector; perform standardization processing on the economic attribute modality historical data to obtain a historical vitality index vector.

[0047] S203. Using the historical emotion feature vector as input and the historical vitality index vector as output, a regression model with an attention fusion mechanism is used for training to obtain a trained emotion-vitality association model.

[0048] S204. Based on the trained emotion-vitality association model, output the correlation coefficient matrix between emotion features and business vitality indicators. The correlation coefficient matrix is ​​used to quantitatively characterize the influence weight of different emotion features on business vitality indicators.

[0049] In this embodiment, the core of step S201 is to prepare historical training data for the construction of the emotion-vitality correlation model. The specific process is as follows: First, acquire multimodal historical data (different from the real-time data in S10) including geospatial modality, textual semantic modality, real-time behavioral modality, and economic attribute modality. Specifically, geospatial modal historical data refers to spatial data such as historical business district locations, past distribution of surrounding facilities, and historical site selections of competing commercial entities; textual semantic modal historical data refers to textual data such as past user reviews, historical social media business-related content, and past public opinion records; real-time behavioral modal historical data refers to behavioral data such as historical consumer check-in records, past traffic flow correlation data, and past store visit statistics; and economic attribute data... Modal historical data refers to economic data such as historical turnover, past customer flow statistics, and historical rental records. This multimodal historical data then undergoes data cleaning and anonymization. This part of the operation (such as missing value removal, outlier filtering, and de-identification of user privacy information) is completely consistent with the corresponding operation in step S10, and will not be elaborated here. After cleaning and anonymization, the data is structured and integrated (e.g., summarizing historical data from Q1 2023 by natural week to form 13 weekly data units, unifying the time granularity of historical data to "week," and standardizing field formats while establishing a correlation index through "historical date + business district ID"). This integration logic is also consistent with step S10, ultimately resulting in a structured historical multimodal dataset. This step, by processing historical multimodal data, provides a reliable foundation and suitable data for model training, ensuring the integrity and standardization of model training data, for subsequently extracting "historical sentiment features" and "historical vitality indicators" from historical data.

[0050] In this embodiment, the core of step S202 is to extract the core input and output data required for training the emotion-vitality association model from the structured historical multimodal data. The specific process is as follows: First, from the structured historical multimodal dataset obtained in step S201, two types of key data are selected: "textual semantic modality historical data" (the sole source of emotion features) and "economic attribute modality historical data" (the source of vitality indicators) that are in the same time interval as the text data, ensuring that the two correspond one-to-one in the spatiotemporal dimension (e.g., user review text of a certain week matches the sales volume and customer flow data of the same business district in the same period); then, emotion features are extracted from the textual semantic modality historical data, and natural language processing algorithms (such as TextCNN, BERT model) are used to perform semantic parsing on historical user reviews, social media content, and other texts to identify the emotional tendencies (such as positive and negative) in the text. The system analyzes historical economic data, including its positive, negative, and neutral aspects, and its intensity (e.g., "strongly recommended" corresponds to a high positive sentiment value). This information is then transformed into numerical historical emotional feature vectors (e.g., [0.8, 0.1, 0.1] represents high positive, low negative, and low neutral sentiment). Simultaneously, the system standardizes historical economic attribute modal data from the same period, using Min-Max normalization and Z-score standardization to eliminate dimensional differences between different vitality indicators (e.g., the difference between "ten thousand yuan" of revenue and "number of visitors"). Historical daily average revenue, daily average visitor flow, and average transaction value are transformed into historical vitality indicator vectors with a unified numerical range. Finally, paired "historical emotional feature vector - historical vitality indicator vector" are obtained, providing clear input (emotional features) and supervised output (vitality indicators) for model training in subsequent step S203, ensuring the model can learn the quantitative correlation between the two.

[0051] In this embodiment, the core of step S203 is to establish a quantitative correlation between historical emotion features and historical vitality indicators through specific model training, thereby obtaining an emotion-vitality correlation model that can be used for subsequent weight optimization. The specific process is as follows: First, determine the input and output data for model training. Use the historical emotion feature vector obtained in step S202 as the model input (representing key information of the emotion dimension), and use the historical vitality indicator vector of the same period as the model output (representing the quantitative result of business vitality). Divide the paired "input-output" vectors into a training set (for model learning) and a validation set (for evaluating training effectiveness) according to a preset ratio. For example, divide 1000 sets of "historical emotion feature vectors - historical vitality indicator vectors" in an 8:2 ratio, with 800 sets used for model parameter learning and 200 sets used to validate the model's prediction accuracy. Then, use a regression model with an attention fusion mechanism for training. The regression model is responsible for learning the continuous value mapping relationship between emotional features and vitality indicators (adapting to the prediction needs of continuous vitality indicators such as turnover and customer flow). The attention mechanism is used to dynamically allocate the weight of each emotional feature (for example, the keyword "poor service" in negative emotions has a greater impact on vitality, so the mechanism will assign it a higher weight to highlight the role of key emotional factors). During training, the model parameters are continuously adjusted through the gradient descent algorithm to minimize the error (such as mean squared error) between the predicted vitality indicator and the actual historical vitality indicator. When the error of the validation set stabilizes within the preset threshold or no longer decreases, training stops. Finally, the trained emotion-vitality association model is obtained. This model has the ability to output the corresponding predicted value of the vitality indicator based on the input emotion feature vector. At the same time, the association weight logic between emotion features and vitality indicators has been stored internally, providing the model basis for the subsequent step S204 to output the association coefficient matrix.

[0052] In this embodiment, the core of step S204 is to extract quantitative correlation information from the trained emotion-vitality correlation model to obtain a correlation coefficient matrix that can be directly used for subsequent feature weight optimization. The specific process is as follows: First, based on the trained emotion-vitality correlation model obtained in step S203, this model has learned the influence logic of different emotional features on the vitality index through the fusion attention mechanism during the training process, and internally stores the corresponding training parameters (such as attention weight values ​​and regression model coefficients); then, these key parameters are extracted from the model and matrix-integrated—the "different emotional features" (such as "comfortable environment" and "high cost performance" in positive emotions, and negative emotions) are matrix-integrated. The model uses "inefficient service" and "parking difficulties" as the row dimensions of a matrix, and "different business vitality indicators" (such as average daily revenue, average daily customer flow, and merchant survival rate) as the column dimensions. The influence of each emotional feature on each vitality indicator in the model parameters is converted into matrix element values ​​(positive values ​​represent positive influence, negative values ​​represent negative influence, and the absolute value represents the intensity of influence). The final output is a correlation coefficient matrix between emotional features and business vitality indicators. This matrix quantitatively represents the specific influence weight of each emotional feature on each vitality indicator (e.g., the weight of the "comfortable environment" emotional feature corresponding to the "average daily customer flow" indicator is 0.6, indicating a strong positive impact on customer flow). This step transforms the internal correlation logic of the model into an intuitive matrix form, providing a clear quantitative basis for optimizing the emotion-related weights of the core feature vector of business vitality in step S20, ensuring that the influence of emotion on business vitality can be precisely controlled.

[0053] In one embodiment, step S20 specifically includes the following steps:

[0054] S211. Based on the feature extraction algorithm matching each modal data, feature extraction is performed on the geospatial modal data, text semantic modal data, real-time behavioral modal data and economic attribute modal data in the structured multimodal dataset, respectively, to obtain geospatial feature vector, emotion feature vector, real-time behavioral feature vector and economic attribute feature vector;

[0055] S212. Employ a cross-modal attention mechanism to perform multimodal dynamic weighted fusion of the geospatial feature vector, emotion feature vector, real-time behavior feature vector, and economic attribute feature vector to obtain an initial fused feature vector.

[0056] S213. Input the initial fusion feature vector and the correlation coefficient matrix into the emotion-vitality correlation model. The emotion-vitality correlation model adjusts the fusion weights of the emotion feature vector, geospatial feature vector, real-time behavior feature vector and economic attribute feature vector in the initial fusion feature vector based on the correlation coefficient matrix.

[0057] S214. Output the initial fusion feature vector after weight adjustment to obtain the core feature vector of business vitality.

[0058] In this embodiment, the core of step S211 is to perform targeted feature extraction on the four types of modal data in the structured multimodal dataset, providing standardized single-modal base vectors for subsequent multimodal fusion. The specific process is as follows: First, the processing objects are defined—geospatial modal data, text semantic modal data, real-time behavioral modal data, and economic attribute modal data in the structured multimodal dataset obtained in step S10. For each type of modal data, a feature extraction algorithm matching its data characteristics is adopted to ensure that the extracted features can accurately represent the core information of each modality. Among them, for geospatial modal data (such as the latitude and longitude of business districts, and the distribution of surrounding transportation stations / residential buildings), geographic information encoding algorithms (such as GeoHash encoding and coordinate normalization processing) are used to extract spatial correlation features, and the spatial location information is converted into numerical geospatial feature vectors (such as the vector dimension can include quantitative indicators such as "distance from business district to subway station" and "number of residential buildings within 3 kilometers"). For text semantic modal data (such as real-time user reviews and social media business-related content), natural language processing algorithms (such as BERT pre-trained model and TextCNN convolutional neural network) are used to extract semantic and... Emotional information is processed by converting text content into emotional feature vectors containing emotional tendencies and key semantics (e.g., vector elements can correspond to "positive emotion intensity," "negative keyword frequency," etc.). For real-time behavioral modal data (e.g., real-time consumer check-in records, associated traffic flow), time-series feature extraction algorithms (e.g., sliding window statistics, short-time Fourier transform) are used to extract dynamic behavioral features, converting real-time behavioral data into real-time behavioral feature vectors containing indicators such as "hourly check-in frequency" and "traffic change rate." For economic attribute modal data (e.g., real-time sales revenue, instantaneous customer flow statistics), a... Statistical feature extraction algorithms (such as real-time average calculation and month-on-month growth rate calculation) extract economic vitality features, transforming economic data into economic attribute feature vectors containing indicators such as "real-time daily average turnover" and "peak passenger flow". Finally, four types of structured single-modal feature vectors are obtained. This step ensures that each type of feature vector can completely retain the core information of the corresponding modality through precise "modality-algorithm matching", avoiding feature loss or redundancy, laying the data foundation for the cross-modal dynamic fusion in the subsequent step S212, and solving the problem of one-sided feature representation caused by traditional single algorithms processing multimodal data.

[0059] In this embodiment, the core of step S212 is to integrate four types of single-modal feature vectors through dynamic weighting to obtain an initial fused feature vector that can initially reflect the synergistic effect of multimodal interactions. The specific process is as follows: First, determine the fusion objects—the geospatial feature vector, emotion feature vector, real-time behavior feature vector, and economic attribute feature vector obtained in step S211. These four types of vectors represent the impact of different dimensions on business vitality, and synergistic association needs to be achieved through a cross-modal attention mechanism. The core of the cross-modal attention mechanism is to dynamically allocate the weights of each modality based on the current scene. The mechanism will first calculate the correlation between each type of feature vector and the business vitality target (i.e., attention). The higher the correlation (e.g., the attention score of the real-time behavior feature vector is higher on weekends than on weekdays, and the corresponding weight is higher), and the total weight is normalized to 1. Then, the four types of feature vectors are weighted and summed according to this dynamic weight. For example, the weight of the geospatial feature vector is 0.2, the weight of the emotion feature vector is 0.3, the weight of the real-time behavior feature vector is 0.3, and the weight of the economic attribute feature vector is 0.2. After multiplying each vector element by its corresponding weight and summing them, an initial fused feature vector with unified dimensions is obtained. This vector initially integrates multimodal information, but does not include the quantitative correlation logic between emotion and vitality, which needs to be further optimized in subsequent steps. This step solves the problem that traditional fixed-weight fusion cannot adapt to scene changes through dynamic weight allocation, making the initial fusion result more in line with the actual influence intensity of each modality in different times and spaces, and providing a basis for the fine-tuning of weights in step S213.

[0060] In this embodiment, the core of step S213 is to refine the modal weights of the initial fusion feature vector based on the quantitative correlation between emotion and vitality, so that the fusion result is more in line with the actual impact mechanism of business vitality. The specific process is as follows: First, clarify the input data—the initial fusion feature vector obtained in step S212 (including the preliminary weighted fusion result of geospatial, emotional, real-time behavior, and economic attribute features), and the correlation coefficient matrix output in step S204 (quantitatively representing the influence weight of different emotional features on business vitality indicators); input these two types of data into the emotion-vitality correlation model trained in step S203. The model will first analyze the quantitative correlation logic between emotional features and vitality indicators in the correlation coefficient matrix (such as "positive emotional features on customer flow indicators"). (The influence weight is set to 0.5). Based on this logic, the fusion weights of the four single-modal feature vectors in the initial fusion feature vector are adjusted: on the one hand, the components in the emotion feature vector that are strongly correlated with vitality are strengthened (such as increasing the proportion of high-weight positive emotion features); on the other hand, the weights of geospatial, real-time behavior, and economic attribute feature vectors are adjusted in conjunction (such as appropriately reducing the weights of features in other modalities that are weakly correlated with emotion when the emotion feature has a significant impact on vitality, to avoid interference). During the adjustment process, the model ensures that the total weight remains normalized, and the change in the weight of each modality is positively correlated with the influence strength in the correlation coefficient matrix (such as when the influence weight of an emotion feature on the vitality index is 0.6, the increase in the weight of its corresponding modality is higher than when the influence weight is 0.3). This step, by introducing the quantitative correlation logic of emotion and vitality, solves the problem that the initial fusion only relies on the dynamics of the scene and does not fully consider the intrinsic correlation between emotion and vitality, so that the weights of each modality more accurately reflect the actual contribution to business vitality, and provides a guarantee for the output of high-quality core feature vectors in step S214.

[0061] In this embodiment, the core of step S214 is to output the fused feature vector after refined weight adjustment, obtaining a core feature vector of business vitality that can be directly used for subsequent prediction. The specific process is as follows: The initial fused feature vector after weight adjustment by the sentiment-vitality correlation model in step S213 is received. This vector integrates the core information of geospatial feature vectors, sentiment feature vectors, real-time behavioral feature vectors, and economic attribute feature vectors. The weights of each modality feature are adapted to dynamic scenario requirements (such as time periods and regional differences) and conform to the quantitative correlation logic between sentiment and business vitality (such as higher weight sentiment features corresponding to a higher proportion). The adjusted vector is then format-verified (ensuring consistent dimensions and compliant numerical ranges). After confirmation, it is directly output, thus obtaining the core feature vector of business vitality. This core feature vector comprehensively and accurately represents the multi-dimensional key factors affecting business vitality, avoiding the one-sidedness of single-modality features and highlighting the core influencing components through weight optimization. This provides high-quality input for the efficient computation of the spatiotemporal sequence prediction model in step S30, ensuring that subsequent prediction results truly reflect the dynamic changes in business vitality.

[0062] In one embodiment, step S30, the construction step of the pre-built spatiotemporal sequence prediction model includes:

[0063] S301. Obtain historical core feature data of business activity and actual business activity indicators for the corresponding time period. The historical core feature data of business activity is obtained based on a structured historical multimodal dataset through single-modal feature extraction, cross-modal attention fusion, and emotion-activity correlation model calculation.

[0064] S302. Decompose the historical commercial vitality core feature data into spatiotemporal dimension features to obtain time series features and spatial correlation features. The time series features are used to characterize the temporal change pattern of commercial vitality, and the spatial correlation features are used to characterize the spatial interaction relationship between different business districts.

[0065] S303. Construct a spatiotemporal prediction network that integrates emotion and vitality correlation. The spatiotemporal prediction network includes a temporal modeling unit, a spatial correlation modeling unit, and a correlation fusion layer. The temporal modeling unit is used to process time series features to capture the time dependence of business vitality. The spatial correlation modeling unit is used to process spatial correlation features to capture the spatial correlation between business districts. The correlation fusion layer introduces a correlation coefficient matrix between emotion features and business vitality indicators to dynamically adjust the weight distribution of emotion features in spatiotemporal modeling.

[0066] S304. Using the time series features and spatial correlation features as inputs, and the actual business vitality index of the corresponding time period as a reference benchmark, train the spatiotemporal prediction network that integrates sentiment-vitality correlation to obtain a trained spatiotemporal sequence prediction model. The spatiotemporal sequence prediction model is used to output the business vitality prediction results within a future preset time period.

[0067] In this embodiment, the core of step S301 is to prepare the "input-supervised output" data pairs required for training the spatiotemporal sequence prediction model. The specific process is as follows: First, it is clear that two types of key historical data need to be obtained—historical business vitality core feature data and the actual business vitality index of the corresponding time period. Among them, the generation process of historical business vitality core feature data is consistent with the multimodal processing logic above. It needs to be based on the structured historical multimodal dataset obtained in step S201, and then perform the single-modal feature extraction consistent with S211 (such as extracting historical geospatial features using GeoHash encoding and extracting historical sentiment features using the BERT model), the cross-modal attention fusion consistent with S212 (dynamically allocating the weights of each modality according to the historical scene), and the calculation and optimization of the sentiment-vitality association model trained in S203 (adjusting the weights in combination with the association coefficient matrix in S204) to finally obtain the structured historical business vitality data that matches the historical time period. The process begins with obtaining core business activity data (e.g., the business district integration feature vector for each week in Q2 2023); next, actual business activity indicators for the corresponding time period are acquired. These indicators are quantitative data that directly reflect the historical level of business activity, such as the historical average daily customer flow, actual turnover, merchant sales per square meter, and consumption frequency for each week. It is also necessary to ensure that the indicator time period is completely aligned with the historical core business activity data (e.g., the core feature data for the first week of June 2023 is matched with the actual average daily customer flow indicator for that week). Finally, paired training data of "historical core business activity data - corresponding actual business activity indicators" is formed. This step generates input data by reusing the mature multimodal processing flow described above, and uses real historical indicators as a supervision benchmark, which solves the problem of "insufficient input data quality and missing supervision signals" in the training of spatiotemporal sequence prediction models. This provides standardized and highly reliable basic data for the feature decomposition in step S302 and the model training in step S304.

[0068] In this embodiment, the core of step S302 is to split the historical business vitality core feature data according to the "time-space" dual dimension, providing targeted input features for the subsequent multidimensional accurate modeling of the spatiotemporal prediction network. The specific process is as follows: First, the processing object is identified—the historical business vitality core feature data obtained in step S301 (historical feature vectors that have been integrated with multimodal information and sentiment association optimization). The decomposition logic is based on the essential characteristic of business vitality being "affected by both dynamic changes over time and spatial interaction," and features that can independently represent the two types of patterns are extracted respectively. Among them, the historical business vitality core feature data is decomposed in the time dimension to obtain time series features: First, the historical core feature data is arranged in order according to a preset time granularity (such as day, week) (such as the core features of January to March 2023 are sorted by natural day). Then, time series feature engineering methods (such as sliding window statistics, STL trend decomposition) are used to extract the time change patterns, such as capturing "the vitality fluctuation during the peak consumption period from 10:00 to 22:00 during the day", "the difference in vitality between weekdays and weekends", "the decline trend of vitality peak before and after holidays", etc., and these patterns are transformed into numerical time series features (such as the vector dimension can contain " Quantitative indicators such as "average activity over the past 7 days," "month-on-month growth rate of activity compared to the same period yesterday," and "monthly variance of activity fluctuation" are used to ensure that this feature accurately reflects the dynamic evolution of commercial activity over time. Simultaneously, the core characteristics of historical commercial activity are spatially decomposed to obtain spatial correlation features: first, historical core characteristic data of different business districts within the same time period are correlated (e.g., the core characteristics of business district A and business districts B and C within a 3-kilometer radius in the first week of February 2023); then, spatial analysis methods (such as Pearson correlation calculation and spatial weight matrix construction) are used to extract the inter-business district correlations. The dynamic relationships, such as identifying the "diversion effect of increased customer traffic in business district A on customer traffic in business district B" and the "vitality radiation intensity of a regional commercial center business district on secondary business districts," are transformed into numerical spatial correlation features (e.g., the vector dimension can include quantitative indicators such as "the correlation coefficient of vitality between the target business district and three adjacent business districts" and "the influence weight of the regional core business district on the vitality of the target business district"). This ensures that the feature can accurately reflect the effect of spatial linkage between different business districts on commercial vitality. Ultimately, independent time series features and spatial correlation features are obtained, with each type of feature focusing on a single-dimensional pattern and having no information overlap. This step, through "spatiotemporal dimension decomposition," solves the problem of "mutual interference between time patterns and spatial correlations, and insufficient modeling accuracy" caused by the traditional prediction model's mixed processing of spatiotemporal features. This allows the subsequent time series modeling unit and spatial correlation modeling unit in step S303 to specifically process the corresponding features, maximizing the capture of core information in the spatiotemporal dimension.

[0069] In this embodiment, the core of step S303 is to construct a predictive network structure that can simultaneously capture spatiotemporal patterns and integrate the logic of emotion-vitality correlation, providing a basic framework for subsequent model training. The specific process is as follows: First, the overall network architecture is defined—a spatiotemporal predictive network that integrates emotion-vitality correlation. This network is composed of three parts working together: a temporal modeling unit, a spatial correlation modeling unit, and a correlation fusion layer. Each unit has a clear division of labor and achieves information interaction through the correlation fusion layer. Among them, the temporal modeling unit is specifically used to process the time series features obtained in step S302. Its core function is to capture the time dependence of business vitality (such as the correlation of short-term intraday fluctuations, medium-term weekly cycles, and long-term seasonal trends). Technically, Long Short-Term Memory (LSTM) networks, temporal attention mechanisms, or Transformer time series modules can be used. By recursively processing the ordered input of time series features (such as the gating mechanism of LSTM), time-related information such as "the influence weight of the previous period's vitality on the next period" and "the recurrence pattern of historical vitality patterns" can be extracted to ensure accurate modeling of the dynamic dependence of commercial vitality over time. The spatial correlation modeling unit is specifically used to process the spatial correlation features obtained in step S302. Its core function is to capture the spatial correlation between different business districts (such as the mutual support of customer flow between adjacent business districts and the radiation effect of the core business district on secondary business districts). Technically, graph neural networks (GNNs), spatial attention mechanisms, or spatial convolutional neural networks (CNNs) can be used. By treating business districts as graph nodes and spatial correlation features as edge weights (such as the adjacency matrix of a GNN), spatial correlation information such as "the intensity of the impact of changes in the vitality of business district A on business district B" and "the coefficient of synergy / competition between multiple business districts" can be calculated to ensure accurate modeling of the effect of spatial interaction between business districts on business vitality. The correlation fusion layer is the core optimization module of the network. Its core function is to introduce the correlation coefficient matrix between the emotional features and the business vitality index output in step S204 and dynamically adjust the weight distribution of emotional features in spatiotemporal modeling.

[0070] Specifically, the fusion layer analyzes the "influence strength of emotional features on vitality indicators" in the correlation coefficient matrix (e.g., the influence weight of positive emotional features on passenger flow is 0.6). When a certain type of emotional feature has a significant impact, its proportion in the temporal modeling unit (e.g., strengthening the weight of time patterns with emotional associations) and the spatial association modeling unit (e.g., highlighting spatial interaction relationships driven by emotions) is increased; conversely, its weight is reduced to ensure that the influence of emotional factors matches the actual correlation strength and to avoid irrelevant emotional features interfering with the spatiotemporal pattern modeling. The three units integrate information through the correlation fusion layer: the outputs of the temporal modeling unit and the spatial association modeling unit are first adjusted by the fusion layer according to the emotional association weights, and then summarized into a unified spatiotemporal feature representation, forming a complete prediction network logic. This step, through "modular modeling + emotional association fusion," solves the problem that traditional spatiotemporal prediction models only focus on spatiotemporal patterns and ignore the quantitative impact of emotions on vitality, enabling the network to simultaneously capture spatiotemporal dynamics and emotional associations, providing more realistic structural support for model training in step S304.

[0071] In this embodiment, the core of step S304 is to obtain a model that can accurately output future business vitality prediction results by training and optimizing a spatiotemporal prediction network that integrates emotion and vitality correlation. The specific process is as follows: First, the input and reference benchmark for model training are defined—the time series features and spatial correlation features obtained in step S302 are used as network inputs (sent to the temporal modeling unit and the spatial correlation modeling unit respectively), and the actual business vitality indicators (such as historical average daily passenger flow and actual turnover) obtained in step S301 are used as the reference benchmark for supervised output to ensure that the input features and supervised indicators are strictly aligned in the spatiotemporal dimension (such as the time / space features of the second week of May 2023, matching the actual passenger flow indicators of that week); then, the input features and reference benchmark are divided into training set and validation set according to a preset ratio (such as an 8:2 ratio, 80% of the data is used for parameter learning, and 20% of the data is used to evaluate generalization ability), and network training is started: the temporal modeling unit processes time series features to capture time dependence, and the spatial correlation modeling unit processes spatial correlation features. To capture spatial correlation, the correlation fusion layer dynamically adjusts the weights of emotional features based on the correlation coefficient matrix in step S204 (e.g., when the weight of "good service" emotion on customer flow is 0.5, the spatiotemporal feature proportion of this emotional correlation is strengthened). During training, the deviation between the predicted vitality index and the actual index output by the network is calculated through the loss function (e.g., mean squared error MSE). The network parameters (including the recursive coefficients of the temporal units, the correlation weights of the spatial units, and the emotion adjustment coefficients of the fusion layer) are adjusted by backpropagation using a gradient descent optimizer (e.g., Adam). The iteration continues until the loss value of the validation set stabilizes within the preset threshold (e.g., MSE < 0.01) and no longer decreases. Finally, the trained spatiotemporal sequence prediction model is obtained. This model has learned the comprehensive logic of time change patterns, spatial correlations, and emotional-vitality quantitative correlations. It can receive new time series features and spatial correlation features and output the business vitality prediction results (e.g., predicted customer flow, predicted turnover, and other quantitative indicators) for the next preset period (e.g., the next 1 day, 3 days, and 7 days). This step, through "joint input of spatiotemporal features + dynamic weighting of emotion association + supervised training with actual indicators," solves the problem of "ignoring the influence of emotion and insufficient learning of spatiotemporal patterns" in the training of traditional spatiotemporal prediction models. It ensures that the model can adapt to complex scenarios involving temporal dynamics, spatial linkage, and emotion-driven factors, providing high-precision model support for the business vitality prediction in step S30.

[0072] In one embodiment, step S30 specifically includes the following steps:

[0073] S311. Real-time spatiotemporal feature extraction is performed on the core feature vector of commercial vitality to obtain the current time series feature and the current spatial correlation feature. The current time series feature is used to characterize the time trend feature of current commercial vitality, and the current spatial correlation feature is used to characterize the real-time spatial interaction feature between different business districts.

[0074] S312. Input the correlation coefficient matrix of the current time series features, current spatial correlation features, and sentiment features with the business vitality index into the trained spatiotemporal sequence prediction model. The current time series features are input into the time series modeling unit, the current spatial correlation features are input into the spatial correlation modeling unit, and the correlation coefficient matrix is ​​input into the correlation fusion layer.

[0075] S313. The temporal modeling unit is used to process the current time series features to capture the real-time time trend. The spatial correlation modeling unit is used to process the current spatial correlation features to capture the real-time spatial interaction. The correlation fusion layer is used to dynamically adjust the influence weight of the emotion features in the real-time time trend and the real-time spatial interaction based on the correlation coefficient matrix to obtain the spatiotemporal features of the fused emotion-vitality correlation.

[0076] S314. Based on the spatiotemporal features of the fusion of emotion and vitality correlation, the trained spatiotemporal sequence prediction model outputs the prediction results of business vitality within a preset future time period.

[0077] In this embodiment, the core of step S311 is to extract real-time spatiotemporal dimension features from the core feature vector of business vitality, providing targeted input for the trained spatiotemporal sequence prediction model. The specific process is as follows: First, the processing object is defined—the core feature vector of business vitality obtained in step S214 (a real-time feature vector that has integrated multimodal information including geospatial, emotional, real-time behavioral, and economic attributes, and optimized through emotional-vitality correlation). The extraction logic is based on the characteristic that "the dynamic changes in real-time business vitality are simultaneously affected by the current time trend and spatial interaction," extracting current time series features and current spatial correlation features that can represent real-time patterns respectively; among them, the core feature vector of business vitality is subjected to real-time time dimension... Feature extraction to obtain current time series features: First, the core feature vectors are arranged in order according to real-time time granularity (e.g., hour, half-day) (e.g., core features from 0:00 to 24:00 of the day are recorded sequentially by hour). Then, real-time time series feature engineering methods (e.g., rolling window statistics, real-time trend slope calculation) are used to extract current time trend features, such as capturing "the rate of increase / decrease of business activity in the past 2 hours", "the difference in activity between the current period and today's morning peak period", and "the increase in real-time cumulative activity compared to the same period yesterday", etc. These real-time trends are transformed into numerical current time series features (e.g., the vector dimension can include "average activity in the past 1 hour", "activity fluctuation value in the past 5 minutes", "cumulative activity in the current day"). Quantitative indicators such as "month-on-month growth rate of business activity" are used to ensure that this feature accurately reflects the real-time trend of current business activity over time. Simultaneously, real-time spatial dimension feature extraction is performed on the core feature vector of business activity to obtain current spatial correlation features: first, the core feature vectors of business activity in different business districts within the same real-time period are correlated (e.g., at 10:00 AM, the core features of the target business district and three adjacent business districts within a 5-kilometer radius). Then, real-time spatial correlation analysis methods (such as real-time Pearson correlation coefficient calculation and dynamic spatial weight updates) are used to extract current spatial interaction features, such as identifying "the real-time diversion ratio of current business district customer flow growth to adjacent business districts" and "the impact of the current vitality of the regional business center on the target business district." The real-time radiation intensity of the business district and the synergistic / competitive relationship of current consumption types in multiple business districts are transformed into numerical current spatial correlation features (such as the vector dimension containing quantitative indicators such as "real-time vitality correlation coefficient between the target business district and adjacent business districts" and "influence weight of the current core business district on the target business district"). This ensures that the feature can accurately reflect the effect of real-time spatial interaction between different business districts on commercial vitality. Finally, independent current time series features and current spatial correlation features are obtained. The two types of features focus on real-time time trends and spatial interactions, respectively, and are consistent with the historical spatiotemporal features extracted in step S302 in terms of dimension and format (to facilitate the model to reuse the patterns learned from history).This step, through "real-time spatiotemporal feature decomposition," solves the problem in traditional real-time prediction where the direct use of fused features leads to "mutual interference between real-time temporal trends and spatial interactions, making it difficult for the model to capture dynamic patterns." It provides accurate and suitable real-time input data for the subsequent step S312, which inputs features into the spatiotemporal sequence prediction model.

[0078] In this embodiment, the core of step S312 is to accurately input the real-time extracted spatiotemporal features and emotional correlation logic into the corresponding module of the trained spatiotemporal sequence prediction model, providing an adapted input stream for the model's real-time operation. The specific process is as follows: First, clarify the input data and corresponding receiving module—the current time series features and current spatial correlation features obtained in step S311, as well as the correlation coefficient matrix between the emotional features and the business vitality index output in step S204, are jointly input into the trained spatiotemporal sequence prediction model in step S304. The three types of data need to be adapted according to the model's preset interface format (such as feature dimensions and numerical ranges being consistent with those used during training). Among them, the current time series features (representing the time trend of current business vitality, such as "the rate of increase in vitality in the last 2 hours") are directed into the model's time series modeling unit. Since this unit has learned the processing logic of time series features during the training phase (such as capturing time dependencies through LSTM), its time regularity modeling can be directly reused in real-time input. The model's spatial correlation modeling unit is directly input with the current spatial correlation features (representing the spatial interaction between current business districts, such as "the real-time vitality correlation coefficient between the target business district and adjacent business districts"). Since this unit has already mastered the processing logic of spatial correlation features during the training phase (e.g., capturing spatial correlation through GNN), its spatial interaction modeling capabilities can be directly reused with real-time input. The correlation coefficient matrix between emotional features and business vitality indicators (quantitatively representing the weight of emotion on vitality, such as "positive emotion has a weight of 0.6 on customer flow indicators") is directly input with the model's correlation fusion layer. Since this layer has already learned the logic of adjusting emotional weights based on the matrix during the training phase, real-time input ensures that its dynamic adjustment rules are consistent with historical correlation patterns. During the input process, it is necessary to ensure that the spatiotemporal benchmarks of the three types of data are strictly aligned (e.g., "time series features at 10:00," "spatial correlation features at 10:00," and "concurrently valid correlation coefficient matrix" are input simultaneously) to avoid model calculation deviations due to temporal or spatial misalignment. This step, through precise "feature-module" matching, solves the problem of "processing logic confusion and decreased prediction accuracy" caused by the mismatch between real-time input data and the model's internal structure. It ensures that the model can efficiently reuse the spatiotemporal patterns and emotion association logic learned during the training phase, providing a stable input foundation for real-time feature processing in step S313.

[0079] In this embodiment, the core of step S313 is to collaboratively process real-time features through various units of the model and incorporate emotion-vitality correlation logic to obtain features that comprehensively reflect real-time spatiotemporal patterns and emotional influences. The specific process is as follows: First, the parallel processing mechanism of the three major units of the model is activated, and targeted operations are performed on the input features based on their respective functional positioning. Then, information integration is achieved through the correlation fusion layer. Among them, the time series modeling unit (such as LSTM, time attention module) processes the input current time series features (such as "vitality increase rate in the past 2 hours" and "cumulative month-on-month growth rate of vitality on the current day") in real time: through recursive operations or attention weight allocation, it captures the time trend pattern of current business vitality, such as recognizing... Instead of focusing on whether the current activity growth between 10:00 and 12:00 continues the morning peak trend, or whether real-time hourly activity fluctuations conform to historical patterns for the same period (e.g., the same day last week), these real-time time trends are transformed into structured time feature representations to ensure accurate depiction of the dynamic evolution of current activity over time. Simultaneously, spatial correlation modeling units (such as GNNs and spatial convolution modules) process the input current spatial correlation features (such as the real-time activity correlation coefficient between the target business district and adjacent business districts, and the real-time radiation weight of the regional core business district) in real time. Through graph node operations or spatial weight calculations, the spatial interaction relationships between different business districts are captured, for example, determining whether the current growth in customer flow in the target business district is... Whether it triggers the diversion of customer flow to adjacent business districts, and whether the real-time high activity of the regional business center has a positive radiating effect on surrounding secondary business districts, these real-time spatial interactions are transformed into structured spatial feature representations to ensure accurate depiction of the spatial linkage effect between current business districts. Subsequently, the correlation and fusion layer dynamically adjusts the weights of the real-time time trend features output by the time series unit and the real-time spatial interaction features output by the spatial unit based on the correlation coefficient matrix of the input emotional features and business activity indicators (such as "the impact of negative emotions on the average daily customer flow on weekends is weighted at -0.4", and "the impact of positive emotions on the frequency of consumption is weighted at 0.5"). If negative emotions have a significant impact during this period, the weight of the time trend feature associated with negative emotions will be increased, while irrelevant time components will be weakened. If there is "mutual support between business districts driven by positive user reviews" in the current spatial interaction (the weight of positive emotions promoting spatial linkage in the correlation coefficient matrix is ​​0.3), the weight of the spatial interaction feature associated with positive emotions will be increased to highlight the spatial effect driven by emotions. Finally, the fusion layer will integrate the real-time time trend feature after weight adjustment with the real-time spatial interaction feature to form a unified dimension of fused emotional-vitality correlation spatiotemporal features (such as a vector that includes both "weekend midday vitality continuation trend (negative emotion weight increase)" and "mutual support effect between business districts driven by positive emotions").This step, through "unit-based modeling + dynamic optimization of emotion weights," solves the problem of traditional real-time prediction focusing only on spatiotemporal patterns and ignoring the impact of emotions on real-time vitality. It ensures that spatiotemporal features can simultaneously reflect dynamic patterns and emotion correlations, providing high-quality feature support for the prediction results output in step S314.

[0080] In this embodiment, the core of step S314 is to generate a business vitality prediction result for a future preset time period based on the comprehensive spatiotemporal features that integrate emotion-vitality correlation. The specific process is as follows: First, clarify the input features and model foundation—using the spatiotemporal features that integrate emotion-vitality correlation (including comprehensive information on real-time time trends, spatial interaction, and emotion weight optimization) obtained in step S313 as input, and relying on the spatiotemporal sequence prediction model trained in step S304 (which has learned historical spatiotemporal patterns and the quantitative correlation logic of emotion-vitality) to perform prediction calculations; after receiving the integrated features, the model performs the final mapping of the features through the output layer (such as a fully connected neural network layer): transforming the "spatiotemporal features that integrate emotion-vitality correlation" into a quantitative indicator of business vitality for the corresponding future preset time period (such as the next 24 hours, 3 days, or 7 days, which can be configured according to actual needs), for example, predicting the next day. The model predicts peak hourly customer traffic, average daily revenue for the next 3 days, and consumption frequency trends for the next 7 days. During the prediction process, the model reuses data learned during the training phase, including "time trend continuity" (e.g., the impact of the current midday activity trend on the evening), "spatial interaction transmission" (e.g., the radiating effect of high activity in adjacent business districts on the future of the target business district), and "emotional association driving force" (e.g., the positive pull weight of current high positive sentiment on future customer traffic). This ensures that the prediction results are adapted to both real-time dynamics and historical patterns. The final output of the business activity prediction results is structured quantitative data (e.g., in vector or tabular form, containing information such as "time period - predicted value - confidence level", for example, "predicted customer traffic of 5,000 people from 18:00 to 20:00 in the next 24 hours, confidence level 92%), and the results must match the decision-making needs of the actual business scenario (e.g., to allow business district operators to adjust manpower, inventory, or marketing activities). This step, through the collaborative computation of "fusion features + training model", solves the prediction bias problem caused by "extrapolating only based on historical patterns and ignoring the influence of real-time emotions and spatiotemporal interactions" in traditional real-time prediction. It ensures that the output of future business vitality results can not only reflect real-time dynamic changes, but also fit the intrinsic correlation logic between emotions and vitality, providing a high-precision quantitative basis for subsequent business decisions.

[0081] In one embodiment, step S40 specifically includes the following steps:

[0082] S401. Obtain real-time updated data and the business activity prediction results, wherein the real-time updated data includes real-time multimodal feature data and the actual business activity indicators for the corresponding time period.

[0083] In this embodiment, the core of step S401 is to collect basic data for the subsequent construction of a two-dimensional evaluation index system. The specific process is as follows: First, it is clear that two types of key data need to be acquired—real-time updated data and business vitality prediction results. The two types of data need to be strictly aligned in the spatiotemporal dimensions to ensure the accuracy of the subsequent evaluation. Among them, the real-time updated data includes two parts: one is real-time multimodal feature data, which is the real-time dynamic update version of the "structured multimodal dataset" in step S10 (such as real-time user evaluations, real-time check-in records of business districts, instant traffic flow data, etc.), which is consistent with the multimodal data source for feature extraction in S211, focusing only on "current or recent real-time data" in the time dimension; the other is the actual business vitality index for the corresponding time period, which is quantitative data that can reflect the true level of current business vitality (such as actual customer traffic, real-time sales, instant consumption frequency, etc. from 10:00 to 11:00 on the same day), which is consistent with the type of "actual business vitality index" in S301. The time frame is limited to "the period that matches the real-time multimodal feature data and the business activity prediction results" (e.g., if the prediction is "business activity in the next 24 hours", the actual indicator must include "real-time actual data of the hour before the prediction period" as the evaluation benchmark). The business activity prediction results directly use the output data of step S314, that is, the "business activity prediction results within the future preset period" output by the trained spatiotemporal sequence prediction model (e.g., "passenger flow prediction value per hour in the next 24 hours", "average daily turnover prediction value in the next 3 days", etc.). It is necessary to ensure that the time period and indicator type (e.g., passenger flow, turnover) of the prediction results completely correspond to the "actual business activity indicators" in the real-time updated data (e.g., if the prediction is "passenger flow in the next 24 hours", the actual indicator must include "real-time passenger flow before the prediction start time" for subsequent comparison and evaluation). Finally, a dataset of "real-time multimodal feature data + corresponding actual business activity indicators + business activity prediction results for the matching period" is formed. This step, by integrating real-time data and predicted data, solves the problem of "lack of benchmark and comparative data" in subsequent evaluations, and provides complete and aligned basic data support for step S402 to construct a quantitative and qualitative dual-dimensional evaluation index system.

[0084] S402. Based on the real-time updated data and business vitality prediction results, construct a quantitative and qualitative dual-dimensional evaluation index system. The quantitative indexes are used to characterize the numerical evaluation results of business vitality, while the qualitative indexes are used to characterize the descriptive evaluation results of business vitality. Specifically, this includes the following steps:

[0085] S4021. Based on the actual business vitality indicators and business vitality prediction results in the real-time updated data, construct quantitative dimension indicators, including prediction deviation rate, vitality growth rate and sentiment correlation influence coefficient.

[0086] S4022. Based on the numerical range of quantitative dimension indicators and the sentiment features in real-time multimodal feature data, construct qualitative dimension indicators, including business vitality level, sentiment-related risk type, and prediction reliability label.

[0087] S4023. By using indicator mapping rules, quantitative dimension indicators are associated with qualitative dimension indicators to form a dual-dimensional evaluation indicator system of quantitative and qualitative dimensions.

[0088] In this embodiment, the core of step S402 is to build upon the real-time updated data (including real-time multimodal feature data and actual business vitality indicators for the corresponding time period) and business vitality prediction results obtained in step S401. Through a coherent process of "quantitative indicator construction - qualitative indicator construction - indicator correlation," a dual-dimensional indicator system that can simultaneously cover numerical and descriptive assessments is established. This solves the problem that traditional single quantitative assessments "can only present numbers, lack contextual interpretation, and are difficult to support business decisions." Specifically, this is executed step by step through S4021 to S4023:

[0089] For step S4021, the core is to generate directly quantifiable assessment indicators by "comparing actual vitality indicators with predicted results, analyzing dynamic data changes, and calculating sentiment correlations," thus providing an objective numerical basis for subsequent qualitative analysis. The specific process is as follows:

[0090] Using the actual business activity indicators (e.g., 5000 actual visitors between 10:00 and 11:00 AM) and business activity prediction results (e.g., 4800 predicted visitors for the same period) from the real-time updated data in step S401 as core data, and also referencing the correlation coefficient matrix between sentiment features and business activity indicators output in step S204 (e.g., the weight of positive sentiment on visitor flow is 0.6), three types of directly calculable quantitative indicators are constructed:

[0091] 1) Prediction deviation rate: assesses the degree of deviation between the prediction result and the actual situation. The calculation formula is: "|actual vitality index - predicted vitality index| / actual vitality index × 100%"; for example: (5000-4800) / 5000×100%=4%, the smaller the value, the higher the prediction accuracy.

[0092] 2) Vitality Growth Rate: This assesses the real-time trend of business vitality. The formula is: "(Actual vitality index in the current period - Actual vitality index in the previous period) / Actual vitality index in the previous period × 100%"; For example: if the customer flow in the previous period was 4,500 people and the current period is 5,000 people, the growth rate is (5,000-4,500) / 4,500 × 100%≈11.1%, and a positive number represents an increase in vitality.

[0093] 3) Emotional Correlation Influence Coefficient: This assesses the actual driving effect of emotions on current vitality. The calculation logic is: "Real-time emotional feature intensity × influence weight of the corresponding emotion in the correlation coefficient matrix"; for example, if the real-time positive emotion intensity is 0.8, the corresponding weight is 0.6, and the coefficient is 0.8 × 0.6 = 0.48. The larger the value, the stronger the positive emotion driving force.

[0094] Ultimately, three types of quantitative indicators were formed that can be directly used for numerical comparison.

[0095] For step S4022, the core is to "transform the abstract quantitative indicators of S4021 into scenario-based descriptions" and combine them with real-time sentiment characteristics to supplement the interpretation of risks and reliability, making the assessment results more aligned with business decisions. The specific process is as follows:

[0096] Based on the numerical range of the quantitative indicators obtained in S4021, and combined with the emotional features in the real-time multimodal feature data in step S401 (such as a real-time negative emotion ratio of 25%), the abstract quantitative results are transformed into a scenario-based description: the business vitality level corresponds to the vitality growth rate range, such as a growth rate > 10% is defined as "high growth level", and a growth rate of 11.1% corresponds to this level; the emotion association risk type combines the emotion association influence coefficient and real-time emotion features, such as a coefficient < 0.2 and a negative emotion ratio > 20% is defined as "service complaint risk", while a coefficient of 0.3 and a negative emotion ratio of 15% correspond to "no significant risk"; the prediction reliability label corresponds to the prediction deviation rate range, such as a deviation rate < 5% is defined as "high reliability", and a deviation rate of 4% corresponds to this label, thus forming a qualitative indicator that can be interpreted intuitively.

[0097] For step S4023, the core is to bind quantitative indicators with qualitative indicators through preset rules, so as to achieve the synergy of "numerical support for description and description for interpretation of numerical values", avoiding fragmented evaluation. The specific process is as follows:

[0098] First, establish pre-defined indicator mapping rules, clarifying that "one quantitative indicator range corresponds to a unique qualitative indicator result." For example, "vitality growth rate 11.1% (>10%)" corresponds to "business vitality level: high growth level," "prediction deviation rate 4% (<5%)" corresponds to "prediction reliability label: high reliability," and "emotional correlation influence coefficient 0.3 (0.2~0.5) + negative emotion proportion 15%" corresponds to "emotional correlation risk type: no significant risk." Then, bind the quantitative indicators of S4021 with the qualitative indicators of S4022 one by one according to these rules to form paired indicators of "quantitative value + qualitative description." Finally, integrate them into a unified quantitative and qualitative dual-dimensional evaluation indicator system, providing logically coherent and format-adapted indicator inputs for the real-time calculation and analysis in the subsequent step S403.

[0099] As needed, the indicator mapping rules in step S4023 are constructed as follows:

[0100] S40231. Based on the numerical distribution characteristics of quantitative dimension indicators, establish classification mapping rules:

[0101] Map the prediction deviation rate range to the prediction reliability label;

[0102] Map the vitality growth rate range to the business vitality level;

[0103] Map the range of emotion-related impact coefficients to emotion-related risk types;

[0104] S40232. Introduce emotion features from real-time multimodal feature data to construct emotion association modulation rules:

[0105] When the proportion of negative emotion characteristics exceeds a preset threshold, the range threshold for mapping quantitative indicators to qualitative labels is tightened.

[0106] S40233. Based on error feedback from historical evaluation results, construct dynamic threshold update rules:

[0107] Regularly optimize the numerical range thresholds in the classification mapping rules to adapt the mapping rules to the long-term changing trends of business activity;

[0108] S40234. Integrate the classification mapping rule, emotion association adjustment rule and dynamic threshold update rule to obtain the indicator mapping rule.

[0109] It should be noted that:

[0110] Step S40231 focuses on constructing classification mapping rules. The core is to establish a basic correspondence between "quantitative intervals and qualitative labels" based on the numerical distribution characteristics of quantitative indicators (such as the prediction deviation rate being concentrated in the range of 0-15% and the vitality growth rate being in the range of -5%-20% in historical data). The prediction deviation rate is divided into three intervals: <5%, 5%-10%, and >10%, which are mapped to the prediction reliability labels of "high reliability", "medium reliability", and "low reliability", respectively. The vitality growth rate is divided into three intervals: >10%, 0-10%, and <0, which are mapped to the business vitality levels of "high growth", "stable", and "declining", respectively. The sentiment correlation influence coefficient is divided into three intervals: >0.5, 0.2-0.5, and <0.2, which are initially mapped to the sentiment correlation risk types of "vitality enhancement opportunities", "no significant risks", and "potential risks", providing a basic framework for indicator correlation.

[0111] Step S40232 aims to introduce dynamic adjustment of emotional features. The core is to combine emotional features (especially the proportion of negative emotions) in real-time multimodal feature data to construct emotional association adjustment rules: preset a threshold for the proportion of negative emotions (e.g., 20%). When the proportion of negative emotions in real time exceeds this threshold, tighten the range threshold for mapping quantitative indicators to qualitative labels. For example, the original "vitality growth rate > 10% → high growth level" is adjusted to "> 12% → high growth level" to avoid misjudging "high growth" due to the fermentation of negative emotions. If the proportion of negative emotions is lower than the threshold, the original range is maintained to ensure that the rules can adapt to the real-time impact of emotions on business vitality.

[0112] Step S40233 focuses on optimizing rules through historical feedback. The core is to construct dynamic threshold update rules based on error data of historical evaluation results (such as the case of "prediction deviation rate of 4% was marked as high reliability but the actual deviation was significant" in the past 3 months): regularly (such as weekly) statistically analyze the matching error between quantitative indicator ranges and qualitative labels. When the error rate of a certain range exceeds the preset value (such as 15%), the threshold of that range is adjusted. For example, in the past, "the actual reliability of prediction deviation rate of 5% was lower than expected", so the threshold of the "high reliability" range is tightened from <5% to <4%, so that the rules can adapt to the long-term changing trend of business vitality (such as changes in consumption habits in business districts and seasonal fluctuations).

[0113] Step S40234 involves the integration and implementation of the rules. The core of this step is to integrate the classification mapping rules (basic framework) of S40231, the emotion association adjustment rules (real-time adaptation) of S40232, and the dynamic threshold update rules (long-term optimization) according to priority. First, the dynamically updated basic range is applied, and then adjustment rules are triggered based on real-time emotion characteristics, ultimately forming a complete indicator mapping rule that is "basically stable + real-time adjustable + long-term adaptable." For example, if the "high reliability" range after dynamic updating is <4%, and the real-time negative emotion percentage is 18% (not exceeding the 20% threshold), then "prediction deviation rate 3.5%" is directly mapped to "high reliability." If the negative emotion percentage is 25% (exceeding the threshold), it is temporarily tightened to <3%, and 3.5% is mapped to "medium reliability." This integrated rule ensures the stability of indicator associations while responding to emotion fluctuations and long-term trends, providing a precise and dynamic basis for the indicator association in S4023.

[0114] S403. Input the quantitative and qualitative indicators into the stream processing framework, and perform real-time calculation and correlation analysis on the two types of indicators through the stream processing framework. Specifically, this includes the following steps:

[0115] S4031. Integrate quantitative and qualitative indicators into the stream processing framework in real time and establish a mapping relationship between the indicators and the framework's computing nodes.

[0116] S4032. Through the numerical computation node of the stream processing framework, the quantization dimension index is calculated in real time to obtain the real-time calculation result of the quantization index.

[0117] S4033. Through the rule matching node of the stream processing framework, the qualitative dimension indicators are matched with labels in real time based on the preset evaluation rules to obtain the real-time processing results of the qualitative indicators.

[0118] S4034. Through the correlation analysis node of the stream processing framework, based on the index mapping rules, the real-time calculation results of quantitative indicators and the real-time processing results of qualitative indicators are correlated in real time to obtain the two-dimensional index correlation analysis results.

[0119] S4035 outputs real-time calculation results of quantitative indicators, real-time processing results of qualitative indicators, and correlation analysis results of two-dimensional indicators as the calculation and analysis results of the stream processing framework.

[0120] In this embodiment, the core of step S403 is to build upon the quantitative and qualitative dual-dimensional indicators constructed in step S402. Through a stream processing framework, it achieves integrated processing of "real-time computation + rule matching + correlation analysis," solving the problems of "analysis lag and separation of quantitative and qualitative indicators" in traditional batch processing. This provides real-time, interconnected analytical support for the subsequent output of dynamic evaluation results in S404. Specifically, this is executed step-by-step through S4031 to S4035:

[0121] For step S4031, the core is to establish the correspondence between the two-dimensional indicators and the computing nodes of the stream processing framework, ensuring that the indicators can be accurately connected to the framework and processed in a targeted manner. The specific process is as follows: First, clarify the selection of the stream processing framework (such as Flink or Spark Streaming, which needs to support low-latency real-time data processing). Then, based on the characteristics of quantitative dimension indicators (numerical) and qualitative dimension indicators (descriptive), establish the mapping rules of "indicator-node"—map quantitative dimension indicators (such as real-time prediction deviation rate and vitality growth rate) to the framework's "numerical computing node" (responsible for numerical calculation), and map qualitative dimension indicators (such as business vitality level and prediction reliability label) to the framework's "rule matching node" (responsible for label matching). At the same time, reserve the interface of "association analysis node" for subsequent linkage between the two types of indicators; for example, connect "prediction deviation rate updated every 5 minutes" to the numerical computing node and "prediction reliability label for the corresponding time period" to the rule matching node to ensure that the functions of the indicators and nodes are compatible and avoid data flow misalignment.

[0122] For step S4032, the core is to perform real-time calculations on quantitative dimensional indicators through the numerical calculation nodes of the stream processing framework to generate quantitative results that can be used for correlation analysis. The specific process is as follows: The numerical calculation nodes dynamically calculate the quantitative indicators that are accessed in real time based on preset calculation rules (which need to match the business evaluation requirements). For the prediction deviation rate, the "mean and maximum deviation rate in the past hour" are calculated (e.g., mean deviation rate of 3.8% and maximum deviation rate of 4.5% in the past hour) to evaluate the stability of short-term prediction accuracy. For the vitality growth rate, the "month-on-month change in growth rate over the past three periods" is calculated (e.g., growth rate of 11.1% in the previous period and 9.8% currently, a month-on-month decrease of 1.3 percentage points) to capture the fluctuations in vitality trends. For the sentiment correlation influence coefficient, the "weighted average of the coefficient in the past half hour" is calculated (e.g., under the dominance of positive sentiment, the average coefficient is 0.42) to quantify the continuous impact of sentiment on vitality. Finally, the structured real-time calculation results of the quantitative indicators are output, ensuring that each result has a timestamp (e.g., "mean deviation rate of 3.8% in the period from 10:05 to 10:10") to facilitate subsequent alignment with qualitative indicators.

[0123] For step S4033, the core is to use the rule matching node of the stream processing framework to perform real-time label verification and supplementation of qualitative dimension indicators based on preset evaluation rules, and generate standardized qualitative processing results. The specific process is as follows: the preset evaluation rules must be consistent with the qualitative indicator definition in S4022. The rule matching node performs two processing on the real-time accessed qualitative indicators: first, label validity verification, to determine whether the qualitative indicator matches the quantitative indicator range (e.g., "prediction deviation rate 3.8%" corresponds to the "high reliability" label, and if the accessed label is "medium reliability", it will be automatically corrected); second, label supplementation explanation, adding scenario-based interpretation in combination with real-time multimodal feature data (e.g., "business vitality level: high growth level" is supplemented with "driven by weekend consumption peak"). For example, for "prediction reliability label: high reliability", it is verified that its corresponding quantitative indicator "deviation rate 3.8%" is in the <5% range, and then the explanation "the deviation rate fluctuation is small in the past 1 hour, and the prediction stability is strong" is added. Finally, the real-time processing result of the qualitative indicators with verification results and supplementary explanations is output to ensure the accuracy and richness of the qualitative description.

[0124] For step S4034, the core is to achieve real-time linkage between quantitative calculation results and qualitative processing results through the correlation analysis node of the stream processing framework, relying on the indicator mapping rules of S4023. The specific process is as follows: The correlation analysis node first obtains the real-time calculation results of the quantitative indicators in S4032 (such as "average deviation rate of 3.8% in the past hour" and "month-on-month decrease of 1.3% in the growth rate of the past 3 periods") and the real-time processing results of the qualitative indicators in S4033 (such as "prediction reliability: high reliability (small fluctuations)" and "business vitality level: high growth level (weekend driven)"). Then, based on the indicator mapping rules, it establishes a one-to-one association between "quantitative value - qualitative description" - linking "average deviation rate of 3.8%" with "high" and "qualitative description". The correlation between "reliable (small fluctuations)" and "prediction accuracy: average deviation rate of 3.8% in the past hour (high reliability, small fluctuations)" is formed. The correlation between "growth rate decreased by 1.3% month-on-month" and "high growth level (weekend driven)" is formed. The correlation between "growth rate decreased by 1.3% month-on-month in the past 3 periods (still high growth level, supported by weekend consumption peak)" is formed. At the same time, the correlation between the sentiment correlation coefficient (average 0.42) and the "no significant risk" label is formed. The correlation between "sentiment influence: average coefficient of 0.42 in the past half hour (no significant risk, positive sentiment accounts for 65%)" is formed. Finally, the correlation analysis results of multiple sets of linked two-dimensional indicators are obtained, which solves the problem of "independent and unlinked interpretation" of the two types of indicators.

[0125] For step S4035, the core is to integrate the output results of each node in the stream processing framework to form a complete calculation and analysis dataset. The specific process is as follows: the real-time calculation results of quantitative indicators in S4032 (such as the average deviation rate and the month-on-month growth rate), the real-time processing results of qualitative indicators in S4033 (such as the verified labels and supplementary explanations), and the correlation analysis results of dual-dimensional indicators in S4034 (such as the "quantitative + qualitative" linkage conclusion) are integrated in the format of "time period - indicator type - result content". For example, "10:00-11:00 time period: quantitative results (average deviation rate 3.8%, month-on-month growth rate -1.3%); qualitative results (high reliability, high growth level); correlation results (deviation rate 3.8% → high reliability, month-on-month growth rate decreased by 1.3% but is still high growth level)". Finally, this integrated dataset is output as the calculation and analysis result of the stream processing framework, ensuring that the data structure is unified and the content is complete, providing a direct basis for generating the dynamic evaluation results of business vitality in step S404.

[0126] S404. Based on the computational analysis results of the aforementioned stream processing framework, a dynamic evaluation result of business vitality is obtained.

[0127] In this embodiment, the core of step S404 is to follow up on the "quantitative calculation results + qualitative processing results + two-dimensional correlation analysis results" output by the stream processing framework in step S403. Through the process of "data integration → multi-dimensional decomposition → dynamic conclusion generation", fragmented analysis data is transformed into dynamic assessment results of business vitality that can directly support business decisions, solving the problems of traditional assessments such as "fragmented results, lack of systematic interpretation, and inability to reflect changes in vitality in real time". The specific execution process is as follows:

[0128] First, the spatiotemporal dimensions of the analysis results are integrated. Based on the calculation and analysis results output in step S403, the data is categorized according to "time granularity (e.g., hourly, daily) + business district scope (e.g., target business district, regional business district cluster)" to ensure the consistency of the spatiotemporal benchmark of the evaluation results. For example, the quantitative results of "10:00-11:00 business district A (average deviation rate 3.8%, month-on-month growth rate -1.3%), qualitative results (high reliability, high growth level), and correlation results (deviation rate 3.8% → high reliability, month-on-month growth rate decreased by 1.3% but still high growth level)" are integrated into "business district A 10-11 hour vitality analysis package" to avoid evaluation confusion caused by spatiotemporal misalignment.

[0129] Next, a multi-dimensional analysis was conducted, focusing on the core concerns of business vitality assessment (prediction accuracy, vitality trend, and sentiment impact). The results were interpreted in depth using a two-dimensional correlation: In the prediction accuracy dimension, the credibility of the prediction was judged by "quantitative deviation rate mean + qualitative reliability label." For example, for business district A10-11, the mean deviation rate was 3.8% (<5%) and the label was "high reliability," which can be interpreted as "the current business vitality prediction results are highly credible and the risk of deviation is low." In the vitality trend dimension, the dynamics of vitality were judged by "quantitative growth rate change + qualitative vitality level." For example, a 1.3% month-on-month decrease in growth rate but still maintaining a "high growth level" can be interpreted as "although vitality has slightly declined, it is still in a high growth range overall, and there is no trend of decline." In the sentiment impact dimension, the driving role of sentiment was judged by "quantitative sentiment correlation coefficient + qualitative risk type." For example, a mean coefficient of 0.42 and a label of "no significant risk" can be interpreted as "positive sentiment still provides positive support for current vitality, and there is no risk of vitality fluctuations caused by negative sentiment."

[0130] Finally, a dynamic evaluation conclusion is generated, integrating the multi-dimensional analysis results into a structured dynamic evaluation report. The report must include four core modules: "Current vitality status (e.g., high growth level), key influencing factors (e.g., weekend consumption peak support, positive sentiment driving), potential risk warnings (e.g., no significant risk / need to pay attention to rising negative sentiment), and short-term trend prediction (e.g., high growth may continue in the next 1-2 hours)." The conclusion must be updated synchronously with the real-time analysis results of the streaming processing framework (e.g., the evaluation report is refreshed every hour). For example, the dynamic evaluation result for business district A10-11 is: "Current business vitality level: high growth level; core support: increased weekend consumer traffic, positive sentiment accounts for 65%; prediction credibility: high reliability (deviation rate 3.8%); short-term prediction: high growth may continue from 11-12, need to pay attention to changes in midday peak traffic; risk warning: no significant sentiment or prediction risk."

[0131] This step, through systematic integration and interpretation, transforms the technical analysis results of the stream processing framework into decision-making basis that can be directly applied by the business side, solving the problem of "disconnect between technical data and business decisions". The final output of the dynamic assessment results of business vitality can be directly used by business district operators to adjust human resource allocation, marketing activities or merchant cooperation strategies, so as to realize real-time monitoring and dynamic management of business vitality.

[0132] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0133] In one embodiment, a spatiotemporal prediction and dynamic evaluation system for business vitality based on multimodal feature fusion is provided. This system corresponds to the spatiotemporal prediction and dynamic evaluation method for business vitality based on multimodal feature fusion described in the previous embodiment. The system includes:

[0134] The data processing module is used to collect multimodal heterogeneous data including geospatial modality, text semantic modality, real-time behavioral modality and economic attribute modality, and to perform data cleaning and desensitization processing on the multimodal heterogeneous data to obtain a structured multimodal dataset.

[0135] The association calculation module is used to extract features from each single modality data in the structured multimodal dataset based on the corresponding feature extraction algorithm, obtain each single modality feature vector, perform feature fusion on each single modality feature vector through a cross-modal attention mechanism, and calculate the core feature vector of business vitality based on the pre-built emotion-vitality association model.

[0136] The result prediction module is used to input the core feature vector of business vitality into a pre-constructed spatiotemporal sequence prediction model, and output the prediction result of business vitality within a future preset time period through the spatiotemporal sequence prediction model.

[0137] The dynamic evaluation module is used to acquire real-time updated data and business vitality prediction results. Based on the real-time updated data and business vitality prediction results, a quantitative and qualitative dual-dimensional evaluation index system is constructed. The evaluation index system is calculated in real time through a stream processing framework to obtain dynamic evaluation results.

[0138] The results output module is used to display the dynamic evaluation results through a visual interactive panel and output a structured evaluation report containing the business vitality prediction results and the dynamic evaluation results.

[0139] Specific limitations regarding the spatiotemporal prediction and dynamic evaluation system for business vitality based on multimodal feature fusion can be found in the limitations of the method for spatiotemporal prediction and dynamic evaluation of business vitality based on multimodal feature fusion described above, and will not be repeated here. Each module in the aforementioned spatiotemporal prediction and dynamic evaluation system for business vitality based on multimodal feature fusion can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0140] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows. Figure 2 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database is used for data storage, data processing, and data analysis. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a spatiotemporal prediction and dynamic evaluation method for business vitality based on multimodal feature fusion.

[0141] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a method for spatiotemporal prediction and dynamic evaluation of business vitality based on multimodal feature fusion.

[0142] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements a method for spatiotemporal prediction and dynamic evaluation of business vitality based on multimodal feature fusion.

[0143] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0144] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0145] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for spatiotemporal prediction and dynamic evaluation of business vitality based on multimodal feature fusion, characterized in that, Includes the following steps: Collect multimodal heterogeneous data including geospatial modality, text semantic modality, real-time behavioral modality, and economic attribute modality; perform data cleaning and desensitization processing on the multimodal heterogeneous data to obtain a structured multimodal dataset. Feature extraction is performed on each unimodal data in the structured multimodal dataset based on the corresponding feature extraction algorithm to obtain unimodal feature vectors. Feature fusion is then performed on these unimodal feature vectors using a cross-modal attention mechanism. Finally, a core feature vector of business vitality is calculated based on a pre-built emotion-vitality association model. The construction steps of the pre-built emotion-vitality association model include: Acquire multimodal historical data containing geospatial modality, textual semantic modality, real-time behavioral modality, and economic attribute modality; perform data cleaning and desensitization processing on the multimodal historical data to obtain a structured historical multimodal dataset; Historical text semantic modality data and historical economic attribute modality data within the corresponding time period are extracted from the structured historical multimodal dataset; sentiment features are extracted from the historical text semantic modality data to obtain a historical sentiment feature vector; and standardization processing is performed on the historical economic attribute modality data to obtain a historical vitality index vector. Using the historical emotion feature vector as input and the historical vitality index vector as output, a regression model with an attention fusion mechanism is used for training to obtain a trained emotion-vitality association model. Based on the trained emotion-vitality association model, the correlation coefficient matrix between emotion features and business vitality index is output. The correlation coefficient matrix is ​​used to quantitatively characterize the influence weight of different emotion features on business vitality index. The core feature vector of business vitality is input into a pre-constructed spatiotemporal sequence prediction model, and the spatiotemporal sequence prediction model outputs the prediction result of business vitality in a future preset time period. The process involves acquiring real-time updated data and business activity prediction results, constructing a quantitative and qualitative dual-dimensional evaluation index system based on these results, and performing real-time calculations on the evaluation index system using a stream processing framework to obtain dynamic evaluation results. Specifically, this includes the following steps: Obtain real-time updated data and the business activity prediction results, wherein the real-time updated data includes real-time multimodal feature data and actual business activity indicators for the corresponding time period; Based on the real-time updated data and business vitality prediction results, a quantitative and qualitative dual-dimensional evaluation index system is constructed. The quantitative dimension index is used to characterize the numerical evaluation results of business vitality, while the qualitative dimension index is used to characterize the descriptive evaluation results of business vitality. Quantitative and qualitative indicators are input into a stream processing framework, which performs real-time calculations and correlation analysis on the two types of indicators. Based on the computational analysis results of the aforementioned stream processing framework, a dynamic evaluation result of business vitality is obtained; The dynamic evaluation results are displayed through a visual interactive panel, and a structured evaluation report containing the business vitality prediction results and dynamic evaluation results is output. The construction of the quantitative and qualitative dual-dimensional evaluation index system is specifically as follows: Based on the actual business vitality indicators and business vitality prediction results in the real-time updated data, a quantitative dimension indicator is constructed, which includes the prediction deviation rate, vitality growth rate and sentiment correlation coefficient. Based on the numerical range of the quantitative dimension indicators and the emotional features in the real-time multimodal feature data, a qualitative dimension indicator is constructed, which includes business vitality level, emotional association risk type and prediction reliability label. The quantitative and qualitative indicators are associated through indicator mapping rules to form a dual-dimensional evaluation indicator system of quantitative and qualitative dimensions. The steps for constructing the indicator mapping rules are as follows: Based on the numerical distribution characteristics of quantitative indicators, classification mapping rules are established: the prediction deviation rate range is mapped to the prediction reliability label; the vitality growth rate range is mapped to the business vitality level; and the sentiment correlation influence coefficient range is mapped to the sentiment correlation risk type. By introducing emotional features from real-time multimodal feature data, an emotion association regulation rule is constructed: when the proportion of negative emotional features exceeds a preset threshold, the interval threshold for mapping quantitative indicators to qualitative labels is tightened. Based on error feedback from historical evaluation results, a dynamic threshold update rule is constructed: the numerical range threshold in the classification mapping rule is periodically optimized to adapt the mapping rule to the long-term changing trend of business vitality. By integrating the classification mapping rule, the emotion association modulation rule, and the dynamic threshold update rule, the indicator mapping rule is obtained.

2. The method for spatiotemporal prediction and dynamic evaluation of business vitality based on multimodal feature fusion as described in claim 1, characterized in that, The steps involved in extracting features from each unimodal data in the structured multimodal dataset using a corresponding feature extraction algorithm to obtain unimodal feature vectors, fusing these unimodal feature vectors through a cross-modal attention mechanism, and calculating the core feature vector of business vitality based on a pre-built sentiment-vitality association model. These steps specifically include the following: Based on the feature extraction algorithm that matches each modality of data, features are extracted from the geospatial modality data, text semantic modality data, real-time behavioral modality data and economic attribute modality data in the structured multimodal dataset, respectively, to obtain geospatial feature vectors, sentiment feature vectors, real-time behavioral feature vectors and economic attribute feature vectors; A cross-modal attention mechanism is used to perform multimodal dynamic weighted fusion of the geospatial feature vector, emotion feature vector, real-time behavior feature vector, and economic attribute feature vector to obtain an initial fused feature vector; The initial fusion feature vector and the correlation coefficient matrix are input into the emotion-vitality correlation model. The emotion-vitality correlation model adjusts the fusion weights of the emotion feature vector, geospatial feature vector, real-time behavior feature vector and economic attribute feature vector in the initial fusion feature vector based on the correlation coefficient matrix. The initial fusion feature vector after weight adjustment is output to obtain the core feature vector of business vitality.

3. The method for spatiotemporal prediction and dynamic evaluation of business vitality based on multimodal feature fusion as described in claim 1, characterized in that, The construction steps of the pre-built spatiotemporal sequence prediction model include: The core features of historical business activity and the actual business activity indicators for the corresponding time periods are obtained. The core features of historical business activity are obtained by extracting single-modal features, cross-modal attention fusion and calculating the emotion-activity correlation model based on a structured historical multimodal dataset. The core feature data of historical business vitality are decomposed into spatiotemporal features to obtain time series features and spatial correlation features. The time series features are used to characterize the temporal variation pattern of business vitality, and the spatial correlation features are used to characterize the spatial interaction relationship between different business districts. A spatiotemporal prediction network integrating emotion and vitality is constructed. The spatiotemporal prediction network includes a temporal modeling unit, a spatial correlation modeling unit, and a correlation fusion layer. The temporal modeling unit is used to process time series features to capture the time dependence of business vitality. The spatial correlation modeling unit is used to process spatial correlation features to capture the spatial correlation between business districts. The correlation fusion layer introduces a correlation coefficient matrix between emotion features and business vitality indicators to dynamically adjust the weight distribution of emotion features in spatiotemporal modeling. Using the time series features and spatial correlation features as inputs, and the actual business vitality index of the corresponding time period as a reference benchmark, the spatiotemporal prediction network that integrates sentiment-vitality correlation is trained to obtain a trained spatiotemporal sequence prediction model. The spatiotemporal sequence prediction model is used to output the business vitality prediction results within a future preset time period.

4. The method for spatiotemporal prediction and dynamic evaluation of business vitality based on multimodal feature fusion as described in claim 3, characterized in that, The step of inputting the core feature vector of business vitality into a pre-constructed spatiotemporal sequence prediction model and outputting the prediction result of business vitality within a preset future time period through the spatiotemporal sequence prediction model specifically includes the following steps: Real-time spatiotemporal feature extraction is performed on the core feature vector of business vitality to obtain the current time series features and the current spatial correlation features. The current time series features are used to characterize the time trend features of current business vitality, and the current spatial correlation features are used to characterize the real-time spatial interaction features between different business districts. The correlation coefficient matrix of the current time series features, current spatial correlation features, and sentiment features with the business vitality index is input into the trained spatiotemporal sequence prediction model. The current time series features are input into the time series modeling unit, the current spatial correlation features are input into the spatial correlation modeling unit, and the correlation coefficient matrix is ​​input into the correlation fusion layer. The temporal modeling unit is used to process the current time series features to capture real-time time trends, the spatial correlation modeling unit is used to process the current spatial correlation features to capture real-time spatial interactions, and the correlation fusion layer is used to dynamically adjust the influence weight of emotional features in real-time time trends and real-time spatial interactions based on the correlation coefficient matrix to obtain spatiotemporal features that fuse emotion-vitality correlation. Based on the spatiotemporal characteristics of the fusion of emotion and vitality, the trained spatiotemporal sequence prediction model outputs the prediction results of business vitality within a preset future time period.

5. A spatiotemporal prediction and dynamic evaluation system for business vitality based on multimodal feature fusion, used to implement the steps of the spatiotemporal prediction and dynamic evaluation method for business vitality based on multimodal feature fusion as described in any one of claims 1-4, characterized in that, include: The data processing module is used to collect multimodal heterogeneous data including geospatial modality, text semantic modality, real-time behavioral modality and economic attribute modality, and to perform data cleaning and desensitization processing on the multimodal heterogeneous data to obtain a structured multimodal dataset. The association calculation module is used to extract features from each single modality data in the structured multimodal dataset based on the corresponding feature extraction algorithm, obtain each single modality feature vector, perform feature fusion on each single modality feature vector through a cross-modal attention mechanism, and calculate the core feature vector of business vitality based on the pre-built emotion-vitality association model. The result prediction module is used to input the core feature vector of business vitality into a pre-constructed spatiotemporal sequence prediction model, and output the prediction result of business vitality within a future preset time period through the spatiotemporal sequence prediction model. The dynamic evaluation module is used to acquire real-time updated data and business vitality prediction results. Based on the real-time updated data and business vitality prediction results, a quantitative and qualitative dual-dimensional evaluation index system is constructed. The evaluation index system is calculated in real time through a stream processing framework to obtain dynamic evaluation results. The results output module is used to display the dynamic evaluation results through a visual interactive panel and output a structured evaluation report containing the business vitality prediction results and the dynamic evaluation results.

6. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the spatiotemporal prediction and dynamic evaluation method for business vitality based on multimodal feature fusion as described in any one of claims 1-4.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the spatiotemporal prediction and dynamic evaluation method for business vitality based on multimodal feature fusion as described in any one of claims 1-4.

Images