Commercial vitality prediction and business district evaluation method based on multi-modal feature fusion

Abstract

The invention provides a business vitality prediction and business district evaluation method based on multi-modal feature fusion, and relates to the technical field of spatial data analysis, and the method comprises the steps: obtaining business POI data and time sequence trajectory data of a target area, extracting the score and emotion features of user comments, and calculating the pedestrian flow distribution based on the time sequence trajectory data; then inputting the features into a recurrent neural network to predict a commercial vitality time sequence, constructing POI data into a knowledge graph, and constructing and optimizing a network reachability matrix in combination with position information and human traffic; and finally, dividing a target region by using a Louvain algorithm, fusing multi-dimensional features by edge computing nodes of each sub-region, dynamically adjusting the weight of an evaluation index through reinforcement learning, and outputting a commercial service evaluation result. By fusing multi-source data and a deep learning method, the life circle business service quality can be evaluated more accurately and comprehensively, and decision support is provided for urban planning and business development.

Description

Technical Field

[0001] The present invention relates to spatial data analysis technology, and in particular to a commercial vitality prediction and business district evaluation method based on multimodal feature fusion. Background Art

[0002] With the acceleration of urbanization, the diversity and spatial distribution of commercial services have an increasingly significant impact on the quality of life of residents. The acquisition and analysis of commercial POI (points of interest) data has become an important means of studying commercial services in the living circle. Existing technologies mainly rely on static data analysis, lacking real-time monitoring and analysis of dynamic human traffic and user behavior, resulting in insufficient accuracy and timeliness of commercial service evaluation. Defects and deficiencies in existing technologies: First, existing technologies often rely on manual review in the classification and labeling of commercial POI data, which is inefficient and prone to subjective bias, affecting the accuracy of the data. Secondly, the existing methods are mostly static in their analysis of human traffic distribution characteristics, failing to effectively capture the dynamic movement trajectory of the crowd, resulting in inaccurate predictions of commercial vitality. Finally, existing technologies lack flexibility in setting the weights of evaluation indicators, and cannot be dynamically adjusted according to real-time data changes, resulting in lags and inadaptability in the evaluation results. Summary of the invention

[0003] The embodiments of the present invention provide a commercial vitality prediction and business district evaluation method based on multimodal feature fusion, which can solve the problems in the prior art.

[0004] According to a first aspect of the embodiments of the present invention, Provides a commercial vitality prediction and business district evaluation method based on multimodal feature fusion, including: Acquire commercial POI data and time series trajectory data of the target area, classify and label the commercial POI data by business type, extract the rating features and sentiment polarity features of user comments, and calculate the flow distribution features based on the time series trajectory data; The scoring features, sentiment polarity features and traffic flow distribution features are input into a recursive neural network to generate a time series of commercial vitality prediction for each region; the commercial POI data is constructed as a knowledge graph node, service complementarity is calculated based on the scoring features and sentiment polarity features, a network accessibility matrix is ​​constructed in combination with location information and traffic flow distribution features, and the network accessibility matrix is ​​optimized through service complementarity to obtain an enhanced network accessibility matrix; The enhanced network reachability matrix is ​​input into the Louvain community discovery algorithm to divide the target area into sub-areas, and edge computing nodes are deployed in each sub-area; each edge computing node inputs the local enhanced network reachability matrix and the business vitality prediction time series into a comparative learning network to generate local business features, and integrates them with the scoring features and sentiment polarity features to obtain comprehensive business features; an evaluation weight adaptive layer is constructed, the comprehensive business features are used as state input, the evaluation index weights are dynamically adjusted through reinforcement learning, and the business service evaluation results are output.

[0005] In an optional embodiment, The step of inputting the scoring feature, the sentiment polarity feature and the traffic flow distribution feature into a recursive neural network to generate a commercial vitality prediction time series for each region includes: The feature interaction matrix is ​​constructed, and the time series features are generated through the bidirectional gated recurrent unit network. The multi-scale decomposition strategy is adopted to predict the trend term, cycle term and residual term respectively through linear projection, Fourier decomposition and multi-layer perceptron. Finally, adaptive fusion is performed based on the reliability score to obtain the business vitality prediction time series, which includes: The scoring features are enhanced to obtain scoring statistical features, including mean, standard deviation, skewness, kurtosis and scoring change rate; the sentiment intensity is calculated based on word frequency and position weight to obtain enhanced sentiment features, including positive sentiment intensity, negative sentiment intensity and sentiment distribution entropy; the time pattern is extracted based on the 24-hour passenger flow distribution characteristics of the target area to obtain passenger flow pattern features, including peak values ​​in the morning, noon and evening periods, valley values ​​in the morning, noon and evening periods, weekday passenger flow ratio, weekend passenger flow ratio and periodicity score; The score statistical features, enhanced emotional features and crowd flow pattern features are used to construct a feature interaction matrix, wherein the feature interaction matrix includes independent representations of the three types of features and Hadamard product interaction representations between two features; the feature interaction matrix is ​​input into a bidirectional gated recurrent unit network with residual connections to generate bidirectional time series features including a forward feature sequence and a reverse feature sequence; Perform multi-scale convolution on the bidirectional time series features to obtain daily scale features, weekly scale features and monthly scale features; perform linear projection based on the monthly scale features to obtain trend item prediction results, perform Fourier decomposition based on the weekly scale features to obtain period item prediction results, input the daily scale features, weekly scale features and monthly scale features into a multi-layer perceptron to obtain residual item prediction results; calculate the fluctuation amplitude and prediction error of the trend item prediction result, the period item prediction result and the residual item prediction result to obtain a reliability score, generate a fusion weight coefficient based on the reliability score, and use the fusion weight coefficient to perform weighted summation on the trend item prediction result, the period item prediction result and the residual item prediction result to obtain a business vitality prediction time series.

[0006] In an optional embodiment, The steps of constructing the commercial POI data as knowledge graph nodes, calculating service complementarity based on rating features and sentiment polarity features, constructing a network accessibility matrix in combination with location information and pedestrian flow distribution features, and optimizing the network accessibility matrix through service complementarity to obtain an enhanced network accessibility matrix include: The commercial POI data is constructed into basic attributes including business type, business area, and business hours, scoring features including taste score, environment score, service score, and cost-effectiveness score, emotional polarity features including positive emotional intensity, negative emotional intensity, and neutral emotional intensity, and knowledge graph nodes including temporal attributes such as average daily order volume, weekend passenger flow, and weekday passenger flow. The knowledge graph nodes are connected to establish spatial relationships based on geographical distance, complementary competitive relationships based on business type, and overlapping relationships based on business hours; The scoring features are weighted based on the scoring dimension weights to obtain scoring complementarity, the emotional complementarity is calculated based on the vector cosine similarity of the emotional polarity features and the difference between the emotional intensity, the time complementarity is calculated based on the difference in business time distribution and the degree of peak passenger flow staggering, and the scoring complementarity, the emotional complementarity and the time complementarity are weightedly integrated to obtain service complementarity; The crowd flow distribution characteristics between adjacent POI nodes are decomposed in time series to obtain periodic crowd flow fluctuation patterns and crowd flow change characteristics. The trajectory stability is calculated by combining the directional entropy and duration of the crowd movement trajectory. The initial network accessibility matrix is ​​constructed based on the overlap degree of the crowd flow fluctuation patterns, the correlation of the crowd flow change characteristics and the trajectory stability. The service complementarity is multiplied by the accessibility strength of the network accessibility matrix, and the result is symmetrically normalized to obtain the enhanced network accessibility matrix.

[0007] In an optional embodiment, The steps of each edge computing node inputting the local enhanced network reachability matrix and the business vitality prediction time series into a comparative learning network to generate local business features, and fusing them with the scoring features and the sentiment polarity features to obtain comprehensive business features include: Through localized processing with spatial constraints and dual-attenuation spatiotemporal data evaluation, a dual-branch contrastive learning network is constructed: the matrix branch uses a graph convolutional network to extract spatial features, and the time branch uses a gated recurrent unit to extract temporal features. Combined with the attention mechanism and multimodal feature interaction, comprehensive business features are obtained, including: The global enhanced network accessibility matrix is ​​localized according to the community discovery results, and the local matrix corresponding to the sub-region is extracted through a selection matrix with spatial constraints. The adaptive weight coefficient is set based on the regional overlap, and the boundary of the local matrix is ​​smoothed. The commercial vitality prediction time series in the sub-region is collected to construct a multi-scale time series sliding window, and the local training data is evaluated and screened using a spatiotemporal data freshness evaluation mechanism based on double decay. A dual-branch contrastive learning network with feature enhancement is constructed, wherein the matrix branch uses a multi-level graph convolutional network to process the local matrix, and extracts hierarchical spatial features through residual connections and attention pooling; the time series branch uses a bidirectional gated recurrent unit to process the business vitality prediction time series, and combines an adaptive frequency domain decomposition module to extract multi-scale time features; the contrastive learning network is optimized by a hierarchical contrast loss function that integrates spatial topology constraints and time series consistency constraints; The dynamic fusion weights of the spatial features and the temporal features are calculated based on the attention mechanism, and the cross-scale feature aggregation module is used to perform multi-level feature fusion to obtain the initial local business features; the momentum update mechanism of the adaptive temperature coefficient is introduced to perform online update on the initial local business features to obtain the local business features; The local commercial features, rating features and sentiment polarity features are mapped to a unified feature space through nonlinear projection to construct a multimodal feature interaction matrix. A hierarchical multi-head attention mechanism is used to model the local and global dependencies between modalities, and the feature distribution is adjusted in combination with a feature recalibration module. The initial comprehensive commercial features are obtained through a residual connection with channel recalibration, and the final comprehensive commercial features are obtained based on an adaptive weight allocation mechanism based on modality importance.

[0008] In an optional embodiment, The steps of constructing an evaluation weight adaptive layer, taking comprehensive business features as state input, dynamically adjusting the evaluation index weights through reinforcement learning, and outputting the business service evaluation results include: Constructing an evaluation state space, combining comprehensive business characteristics, evaluation index weight vectors, differential characteristics of business vitality prediction time series, and business format collaborative development index to form a state vector, wherein the business format collaborative development index is calculated based on the accessibility strength of service complementarity and network accessibility matrix; Designing a multi-level reward function based on the state vector, including using the Huber loss smooth error metric to calculate the evaluation deviation, combining the relative entropy and the mixed norm constraint weight change penalty term to construct an immediate reward function, and the weight change penalty term is adjusted through gradient adaptive adjustment; Based on the time series change characteristics of the business format collaborative development index, a time series reward function is constructed in combination with the dynamic change of the business format distribution entropy; A residual network structure is used to generate reward combination weights, the state vector and the action vector are input into the residual block for feature extraction and mapping, the reward weight is obtained through a softmax function, the instantaneous reward function and the time series reward function are weightedly combined to obtain a total reward function, and a smoothing factor based on adaptive adjustment of signal variance is used for time series smoothing; Construct a dual network structure, in which the value network evaluates the state-action value through the state value function and the advantage function, and the strategy network generates the adjustment amount of the evaluation index weight through the mean network and the variance network; optimize the value network and the strategy network parameters simultaneously based on the temporal difference error to obtain the optimized evaluation index weight vector; The optimized evaluation index weight vector is applied to the comprehensive business characteristics to obtain the business service evaluation results, which include business service quality score, business format development balance index and regional coordinated development level.

[0009] In an optional embodiment, Designing a multi-level reward function based on the state vector includes calculating the evaluation deviation using the Huber loss smooth error metric, constructing an immediate reward function in combination with the weight change penalty term constrained by relative entropy and mixed norm, and the trade-off factor of the weight change penalty term is adaptively adjusted by gradient; Based on the time series change characteristics of the business format collaborative development index, the steps of constructing a time series reward function in combination with the dynamic change of business format distribution entropy include: Construct a multi-dimensional evaluation state vector for the state vector, including the evaluation score and its derivative, weight gradient and environmental characteristics; calculate the evaluation distribution and time series deviation based on wavelet transform; construct an instant reward function including smooth error measurement, weight change penalty and uncertainty measurement; construct a time series reward function based on the time series changes of the business format collaborative development index and distribution entropy, and optimize the trade-off factor through the policy gradient, including: The state vector is expanded to construct a multidimensional evaluation state vector, wherein the multidimensional evaluation state vector includes the evaluation score, evaluation score gradient, evaluation score second-order derivative, evaluation index weight vector gradient and environmental characteristics of the comprehensive business characteristics, wherein the environmental characteristics include time periodicity information and evaluation scene type information; based on the evaluation score, a wavelet transform is used to perform multi-scale decomposition to obtain a trend component and a detail component, and the evaluation distribution is calculated based on the local adaptive kernel density estimation; the trend component and the detail component are used to respectively calculate the time series difference to obtain the trend deviation and the detail deviation, and an adaptive threshold is generated by a multi-layer perceptron; Based on the evaluation index weight vector combined with relative entropy constraints and mixed norm constraints, a weight structure regularization term is constructed; based on the evaluation distribution, an uncertainty measure is calculated; the evaluation index weight vector is adaptively adjusted; based on the Huber loss function, a smooth error measure of the trend deviation and detail deviation is calculated, and a weight change penalty term is constructed, and the weight factor of the weight change penalty term is adaptively adjusted by the gradient; the smooth error measure, the weight change penalty term, and the uncertainty measure are combined to construct an instant reward function; Based on the business format collaborative development index, its time difference and second-order derivative are calculated, and a time series reward function is constructed in combination with the time series difference of the business format distribution entropy. The trade-off factor of the time series reward function is optimized by the policy gradient method.

[0010] In an optional embodiment, A dual network structure is constructed, in which the value network evaluates the state-action value through the state value function and the advantage function, and the strategy network generates the adjustment amount of the evaluation index weight through the mean network and the variance network; the value network and the strategy network parameters are optimized simultaneously based on the temporal difference error, and the steps of obtaining the optimized evaluation index weight vector include: A multi-granularity state encoder is constructed, and an adaptive threshold is designed based on the temporal change rate of the evaluation index. The evaluation index sequence is divided into feature segments of different time scales. The statistical moment and fluctuation characteristics of each feature segment are calculated to construct a multi-scale feature pyramid. The hierarchical attention mechanism is used for feature fusion to obtain a hierarchical state representation. A dual network structure is constructed based on the hierarchical state representation, wherein the value network is composed of a state value function network and an advantage function network, wherein the state value function network evaluates the baseline value of the current state, and the advantage function network evaluates the value improvement of the action relative to the average level, and the state-action value is obtained by combining the state value function network and the advantage function network; a strategy network is constructed based on the state-action value, wherein the strategy network includes a mean network and a variance network, wherein the mean network outputs the expected direction of the evaluation index weight adjustment, and the variance network represents the uncertainty distribution of the weight adjustment, and the adjustment amount of the evaluation index weight is generated by a reparameterization method; Constructing a temporal difference error, taking the difference between the predicted value of the state value function network and the target value calculated based on the advantage function network as the optimization target of the value network; constructing a generalized advantage estimate based on the temporal difference error, calculating the policy gradient of the evaluation index weight adjustment amount in combination with the outputs of the mean network and the variance network, and optimizing the parameters of the value network and the policy network at the same time; Constructing a weight constraint based on the Wasserstein metric, including: constructing a support set for weight adjustment based on the dynamic distribution characteristics of the weight adjustment sequence, calculating the Wasserstein distance between the current weight adjustment amount and the support set, and projecting the adjustment amount of the evaluation index weight to a feasible domain that satisfies the Wasserstein distance constraint through iterative projection; updating the evaluation index weight vector based on the weight constraint, and dynamically adjusting the evaluation system through alternating optimization of the value network and the strategy network.

[0011] According to a second aspect of the embodiments of the present invention, Provides a business vitality prediction and business district evaluation system based on multi-modal feature fusion, including: The first unit is used to obtain commercial POI data and time series trajectory data of the target area, classify and label the commercial POI data, extract the rating features and sentiment polarity features of user comments, and calculate the distribution features of pedestrian flow based on the time series trajectory data; The second unit is used to input the scoring features, sentiment polarity features and traffic flow distribution features into a recursive neural network to generate a time series of commercial vitality prediction for each region; construct the commercial POI data into knowledge graph nodes, calculate service complementarity based on the scoring features and sentiment polarity features, construct a network accessibility matrix based on location information and traffic flow distribution features, and optimize the network accessibility matrix through service complementarity to obtain an enhanced network accessibility matrix; The third unit is used to input the enhanced network reachability matrix into the Louvain community discovery algorithm to divide the target area into sub-areas, and deploy edge computing nodes in each sub-area; each edge computing node inputs the local enhanced network reachability matrix and the business vitality prediction time series into a comparative learning network to generate local business features, and integrates them with the scoring features and sentiment polarity features to obtain comprehensive business features; construct an evaluation weight adaptive layer, take the comprehensive business features as state input, dynamically adjust the evaluation index weights through reinforcement learning, and output the business service evaluation results.

[0012] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0013] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0014] The present invention can accurately evaluate the commercial service quality of the target area and improve the accuracy of commercial vitality prediction through comprehensive analysis of commercial POI data and time series trajectory data. By dynamically adjusting the network edge weights and optimizing the accessibility matrix, the service complementarity between commercial POIs is enhanced, the coordinated development of business formats is promoted, and the overall efficiency of commercial services is improved. By using edge computing nodes and comparative learning networks, local business features can be generated in real time, and dynamic adjustments can be made in combination with comprehensive business features to ensure the adaptability and accuracy of evaluation indicators. The present invention uses a comparative learning network to generate local business features, and dynamically adjusts the evaluation index weights through reinforcement learning, so that the commercial service evaluation can be dynamically adjusted according to the actual situation, which improves the dynamic and adaptability of the evaluation and is closer to the actual development and changes of commercial services. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flowchart of a method for predicting business vitality and evaluating business district based on multimodal feature fusion according to an embodiment of the present invention; Figure 2 This is a comparison chart of the prediction error distribution; Figure 3 This is a performance comparison chart for tracking the business format collaborative development index. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] Combine the following Figure 1-Figure 3 The technical solution of the present invention is described in detail with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0018] Figure 1 FIG. 1 is a flow chart of a method for predicting business vitality and evaluating business districts based on multimodal feature fusion according to an embodiment of the present invention. Figure 1 As shown, the method includes: Acquire commercial POI data and time series trajectory data of the target area, classify and label the commercial POI data by business type, extract the rating features and sentiment polarity features of user comments, and calculate the flow distribution features based on the time series trajectory data; The scoring features, sentiment polarity features and traffic flow distribution features are input into a recursive neural network to generate a time series of commercial vitality prediction for each region; the commercial POI data is constructed as a knowledge graph node, service complementarity is calculated based on the scoring features and sentiment polarity features, a network accessibility matrix is ​​constructed in combination with location information and traffic flow distribution features, and the network accessibility matrix is ​​optimized through service complementarity to obtain an enhanced network accessibility matrix; The enhanced network reachability matrix is ​​input into the Louvain community discovery algorithm to divide the target area into sub-areas, and edge computing nodes are deployed in each sub-area; each edge computing node inputs the local enhanced network reachability matrix and the business vitality prediction time series into a comparative learning network to generate local business features, and integrates them with the scoring features and sentiment polarity features to obtain comprehensive business features; an evaluation weight adaptive layer is constructed, the comprehensive business features are used as state input, the evaluation index weights are dynamically adjusted through reinforcement learning, and the business service evaluation results are output.

[0019] Exemplarily, the technical implementation of the commercial service evaluation method first requires data acquisition and preprocessing. Obtain commercial POI data in the target area, including catering, retail, leisure and entertainment and other business types. Each POI contains basic information such as name, address, longitude and latitude. Obtain time series trajectory data from mobile operators, including user ID, timestamp, longitude and latitude information. Classify and label commercial POIs by business type, using a hierarchical classification system, such as subdividing catering into Chinese food, Western food, fast food and other subcategories. Crawl user comments from platforms such as Dianping, extract rating features in four dimensions: taste, environment, service, and cost-effectiveness, and use sentiment analysis models to calculate the intensity values ​​of three types of sentiment polarity features: positive, negative, and neutral. Based on the time series trajectory data, statistics are made on the average daily, weekend, and weekday traffic distribution characteristics of each POI.

[0020] In order to achieve accurate prediction of business vitality, a recursive neural network model is constructed using the long short-term memory network LSTM. The input layer contains scoring features, sentiment polarity features, and traffic distribution features, with a feature dimension of 11. The number of LSTM units is set to 128, and the number of hidden layers is 2. The training sample is constructed by sliding the time window, with the window size set to 7 days and the step size set to 1 day. The model outputs the predicted value of business vitality in the next 3 days.

[0021] In terms of knowledge graph construction, commercial POI data is constructed as nodes, and node attributes include basic attributes such as business type, business area, business hours, scoring characteristics, emotional polarity characteristics, and temporal attributes of traffic flow distribution. The relationship between nodes includes the spatial relationship of geographical distance, the complementary competitive relationship of business types, and the overlapping relationship of business hours. When calculating service complementarity, the scoring complementarity is calculated based on the weighted weights of the four scoring dimensions, the emotional complementarity is calculated by the cosine similarity and intensity difference of the emotional vector, and the time complementarity considers the differences in business hours distribution and the degree of passenger flow staggering. On this basis, a network accessibility matrix is ​​constructed, combining the physical distance between nodes and the characteristics of traffic flow distribution, and an enhanced network accessibility matrix is ​​obtained through optimization through service complementarity.

[0022] The Louvain algorithm is used for community discovery in regional division, and the optimal number of sub-regions is determined by modularity optimization. For example, the region is divided into 12 sub-regions, and edge computing nodes are deployed in each sub-region. Each node uses a contrastive learning network to extract local business features. The network consists of two parts: an encoder and a contrastive loss. The encoder uses a three-layer graph convolutional network, and the input is the enhanced network accessibility matrix of the sub-region and the business vitality prediction time series. By maximizing the similarity of sample representations in the same sub-region and minimizing the similarity of sample representations in different sub-regions, local business features are learned. The local business features are weightedly fused with the rating features and sentiment polarity features to obtain comprehensive business features.

[0023] The evaluation weight adaptive layer is dynamically adjusted based on reinforcement learning. Comprehensive business features are used as state input, and a multi-level reward function is used to guide weight adjustment. The immediate reward function includes three parts: smooth error measurement, weight change penalty, and uncertainty measurement. The time series reward function is constructed based on the dynamic changes of the business format collaborative development index and distribution entropy. The weight adjustment strategy is updated through the Actor-Critic architecture to achieve adaptive optimization of the evaluation index weight.

[0024] The present invention depicts the complex relationship between commercial POIs through knowledge graphs and integrates multi-dimensional evaluation features to improve evaluation accuracy; the edge computing architecture realizes distributed feature extraction and reduces computational complexity; reinforcement learning ensures that the evaluation index weights dynamically adapt to environmental changes and improves the rationality and timeliness of the evaluation results.

[0025] In an optional embodiment, The step of inputting the scoring feature, the sentiment polarity feature and the traffic flow distribution feature into a recursive neural network to generate a commercial vitality prediction time series for each region includes: The feature interaction matrix is ​​constructed, and the time series features are generated through the bidirectional gated recurrent unit network. The multi-scale decomposition strategy is adopted to predict the trend term, cycle term and residual term respectively through linear projection, Fourier decomposition and multi-layer perceptron. Finally, adaptive fusion is performed based on the reliability score to obtain the business vitality prediction time series, which includes: The scoring features are enhanced to obtain scoring statistical features, including mean, standard deviation, skewness, kurtosis and scoring change rate; the sentiment intensity is calculated based on word frequency and position weight to obtain enhanced sentiment features, including positive sentiment intensity, negative sentiment intensity and sentiment distribution entropy; the time pattern is extracted based on the 24-hour passenger flow distribution characteristics of the target area to obtain passenger flow pattern features, including peak values ​​in the morning, noon and evening periods, valley values ​​in the morning, noon and evening periods, weekday passenger flow ratio, weekend passenger flow ratio and periodicity score; The score statistical features, enhanced emotional features and crowd flow pattern features are used to construct a feature interaction matrix, wherein the feature interaction matrix includes independent representations of the three types of features and Hadamard product interaction representations between two features; the feature interaction matrix is ​​input into a bidirectional gated recurrent unit network with residual connections to generate bidirectional time series features including a forward feature sequence and a reverse feature sequence; Perform multi-scale convolution on the bidirectional time series features to obtain daily scale features, weekly scale features and monthly scale features; perform linear projection based on the monthly scale features to obtain trend item prediction results, perform Fourier decomposition based on the weekly scale features to obtain period item prediction results, input the daily scale features, weekly scale features and monthly scale features into a multi-layer perceptron to obtain residual item prediction results; calculate the fluctuation amplitude and prediction error of the trend item prediction result, the period item prediction result and the residual item prediction result to obtain a reliability score, generate a fusion weight coefficient based on the reliability score, and use the fusion weight coefficient to perform weighted summation on the trend item prediction result, the period item prediction result and the residual item prediction result to obtain a business vitality prediction time series.

[0026] For example, assume that the target area is a commercial district in a city, and it is necessary to predict the commercial activity in the next week.

[0027] Data preparation: First, collect data from the commercial district over the past three months, including user online ratings, user review texts, and pedestrian flow data collected based on decrypted mobile phone signaling data or other sensors.

[0028] Enhanced rating features: Statistical analysis is performed on the collected user rating data to calculate the mean, standard deviation, skewness, kurtosis, and rating change rate. For example, if the rating data for a certain day is [4, 5, 4, 3, 5, 5, 4], the mean is 4.29, the standard deviation is 0.87, the skewness is -0.59, and the kurtosis is -1.06. The rating change rate can be calculated by subtracting the mean rating of yesterday from the mean rating of today.

[0029] Enhance sentiment features: Perform natural language processing on user comment texts. First, construct a sentiment dictionary containing positive and negative sentiment words. Then, calculate the frequency of positive and negative sentiment words in each comment, and assign weights based on the position of the words in the sentence, and finally obtain the positive and negative sentiment intensity of each comment. For example, in the sentence "This store has good product quality and good service attitude", "good product quality" and "good service attitude" are both positive sentiment expressions. After assigning certain weights, calculate the positive sentiment intensity. Finally, calculate the sentiment distribution entropy to measure the distribution of sentiment.

[0030] Crowd flow pattern feature extraction: Analyze the 24-hour crowd flow distribution data and extract time pattern features. For example, a day can be divided into three periods: morning, noon, and evening, and the peak and valley values ​​of the crowd flow in each period are calculated. At the same time, the proportion of weekday crowd flow and weekend crowd flow, as well as the periodic score of crowd flow, are calculated to measure the periodic change pattern of crowd flow.

[0031] Feature interaction matrix construction: The feature interaction matrix is ​​constructed by combining the score statistical features, enhanced sentiment features, and traffic pattern features. This matrix contains independent representations of the three types of features and Hadamard product interaction representations between each pair of features. The Hadamard product refers to the multiplication of the elements at corresponding positions of two matrices. In this way, the interaction relationship between different features can be captured.

[0032] Bidirectional temporal feature generation: The feature interaction matrix is ​​input into a bidirectional gated recurrent unit network with residual connections. Residual connections can effectively alleviate the gradient vanishing problem and improve the model training effect. The bidirectional gated recurrent unit network can capture both forward and reverse temporal information and generate bidirectional temporal features containing forward feature sequences and reverse feature sequences.

[0033] Multi-scale feature extraction: Multi-scale convolution is performed on the bidirectional time series features to obtain daily scale features, weekly scale features, and monthly scale features. Multi-scale convolution can capture feature information of different time granularities.

[0034] Business vitality forecast result generation: linear projection is performed based on monthly scale features to obtain trend item forecast results, Fourier decomposition is performed based on weekly scale features to obtain period item forecast results, and daily scale features, weekly scale features, and monthly scale features are input into a multi-layer perceptron to obtain residual item forecast results. The fluctuation range and forecast error of trend items, period items, and residual items are calculated to obtain reliability scores. Fusion weight coefficients are generated based on the reliability scores, and the fusion weight coefficients are used to perform weighted summation on trend items, period items, and residual items to obtain the final business vitality forecast time series.

[0035] Figure 2 is a comparison chart of prediction error distribution. Figure 2 The error distribution of different prediction methods is shown. The horizontal axis represents the prediction error range, and the vertical axis represents the frequency of each error interval. The black solid line column represents the method of the present invention, the gray dashed line column represents the Transformer method, and the black dotted line column represents the LSTM method. The gray shaded area represents the ±5% error interval. 73% of the data points of the method of the present invention fall within the ±5% error interval, while only 53% and 47% of the comparison methods fall within this interval, which significantly proves the higher stability of the prediction results of the present method. The present invention can more accurately capture the inherent laws of business activities through multi-feature fusion and deep learning models, thereby improving the accuracy of business vitality prediction; the present invention takes into account multiple factors such as ratings, emotions and traffic, and captures the interaction relationship between different features through the feature interaction matrix, so that the prediction results are more interpretable, and can process data in real time and generate prediction results to meet the timeliness requirements of business activities.

[0036] In an optional embodiment, The steps of constructing the commercial POI data as knowledge graph nodes, calculating service complementarity based on rating features and sentiment polarity features, constructing a network accessibility matrix in combination with location information and pedestrian flow distribution features, and optimizing the network accessibility matrix through service complementarity to obtain an enhanced network accessibility matrix include: The commercial POI data is constructed into basic attributes including business type, business area, and business hours, scoring features including taste score, environment score, service score, and cost-effectiveness score, emotional polarity features including positive emotional intensity, negative emotional intensity, and neutral emotional intensity, and knowledge graph nodes including temporal attributes such as average daily order volume, weekend passenger flow, and weekday passenger flow. The knowledge graph nodes are connected to establish spatial relationships based on geographical distance, complementary competitive relationships based on business type, and overlapping relationships based on business hours; The scoring features are weighted based on the scoring dimension weights to obtain scoring complementarity, the emotional complementarity is calculated based on the vector cosine similarity of the emotional polarity features and the difference between the emotional intensity, the time complementarity is calculated based on the difference in business time distribution and the degree of peak passenger flow staggering, and the scoring complementarity, the emotional complementarity and the time complementarity are weightedly integrated to obtain service complementarity; The crowd flow distribution characteristics between adjacent POI nodes are decomposed in time series to obtain periodic crowd flow fluctuation patterns and crowd flow change characteristics. The trajectory stability is calculated by combining the directional entropy and duration of the crowd movement trajectory. The initial network accessibility matrix is ​​constructed based on the overlap degree of the crowd flow fluctuation patterns, the correlation of the crowd flow change characteristics and the trajectory stability. The service complementarity is multiplied by the accessibility strength of the network accessibility matrix, and the result is symmetrically normalized to obtain the enhanced network accessibility matrix.

[0037] Exemplarily, first, construct a knowledge graph of commercial POIs. Collect commercial POI data in the target area, including basic attributes such as the name, business type (such as catering, shopping, entertainment, etc.), business area, business hours, and rating features such as taste rating, environment rating, service rating, and cost-effectiveness rating. At the same time, collect user comment data on each POI, and use sentiment analysis technology to extract sentiment polarity features such as positive sentiment intensity, negative sentiment intensity, and neutral sentiment intensity. In addition, it is also necessary to collect time series attributes such as the average daily order volume, weekend passenger flow, and weekday passenger flow of each POI.

[0038] Each POI is represented as a node in the knowledge graph, and the node contains all the above attributes and features. Multiple relationships are established between nodes. For example, spatial relationships are established based on the geographical distance between POIs. The closer the distance, the stronger the spatial relationship. Complementary and competitive relationships are established based on the business types of POIs. For example, a coffee shop and a dessert shop are complementary, while two coffee shops are competitive. Overlapping relationships are established based on the business hours of POIs. The higher the overlap of business hours, the stronger the overlapping relationship.

[0039] Calculate service complementarity. First, determine the weights of the scoring dimensions based on expert experience or data analysis, such as a weight of 0.4 for taste, 0.3 for environment, 0.2 for service, and 0.1 for cost-effectiveness. Then, perform weighted calculations on the scoring features of each POI to obtain a weighted score. For example, a coffee shop has a taste score of 4.5, an environment score of 4.0, a service score of 3.5, and a cost-effectiveness score of 4.0. Its weighted score is 4.5×0.4+4.0×0.3+3.5×0.2+4.0×0.1=4.1. The score complementarity of two POIs can be measured by calculating the absolute value of the difference in their weighted scores. The smaller the difference, the higher the complementarity.

[0040] Calculate emotional complementarity using emotional polarity features. Represent the emotional polarity features of each POI as a three-dimensional vector, such as (positive emotional intensity, negative emotional intensity, neutral emotional intensity). The emotional complementarity of two POIs can be measured by calculating the cosine similarity and emotional intensity difference between their emotional vectors. The higher the cosine similarity and the smaller the emotional intensity difference, the higher the emotional complementarity.

[0041] In addition, time complementarity is calculated based on the difference in business time distribution and the degree of peak shift in customer flow. The greater the difference in business time distribution and the higher the degree of peak shift in customer flow, the higher the time complementarity. For example, if a coffee shop's main business hours are morning and noon, and a bar's main business hours are evening, their time complementarity is high.

[0042] The score complementarity, emotion complementarity and time complementarity are weighted and integrated to obtain service complementarity. For example, assuming that the weight of score complementarity is 0.5, the weight of emotion complementarity is 0.3, and the weight of time complementarity is 0.2, the service complementarity of two POIs can be obtained by weighted sum of the three.

[0043] Construct an enhanced network accessibility matrix. First, perform time series decomposition on the distribution characteristics of the flow of people between adjacent POI nodes to obtain the periodic flow fluctuation pattern and flow change characteristics. For example, the flow of people data for each day of the week can be decomposed to obtain the daily flow fluctuation pattern and flow change characteristics.

[0044] The trajectory stability is calculated by combining the directional entropy and duration of the crowd movement trajectory. The lower the directional entropy and the longer the duration, the higher the trajectory stability. For example, if the movement direction of people between two POIs is relatively concentrated and the duration is long, the trajectory stability is high.

[0045] The initial network accessibility matrix is ​​constructed based on the degree of overlap of crowd flow fluctuation patterns, the correlation of crowd flow change characteristics, and the stability of trajectories. The higher the degree of overlap of crowd flow fluctuation patterns, the higher the correlation of crowd flow change characteristics, and the higher the trajectory stability, the higher the network accessibility.

[0046] The service complementarity and the accessibility strength of the initial network reachability matrix are multiplied, and the result is symmetrically normalized to obtain the enhanced network reachability matrix.

[0047] Suppose there are two POIs, coffee shop A and dessert shop B. The weighted score of coffee shop A is 4.1, the sentiment vector is (0.8, 0.1, 0.1), and the main business hours are morning and noon. The weighted score of dessert shop B is 4.2, the sentiment vector is (0.7, 0.2, 0.1), and the main business hours are afternoon and evening. Assume that the score complementarity weight is 0.5, the sentiment complementarity weight is 0.3, and the time complementarity weight is 0.2. The service complementarity between coffee shop A and dessert shop B is calculated to be at a high level. Assuming that the initial network accessibility of coffee shop A and dessert shop B is 0.6, the enhanced network accessibility is at a high level.

[0048] The present invention comprehensively considers the multi-dimensional characteristics of POI, including basic attributes, scoring characteristics, sentiment polarity characteristics, time series attributes and pedestrian flow distribution characteristics, and can more accurately characterize the association relationship between POIs and construct a more comprehensive POI knowledge graph; by calculating service complementarity, POIs with complementary relationships can be effectively identified, providing users with more personalized and accurate service recommendations, and improving user experience; the enhanced network accessibility matrix can reflect the accessibility between different POIs, provide data support for urban planning and commercial layout, and can identify potential business opportunities based on the enhanced network accessibility matrix, optimize traffic planning, etc.

[0049] In an optional embodiment, The steps of each edge computing node inputting the local enhanced network reachability matrix and the business vitality prediction time series into a comparative learning network to generate local business features, and fusing them with the scoring features and the sentiment polarity features to obtain comprehensive business features include: Through localized processing with spatial constraints and dual-attenuation spatiotemporal data evaluation, a dual-branch contrastive learning network is constructed: the matrix branch uses a graph convolutional network to extract spatial features, and the time branch uses a gated recurrent unit to extract temporal features. Combined with the attention mechanism and multimodal feature interaction, comprehensive business features are obtained, including: The global enhanced network accessibility matrix is ​​localized according to the community discovery results, and the local matrix corresponding to the sub-region is extracted through a selection matrix with spatial constraints. The adaptive weight coefficient is set based on the regional overlap, and the boundary of the local matrix is ​​smoothed. The commercial vitality prediction time series in the sub-region is collected to construct a multi-scale time series sliding window, and the local training data is evaluated and screened using a spatiotemporal data freshness evaluation mechanism based on double decay. A dual-branch contrastive learning network with feature enhancement is constructed, wherein the matrix branch uses a multi-level graph convolutional network to process the local matrix, and extracts hierarchical spatial features through residual connections and attention pooling; the time series branch uses a bidirectional gated recurrent unit to process the business vitality prediction time series, and combines an adaptive frequency domain decomposition module to extract multi-scale time features; the contrastive learning network is optimized by a hierarchical contrast loss function that integrates spatial topology constraints and time series consistency constraints; The dynamic fusion weights of the spatial features and the temporal features are calculated based on the attention mechanism, and the cross-scale feature aggregation module is used to perform multi-level feature fusion to obtain the initial local business features; the momentum update mechanism of the adaptive temperature coefficient is introduced to perform online update on the initial local business features to obtain the local business features; The local commercial features, rating features and sentiment polarity features are mapped to a unified feature space through nonlinear projection to construct a multimodal feature interaction matrix. A hierarchical multi-head attention mechanism is used to model the local and global dependencies between modalities, and the feature distribution is adjusted in combination with a feature recalibration module. The initial comprehensive commercial features are obtained through a residual connection with channel recalibration, and the final comprehensive commercial features are obtained based on an adaptive weight allocation mechanism based on modality importance.

[0050] Exemplarily, first, the global network accessibility matrix is ​​localized. Assuming that a city is divided into three regions A, B, and C, the global network accessibility matrix describes the convenience of transportation between any two points. In order to obtain the local matrix of region A, it is necessary to extract the connection information between the internal nodes of region A in the global matrix. If region A contains nodes 1, 2, and 3, the local matrix is ​​the submatrix corresponding to rows 1, 2, and 3 and columns 1, 2, and 3 in the global matrix. Considering that there may be overlaps between regions, for example, node 4 belongs to both regions A and B, an adaptive weight coefficient needs to be set. For example, if node 4 spends 70% of its stay time in region A and 30% of its stay time in region B, when constructing the local matrices of A and B, the connection weights associated with node 4 are multiplied by 0.7 and 0.3, respectively. In order to smooth the influence of regional boundaries, the local matrix is ​​subjected to boundary smoothing. For example, a Gaussian kernel is used to perform a convolution operation on the local matrix to make the connection information of the boundary nodes smoother.

[0051] Collect time series data for predicting commercial vitality in the sub-region. Assume that daily passenger flow data for region A is collected within one year to form a time series. In order to construct a multi-scale time series sliding window, the window size can be set to 7 days, 30 days, and 90 days, respectively, representing the changing trends of commercial vitality in a week, a month, and a quarter. In order to evaluate the quality of the data, a spatiotemporal data freshness evaluation mechanism based on double decay is adopted. For example, the closer the data is to the current time, the higher its weight; the closer the data is to the center of the region, the higher its weight. Filter out high-quality local training data based on the evaluation results. Assuming that the passenger flow data near the center of region A is more reliable, it is given a higher weight.

[0052] Construct a feature-enhanced dual-branch contrastive learning network. The matrix branch uses a multi-layer graph convolutional network to process the local matrix. For example, the first-layer graph convolutional network extracts the direct connection information of the nodes in the local matrix, the second-layer graph convolutional network extracts the second-order connection information of the nodes, and so on. Hierarchical spatial features are extracted through residual connections and attention pooling. For example, the attention mechanism can focus on nodes with higher connection weights in the local matrix, thereby extracting more important spatial features. The time series branch uses a bidirectional gated recurrent unit to process the business vitality prediction time series. For example, a bidirectional GRU network is used to capture the front-to-back dependencies of the time series. Combined with the adaptive frequency domain decomposition module, multi-scale time features are extracted. For example, the time series is decomposed into high-frequency and low-frequency parts, representing short-term fluctuations and long-term trends, respectively. The contrastive learning network is optimized by a hierarchical contrastive loss function that integrates spatial topology constraints and temporal consistency constraints. For example, the spatial topology constraint requires that spatially adjacent nodes have similar feature representations, and the temporal consistency constraint requires that temporally continuous moments have similar feature representations.

[0053] Fusion of spatial and temporal features. Use the attention mechanism to calculate the dynamic fusion weights of spatial and temporal features. For example, if the spatial features of a certain time period are more important, a higher weight is given to the spatial features. Use a cross-scale feature aggregation module to perform multi-level feature fusion to obtain the initial local business features. For example, spatial features and temporal features of different scales are spliced ​​to obtain a more comprehensive feature representation. Introduce a momentum update mechanism with an adaptive temperature coefficient to update the initial local business features online to obtain local business features. For example, dynamically adjust the temperature coefficient according to the distribution of training data to make contrastive learning more effective.

[0054] Fusion of local business features, rating features, and sentiment polarity features. Assume that the rating feature is the average merchant rating of region A, and the sentiment polarity feature is the sentiment tendency of comments on region A on social media. Map these three features to a unified feature space through nonlinear projection. For example, use a multi-layer perceptron to map features of different dimensions and value ranges to the same feature space. Construct a multimodal feature interaction matrix. For example, calculate the similarity or correlation between different features and construct an interaction matrix. Use a hierarchical multi-head attention mechanism to model local and global dependencies between modalities. For example, the local attention mechanism focuses on the local interactions between different features, and the global attention mechanism focuses on the global dependencies between different features. Combine feature recalibration modules to adjust feature distribution. For example, use a batch normalization layer to standardize features. Obtain initial comprehensive business features through residual connections with channel recalibration. For example, residual connections can avoid the gradient vanishing problem, and channel recalibration can adjust the importance of different feature channels. Obtain the final comprehensive business features through an adaptive weight allocation mechanism based on modality importance. For example, dynamically adjust the weights of different features according to their contribution to the final prediction result.

[0055] The present invention can more comprehensively characterize the commercial characteristics of a region and avoid the limitations of a single data source by fusing multi-source data, including network accessibility, commercial vitality prediction time series, ratings and sentiment polarity; contrastive learning networks and multi-scale time series analysis can effectively capture potential patterns and laws in the data, improve the robustness and noise resistance of feature representation, and achieve deep fusion of spatial and temporal features, wherein the matrix branch uses a multi-level graph convolutional network to process local matrices, and innovatively introduces residual connections and attention pooling mechanisms to extract hierarchical spatial features; the time series branch uses a bidirectional gated recurrent unit combined with an adaptive frequency domain decomposition module to effectively capture multi-scale temporal dynamic features, and optimizes the network by fusing a hierarchical contrast loss function that combines spatial topological constraints and temporal consistency constraints, which not only ensures that the feature representation maintains the spatial relationship and temporal continuity of the original network structure, but also improves the model performance; the adaptive weight coefficient, attention mechanism and feature recalibration module can dynamically adjust parameters according to the actual situation of the data, optimize the feature fusion process, and thus improve the accuracy of the final comprehensive commercial features.

[0056] In an optional embodiment, The steps of constructing an evaluation weight adaptive layer, taking comprehensive business features as state input, dynamically adjusting the evaluation index weights through reinforcement learning, and outputting the business service evaluation results include: Constructing an evaluation state space, combining comprehensive business characteristics, evaluation index weight vectors, differential characteristics of business vitality prediction time series, and business format collaborative development index to form a state vector, wherein the business format collaborative development index is calculated based on the accessibility strength of service complementarity and network accessibility matrix; Designing a multi-level reward function based on the state vector, including using the Huber loss smooth error metric to calculate the evaluation deviation, combining the relative entropy and the mixed norm constraint weight change penalty term to construct an immediate reward function, and the weight change penalty term is adjusted through gradient adaptive adjustment; Based on the time series change characteristics of the business format collaborative development index, a time series reward function is constructed in combination with the dynamic change of the business format distribution entropy; A residual network structure is used to generate reward combination weights, the state vector and the action vector are input into the residual block for feature extraction and mapping, the reward weight is obtained through a softmax function, the instantaneous reward function and the time series reward function are weightedly combined to obtain a total reward function, and a smoothing factor based on adaptive adjustment of signal variance is used for time series smoothing; Construct a dual network structure, in which the value network evaluates the state-action value through the state value function and the advantage function, and the strategy network generates the adjustment amount of the evaluation index weight through the mean network and the variance network; optimize the value network and the strategy network parameters simultaneously based on the temporal difference error to obtain the optimized evaluation index weight vector; The optimized evaluation index weight vector is applied to the comprehensive business characteristics to obtain the business service evaluation results, which include business service quality score, business format development balance index and regional coordinated development level.

[0057] Exemplarily, an evaluation state space is constructed. The comprehensive business characteristics that describe the business development status, such as traffic, transaction volume, number of stores, consumer evaluation, etc., are combined with the current evaluation index weight vector. In addition, in order to capture the business development trend, the differential characteristics of the business vitality prediction time series are also included in the state space. At the same time, taking into account the interaction between different business formats, the business format collaborative development index is calculated based on the accessibility intensity of the service complementarity and network accessibility matrix, and the index is also used as a component of the state vector. For example, the catering industry and cinemas in a certain area can complement each other, and their network accessibility is also high, so their collaborative development index is also high. Assuming that the comprehensive business characteristics have 5 dimensions, the weight vector has 3 dimensions, the business vitality prediction time series differential characteristics have 2 dimensions, and the business format collaborative development index is 1 dimension, the final state vector dimension is 11.

[0058] Design a multi-level reward function. The design of the reward function aims to guide the reinforcement learning model to learn the best evaluation indicator weights. The reward function is divided into two parts: immediate reward and time-series reward. The immediate reward function consists of two parts: one is the evaluation deviation calculated by a smooth error measurement method similar to Huber loss, such as the difference between the current evaluation result and the expert score; the other is a weight change penalty term combined with relative entropy and mixed norm constraints to limit the weight adjustment range and avoid drastic fluctuations. The trade-off factor of the weight change penalty term is adaptively adjusted by a method similar to gradient descent. The time-series reward function is constructed based on the time-series change characteristics of the business format collaborative development index and the dynamic changes of the business format distribution entropy, aiming to encourage the model to learn the evaluation indicator weights that can promote the long-term collaborative development of the business format. For example, if the business format distribution entropy in a certain area continues to increase, indicating that the business format is more balanced, a higher time-series reward will be given.

[0059] The residual network structure is used to generate the reward combination weights. The state vector and action vector (i.e., the adjustment amount of the evaluation index weight) are input into the residual network for feature extraction and mapping. The output of the residual network is processed by a function similar to softmax to obtain the weights of the immediate reward and the sequential reward. The immediate reward and the sequential reward are multiplied by the corresponding weights and added together to obtain the total reward function. In order to make the reward signal smoother, a smoothing factor based on adaptive adjustment of the signal variance is used for sequential smoothing. For example, if the reward signal fluctuates greatly, a larger smoothing factor is used.

[0060] A dual network structure is constructed for reinforcement learning. The structure includes a value network and a policy network. The value network evaluates the state-action value through the state value function and the advantage function, that is, the value of taking a certain action in the current state. The policy network generates the adjustment amount of the evaluation index weight through the mean network and the variance network as the action output. The model uses an algorithm similar to the temporal difference error to simultaneously optimize the parameters of the value network and the policy network, and finally obtains the optimized evaluation index weight vector.

[0061] The optimized evaluation index weight vector is applied to the comprehensive business characteristics to obtain the business service evaluation results. The evaluation results include multiple dimensions, such as business service quality score, business format development balance index and regional coordinated development level. For example, assuming that the optimized weight vector is [0.5, 0.3, 0.2], and the corresponding evaluation indicators are consumer evaluation, number of stores and transaction amount, the final business service quality score is the weighted average of these three indicators.

[0062] The present invention improves the accuracy and reliability of the evaluation results. By dynamically adjusting the weights of evaluation indicators, it can better adapt to different business environments and characteristics, thereby more accurately reflecting the true level of business services. By designing a multi-level reward function, the model can be guided to learn the weights of evaluation indicators that can promote the long-term coordinated development of business formats, thereby optimizing the business ecology and enhancing overall business vitality. Through the reinforcement learning algorithm, the optimal evaluation indicator weights can be automatically learned without human intervention, thereby improving evaluation efficiency and reducing evaluation costs.

[0063] In an optional embodiment, Designing a multi-level reward function based on the state vector includes calculating the evaluation deviation using the Huber loss smooth error metric, constructing an immediate reward function in combination with the weight change penalty term constrained by relative entropy and mixed norm, and the trade-off factor of the weight change penalty term is adaptively adjusted by gradient; Based on the time series change characteristics of the business format collaborative development index, the steps of constructing a time series reward function in combination with the dynamic change of business format distribution entropy include: Construct a multi-dimensional evaluation state vector for the state vector, including the evaluation score and its derivative, weight gradient and environmental characteristics; calculate the evaluation distribution and time series deviation based on wavelet transform; construct an instant reward function including smooth error measurement, weight change penalty and uncertainty measurement; construct a time series reward function based on the time series changes of the business format collaborative development index and distribution entropy, and optimize the trade-off factor through the policy gradient, including: The state vector is expanded to construct a multidimensional evaluation state vector, wherein the multidimensional evaluation state vector includes the evaluation score, evaluation score gradient, evaluation score second-order derivative, evaluation index weight vector gradient and environmental characteristics of the comprehensive business characteristics, wherein the environmental characteristics include time periodicity information and evaluation scene type information; based on the evaluation score, a wavelet transform is used to perform multi-scale decomposition to obtain a trend component and a detail component, and the evaluation distribution is calculated based on the local adaptive kernel density estimation; the trend component and the detail component are used to respectively calculate the time series difference to obtain the trend deviation and the detail deviation, and an adaptive threshold is generated by a multi-layer perceptron; Based on the evaluation index weight vector combined with relative entropy constraints and mixed norm constraints, a weight structure regularization term is constructed; based on the evaluation distribution, an uncertainty measure is calculated; the evaluation index weight vector is adaptively adjusted; based on the Huber loss function, a smooth error measure of the trend deviation and detail deviation is calculated, and a weight change penalty term is constructed, and the weight factor of the weight change penalty term is adaptively adjusted by the gradient; the smooth error measure, the weight change penalty term, and the uncertainty measure are combined to construct an instant reward function; Based on the business format collaborative development index, its time difference and second-order derivative are calculated, and a time series reward function is constructed in combination with the time series difference of the business format distribution entropy. The trade-off factor of the time series reward function is optimized by the policy gradient method.

[0064] Exemplarily, a multi-dimensional evaluation state vector is constructed. The vector contains multiple dimensions to comprehensively describe the state of the system. These dimensions include: evaluation scores for each business feature (such as customer flow, sales, customer satisfaction, etc.); the rate of change (gradient) and change trend (second-order derivative) of the evaluation score, which are used to reflect the dynamic changes of the evaluation index; the gradient of the evaluation index weight vector, which is used to reflect the changes in the importance of different indicators; and environmental characteristics, such as time periodicity information (such as weekdays / weekends, holidays, etc.) and evaluation scene type information (such as promotions, special events, etc.). For example, the evaluation state vector of a shopping mall may include: sales of clothing stores, sales growth rate, sales growth trend, rate of change of clothing store weights, and information that the day is the weekend.

[0065] Multi-scale decomposition of evaluation scores. Wavelet transform is used to decompose each evaluation score into trend component and detail component. The trend component reflects the long-term trend of the evaluation index, and the detail component reflects the short-term fluctuation of the evaluation index. For example, the trend component of the sales of clothing stores can reflect the overall sales of the store, and the detail component can reflect the sales fluctuation caused by weekend promotions.

[0066] Calculate the evaluation distribution and time series deviation. Based on the local adaptive kernel density estimation method, calculate the probability distribution of the evaluation score. Use the trend component and detail component to calculate the difference with the historical data, and obtain the trend deviation and detail deviation. These deviations can reflect the degree of difference between the current state and the historical state. For example, by comparing the trend component of the current sales of clothing stores with the trend component of the sales in the past month, the trend deviation can be obtained, reflecting the changes in the long-term sales of the store. Generate an adaptive threshold through a multi-layer perceptron to determine whether the deviation exceeds the normal range.

[0067] Construct an immediate reward function. This function consists of three parts: a smooth error metric, a weight change penalty term, and an uncertainty metric. The smooth error metric calculates trend deviation and detail deviation based on the Huber loss function, and is used to measure the gap between the current state and the expected state. The weight change penalty term is based on the evaluation index weight vector, and combines the relative entropy constraint and the mixed norm constraint to construct a weight structure regularization term to limit the magnitude of weight changes and avoid overfitting. The uncertainty metric is calculated based on the evaluation distribution and is used to measure the uncertainty of the state estimate. For example, if the distribution of sales in clothing stores is more dispersed, the uncertainty metric is higher. These three parts are combined to form the final immediate reward function. Among them, the trade-off factor of the weight change penalty term is adjusted through gradient adaptive adjustment to balance the contribution of different parts.

[0068] Construct a temporal reward function. This function is constructed based on the temporal changes of the industry synergy development index and the industry distribution entropy. The temporal difference and second-order derivative of the industry synergy development index and the temporal difference of the industry distribution entropy are calculated to reflect the dynamic changes of the industry combination. For example, if the synergy development index of the catering industry and the entertainment industry continues to rise, it means that the combination of these two industries is good. These indicators are combined to form the final temporal reward function. The trade-off factor of the temporal reward function is optimized by the policy gradient method to maximize the long-term cumulative reward.

[0069] Use reinforcement learning algorithms to optimize the system. Input the constructed multi-dimensional evaluation state vector, immediate reward function, and time series reward function into the reinforcement learning algorithm, such as the policy gradient algorithm. Through continuous iterative optimization, find the best business combination strategy to maximize the long-term benefits of the system.

[0070] Figure 3 The performance comparison chart for the industry collaborative development index tracking is as follows: Figure 3 The dotted line in the figure is the true synergy index, the diamond-marked curve is the traditional weighted method, the circle-marked curve is the present solution, and the curves of each method represent their predicted tracking values. The tracking curve of this solution almost coincides with the true value, especially in areas where the index fluctuates greatly (such as 24-30 months). In terms of average tracking error, this technical solution is 0.057, and the traditional weighted method is 0.251. The tracking accuracy of this technical solution has been improved by 77.3%, which is attributed to the use of wavelet transform and adaptive kernel density estimation. The present invention can more comprehensively capture the changes in the system state through multi-dimensional evaluation state vectors and multi-scale decomposition, thereby improving the sensitivity and accuracy of the reward function; through smooth error metrics, weight change penalty terms and uncertainty metrics, it can effectively suppress the influence of noise and outliers and enhance the robustness of the reward function; by combining the immediate reward function and the time series reward function, it can achieve dynamic optimization of the business format combination, thereby maximizing the long-term benefits of the system.

[0071] In an optional embodiment, A dual network structure is constructed, in which the value network evaluates the state-action value through the state value function and the advantage function, and the strategy network generates the adjustment amount of the evaluation index weight through the mean network and the variance network; the value network and the strategy network parameters are optimized simultaneously based on the temporal difference error, and the steps of obtaining the optimized evaluation index weight vector include: A multi-granularity state encoder is constructed, and an adaptive threshold is designed based on the temporal change rate of the evaluation index. The evaluation index sequence is divided into feature segments of different time scales. The statistical moment and fluctuation characteristics of each feature segment are calculated to construct a multi-scale feature pyramid. The hierarchical attention mechanism is used for feature fusion to obtain a hierarchical state representation. A dual network structure is constructed based on the hierarchical state representation, wherein the value network is composed of a state value function network and an advantage function network, wherein the state value function network evaluates the baseline value of the current state, and the advantage function network evaluates the value improvement of the action relative to the average level, and the state-action value is obtained by combining the state value function network and the advantage function network; a strategy network is constructed based on the state-action value, wherein the strategy network includes a mean network and a variance network, wherein the mean network outputs the expected direction of the evaluation index weight adjustment, and the variance network represents the uncertainty distribution of the weight adjustment, and the adjustment amount of the evaluation index weight is generated by a reparameterization method; Constructing a temporal difference error, taking the difference between the predicted value of the state value function network and the target value calculated based on the advantage function network as the optimization target of the value network; constructing a generalized advantage estimate based on the temporal difference error, calculating the policy gradient of the evaluation index weight adjustment amount in combination with the outputs of the mean network and the variance network, and optimizing the parameters of the value network and the policy network at the same time; Constructing a weight constraint based on the Wasserstein metric, including: constructing a support set for weight adjustment based on the dynamic distribution characteristics of the weight adjustment sequence, calculating the Wasserstein distance between the current weight adjustment amount and the support set, and projecting the adjustment amount of the evaluation index weight to a feasible domain that satisfies the Wasserstein distance constraint through iterative projection; updating the evaluation index weight vector based on the weight constraint, and dynamically adjusting the evaluation system through alternating optimization of the value network and the strategy network.

[0072] Exemplarily, a multi-granularity state encoder is constructed. The encoder is used to convert the time series data of the evaluation index into a hierarchical state representation. Specifically, the time series change rate of each evaluation index is first calculated. According to these change rates, an adaptive threshold is set to divide the evaluation index sequence into different feature segments. The threshold setting method can be adjusted according to the specific application scenario, for example, a multiple of the mean plus the standard deviation can be used, or a method such as quantile can be used. Each feature segment represents the performance of the evaluation index at different time scales. For example, one feature segment may represent short-term fluctuations, and another feature segment may represent long-term trends. For each feature segment, its statistical moments, such as mean, variance, skewness and kurtosis, and fluctuation features, such as amplitude, frequency and period, are calculated. These features are combined into a multi-scale feature pyramid. Finally, a hierarchical attention mechanism is used to fuse the multi-scale feature pyramid to obtain a hierarchical state representation. For example, a bottom-up attention mechanism can be used to first fuse short-term features, and then gradually fuse long-term features, and finally obtain a hierarchical state representation that comprehensively considers information at different time scales. Assume that there are three evaluation indicators A, B and C, and their values ​​in a time period are A=[1, 2, 3, 4, 5], B=[2, 4, 6, 8, 10], and C=[1, 1, 2, 3, 5]. By calculating the rate of change and setting the threshold, A is divided into two segments [1, 2] and [3, 4, 5], B is divided into two segments [2, 4] and [6, 8, 10], and C is divided into two segments [1, 1, 2] and [3, 5]. The mean and variance of each segment are calculated separately, a multi-scale feature pyramid is constructed, and the final state representation is obtained by fusion through the hierarchical attention mechanism.

[0073] A dual network structure is constructed based on hierarchical state representation. The dual network structure consists of a value network and a policy network. The value network is used to evaluate the value of the state and the value of the state-action, and the policy network is used to generate the adjustment amount of the evaluation index weight. The value network consists of a state value function network and an advantage function network. The state value function network evaluates the baseline value of the current state. The advantage function network evaluates the value improvement of taking a certain action relative to the average level in the current state. The output of the state value function network and the output of the advantage function network are combined to obtain the state-action value. The policy network contains a mean network and a variance network. The mean network outputs the expected direction of the evaluation index weight adjustment. The variance network represents the uncertainty distribution of the weight adjustment. The adjustment amount of the evaluation index weight is generated by reparameterization methods, such as sampling from a standard normal distribution and combining the outputs of the mean network and the variance network. Assuming that the state representation is [0.1, 0.2, 0.3], the state value function network outputs 0.5, and the advantage function network outputs [0.1, -0.1, 0.2], then the state-action value is [0.6, 0.4, 0.7]. Assuming the mean network output is [0.01, 0.02, -0.01] and the variance network output is [0.001, 0.002, 0.001], the weight adjustment generated by the reparameterization method is [0.011, 0.022, -0.009].

[0074] Construct the temporal difference error and optimize the network parameters. The difference between the predicted value of the state value function network and the target value calculated based on the advantage function network is used as the optimization target of the value network. Construct the generalized advantage estimate based on the temporal difference error, calculate the policy gradient of the evaluation index weight adjustment amount by combining the output of the mean network and the variance network, and optimize the parameters of the value network and the policy network at the same time.

[0075] Construct a weight constraint based on Wasserstein metric. First, construct a support set for weight adjustment based on the dynamic distribution characteristics of the weight adjustment sequence. Then, calculate the Wasserstein distance between the current weight adjustment and the support set. The Wasserstein distance is used to measure the difference between the two distributions. Through iterative projection, the adjustment of the evaluation index weight is projected to the feasible domain that satisfies the Wasserstein distance constraint. Assuming that the current weight adjustment is [0.011, 0.022, -0.009] and the support set is {[0.01, 0.02, -0.01], [0.012, 0.021, -0.008]}, the final weight adjustment is obtained by calculating the Wasserstein distance and projecting the weight adjustment to the feasible domain. Update the evaluation index weight vector based on the weight constraint, and dynamically adjust the evaluation system through alternating optimization of the value network and the policy network.

[0076] The present invention can more accurately reflect the true value of the evaluated object by dynamically adjusting the weights of the evaluation indicators, avoid the deviation caused by static weight settings, and can adaptively adjust the weights according to the temporal changes of the evaluation indicators, thereby improving the adaptability of the evaluation system to different environments and scenarios. Through the weight constraints of the Wasserstein metric, the amplitude and direction of the weight adjustment can be effectively controlled, the robustness of the evaluation system can be enhanced, and problems such as excessive weight fluctuations can be avoided.

[0077] Exemplarily, this embodiment provides a method for predicting commercial vitality and evaluating business districts based on multimodal feature fusion. Obtain commercial POI data in the target area, and reclassify and label the POIs according to their functions. In this embodiment, POIs are reclassified into the following categories according to functional land use: residential function, office function, commercial service, scientific research and education, medical service, cultural tourism, industrial production, transportation function, etc. At the same time, obtain time series trajectory data, specifically the total number of users staying in each grid per hour.

[0078] Scoring features and sentiment polarity features are extracted from user reviews of commercial POIs. Scoring features include taste scores, environment scores, service scores, cost-effectiveness scores, and other dimensions. When extracting POI information, a weighted form is used to highlight the role of functional determination by assigning higher weights to POI facilities with larger areas, such as large shopping malls and tourist attractions.

[0079] The distribution characteristics of human traffic are calculated based on time series trajectory data. There are significant differences in human activities in different functional areas. For example, the working area shows an inflow during the morning peak and an outflow during the evening peak, while the residential area shows the opposite. In this embodiment, a time clustering algorithm is used to analyze the change pattern of human traffic. The selectable clustering algorithms include: Representative-based clustering: based on kmeans, using Euclidean distance, dynamic time warping (DTW), etc. to measure the similarity of time series; density-based clustering: such as the DBSCAN algorithm; hierarchical methods: such as the BIRCH algorithm based on the CF tree; clustering based on Gaussian mixture distribution; spectral clustering, etc.

[0080] Finally, the POI data is constructed as a knowledge graph node, and the service complementarity is calculated based on the node's rating characteristics and sentiment polarity characteristics. The network accessibility matrix is ​​constructed by combining location information and traffic distribution characteristics. When aggregating POIs, not only the POIs inside the grid are considered, but also the distance function is used to examine the surrounding areas, which has a gradually decreasing effect. Finally, the evaluation index weights are dynamically adjusted through reinforcement learning to output the commercial service evaluation results.

[0081] This embodiment combines static POI data and dynamic crowd flow data, taking into account both spatial distribution and temporal changes, and can more accurately evaluate the commercial service level of the living circle.

[0082] A second aspect of an embodiment of the present invention provides a business vitality prediction and business district evaluation system based on multimodal feature fusion, the system comprising: The first unit is used to obtain commercial POI data and time series trajectory data of the target area, classify and label the commercial POI data, extract the rating features and sentiment polarity features of user comments, and calculate the distribution features of pedestrian flow based on the time series trajectory data; The second unit is used to input the scoring features, sentiment polarity features and traffic flow distribution features into a recursive neural network to generate a time series of commercial vitality prediction for each region; construct the commercial POI data into knowledge graph nodes, calculate service complementarity based on the scoring features and sentiment polarity features, construct a network accessibility matrix based on location information and traffic flow distribution features, and optimize the network accessibility matrix through service complementarity to obtain an enhanced network accessibility matrix; The third unit is used to input the enhanced network reachability matrix into the Louvain community discovery algorithm to divide the target area into sub-areas, and deploy edge computing nodes in each sub-area; each edge computing node inputs the local enhanced network reachability matrix and the business vitality prediction time series into a comparative learning network to generate local business features, and integrates them with the scoring features and sentiment polarity features to obtain comprehensive business features; construct an evaluation weight adaptive layer, take the comprehensive business features as state input, dynamically adjust the evaluation index weights through reinforcement learning, and output the business service evaluation results.

[0083] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0084] A fourth aspect of the embodiments of the present invention is: A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0085] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A commercial vitality prediction and business district evaluation method based on multimodal feature fusion, characterized in that: include: Acquire commercial POI data and time series trajectory data of the target area, classify and label the commercial POI data by business type, extract the rating features and sentiment polarity features of user comments, and calculate the flow distribution features based on the time series trajectory data; The scoring features, sentiment polarity features and traffic flow distribution features are input into a recursive neural network to generate a time series of commercial vitality prediction for each region; the commercial POI data is constructed as a knowledge graph node, service complementarity is calculated based on the scoring features and sentiment polarity features, a network accessibility matrix is ​​constructed in combination with location information and traffic flow distribution features, and the network accessibility matrix is ​​optimized through service complementarity to obtain an enhanced network accessibility matrix; Inputting the enhanced network reachability matrix into the Louvain community discovery algorithm to divide the target area into sub-areas, and deploying edge computing nodes in each sub-area; Each edge computing node compares the local enhanced network reachability matrix with the business vitality prediction time series input into the learning network to generate local business features, and integrates them with the scoring features and sentiment polarity features to obtain comprehensive business features; constructs an evaluation weight adaptive layer, takes the comprehensive business features as state input, dynamically adjusts the evaluation index weights through reinforcement learning, and outputs the business service evaluation results.

2. The method according to claim 1, characterized in that The steps of inputting the scoring features, sentiment polarity features and traffic flow distribution features into a recursive neural network to generate a commercial vitality prediction time series for each region include: The feature interaction matrix is ​​constructed, and the time series features are generated through the bidirectional gated recurrent unit network. The multi-scale decomposition strategy is adopted to predict the trend term, cycle term and residual term respectively through linear projection, Fourier decomposition and multi-layer perceptron. Finally, adaptive fusion is performed based on the reliability score to obtain the business vitality prediction time series, which includes: The scoring features are enhanced to obtain scoring statistical features, including mean, standard deviation, skewness, kurtosis and scoring change rate; the sentiment intensity is calculated based on word frequency and position weight to obtain enhanced sentiment features, including positive sentiment intensity, negative sentiment intensity and sentiment distribution entropy; the time pattern is extracted based on the 24-hour passenger flow distribution characteristics of the target area to obtain passenger flow pattern features, including peak values ​​in the morning, noon and evening periods, valley values ​​in the morning, noon and evening periods, weekday passenger flow ratio, weekend passenger flow ratio and periodicity score; The score statistical features, enhanced emotional features and crowd flow pattern features are used to construct a feature interaction matrix, wherein the feature interaction matrix includes independent representations of the three types of features and Hadamard product interaction representations between two features; the feature interaction matrix is ​​input into a bidirectional gated recurrent unit network with residual connections to generate bidirectional time series features including a forward feature sequence and a reverse feature sequence; Perform multi-scale convolution on the bidirectional time series features to obtain daily scale features, weekly scale features and monthly scale features; perform linear projection based on the monthly scale features to obtain trend item prediction results, perform Fourier decomposition based on the weekly scale features to obtain period item prediction results, input the daily scale features, weekly scale features and monthly scale features into a multi-layer perceptron to obtain residual item prediction results; calculate the fluctuation amplitude and prediction error of the trend item prediction result, the period item prediction result and the residual item prediction result to obtain a reliability score, generate a fusion weight coefficient based on the reliability score, and use the fusion weight coefficient to perform weighted summation on the trend item prediction result, the period item prediction result and the residual item prediction result to obtain a business vitality prediction time series.

3. The method according to claim 1, characterized in that The steps of constructing the commercial POI data as knowledge graph nodes, calculating service complementarity based on rating features and sentiment polarity features, constructing a network accessibility matrix in combination with location information and pedestrian flow distribution features, and optimizing the network accessibility matrix through service complementarity to obtain an enhanced network accessibility matrix include: The commercial POI data is constructed into basic attributes including business type, business area, and business hours, scoring features including taste score, environment score, service score, and cost-effectiveness score, emotional polarity features including positive emotional intensity, negative emotional intensity, and neutral emotional intensity, and knowledge graph nodes including temporal attributes such as average daily order volume, weekend passenger flow, and weekday passenger flow. The knowledge graph nodes are connected to establish spatial relationships based on geographical distance, complementary competitive relationships based on business type, and overlapping relationships based on business hours; The scoring features are weighted based on the scoring dimension weights to obtain scoring complementarity, the emotional complementarity is calculated based on the vector cosine similarity of the emotional polarity features and the difference between the emotional intensity, the time complementarity is calculated based on the difference in business time distribution and the degree of peak passenger flow staggering, and the scoring complementarity, the emotional complementarity and the time complementarity are weightedly integrated to obtain service complementarity; The crowd flow distribution characteristics between adjacent POI nodes are decomposed in time series to obtain periodic crowd flow fluctuation patterns and crowd flow change characteristics. The trajectory stability is calculated by combining the directional entropy and duration of the crowd movement trajectory. The initial network accessibility matrix is ​​constructed based on the overlap degree of the crowd flow fluctuation patterns, the correlation of the crowd flow change characteristics and the trajectory stability. The service complementarity is multiplied by the accessibility strength of the network accessibility matrix, and the result is symmetrically normalized to obtain the enhanced network accessibility matrix.

4. The method according to claim 1, characterized in that: The steps of each edge computing node inputting the local enhanced network reachability matrix and the business vitality prediction time series into a comparative learning network to generate local business features, and fusing them with the scoring features and the sentiment polarity features to obtain comprehensive business features include: Through localized processing with spatial constraints and dual-attenuation spatiotemporal data evaluation, a dual-branch contrastive learning network is constructed: the matrix branch uses a graph convolutional network to extract spatial features, and the time branch uses a gated recurrent unit to extract temporal features. Combined with the attention mechanism and multimodal feature interaction, comprehensive business features are obtained, including: The global enhanced network accessibility matrix is ​​localized according to the community discovery results, and the local matrix corresponding to the sub-region is extracted through a selection matrix with spatial constraints. The adaptive weight coefficient is set based on the regional overlap, and the boundary of the local matrix is ​​smoothed. The commercial vitality prediction time series in the sub-region is collected to construct a multi-scale time series sliding window, and the local training data is evaluated and screened using a spatiotemporal data freshness evaluation mechanism based on double decay. A dual-branch contrastive learning network with feature enhancement is constructed, wherein the matrix branch uses a multi-level graph convolutional network to process the local matrix, and extracts hierarchical spatial features through residual connections and attention pooling; the time series branch uses a bidirectional gated recurrent unit to process the business vitality prediction time series, and combines an adaptive frequency domain decomposition module to extract multi-scale time features; the contrastive learning network is optimized by a hierarchical contrast loss function that integrates spatial topology constraints and time series consistency constraints; The dynamic fusion weights of the spatial features and the temporal features are calculated based on the attention mechanism, and the cross-scale feature aggregation module is used to perform multi-level feature fusion to obtain the initial local business features; the momentum update mechanism of the adaptive temperature coefficient is introduced to perform online update on the initial local business features to obtain the local business features; The local commercial features, rating features and sentiment polarity features are mapped to a unified feature space through nonlinear projection to construct a multimodal feature interaction matrix. A hierarchical multi-head attention mechanism is used to model the local and global dependencies between modalities, and the feature distribution is adjusted in combination with a feature recalibration module. The initial comprehensive commercial features are obtained through a residual connection with channel recalibration, and the final comprehensive commercial features are obtained based on an adaptive weight allocation mechanism based on modality importance.

5. The method according to claim 1, characterized in that The steps of constructing an evaluation weight adaptive layer, taking comprehensive business features as state input, dynamically adjusting the evaluation index weights through reinforcement learning, and outputting the business service evaluation results include: Constructing an evaluation state space, combining comprehensive business characteristics, evaluation index weight vectors, differential characteristics of business vitality prediction time series, and business format collaborative development index to form a state vector, wherein the business format collaborative development index is calculated based on the accessibility strength of service complementarity and network accessibility matrix; Designing a multi-level reward function based on the state vector, including using the Huber loss smooth error metric to calculate the evaluation deviation, combining the relative entropy and the mixed norm constraint weight change penalty term to construct an immediate reward function, and the weight change penalty term is adjusted through gradient adaptive adjustment; Based on the time series change characteristics of the business format collaborative development index, a time series reward function is constructed in combination with the dynamic change of the business format distribution entropy; A residual network structure is used to generate reward combination weights, the state vector and the action vector are input into the residual block for feature extraction and mapping, the reward weight is obtained through a softmax function, the instantaneous reward function and the time series reward function are weightedly combined to obtain a total reward function, and a smoothing factor based on adaptive adjustment of signal variance is used for time series smoothing; Construct a dual network structure, in which the value network evaluates the state-action value through the state value function and the advantage function, and the strategy network generates the adjustment amount of the evaluation index weight through the mean network and the variance network; optimize the value network and the strategy network parameters simultaneously based on the temporal difference error to obtain the optimized evaluation index weight vector; The optimized evaluation index weight vector is applied to the comprehensive business characteristics to obtain the business service evaluation results, which include business service quality score, business format development balance index and regional coordinated development level.

6. The method according to claim 5, characterized in that Designing a multi-level reward function based on the state vector, including calculating the evaluation deviation using a smooth error metric of Huber loss, and constructing an immediate reward function in combination with a weight change penalty term constrained by relative entropy and mixed norm, wherein the trade-off factor of the weight change penalty term is adaptively adjusted by gradient; Based on the time series change characteristics of the business format collaborative development index, the steps of constructing a time series reward function in combination with the dynamic change of the business format distribution entropy include: Construct a multi-dimensional evaluation state vector for the state vector, including the evaluation score and its derivative, weight gradient and environmental characteristics; calculate the evaluation distribution and time series deviation based on wavelet transform; construct an instant reward function including smooth error measurement, weight change penalty and uncertainty measurement; construct a time series reward function based on the time series changes of the business format collaborative development index and distribution entropy, and optimize the trade-off factor through the policy gradient, including: The state vector is expanded to construct a multidimensional evaluation state vector, wherein the multidimensional evaluation state vector includes the evaluation score, evaluation score gradient, evaluation score second-order derivative, evaluation index weight vector gradient and environmental characteristics of the comprehensive business characteristics, wherein the environmental characteristics include time periodicity information and evaluation scene type information; based on the evaluation score, a wavelet transform is used to perform multi-scale decomposition to obtain a trend component and a detail component, and the evaluation distribution is calculated based on the local adaptive kernel density estimation; the trend component and the detail component are used to respectively calculate the time series difference to obtain the trend deviation and the detail deviation, and an adaptive threshold is generated by a multi-layer perceptron; Based on the evaluation index weight vector combined with relative entropy constraints and mixed norm constraints, a weight structure regularization term is constructed; based on the evaluation distribution, an uncertainty measure is calculated; the evaluation index weight vector is adaptively adjusted; based on the Huber loss function, a smooth error measure of the trend deviation and detail deviation is calculated, and a weight change penalty term is constructed, and the weight factor of the weight change penalty term is adaptively adjusted by the gradient; the smooth error measure, the weight change penalty term, and the uncertainty measure are combined to construct an instant reward function; Based on the business format collaborative development index, its time difference and second-order derivative are calculated, and a time series reward function is constructed in combination with the time series difference of the business format distribution entropy. The trade-off factor of the time series reward function is optimized by the policy gradient method.

7. The method according to claim 5, characterized in that Construct a dual network structure, in which the value network evaluates the state-action value through the state value function and the advantage function, and the strategy network generates the adjustment of the evaluation index weight through the mean network and the variance network; The steps of simultaneously optimizing the value network and the policy network parameters based on the temporal difference error to obtain the optimized evaluation index weight vector include: A multi-granularity state encoder is constructed, and an adaptive threshold is designed based on the temporal change rate of the evaluation index. The evaluation index sequence is divided into feature segments of different time scales. The statistical moment and fluctuation characteristics of each feature segment are calculated to construct a multi-scale feature pyramid. The hierarchical attention mechanism is used for feature fusion to obtain a hierarchical state representation. A dual network structure is constructed based on the hierarchical state representation, wherein the value network is composed of a state value function network and an advantage function network, wherein the state value function network evaluates the baseline value of the current state, and the advantage function network evaluates the value improvement of the action relative to the average level, and the state-action value is obtained by combining the state value function network and the advantage function network; a strategy network is constructed based on the state-action value, wherein the strategy network includes a mean network and a variance network, wherein the mean network outputs the expected direction of the evaluation index weight adjustment, and the variance network represents the uncertainty distribution of the weight adjustment, and the adjustment amount of the evaluation index weight is generated by a reparameterization method; Constructing a temporal difference error, taking the difference between the predicted value of the state value function network and the target value calculated based on the advantage function network as the optimization target of the value network; constructing a generalized advantage estimate based on the temporal difference error, calculating the policy gradient of the evaluation index weight adjustment amount in combination with the outputs of the mean network and the variance network, and optimizing the parameters of the value network and the policy network at the same time; Constructing a weight constraint based on the Wasserstein metric, including: constructing a support set for weight adjustment based on the dynamic distribution characteristics of the weight adjustment sequence, calculating the Wasserstein distance between the current weight adjustment amount and the support set, and projecting the adjustment amount of the evaluation index weight to a feasible domain that satisfies the Wasserstein distance constraint through iterative projection; updating the evaluation index weight vector based on the weight constraint, and dynamically adjusting the evaluation system through alternating optimization of the value network and the strategy network.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images