A location-aware dynamic rights distribution method, medium, and system

Abstract

This invention provides a location-aware dynamic benefits distribution method, medium, and system, belonging to the field of dynamic benefits distribution technology. First, it collects real-time consumer location data via mobile terminals to establish a location trajectory map and calculates a location dwell intensity index. Then, it acquires the locations of merchants within a business district to construct a location matrix and combines this with wireless network identifiers to determine the location of the merchant where the consumer is located, generating a location correlation. Next, it uses a prediction model based on a multi-head spatiotemporal attention network structure to analyze consumers' historical consumption data and behavioral characteristics, and calculates a merchant attractiveness index by combining the location dwell intensity index and location correlation. Further, it adjusts the time sensitivity coefficient through a consumer value judgment function and dynamically adjusts the benefits distribution threshold based on the business district's pedestrian density factor. Finally, it selects suitable benefits types based on consumer profiles and accurately pushes them, achieving location-aware dynamic and precise benefits distribution.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of dynamic benefit delivery, and particularly relates to a dynamic benefit delivery method based on location awareness, a medium and a system. BACKGROUND

[0002] With the rapid development of mobile internet technology, location-aware services have been widely used in the field of business marketing. The traditional benefit delivery method mainly relies on the static portrait features, historical consumption data or simple location information of consumers, and triggers the benefit delivery mechanism through preset rules. The common benefit delivery systems in the current market mostly use delivery strategies based on member levels, consumption amounts or fixed time nodes, such as holiday promotions, member birthday discounts, etc. These methods have formed a relatively mature pattern in practical application. However, the traditional benefit delivery method has obvious defects: on the one hand, the preset rules lack flexibility and cannot dynamically adjust the benefit delivery strategy according to the real-time behavior and location changes of consumers; on the other hand, the simple location information application only considers whether the consumer enters a specific area, ignoring the location stay duration, frequency, movement trajectory and other deep-level space-time behavior characteristics, resulting in low precision of benefit delivery and low conversion efficiency. Although some systems introduce location factors, they are mostly coarse-grained judgments and do not fully utilize fine-grained location data to mine the potential needs of consumers. There is a technical problem in the prior art that dynamic and precise benefit delivery cannot be realized based on real-time location data and space-time behavior characteristics of consumers. This problem leads to low efficiency of benefit resource allocation, making it difficult to accurately push related benefits to consumers with actual consumption intention at the right time, and reduces the marketing effect and user experience. SUMMARY

[0003] Therefore, the application provides a dynamic benefit delivery method based on location awareness, a medium and a system, which can solve the technical problem that the prior art often only delivers consumer benefits based on consumer portraits or programmer presets, and cannot realize dynamic benefit delivery.

[0004] The application is implemented as follows: the first aspect of the application provides a dynamic benefit delivery method based on location awareness, which comprises the following steps: collecting real-time location data of consumers to establish a location trajectory graph; obtaining stay time and stay point location coordinates to calculate a location stay intensity index; obtaining merchant location coordinates to construct a commercial district location matrix and calculate a distance vector; collecting wireless network identifiers and combining the distance vector to determine the spatial location of the merchant and generate a location correlation degree; using a space-time consumption behavior prediction model to analyze consumption behavior characteristics and establish a consumption tendency model; calculating the attraction index of the merchant and the matching degree of the consumption tendency model, and then performing weighted calculation; setting a benefit delivery threshold; calculating a judgment value through a consumer value judgment function and adjusting a time sensitivity coefficient; selecting a benefit type suitable for the consumer to determine the benefit content; pushing the benefit information to the consumer and recording the use situation to update the model.

[0005] The position trajectory diagram refers to a continuous trajectory data set formed by recording a plurality of position point coordinates of a consumer within a certain time period, and is used for analyzing the moving mode and staying habit of the consumer.

[0006] In the step of collecting real-time position data of the consumer, a time sampling interval is set as a preferred sampling interval according to a position data update frequency, and the preferred sampling interval refers to an optimal time sampling unit automatically set by the system according to the position data update frequency, and is used for balancing the position data collection accuracy and system resource consumption.

[0007] The position staying intensity index refers to a comprehensive quantitative index of the staying time and frequency of the consumer in a certain area, and the longer the staying time and the higher the frequency, the greater the position staying intensity index.

[0008] The distance vector refers to a multi-dimensional vector formed by the spatial distance from the current position of the consumer to each merchant position point in the commercial district, and is used for judging the physical proximity of the consumer to each merchant; and the position correlation degree refers to a quantitative index of the close spatial relationship between the consumer and a certain merchant calculated according to the wireless network signal strength and the distance vector.

[0009] The consumption tendency model refers to a prediction model constructed based on the historical consumption behavior, preference characteristics and time regularity of the consumer, and is used for evaluating the probability of the consumer generating a consumption behavior in a certain scenario; and the merchant attraction index refers to a quantitative index of the attraction degree of a merchant to a certain consumer calculated in combination with the position staying intensity index and the position correlation degree.

[0010] The benefit issuance threshold refers to the minimum standard value triggering the benefit issuance mechanism, and when the weighted calculation result exceeds the benefit issuance threshold, the system determines that the consumer has a high consumption possibility; and the benefit issuance threshold parameter is dynamically adjusted by the spatio-temporal consumption behavior prediction model according to the commercial district crowd density factor.

[0011] The commercial district crowd density factor refers to the crowd flow per unit area in the commercial district within a certain time period, and is used for reflecting the busy degree and potential consumption activity of the commercial district.

[0012] The consumer value judgment function refers to a mathematical function for calculating the consumer value score by comprehensively evaluating the historical value, current position information and consumption intention of the consumer, and the judgment value is divided into three intervals, i.e., a high value interval, a medium value interval and a low value interval.

[0013] The time sensitivity coefficient refers to a parameter in the spatio-temporal consumption behavior prediction model for measuring the sensitivity of the consumer to the time factor, and the time sensitivity coefficient increases when the judgment value is in the high value interval, and the time sensitivity coefficient decreases when the judgment value is in the low value interval.

[0014] The specific structure of the spatio-temporal consumption behavior prediction model is a multi-head spatio-temporal attention network structure, including a spatial feature extraction layer, a time series encoding layer, a multi-head attention layer and a prediction output layer, wherein the spatial feature extraction layer adopts a graph convolution network to capture spatial relationship features. The time series encoding layer adopts a long short-term memory network to process historical behavior time series data, the multi-head attention layer combines the position residence intensity index and the historical behavior sequence to generate an attention score matrix, and the prediction output layer combines all features to output a final consumption behavior prediction result. The steps for establishing the training data set of the spatio-temporal consumption behavior prediction model include collecting historical position trajectory data, consumption behavior records, wireless network connection records and equity use data, performing time alignment processing on the data to form a spatio-temporal sequence sample, and labeling whether a consumer produces an actual consumption behavior as a label.

[0015] The second aspect of the present application provides a computer readable storage medium, the computer readable storage medium stores program instructions, the program instructions are used to execute the above-mentioned position-aware dynamic equity distribution method when running in the computer.

[0016] The third aspect of the present application provides a position-aware dynamic equity distribution system, including the above-mentioned computer readable storage medium, the system is any one of a computer, a server and a single-chip microcomputer, the computer readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer readable storage medium.

[0017] The application realizes deep mining and analysis of real-time location data and space-time behavior characteristics of consumers by constructing a multi-head space-time attention network model. The method collects and analyzes space-time characteristics such as consumer location trajectory, residence time, residence intensity, and accurately locates the consumer's location in the merchant by combining wireless network connection information, and then calculates the location correlation degree and the merchant attraction index. Compared with the traditional benefit distribution method, the application breaks through the limitation of static rules and can dynamically adjust the benefit distribution strategy according to the real-time location change and space-time behavior characteristics of consumers. Through the comprehensive evaluation of multi-dimensional indexes such as location residence intensity index, location correlation degree and consumption tendency model, the system can accurately identify the current consumption willingness and potential demand of consumers, greatly improving the accuracy and conversion efficiency of benefit distribution. At the same time, the application introduces the commercial circle flow density factor to dynamically adjust the benefit distribution threshold, so that the system can adaptively optimize the benefit distribution strategy according to the real-time environment of the commercial circle. The application successfully solves the technical problem that the existing technology cannot realize dynamic and accurate benefit distribution based on real-time location data and space-time behavior characteristics of consumers. Through the organic combination of space-time consumption behavior prediction model and multi-dimensional location perception mechanism, the efficient allocation and accurate placement of benefit resources are realized, which improves the benefit utilization rate and consumption conversion rate, and significantly improves the consumer experience. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A flowchart of the method of the application. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical scheme and advantages of the embodiments of the application more clear, the technical scheme in the embodiments of the application will be described clearly and completely with reference to the drawings in the embodiments of the application.

[0020] As shown in Figure 1 , it is a flowchart of a dynamic benefit distribution method based on location perception provided by the first aspect of the application. The method includes the following steps:

[0021] S01, collecting real-time location data of consumers through a mobile terminal and establishing a location trajectory graph, setting a time sampling interval as a preferred sampling interval according to the location data update frequency;

[0022] S02, obtaining the residence time and residence point location coordinates of the consumer in the commercial circle based on the preferred sampling interval, and calculating the location residence intensity index;

[0023] S03, obtaining the location coordinates of all merchants in the commercial circle, constructing a commercial circle location matrix, and calculating the distance vector between the consumer residence point and each merchant;

[0024] S04, collect the wireless network identifier connected by the consumer's mobile terminal, determine the merchant space position where the consumer is located combined with the distance vector, and generate the position correlation degree;

[0025] S05, analyze the consumer's historical consumption data and behavior characteristics by using the spatio-temporal consumption behavior prediction model, and establish a consumption tendency model combined with the current merchant promotion activity information;

[0026] S06, combine the position residence intensity index with the position correlation degree, input into the spatio-temporal consumption behavior prediction model, calculate the merchant attraction index, and perform weighted calculation with the matching degree of the consumption tendency model;

[0027] S07, set the benefit issuance threshold according to the weighted calculation result, and the benefit issuance threshold parameter is dynamically adjusted by the spatio-temporal consumption behavior prediction model according to the commercial circle flow density factor;

[0028] S08, calculate the judgment value through the consumer value judgment function combined with the matching degree of the consumption tendency model, the position correlation degree and the merchant attraction index, and adjust the time sensitivity coefficient in the spatio-temporal consumption behavior prediction model according to the judgment value;

[0029] S09, select the benefit type suitable for the consumer according to the consumer portrait, and determine the benefit content and validity period;

[0030] S10, accurately push the benefit information to the consumer through the mobile terminal, and record the benefit use situation to update the consumption tendency model and the spatio-temporal consumption behavior prediction model.

[0031] Among them, the position trajectory diagram refers to a continuous trajectory data set formed by recording the coordinates of multiple position points of the consumer within a certain period of time, which is used to analyze the consumer's movement mode and residence habit.

[0032] Among them, the preferred sampling interval refers to the optimal time sampling unit automatically set by the system according to the position data update frequency, which is used to balance the position data collection accuracy and system resource consumption.

[0033] Among them, the position residence intensity index refers to a comprehensive quantitative index of the consumer's residence time and frequency in a certain area, and the longer the residence time and the higher the frequency, the larger the position residence intensity index.

[0034] Among them, the distance vector refers to a multi-dimensional vector composed of the spatial distance from the consumer's current position to each merchant position point in the commercial circle, which is used to judge the physical proximity between the consumer and each merchant.

[0035] Among them, the position correlation degree refers to a quantitative index of the close degree of the consumer's spatial contact with a certain merchant calculated according to the wireless network signal strength and the distance vector.

[0036] wherein the consumption propensity model refers to a prediction model based on historical consumption behaviors, preference characteristics and time rules of consumers, for evaluating the probability of consumers generating consumption behaviors in a certain scenario.

[0037] wherein the merchant attraction index refers to a quantitative index of the attraction degree of a merchant to a consumer calculated in combination with the location residence intensity index and the location correlation degree.

[0038] wherein the benefit issuance threshold refers to the minimum standard value triggering the benefit issuance mechanism, and when the weighted calculation result exceeds the benefit issuance threshold, the system determines that the consumer has a high consumption possibility.

[0039] wherein the consumer portrait refers to a user feature description containing multi-dimensional data such as consumer demographic characteristics, consumption habits, preferences and behavior patterns.

[0040] wherein the commercial district flow density factor refers to the flow per unit area in a commercial district within a certain period of time, for reflecting the busy degree and potential consumption activity of the commercial district.

[0041] wherein the consumer value determination function refers to a mathematical function for calculating the consumer value score by comprehensively evaluating the historical value of the consumer, the current location information and the consumption intention, and the determination value is divided into three intervals of high value interval, medium value interval and low value interval.

[0042] wherein the time sensitivity coefficient refers to a parameter in the spatio-temporal consumption behavior prediction model for measuring the sensitivity of the consumer to the time factor, and the time sensitivity coefficient increases when the determination value is in the high value interval, and the time sensitivity coefficient decreases when the determination value is in the low value interval.

[0043] The specific structure of the spatio-temporal consumption behavior prediction model is a multi-head spatio-temporal attention network structure, which includes a spatial feature extraction layer, a time series encoding layer, a multi-head attention layer and a prediction output layer. The spatial feature extraction layer adopts a graph convolution network to capture the spatial relationship characteristics of the consumer in the commercial district, the time series encoding layer adopts a long short-term memory network to process the historical behavior time series data of the consumer, the multi-head attention layer combines the location residence intensity index and the historical behavior sequence of the consumer to generate an attention score matrix for weighted processing of the spatio-temporal features, and the prediction output layer combines all the features to output the final consumption behavior prediction result. The overall parameter quantity of the model is 120,000 parameters, the hidden layer dimension is 256, the number of attention heads is 8, and the time sensitivity coefficient is used to adjust the time decay rate of the long short-term memory unit in the time series encoding layer.

[0044] The step of establishing the training data set of the spatio-temporal consumption behavior prediction model specifically comprises collecting historical consumer location trajectory data in a commercial district, consumption behavior records, wireless network connection records and benefit usage data, performing time alignment processing on the data to form spatio-temporal sequence samples, labeling whether the consumer produces actual consumption behavior in each spatio-temporal sequence sample as a label, performing stratified sampling on the samples according to commercial district types, time periods and consumer portrait features to ensure sample balance, extracting fixed-length training samples from long-time sequence data using a sliding window technique, performing normalization processing on location data, performing one-hot encoding on category features, and finally constructing a structured training data set containing spatio-temporal feature vectors and consumption behavior labels. The total amount of training data is not less than 10 million spatio-temporal sequence records, and covers at least 100 different types of commercial district scenes.

[0045] The step of training the spatio-temporal consumption behavior prediction model specifically comprises first training on a large-scale general commercial district consumer behavior data set, learning a general representation of consumer spatio-temporal behavior using a contrastive learning method, using a self-supervised learning method to construct positive and negative sample pairs in the training stage, setting the learning rate to 0.001, the batch size to 256, and the number of training rounds to 50, then fine-tuning for the to-be-tested commercial district scene data, using a cross-entropy loss function in the fine-tuning stage, reducing the learning rate to 0.0001, using a cosine annealing learning rate scheduling strategy, and setting the number of training rounds to 20. At the same time, an early stopping mechanism is introduced to prevent overfitting. In the fine-tuning process, the benefit conversion rate is used as the main evaluation index. The multi-head attention layer parameters and the initial value of the time sensitivity coefficient are optimized through a grid search method. Finally, the model parameters with the highest benefit conversion rate on the validation set are selected as the final model.

[0046] The specific implementation of the above steps is described in detail below.

[0047] The specific implementation of step S01 is to first collect the real-time location information of the consumer by the global positioning system module of the mobile terminal, including longitude, latitude and positioning accuracy and other parameters, to reduce the positioning error by using a local weighted scatter smoothing algorithm to smooth the collected original location data. The kernel function used in the smoothing process is a Gaussian kernel function, and the neighborhood size is set to about 5 sampling points. Then the processed location data is stored in chronological order to form a location trajectory graph, which is stored in a directed graph structure, with nodes representing location coordinate points and edges representing the space-time conversion relationship between adjacent location points. Next, the system dynamically calculates the optimal sampling interval according to the update frequency of the location data, and the specific method is to count the location data upload frequency in the last 30 minutes. If the upload frequency is higher than once every 5 seconds, the preferred sampling interval is set to 10 seconds; if the upload frequency is between every 5 to 30 seconds, the preferred sampling interval is set to 30 seconds; if the upload frequency is lower than once every 30 seconds, the preferred sampling interval is set to 60 seconds. The role of this step is to obtain the real-time location data of the consumer and optimize the data collection frequency to provide basic data support for subsequent analysis, while balancing the accuracy of location information and system resource consumption.

[0048] The specific implementation of step S02 is to resample the consumer location data based on the preferred sampling interval, and to identify the residence points in the consumer location trajectory by using an improved DBSCAN density clustering algorithm. The neighborhood radius in the clustering parameter is set to 25 meters, and the minimum number of points is set to 3 consecutive sampling points. Then the total duration of the consumer's stay at each residence point is calculated, and the center coordinates of each residence point are recorded. Temporary residence points with a residence time of less than 2 minutes are filtered. Then the location residence intensity index is calculated according to the residence time and the residence frequency, and the calculation formula is: location residence intensity index = T i × F i × W t , where T i represents the cumulative residence time at the location (unit: minutes), F i represents the residence frequency at the location in the last 7 days, and W t represents the time weight factor, with a value of 1.0 on weekdays and a value of 1.2 on weekends. The role of this step is to identify the main residence location of the consumer in the business district and quantify the residence intensity, providing a time dimension quantitative index for evaluating the consumer's interest in the business.

[0049] The specific implementation of step S03 is to extract the location coordinate information of all merchants from the business district information database, including longitude, latitude, floor, area code and other data, to construct a business district location matrix. Then map the merchant location information to a unified coordinate system, use the Haversine formula to calculate the distance between two points on the earth's surface, and introduce a floor spacing correction factor for merchants located in different floors in the same building, with a vertical distance of 4 meters between each floor. Then calculate the distance vector between the consumer's current residence point and the location of each merchant in the business district. The distance vector includes two components: planar distance and vertical distance. Planar distance represents the straight-line distance on the horizontal projection plane, and vertical distance represents the height difference. The role of this step is to construct a spatial relationship model between the consumer and each merchant in the business district, providing quantitative data in the spatial dimension for subsequent location relevance analysis.

[0050] The specific implementation of step S04 is to scan the available wireless networks around the mobile terminal through the wireless network interface, collect wireless network identifiers (SSID), signal strength (RSSI) and MAC address information, and construct a wireless network signal fingerprint. Then match the collected wireless network signal fingerprint with the business district wireless network database, which pre-stores the wireless network information and signal coverage range of each merchant in the business district. Then use the weighted k-nearest neighbor algorithm to consider the wireless network matching degree and distance vector, calculate the location relevance of the consumer and each merchant, and the calculation formula is location relevance = α × (1-D i / D max ) + β × S i / S max , where D i represents the distance vector module length between the consumer and merchant i, D max represents the set maximum effective distance threshold (set to 100 meters), S i represents the wireless network signal strength associated with merchant i, S max represents the theoretical maximum signal strength, and the weight coefficients α and β are set to 0.6 and 0.4 respectively. The role of this step is to combine wireless network signal information and spatial distance information to accurately determine the spatial location of the consumer, improving indoor positioning accuracy.

[0051] The specific implementation of step S05 is to extract the consumer historical consumption records from the user database, including consumption time, consumption amount, consumption merchant category and other information, apply the association rule mining algorithm to analyze the cross-merchant consumption pattern of the consumer, and set the support threshold to 0.05 and the confidence threshold to 0.6 in the association rule mining. Then, combined with the time series analysis method, the time law of the consumer's consumption behavior is identified, and the Fourier transform is used to extract the periodic pattern in the consumption time series, focusing on analyzing the periodic characteristics of the three time scales of day, week and month. Then, the merchant promotion activity information is integrated, including activity type, discount range, activity period and the like, to build a consumption tendency model, and a probability prediction model of the consumer's consumption behavior in a certain scenario is established by using the logistic regression algorithm. The input features of the model include the consumer's historical consumption frequency, average consumption amount, merchant category preference score, current time period historical consumption probability and current merchant promotion activity attraction score. The role of this step is to analyze the consumer's historical behavior characteristics and the current business environment to establish a consumer consumption tendency prediction model, providing a behavior prediction basis for precise rights and interests placement.

[0052] The specific implementation of step S06 is to input the position residence intensity index and the position correlation degree as spatiotemporal features into the spatiotemporal consumption behavior prediction model. The model uses a multi-head self-attention mechanism to process the interaction between the position features and the time features, and the dimensions of the query matrix, key matrix and value matrix in the attention calculation are all set to 256. Then, the model outputs the merchant's attraction prediction score for the consumer, i.e., the merchant attraction index, with an index value range of 0 to 1, and the larger the value, the stronger the attraction. Then, the weighted sum of the merchant attraction index and the consumption tendency model matching degree is calculated, and the calculation formula is weighted calculation result = γ x merchant attraction index + (1-γ) x consumption tendency model matching degree, where the weight coefficient γ is dynamically adjusted according to the consumer's position sensitivity. The γ value of the consumer with high position sensitivity is set to 0.7, the γ value of the consumer with medium position sensitivity is set to 0.5, and the γ value of the consumer with low position sensitivity is set to 0.3. The role of this step is to comprehensively evaluate the spatial behavior characteristics and consumption tendency characteristics of the consumer to generate a comprehensive consumption possibility score, providing a quantitative basis for rights and interests placement decisions.

[0053] The specific implementation of step S07 is based on the current business circle flow density data obtained by the business circle real-time monitoring system, the flow density is expressed by the average number of people per unit area (100 square meters), and the flow density is divided into three levels: low density (less than 5 people / 100 square meters), medium density (5-15 people / 100 square meters), and high density (more than 15 people / 100 square meters). Then, the benefit distribution threshold is dynamically adjusted according to the business circle flow density. In the low density state, the benefit distribution threshold is set to 0.7; in the medium density state, the benefit distribution threshold is set to 0.6; in the high density state, the benefit distribution threshold is set to 0.5. Then, the spatio-temporal consumption behavior prediction model continuously updates the model parameters through online learning, adjusts the benefit distribution threshold according to the real-time feedback of the benefit conversion rate, and balances the exploration and utilization of the benefit distribution by using the ε-greedy strategy in reinforcement learning. The initial value of ε is set to 0.2, and it gradually decreases to 0.05 as the model learning process progresses. The role of this step is to dynamically adjust the benefit distribution standard according to the real-time state of the business circle, optimize the benefit distribution strategy under different busy degrees of the business circle, and improve the adaptability of the system.

[0054] The specific implementation of step S08 is to construct a consumer value judgment function, the function expression is judgment value = w1 × consumption tendency model matching degree + w2 × location correlation degree + w3 × merchant attraction index, the weight coefficients w1, w2 and w3 are determined according to the historical value level of consumers. The three weights of high-value consumers are 0.5, 0.2 and 0.3 respectively, the three weights of medium-value consumers are 0.4, 0.3 and 0.3 respectively, and the three weights of low-value consumers are 0.3, 0.4 and 0.3 respectively. Then, according to the judgment value, the consumers are divided into three value intervals: the judgment value greater than 0.8 is the high-value interval, the judgment value between 0.5 and 0.8 is the medium-value interval, and the judgment value less than 0.5 is the low-value interval. Then, according to the value interval in which the consumer is located, the time sensitivity coefficient in the spatio-temporal consumption behavior prediction model is adjusted. When the judgment value is in the high-value interval, the time sensitivity coefficient increases from the baseline value 1.0 to 1.5; when the judgment value is in the medium-value interval, the time sensitivity coefficient remains unchanged at the baseline value 1.0; when the judgment value is in the low-value interval, the time sensitivity coefficient decreases from the baseline value 1.0 to 0.7. The role of this step is to evaluate the value of consumers and dynamically adjust the weight of time factor in the prediction model, and to adopt different time-sensitive strategies for different value consumers.

[0055] The specific implementation of step S09 is to select suitable benefit types according to the consumer portrait information, which includes multi-dimensional characteristics such as consumer age range, gender, consumption capacity level, interest preferences, etc. Then, a collaborative filtering recommendation algorithm is used to recommend the most suitable benefit type for the target consumer based on the benefit response history of similar consumers. The algorithm calculates the similarity between consumers using cosine similarity, and the similarity threshold is set to 0.7. Next, the benefit content details are determined, including discount rate, discount amount, usage threshold, etc. For high-value interval consumers, the benefit intensity is set to 25% to 30% of the product price; for medium-value interval consumers, the benefit intensity is set to 15% to 25% of the product price; for low-value interval consumers, the benefit intensity is set to 5% to 15% of the product price. Finally, the benefit validity period is set, which is usually 24 hours, and can be extended to 72 hours for special promotion activities, and shortened to 2 hours for time-limited purchase benefits. The role of this step is to individualize the benefit content according to consumer characteristics, ensuring that the benefit content is highly matched with consumer demand, and improving the benefit usage rate.

[0056] The specific implementation of step S10 is to build a mobile terminal push engine, supporting multiple push methods, including in-app messages, SMS, and push notifications. The system selects the optimal push method based on the historical response rate of consumers to different push channels. Then, a benefit information presentation template is designed, including merchant information, benefit details, validity period, and usage instructions. The template design follows the principles of clear information hierarchy and prominent key information. Next, the benefit push operation is performed, and the system records the push time, push channel, arrival status, etc., and sets the minimum push interval to no more than 2 times per 4 hours for the same consumer. Finally, the benefit usage is tracked, including viewing time, activation status, usage time, consumption amount, etc. The benefit conversion rate (actual benefit usage quantity divided by total benefit push quantity) is calculated, and the benefit usage data is fed back to the consumption tendency model and the spatio-temporal consumption behavior prediction model to update the model parameters. The gradient descent method is used to optimize the model weights, with a learning rate of 0.01, and the model is updated once every 100 new data collected. The role of this step is to complete the precise push of benefit information and continuously optimize the model effect through feedback loops, forming a closed-loop optimization mechanism to continuously improve the precision of benefit placement.

[0057] The detailed structure of the spatio-temporal consumption behavior prediction model is a multi-head spatio-temporal attention network structure, which consists of four main components. The first part is the spatial feature extraction layer, which uses a graph convolution network to process the spatial relationship data of consumers in the shopping district. The specific implementation includes two layers of graph convolution, with an output feature dimension of 128 and a ReLU activation function. The graph convolution operation is approximated using Chebyshev polynomials with an order of 2 to capture the spatial topological relationship between the consumer's location and the merchant's location. The second part is the time series encoding layer, which uses a long short-term memory network (LSTM) to process the consumer's historical behavior time series data. The LSTM unit contains three gating mechanisms: forget gate, input gate, and output gate. The hidden state dimension is 256, the input sequence length is set to the behavior data of the last 30 days, the sliding window size is 7 days, and the time step is 1 day. A time decay factor is introduced for different time historical data, and the decay rate is controlled by the time sensitivity coefficient. The third part is the multi-head attention layer, which combines the location residence intensity index with the consumer's historical behavior sequence to generate an attention score matrix for weighted processing of spatio-temporal features. The number of attention heads is set to 8, the output dimension of each attention head is 32, the combined feature dimension is 256, and the attention score calculation uses a scaled dot-product attention mechanism with a scaling factor of where d k is the dimension of the key vector. The fourth part is the prediction output layer, which uses a fully connected network structure with two hidden layers with dimensions of 128 and 64, respectively. The activation function is ReLU, and the last layer uses a Sigmoid activation function to output the final consumption behavior probability prediction result, with a value range of 0 to 1. The total number of model parameters is 120,000, of which the spatial feature extraction layer has approximately 33,000 parameters, the time series encoding layer has approximately 52,000 parameters, the multi-head attention layer has approximately 21,000 parameters, and the prediction output layer has approximately 14,000 parameters.

[0058] The training data set of the spatio-temporal consumption behavior prediction model includes the following steps: first, collecting the historical consumer location trajectory data in the business circle from the business circle monitoring system database, collecting data for nearly 2 years, covering different seasons and holiday characteristics. Then integrate consumption behavior records, wireless network connection records and benefit use data, extract from transaction system, network management system and marketing system respectively, align the timestamps of these heterogeneous data, unify the time accuracy to minute level, form spatio-temporal sequence samples. Then label whether the consumer produces actual consumption behavior for each spatio-temporal sequence sample as a label, positive samples are defined as records of consumption behavior in related merchants within 4 hours after location trajectory record, negative samples are defined as records without consumption behavior. Then stratified sampling is performed according to business circle type (such as shopping center, pedestrian street, commercial complex, etc.), time period (such as weekdays, weekends, holidays and different time periods), consumer portrait characteristics (such as age group, consumption ability level, etc.), to ensure balanced number of samples of each type, each type contains at least 100,000 records. Then use sliding window technology to extract fixed length training samples from long time series data, window length is set to 7 days, sliding step is 1 day, that is, each sample contains 7 consecutive days of spatio-temporal behavior data of the consumer and the corresponding label. Finally, normalize the location data, convert latitude and longitude coordinates to standardized coordinates relative to the center point of the business circle, and convert category features (such as merchant type, consumer preference, etc.) to numerical feature vectors through one-hot encoding, to build a structured training data set containing spatio-temporal feature vectors and consumption behavior labels. The total amount of training data reaches 12 million spatio-temporal sequence records, covering 120 different types of business circle scenes, and the data set is divided into training set, validation set and test set in the ratio of 8:1:1.

[0059] Among them, the time sensitivity coefficient is the core adjustment parameter in the prediction model, which significantly improves the accuracy and flexibility of the model by dynamically adapting to the time preferences of different value consumers. For high-value interval consumers, increasing the time sensitivity coefficient can make the model pay more attention to recent behavior, quickly capture the immediate demand changes and consumption intentions of such consumers, and thus provide more timely benefit push. For low-value interval consumers, reducing the time sensitivity coefficient makes the model consider more long-term stable behavior patterns, avoiding being disturbed by short-term fluctuations, and more comprehensively evaluating their consumption tendency. This differentiated time weight adjustment mechanism enables the prediction model to intelligently balance immediacy and stability, providing personalized benefit distribution strategies for consumers at different value levels, improving the conversion efficiency of high-value consumers, and optimizing the long-term cultivation effect on low-value consumers.

[0060] Among them, the multi-head attention layer constructs a deep correlation analysis mechanism of space-time characteristics by fusing the position residence intensity index and the historical behavior sequence. This layer structure can simultaneously pay attention to information features in multiple dimensions, automatically identify the internal relationship between the consumer's location behavior and the historical consumption mode. When the consumer shows high residence intensity near a certain merchant, the attention mechanism will automatically enhance the weight of the historical behavior related to this type of merchant, forming a more accurate consumption intention judgment. The multi-head design further improves the ability of the model to capture complex patterns, enabling the system to simultaneously focus on multiple behavior characteristics of consumers, such as sensitivity to promotions, brand loyalty, and time regularity. The attention score matrix, as the core of the weighting process, dynamically adjusts the importance of different spatio-temporal features, enabling the model to accurately distinguish between key signals and noise, and extract the most predictive feature combination, thereby significantly improving the relevance and effectiveness of the benefit push.

[0061] The second aspect of the present application provides a computer readable storage medium, the computer readable storage medium stores program instructions, the program instructions are used to execute the above-mentioned location-aware dynamic benefit delivery method when running in the computer.

[0062] The third aspect of the present application provides a location-aware dynamic benefit delivery system, which includes the above-mentioned computer readable storage medium, the system is any one of a computer, a server, and a single-chip microcomputer, the computer readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer readable storage medium.

[0063] Specifically, the principle of the present application is: the core technical principle of the present application is to deeply integrate location-aware technology and space-time consumption behavior prediction model, and to build a complete dynamic benefit delivery technology system. First, the system collects real-time location data of consumers through mobile terminals, establishes a refined location trajectory graph, and calculates a location residence intensity index based on a preferred sampling interval, thereby quantifying the willingness of consumers to stay in a specific area. This index provides a direct quantitative representation of the consumer's interest in the area, and becomes an important indicator for predicting consumer potential.

[0064] Secondly, the system constructs a business circle location matrix by obtaining the location coordinates of all merchants in the business circle, calculates the distance vector in combination with the location of the consumer's residence point, and further improves the location determination accuracy by using wireless network identifiers to generate a location correlation degree. This multi-source location information fusion method significantly improves the identification accuracy of the spatial location of the consumer's merchant, solving the problem of insufficient accuracy of traditional GPS positioning in indoor scenes.

[0065] Thirdly, the application designs a spatio-temporal consumption behavior prediction model based on a multi-head spatio-temporal attention network, which contains four key components: a spatial feature extraction layer, a time series encoding layer, a multi-head attention layer and a prediction output layer. The spatial feature extraction layer uses a graph convolution network to capture the spatial relationship features of consumers in a commercial district; the time series encoding layer processes historical behavior time series data through a long short-term memory network; the multi-head attention layer combines the location residence intensity index with the historical behavior sequence to form an attention score matrix, realizing dynamic weighting of spatio-temporal features; the prediction output layer integrates all features to generate the final prediction result. The model is trained on a large-scale dataset through contrastive learning and self-supervised learning methods, and has strong spatio-temporal feature representation ability and consumption behavior prediction ability.

[0066] In addition, the application introduces a consumer value judgment function, calculates a judgment value by comprehensively evaluating the current state of the consumer, and dynamically adjusts the time sensitivity coefficient accordingly, so that the model can adjust the time decay rate of historical data according to different value consumers. At the same time, the system dynamically adjusts the benefit distribution threshold based on the commercial district flow density factor, realizing the adaptive optimization of the benefit distribution mechanism. This multi-level dynamic adjustment mechanism enables the system to maintain efficient operation in complex and variable business environments, ensuring optimal allocation of benefit resources.

[0067] Through the organic combination of the above technical principles, the application realizes deep mining and analysis of consumer real-time location data and spatio-temporal behavior features, accurately identifies the current consumption willingness and potential demand of consumers, thereby solving the core technical problem of the prior art that cannot realize dynamic and accurate benefit distribution.

[0068] A specific embodiment 1 of the application is provided below, and the specific implementation of each step in embodiment 1 is described in detail as follows.

[0069] The specific implementation of step S01 is to first collect consumer real-time location information through the global positioning system module of the mobile terminal, including longitude, latitude and positioning accuracy parameters, and use a local weighted scatter smoothing algorithm to smooth the collected original location data to reduce positioning errors. When smoothing, the kernel function is a Gaussian kernel function, and the kernel function expression is:

[0070]

[0071] In the formula, K(d) is the kernel function value, d is the distance from the current point to the neighborhood point, and σ is the smoothing parameter, which controls the smoothing degree and takes a value range of 10-30 meters. The mathematical expression of smoothing is:

[0072]

[0073] In the formula, P sPsmoothed i Poriginal i n is the neighborhood size, set as 5 nearest sample points. Then the processed location data is stored in time order to form a location trajectory graph, which is stored in a directed graph structure G = (V, E), where V represents the set of location coordinate points, and E represents the set of spatio-temporal conversion relationships between adjacent location points. Next, the system dynamically calculates the optimal sampling interval according to the update frequency of the location data, and the specific method is to count the location data upload frequency in the last 30 minutes, and the sampling interval calculation formula is:

[0074]

[0075] T sampling preferred sampling interval, unit: second, f update location data update frequency, unit: hertz. The role of this step is to obtain real-time location data of consumers and optimize the data collection frequency, providing basic data support for subsequent analysis, while balancing the accuracy of location information and system resource consumption.

[0076] The specific implementation of step S02 is to resample the consumer location data based on the preferred sampling interval, and to identify the stay points in the consumer location trajectory by applying the improved density clustering algorithm. The neighborhood radius ∈ in the clustering parameter is set to 25 meters, and the minimum point number MinPts is set to 3 consecutive sample points. The improved density clustering algorithm introduces a time continuity constraint, and the clustering condition expression is:

[0077]

[0078] C is the stay point cluster, P is the set of location points, N ∈ (p i ) represents the set of points within the ∈ neighborhood of point p i , T j (p stay ) represents the timestamp of point p i , and ΔT is the time threshold, set to 120 seconds. Then calculate the total stay time of the consumer at each stay point, record the center coordinates of each stay point, and filter the temporary stay points with a stay time of less than 2 minutes. Next, calculate the location stay intensity index according to the stay time and stay frequency, and the calculation formula is:

[0079] I stay = T i × F i × W t ;

[0080] I stay represents the location stay intensity index, T iF represents the cumulative residence time at the location, in minutes, F i W represents the residence frequency at the location in the past 7 days, W t W represents the time weight factor, which is 1.0 for weekdays and 1.2 for weekends. The purpose of this step is to identify the main residence location of the consumer in the business district and quantify the residence intensity, providing a time dimension quantitative index for evaluating the consumer's interest in the business.

[0081] The specific implementation of step S03 is to extract the location coordinate information of all businesses from the business district information database, including longitude, latitude, floor, area code and other data, to construct a business district location matrix M loc , the matrix elements are:

[0082]

[0083] In the formula, lon i , lat i , floor i , code i respectively represent the longitude, latitude, floor and area code of the i-th business, and n is the total number of businesses in the business district. Then map the business location information to a unified coordinate system, and use the Haversine formula to calculate the distance between two points on the earth's surface, the calculation formula is:

[0084]

[0085] In the formula, d is the distance between two points, r is the radius of the earth, which is 6371 kilometers, φ1 and φ2 are the latitudes (in radian) of the two points, and λ1 and λ2 are the longitudes (in radian) of the two points. For businesses in the same building but on different floors, introduce a floor distance correction factor, and set the vertical distance between each floor to 4 meters. The corrected distance calculation formula is:

[0086]

[0087] In the formula, d adj is the corrected distance, and floor1 and floor2 are the floor numbers of the two points. Then calculate the distance vector between the consumer's current residence point and the location of each business in the business district, the distance vector is composed of two components: the horizontal distance d plane and the vertical distance d vert .

[0088]

[0089] In the formula, d plane represents the straight-line distance on the horizontal projection plane, and d vertThe height difference is represented. The role of this step is to build a spatial relationship model between the consumer and each merchant in the business district, and to provide quantitative data of spatial dimension for subsequent location relevance analysis.

[0090] The specific implementation of step S04 is to scan the available wireless networks around the mobile terminal through the wireless network interface, collect information such as wireless network identifier, signal strength, and media access control address, and build a wireless network signal fingerprint. Then match the collected wireless network signal fingerprint with the business district wireless network database, which pre-stores the wireless network information and signal coverage range of each merchant in the business district. Then use the weighted k-nearest neighbor algorithm to consider the wireless network matching degree and distance vector comprehensively, calculate the location relevance of the consumer and each merchant, and the calculation formula is:

[0091]

[0092] In the formula, R loc represents the location relevance, the value range is 0-1, D i represents the distance vector module length between the consumer and the merchant i, and the calculation formula is D max represents the set maximum effective distance threshold, which is set to 100 meters, S i represents the wireless network signal strength associated with the merchant i, the unit is decibel milliwatt, S max represents the theoretical maximum signal strength, the value is 0 decibel milliwatt, and the weight coefficients a and b are set to 0.6 and 0.4 respectively. The role of this step is to combine the wireless network signal information and the spatial distance information to accurately determine the spatial location of the consumer in the merchant, and to improve the indoor positioning accuracy.

[0093] The specific implementation of step S05 is to extract the consumer's historical consumption records from the user database, including consumption time, consumption amount, and consumption merchant category information, and to analyze the consumer's cross-merchant consumption pattern by applying the association rule mining algorithm. The support threshold in the association rule mining is set to 0.05, and the confidence threshold is set to 0.6. The mathematical expression of the association rule generation is:

[0094]

[0095] In the formula, represents the confidence of the association rule , Supp(X∪Y) represents the support of the item set X and Y appearing at the same time, and Supp(X) represents the support of the item set X. Then combine the time series analysis method to identify the time regularity of the consumer's consumption behavior, and use Fourier transform to extract the periodic pattern in the consumption time series. The mathematical expression of the Fourier transform is:

[0096]

[0097] where X(k) is the frequency domain representation, x(n) is the time domain sequence, N is the sequence length, and j is the imaginary unit. The periodic characteristics of the three time scales, i.e., intra-day, intra-week, and intra-month, are analyzed. Then, the merchant promotion information, including the activity type, discount rate, and activity period, is integrated to construct a consumption tendency model. The probability prediction model of consumer behavior in a certain scenario is established using the logistic regression algorithm, and the mathematical expression is as follows:

[0098]

[0099] where P(Y=1|X) represents the probability of the occurrence of consumer behavior under the condition of a given feature vector X, X1, X2,..., X n are input features, including consumer historical consumption frequency, average consumption amount, merchant category preference score, current period historical consumption probability, and current merchant promotion attraction score, θ0, θ1, θ2,..., θ n are model parameters, which are solved by the maximum likelihood estimation method. The purpose of this step is to establish a consumer consumption tendency prediction model by analyzing the consumer historical behavior characteristics and the current business environment, and to provide a behavior prediction basis for the precise placement of rights and interests.

[0100] The specific implementation of step S06 is to input the position residence intensity index and the position correlation degree as spatiotemporal features into the spatiotemporal consumption behavior prediction model. The model uses a multi-head self-attention mechanism to process the interaction between the position features and the time features. The dimensions of the query matrix, key matrix, and value matrix in the attention calculation are all set to 256. The mathematical expression of the multi-head attention mechanism is as follows:

[0101] MultiHead(Q, K, V) = Concat(head1, head2,..., head h )W O ;

[0102] where MultiHead(Q, K, V) represents the output of the multi-head attention, head i represents the output of the i-th attention head, W O is the output projection matrix, and h is the number of attention heads, which is set to 8. The calculation formula of each attention head is as follows:

[0103] head i = Attention(QW i Q , KW i K , VW i V );

[0104] wherein W i Q , W i K , W i V are the projection matrices of the query, key, value of the i-th attention head, respectively. The attention function adopts the scaled dot-product attention, and the calculation formula is as follows:

[0105]

[0106] wherein d k is the dimension of the key vector, which is set to 256. Then the model outputs the prediction score of the attraction of the merchant to the consumer, i.e., the merchant attraction index, and the index value ranges from 0 to 1, and the larger the value is, the stronger the attraction is. Then the weighted sum of the merchant attraction index and the matching degree of the consumer tendency model is calculated, and the calculation formula is as follows:

[0107] R weighted = γ × I attr + (1-γ) × M consumption ;

[0108] wherein R weighted is the weighted calculation result, I attr is the merchant attraction index, M consumption is the matching degree of the consumer tendency model, and γ is the weight coefficient, which is dynamically adjusted according to the location sensitivity of the consumer. The γ value of the consumer with high location sensitivity is set to 0.7, the γ value of the consumer with medium location sensitivity is set to 0.5, and the γ value of the consumer with low location sensitivity is set to 0.3. The role of this step is to comprehensively evaluate the spatial behavior characteristics and consumer tendency characteristics of the consumer, and to generate a comprehensive score of the consumer possibility, thereby providing a quantitative basis for the benefit issuance decision.

[0109] The specific implementation of step S07 is to obtain the current business circle crowd density data based on the business circle real-time monitoring system, the crowd density is represented by the average number of people per unit area, and the crowd density is divided into three levels of low density, medium density and high density. Then the benefit issuance threshold is dynamically adjusted according to the crowd density of the business circle, and the threshold adjustment formula is as follows:

[0110]

[0111] wherein T benefits is the benefit issuance threshold, and ρ crowd is the crowd density of the business circle, which is in the unit of people per 100 square meters. Then the spatio-temporal consumer behavior prediction model continuously updates the model parameters through online learning, adjusts the benefit issuance threshold according to the real-time feedback of the benefit conversion rate, and balances the exploration and utilization of the benefit issuance by using the ε-greedy strategy in reinforcement learning, and the action selection probability is as follows:

[0112]

[0113] In the formula, P(a) represents the probability of selecting action a, Q(s, a) represents the value function of action a in state s, |A| represents the size of the action space, and the value of ε is initially set to 0.2 and gradually reduced to 0.05 as the model learning process progresses. The role of this step is to dynamically adjust the benefit distribution standard according to the real-time state of the business circle, optimize the benefit distribution strategy under different busy degrees of the business circle, and improve the adaptability of the system.

[0114] The specific implementation of step S08 is to construct a consumer value judgment function, and the function expression is:

[0115] V judge = w1 x M consumption + w2 x R loc + w3 x I attr ;

[0116] In the formula, V judge is the judgment value, M consumption is the matching degree of consumption tendency model, R loc is the location correlation degree, I attr is the merchant attraction index, and w1, w2 and w3 are weight coefficients determined according to the historical value level of consumers. The three weights of high-value consumers are 0.5, 0.2 and 0.3 respectively, the three weights of medium-value consumers are 0.4, 0.3 and 0.3 respectively, and the three weights of low-value consumers are 0.3, 0.4 and 0.3 respectively. Then, according to the judgment value, the consumers are divided into three value intervals: the judgment value greater than 0.8 is the high-value interval, the judgment value between 0.5 and 0.8 is the medium-value interval, and the judgment value less than 0.5 is the low-value interval. Then, according to the value interval in which the consumer is located, the time sensitivity coefficient in the spatio-temporal consumption behavior prediction model is adjusted, and the adjustment formula is:

[0117]

[0118] In the formula, λ time is the time sensitivity coefficient, and V judge is the judgment value. The role of this step is to evaluate the value of the consumer and dynamically adjust the weight of the time factor in the prediction model, and to adopt different time-sensitive strategies for different value consumers.

[0119] The specific implementation of step S09 is to select the appropriate benefit type according to the consumer portrait information, which includes consumer age, gender, consumption ability level, interest preference and other multi-dimensional characteristics. Then, a collaborative filtering recommendation algorithm is used to recommend the most suitable benefit type for the target consumer based on the benefit response history of similar consumers. The algorithm uses cosine similarity to calculate the similarity between consumers, and the calculation formula is:

[0120]

[0121] where similarity(A, B) represents the similarity between consumers A and B, A i and B i represent the values of consumer A and B in the i-th feature dimension, n is the total number of feature dimensions, and the similarity threshold is set to 0.7. Then the benefit content details are determined, including the discount rate, the discount amount, the use threshold, etc. For high-value interval consumers, the benefit intensity is set to 25% to 30% of the product price; for medium-value interval consumers, the benefit intensity is set to 15% to 25% of the product price; and for low-value interval consumers, the benefit intensity is set to 5% to 15% of the product price. Finally, the benefit validity period is set, which is usually 24 hours, can be extended to 72 hours for special promotion activities, and shortened to 2 hours for time-limited purchase benefits. The role of this step is to design the benefit content according to the consumer characteristics, to ensure that the benefit content is highly matched with the consumer demand, and to improve the benefit usage rate.

[0122] The specific implementation of step S10 is to build a mobile terminal push engine to support multiple push methods, including in-app messages, SMS, and push notifications, etc. The system selects the optimal push method according to the historical response rate of consumers to different push channels. Then the benefit information presentation template is designed, including merchant information, benefit details, validity period, and usage instructions, etc. The template design follows the principles of clear information hierarchy and key information prominence. Then the benefit push operation is performed, the system records the push time, push channel, arrival state, etc. and sets the minimum push interval to no more than 2 times per 4 hours for the same consumer. Finally, the benefit usage is tracked and recorded, including viewing time, activation status, usage time, consumption amount, etc. The benefit conversion rate is calculated, and the calculation formula is:

[0123]

[0124] where CR benefits represents the benefit conversion rate, N used represents the actual number of benefits used, and N pushed represents the total number of benefits pushed. The benefit usage data is fed back to the consumption tendency model and the spatio-temporal consumption behavior prediction model, the model parameters are updated, the gradient descent method is used to optimize the model weights, and the update formula is:

[0125]

[0126] where θ j represents the model parameters, α represents the learning rate, which is set to 0.01, J(θ) represents the loss function, and represents the loss function with respect to the parameter θ jThe partial derivative of the model is updated once every 100 new data collected. The purpose of this step is to complete the accurate push of equity information and continuously optimize the model effect through the feedback loop to form a closed-loop optimization mechanism and continuously improve the accuracy of equity investment.

[0127] The detailed structure of the spatio-temporal consumption behavior prediction model is a multi-head spatio-temporal attention network structure, which consists of four main components. The first part is the spatial feature extraction layer, which uses a graph convolution network to process the spatial relationship data of consumers in the business circle. The mathematical expression of the graph convolution operation is:

[0128]

[0129] In the formula, H (l) represents the feature matrix of the lth layer, is the adjacency matrix after adding a self-loop, is the degree matrix of , W (l) is the weight parameter matrix of the lth layer, and σ is the activation function, which is selected as the rectified linear unit function. The second part is the time series encoding layer, which uses a long short-term memory network to process the time series data of consumer historical behavior. The mathematical expression of the long short-term memory unit is:

[0130] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0131] i t =σ(W i ·[h t-1 ,x t ]+b i );

[0132]

[0133] o t =σ(W o ·[h t-1 ,x t ]+b o );

[0134] h t =o t ×tanh(C t );

[0135] In the formula, f t , i t , o t represent the forget gate, input gate and output gate respectively, C t represents the cell state, and h tIndicates the hidden state, x t Let W represent the input sequence. f W i W C W o Let b be the weight matrix. f b i b C b o The bias vector is σ, where σ is the sigmoid activation function and tanh is the hyperbolic tangent activation function. The third part is the multi-head attention layer, which combines the location dwell intensity index with the consumer's historical behavior sequence to generate an attention score matrix for weighted processing of spatiotemporal features. The fourth part is the prediction output layer, which uses a fully connected network structure. The output layer activation function is the sigmoid function, calculated as follows:

[0136]

[0137] In the formula, h represents the final predicted output. o W represents the output feature of the previous layer. o and b o Here, σ represents the weight matrix and bias vector of the output layer, respectively, and σ is the sigmoid activation function. The entire model has 120,000 parameters, of which the spatial feature extraction layer has approximately 33,000 parameters, the time series encoding layer has approximately 52,000 parameters, the multi-head attention layer has approximately 21,000 parameters, and the prediction output layer has approximately 14,000 parameters.

[0138] The specific function of the time sensitivity coefficient is the time sensitivity coefficient λ. time The modified forgetting gate calculation formula, used to adjust the rate of decay of historical data in the Long Short-Term Memory (LSTM) network, is as follows:

[0139] f t =σ(W f ·[h t-1 x t ]+b f )×λ time ;

[0140] In the formula, f t λ represents the output value of the forget gate. time This represents the time sensitivity coefficient. When λ... time As the value of λ increases, the output value of the forget gate increases, the model's forgetting speed of historical data slows down, and more historical information is retained; when λ increases... timeWhen the value decreases, the forgetting gate output value decreases, the forgetting speed of the model to the historical data increases, and more attention is paid to the recent data. By dynamically adjusting the time sensitivity coefficient, the system can adjust the sensitivity of the model to the time series data according to the consumer value interval, and provide more accurate consumer behavior prediction for consumers at different value levels.

[0141] The training data set establishment step of the spatio-temporal consumer behavior prediction model includes first collecting historical consumer location trajectory data in the commercial district from the commercial district monitoring system database, collecting data for nearly 2 years, covering different seasons and holiday characteristics. Then integrate consumer behavior records, wireless network connection records and benefit use data, extract from transaction system, network management system and marketing system respectively, align the timestamps of these heterogeneous data, unify the time accuracy to minute level, form spatio-temporal sequence samples. Then label whether the consumer produces actual consumer behavior for each spatio-temporal sequence sample as a label, the positive sample is defined as the record of consumer behavior in the relevant merchant within 4 hours after the location trajectory record, and the negative sample is defined as the record without consumer behavior. Then sample the samples according to the commercial district type (such as shopping center, pedestrian street, commercial complex, etc.), time period (such as weekday, weekend, holiday and different time period), consumer portrait feature (such as age group, consumer ability level, etc.), ensure the balanced number of each type of sample, each type contains at least 100,000 records. The mathematical expression of stratified sampling is:

[0142]

[0143] In the formula, n h represents the number of samples to be extracted in the hth layer, n represents the total sample demand, N h represents the total number of samples in the hth layer, and N represents the total number of samples. Then use the sliding window technique to extract fixed length training samples from long time series data, the window length is set to 7 days, and the sliding step is 1 day, that is, each sample contains 7 consecutive days of spatio-temporal behavior data of the consumer and the corresponding label. Finally, normalize the location data, convert the latitude and longitude coordinates to standardized coordinates relative to the center point of the commercial district, and the normalization formula is:

[0144]

[0145] In the formula, X norm represents the normalized coordinate value, X represents the original coordinate value, X min and X maxrespectively represent the minimum and maximum values of the coordinates. The one-hot encoding conversion of the category features (such as merchant types, consumer preferences, etc.) is converted into a numerical feature vector, a structured training data set containing the spatio-temporal feature vector and the consumption behavior label is constructed, the total amount of training data reaches 12 million spatio-temporal sequence records, covering 120 different types of commercial district scenes, and the data set is divided into a training set, a validation set and a test set according to a ratio of 8:1:1.

[0146] The steps of training the spatio-temporal consumption behavior prediction model specifically include first training on a large-scale general commercial district consumer behavior data set, learning the general representation of consumer spatio-temporal behavior by using a contrastive learning method, and using a self-supervised learning method to construct positive and negative sample pairs in the training stage. The loss function of the contrastive learning is:

[0147]

[0148] In the formula, L contrastive represents the contrastive loss, sim (h i , h j ) represents the similarity between the two sample representations, the cosine similarity is used for calculation, τ represents the temperature parameter, which is set to 0.07, and 1 [k≠i] is an indicator function, which takes a value of 1 when k≠i, and otherwise takes a value of 0. The learning rate is set to 0.001, the batch size is 256, and the number of training rounds is 50. Then fine-tune the data of the to-be-tested commercial district scene, and use the cross-entropy loss function in the fine-tuning stage, the calculation formula is:

[0149]

[0150] In the formula, L CE represents the cross-entropy loss, N represents the number of samples, y i represents the true label of the i-th sample, represents the prediction probability of the i-th sample. The learning rate is reduced to 0.0001, and the cosine annealing learning rate scheduling strategy is used, and the learning rate adjustment formula is:

[0151]

[0152] In the formula, η t represents the learning rate of the t-th round, η min and η maxrespectively, are set to 0.00001 and 0.0001, T represents the total number of training rounds, which is set to 20 rounds. At the same time, an early stopping mechanism is introduced to prevent overfitting, and the early stopping condition is set to 5 consecutive rounds of validation set loss no longer decreasing. In the fine-tuning process, the equity conversion rate is used as the main evaluation index, and the initial values of the multi-head attention layer parameters and the time sensitivity coefficient are optimized through the grid search method. The parameter search space is: the number of attention heads h e {4, 8, 16}, the attention layer dimension d e {64, 128, 256}, and the initial value of the time sensitivity coefficient λ initial e {0.8, 1.0, 1.2}. Finally, the model parameters with the highest equity conversion rate on the validation set are selected as the final model.

[0153] In this embodiment, the detailed mechanism of the multi-head attention layer is to combine the location residence intensity index with the consumer historical behavior sequence to generate an attention score matrix, and to perform weighted processing on the space-time features. This mechanism first converts the input features through linear projection into query, key and value three parts, each attention head independently calculates the attention score to form a weighted feature representation. This multi-head mechanism enables the model to focus on feature interactions in different subspaces at the same time, capturing the association between location residence patterns and historical behaviors from multiple perspectives. The location residence intensity index serves as a key input for the query vector, guiding the model to focus on historical consumption patterns related to the current spatial behavior; while the historical behavior sequence constitutes the key-value pair, providing a knowledge base of consumer behavior in the time dimension. The calculation process of the attention score matrix is essentially measuring the relevance strength between the current location behavior and the historical consumption patterns, and the historical behavior patterns with high relevance receive higher weights, thereby having a greater impact on the final prediction result. In this way, the model can effectively integrate features in the spatial and temporal dimensions, identify the consumption intentions that consumers may have when they reside in a particular location, and form a more accurate basis for predicting consumer behavior. The technical effect of the time sensitivity coefficient is to dynamically adjust the sensitivity of the model to time series data, achieving differentiated prediction strategies for different value consumers. For high-value interval consumers, increasing the time sensitivity coefficient makes the model retain more historical consumption information, forming a more complete understanding of behavior, and enabling it to capture the complex and diverse consumption patterns and long-term stable brand preferences of such consumers, avoiding prediction bias caused by short-term behavior fluctuations, and improving the accuracy and personalization of high-value consumer interest promotion. For medium-value interval consumers, maintaining the standard time sensitivity coefficient balances the consideration of historical behavior and recent changes, focusing on their stable consumption habits without ignoring the evolution trend of consumption patterns. For low-value interval consumers, reducing the time sensitivity coefficient makes the model focus more on recent behavior data, quickly responding to possible changes in consumption intentions of such consumers, and grasping rare consumption opportunities to improve conversion probability through timely interest stimulation. This time dimension adaptive adjustment mechanism based on consumer value stratification significantly improves the model's understanding depth and prediction accuracy for different types of consumers, providing more scientific algorithm support for precise marketing.

[0154] For better understanding and implementation of the present application, the following provides an embodiment 2 of a specific application scenario of the present application: researchers implemented a location-aware based dynamic benefit issuing system in a large-scale comprehensive shopping center with an area of 45000 square meters. The shopping center contains 200 merchants, distributed in 5 floors, with an average daily passenger flow of about 25000 people. The system realizes accurate benefit push by obtaining real-time location data of consumers and combining historical consumption behavior analysis. The experiment lasted for 90 days, with a total of 15000 consumers participating, and the system cumulatively pushed 78540 times of benefit information. The benefit issuing in this embodiment refers to, for example, coupons, discount coupons, and cash vouchers of shops, etc. sent to consumers through mobile terminals

[0155] The system first collects consumer location data through mobile terminals, applies a local weighted scatter smoothing algorithm for processing, and sets the smoothing parameter σ to 15 meters and the neighborhood size to 5 sampling points. According to the statistical distribution of data update frequency, as shown in Table 1:

[0156] Table 1 Position data update frequency distribution table

[0157] Update frequency interval Consumer proportion Set sampling interval > 0.2 Hz (within 5 seconds) 42.7％ 10 seconds 0.033-0.2 Hz (5-30 seconds) 38.5％ 30 seconds <0.033 Hz (30 seconds or more) 18.8％ 60 seconds

[0158] The system applies an improved density clustering algorithm to the collected location data for resident point identification, with the parameter settings of neighborhood radius ∈ = 25 meters, minimum point number MinPts = 3, and time threshold ΔT = 120 seconds. Through analysis, the average residence time statistics of consumers in different types of areas are obtained, as shown in Table 2:

[0159] Table 2 Average residence time statistics of consumers in different areas

[0160] Area type Average dwell time (minutes) Dwell frequency (within 7 days) Location dwell intensity index Dining area 42.8 3.2 136.96 Clothing area 28.6 2.7 77.22 Electronic product area 35.4 1.9 67.26 Supermarket area 38.7 3.5 135.45 Cinema area 126.5 1.2 151.8 Leisure and entertainment area 56.3 2.4 135.12

[0161] The system calculates the position correlation degree according to the distance vector between the residence position of the consumer and each merchant and the wireless network connection condition. In the experiment, the weight coefficients α and β are set to 0.6 and 0.4 respectively. Combined with historical consumption data analysis, a consumption tendency model is constructed, and the comparison of the model prediction results with the actual consumption situation is shown in Table 3:

[0162] Table 3 Comparison table of consumption tendency model prediction results

[0163] Consumer grouping Sample size Model predicted consumption probability Actual consumption proportion Prediction accuracy rate High-value consumer 2850 0.687 0.712 92.3％ Medium-value consumer 8420 0.437 0.452 89.6％ Low-value consumer 3730 0.183 0.195 86.2％

[0164] According to the output of the spatio-temporal consumption behavior prediction model, the system sets a dynamic benefit issuing threshold based on the density of commercial circle passenger flow, and the multi-head attention layer in the spatio-temporal consumption behavior prediction model is set to 8 attention heads, with an output dimension of 32 for each attention head. Under different passenger flow density conditions, the benefit issuing situation of the system is shown in Table 4:

[0165] Table 4 Statistics of benefit distribution under different human flow densities

[0166]

[0167] The system realizes differentiated prediction and push strategies for consumers in different value intervals by adjusting the time sensitivity coefficient. During the experiment, the benefit response of the three types of consumers is shown in Table 5:

[0168] Table 5 Benefit response of consumers in different value intervals

[0169]

[0170] The traditional business benefit distribution method mainly relies on fixed rules and time nodes, such as membership level, birthday, holiday, etc., and lacks real-time location and behavior perception ability. This method often leads to mismatch between benefits and current consumer needs, with a generally low usage rate of 10%-15%. The dynamic benefit distribution method based on location perception implemented by the present application captures consumer location behavior in real time, analyzes their residence patterns, combines historical consumption habits and merchant characteristics, and establishes a spatio-temporal consumption behavior prediction model to achieve precise benefit distribution.

[0171] The experimental results show that the method of the present application improves the overall benefit usage rate to 24.864% (weighted average of Table 4 data), which is about 100% higher than the traditional method. Especially for high-value consumer groups, the benefit usage rate reaches 45.8%, with an increase of more than 200%. At the same time, by introducing the dynamic adjustment mechanism of the time sensitivity coefficient, the system can adopt differentiated time-sensitive strategies according to the consumer value interval, effectively improving the prediction accuracy of the model, with a prediction accuracy of 92.3% for high-value consumers. In addition, the dynamic threshold adjustment mechanism based on the flow density of the business district enables the system to adaptively respond to changes in passenger flow in different scenarios, achieving a maximum conversion rate of 36.0% in high-density flow environments. Compared with the traditional fixed rule benefit distribution method, the method of the present application realizes the transformation from "goods-centered" to "consumer-centered", greatly improving the accuracy and effectiveness of benefit distribution.

[0172] It should be noted that the variables involved in the present application are explained in detail as shown in Tables 6 and 7.

[0173] Table 6 Variable Explanation Table (First Part)

[0174]

[0175]

[0176] Table 7 Variable Explanation Table (Second Part)

[0177]

[0178]

[0179] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for dynamic benefit delivery based on location awareness, the method comprising: The application comprises the following steps: Collecting real-time location data of consumers to establish a location trajectory graph; obtaining residence time and residence point location coordinates to calculate a location residence intensity index; Obtaining location coordinates of a merchant to construct a commercial district location matrix and calculate a distance vector; Collecting wireless network identifiers connected by a consumer's mobile terminal, determining the spatial location of the consumer in the merchant space in combination with the distance vector, and generating a location correlation degree; analyzing consumer behavior characteristics by using a spatio-temporal consumer behavior prediction model to establish a consumer tendency model; combining the location residence intensity index and the location correlation degree, inputting them into the spatio-temporal consumer behavior prediction model, calculating a merchant attraction index, and performing weighted calculation with the matching degree of the consumer tendency model; setting a benefit distribution threshold according to the weighted calculation result, and dynamically adjusting the benefit distribution threshold parameter by the spatio-temporal consumer behavior prediction model according to the commercial district flow density factor; calculating a judgment value by a consumer value judgment function in combination with the matching degree of the consumer tendency model, the location correlation degree, and the merchant attraction index, and adjusting the time sensitivity coefficient in the spatio-temporal consumer behavior prediction model according to the judgment value; selecting a benefit type suitable for the consumer to determine the benefit content; pushing the benefit information to the consumer and recording the use condition to update the model; The specific structure of the spatio-temporal consumer behavior prediction model is a multi-head spatio-temporal attention network structure, which comprises a spatial feature extraction layer, a time series encoding layer, a multi-head attention layer, and a prediction output layer. The spatial feature extraction layer adopts a graph convolution network to capture the spatial relationship characteristics of consumers in a commercial district. The time series encoding layer adopts a long short-term memory network to process historical behavior time series data of consumers. The multi-head attention layer combines the location residence intensity index and the historical behavior sequence of consumers to generate an attention score matrix to perform weighted processing on spatio-temporal features. The prediction output layer combines all features to output the final consumer behavior prediction result. The consumer value judgment function is a mathematical function for calculating the value score of a consumer by comprehensively evaluating the historical value, current location information, and consumption intention of the consumer. The judgment value is divided into three intervals: high value interval, medium value interval, and low value interval. The time sensitivity coefficient is used to adjust the time decay rate of the long short-term memory unit in the time series encoding layer. 2.The location-aware based dynamic benefit delivery method of claim 1, wherein, The location trajectory graph is a continuous trajectory data set formed by recording the coordinates of multiple location points of a consumer within a certain time period, which is used to analyze the movement mode and residence habit of the consumer. 3.The location-aware based dynamic benefit delivery method of claim 2, wherein, In the step of collecting real-time location data of consumers, the time sampling interval is set as the preferred sampling interval according to the location data update frequency. The preferred sampling interval refers to the optimal time sampling unit automatically set by the system according to the location data update frequency, which is used to balance the location data collection accuracy and system resource consumption. 4.The location-aware based dynamic benefit delivery method of claim 3, wherein, The location residence intensity index is a comprehensive quantitative indicator of the residence time and frequency of a consumer in a certain area. The longer the residence time and the higher the frequency, the larger the location residence intensity index.

5. The method of claim 4, wherein, The distance vector refers to a multi-dimensional vector composed of spatial distances from the current location of the consumer to the location points of each merchant in the commercial district, used to determine the physical proximity of the consumer to each merchant; the location correlation degree refers to a quantitative indicator of the close spatial relationship between the consumer and a certain merchant, calculated based on the wireless network signal strength and the distance vector. 6.The location-aware based dynamic benefit delivery method of claim 5, wherein, The consumer tendency model refers to a prediction model constructed based on the historical consumer behavior, preference characteristics and time regularity of the consumer, used to evaluate the probability of the consumer generating a consumer behavior in a certain scenario; the merchant attraction index refers to a quantitative indicator of the attraction degree of a merchant to a certain consumer, calculated in combination with the location residence intensity index and the location correlation degree.

7. The location-aware based dynamic benefit delivery method of claim 6, wherein, The benefit issuance threshold refers to the minimum standard value triggering the benefit issuance mechanism, and when the weighted calculation result exceeds the benefit issuance threshold, the system determines that the consumer has a higher consumer possibility; the benefit issuance threshold parameter is dynamically adjusted by the spatio-temporal consumer behavior prediction model according to the commercial district passenger flow density factor. 8.The location-aware based dynamic benefit delivery method of claim 7, wherein, The commercial district passenger flow density factor refers to the passenger flow per unit area in a commercial district within a certain period of time, used to reflect the busy degree and potential consumer activity of the commercial district.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores program instructions, and when the program instructions are run in the computer, the program instructions are used to execute the method of claim 1-8.

10. A dynamic benefit delivery system based on location awareness, comprising: The system is any one of a computer, a server, and a single-chip microcomputer, and the computer readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer readable storage medium.

Images