A dynamic distribution method in catering precision marketing

Abstract

The application relates to the technical field of Internet advertisement service, and discloses a dynamic distribution method in catering precise marketing, which comprises the following steps: splicing space-time state data, historical consumption mapping data and environmental fluctuation parameters to construct a three-dimensional input feature tensor; separating static history and dynamic high-frequency dimensions through an orthogonal feature decoupling network to generate micro-intention vectors and offset vectors which are mutually orthogonal; calculating an initial matching degree based on a basic distribution benchmark value and the micro-intention vectors; adding a dynamic compensation amount to the initial matching degree according to the offset vectors and a sensitivity constant to obtain a dynamic distribution weight, and cutting off and setting to zero when the weight is negative; rearranging candidate advertisements according to the weight and transmitting a target advertisement set; and the application reduces environmental noise interference, reduces decision lag, improves hit precision and prevents power dissipation.

Description

Technical Field

[0001] This invention belongs to the field of internet advertising service technology, and in particular relates to a dynamic distribution method in precise marketing for the catering industry. Background Technology

[0002] Existing dynamic advertising distribution systems typically establish a matching relationship between user characteristics and candidate advertising content based on users' historical behavior data, search interaction sequences, and static consumption tags. During advertising distribution, the system generally organizes relevant user and advertising characteristics according to a fixed offline period and converts these characteristics into low-dimensional vectors. This type of method usually assumes that users' consumption intentions and short-term decision-making characteristics will not change significantly within a certain period of time, and then completes the matching, sorting, and push of candidate advertising content accordingly. In the context of precise marketing in the catering industry, the effectiveness of advertising is often closely related to the user's current time, location, and surrounding environment. When the real-time geographical location of the target audience, the weather in the area, or traffic congestion suddenly changes, the user's dining needs may also change accordingly. Existing models usually directly merge rapidly changing environmental fluctuation parameters with relatively stable historical preference features, which can easily cause environmental changes to interfere with the original matching relationship. This results in the feature extraction results not reflecting the user's current needs in a timely manner, and the catering advertisements delivered may deviate from the target audience's current consumption intentions, resulting in the ineffective use of advertising server computing resources and communication network bandwidth.

[0003] Besides data mining methods based on static labels, the industry has also tried to introduce various graph models and algorithm matching software-level methods to improve recommendation performance. However, these control methods also have significant shortcomings in dealing with transient environmental changes. For example, Chinese invention patent CN116739836B discloses a knowledge graph-based catering data analysis method and system. It extracts big data from catering to construct a knowledge graph and combines on-site video facial recognition with historical data to recommend dishes and display marketing information. However, this method is essentially still in the category of static modeling that uses high-dimensional graph retrieval of user historical preferences and relational features. It lacks the mechanism for stripping away and dynamically orthogonally decoupling high-frequency transient environmental fluctuation factors such as weather and traffic in the user's real-time location. In the process of big data and complex graph structure splicing retrieval, high-frequency environmental noise is prone to cross-interference with stable historical preferences, which not only leads to a lengthy feature calculation link and a high degree of response delay, but also increases computing power consumption and server latency in high-concurrency scenarios, making it difficult to meet the objective working conditions of millisecond-level dynamic and accurate distribution.

[0004] Therefore, how to reduce the interference of environmental changes on historical preference characteristics, adjust the advertising distribution results in a timely manner according to environmental changes, and maintain low response latency and stable operation under high concurrency conditions are problems that existing dynamic distribution technologies for catering advertisements need to solve. Summary of the Invention

[0005] This invention aims to solve the problems of lagging feature extraction and excessive computational resource consumption in existing dynamic distribution architectures under transient environmental fluctuations.

[0006] In this technical solution, a dynamic distribution method for precision marketing in the catering industry includes the following steps:

[0007] Step S1: Obtain the real-time spatiotemporal state data, historical consumption mapping data and environmental fluctuation parameters of the target object, and concatenate and construct a three-dimensional input feature tensor in the matrix space;

[0008] Step S2: Input the three-dimensional input feature tensor into the preset orthogonal feature decoupling network, strip away the static historical dimension whose variance change rate is lower than the preset steady-state threshold, extract the dynamic high-frequency dimension data whose variance change rate is higher than the preset steady-state threshold, generate the micro-intent vector, and at the same time extract the numerical sequence representing the magnitude of environmental factor anomalies, generate the offset vector that is linearly orthogonal to the micro-intent vector.

[0009] Step S3: Obtain the basic distribution benchmark value of each candidate restaurant advertisement in the candidate restaurant advertisement content library, calculate the cosine similarity between the micro intent vector and the feature identifier vector of each candidate restaurant advertisement, multiply the cosine similarity by the basic distribution benchmark value, and calculate the initial matching degree of each candidate restaurant advertisement.

[0010] Step S4: Based on the offset vector and the preset sensitivity constant vector, the dynamic compensation amount is added to the initial matching degree to generate the dynamic distribution weight under the current time slice. The sensitivity constant vector is determined by the mapping relationship between the catering category and the environmental parameters. When the dynamic distribution weight is less than 0, the dynamic distribution weight is forcibly truncated to 0, and the data packet of the corresponding candidate catering advertisement content is terminated.

[0011] Step S5: Sort the candidate catering advertisement content library in descending order according to the dynamic distribution weight, select the target advertisement content set, and encapsulate the target advertisement content set into a packet and transmit it to the target's terminal device via the communication network.

[0012] Preferably, the method further includes a high-concurrency adaptive degradation control step: Step S201, when the concurrent request frequency detected by the server cluster exceeds the preset throughput threshold, the dynamic compensation calculation in step S4 is automatically suspended; Step S202, the initial matching degree calculated in step S3 is assigned to the dynamic distribution weight, and the truncation and zeroing step when the dynamic distribution weight is less than 0 is skipped; Step S203, based on the assigned dynamic distribution weight, the descending sorting and distribution of the candidate catering advertisement content library is completed in a degraded manner to keep the latency response of the server cluster within a safe range of less than 10ms.

[0013] Preferably, step S1, which involves acquiring the real-time spatiotemporal state data, historical consumption mapping data, and environmental fluctuation parameters of the target object, and then assembling and constructing a three-dimensional input feature tensor in the matrix space, includes the following steps: Step S11, collecting the real-time geographic location coordinates, timestamps, and retrieval action sequences of the target object as real-time spatiotemporal state data; Step S12, extracting the category click frequency and historical conversion rate of the target object within a preset historical period as historical consumption mapping data; Step S13, collecting the real-time weather index and traffic congestion index of the area where the target object is currently located as environmental fluctuation parameters; Step S14, performing numerical normalization on the real-time spatiotemporal state data, historical consumption mapping data, and environmental fluctuation parameters, and assembling and constructing a fixed-size three-dimensional input feature tensor in the matrix space, wherein the dimension of the real-time spatiotemporal state data is 3, the dimension of the historical consumption mapping data is 4, and the dimension of the environmental fluctuation parameters is 2.

[0014] Preferably, the extraction of the micro-intent vector and offset vector by the orthogonal feature decoupling network in step S2 includes the following steps: Step S21, calculating the variance change rate of the three-dimensional input feature tensor in each dimension data channel; Step S22, comparing the variance change rate with a preset steady-state threshold, stripping the static historical dimensions with variance change rates lower than the preset steady-state threshold, and projecting the extracted dynamic high-frequency dimension data with variance change rates higher than the preset steady-state threshold in a low dimension to generate a micro-intent vector composed of one-dimensional values; Step S23, extracting the numerical sequences of weather variation amplitude and traffic variation amplitude representing the amplitude of environmental factor anomalies, generating an offset vector, and constraining the micro-intent vector and offset vector to satisfy a linear orthogonal relationship with a dot product of 0 in the decoupling space through orthogonal transformation.

[0015] Preferably, the calculation of the initial matching degree in step S3 includes the following steps: Step S31, read the basic distribution benchmark value of the candidate catering advertisement content from the advertising database. The basic distribution benchmark value represents the global historical average conversion rate of the candidate catering advertisement content; Step S32, extract the category attribute feature vector of the candidate catering advertisement content as the feature identifier vector; Step S33, calculate the cosine value of the vector angle between the micro intent vector and the feature identifier vector to obtain the cosine similarity; Step S34, perform a scalar multiplication operation between the cosine similarity and the basic distribution benchmark value to obtain the initial matching degree.

[0016] Preferably, step S4, which involves forcibly truncating and setting the dynamic distribution weight to 0 when it is less than 0, and terminating the packet encapsulation, includes the following steps: Step S41, determining the positive or negative polarity of the calculated dynamic distribution weight; Step S42, triggering a hard zeroing mechanism when the dynamic distribution weight is less than 0, assigning the dynamic distribution weight to 0; Step S43, removing the candidate catering advertisement content with a dynamic distribution weight of 0 from the candidate queue for distribution under the current time slice, and canceling the serialization of the network communication protocol packet corresponding to the candidate catering advertisement content.

[0017] Preferably, step S5, which involves selecting the target advertising content set and transmitting the target advertising content set via a communication network, includes the following steps: Step S51, rearranging the candidate restaurant advertising content library from largest to smallest according to dynamic distribution weights; Step S52, extracting the top 5 candidate restaurant advertising contents after the rearrangement to form the target advertising content set; Step S53, concatenating the data payload of the target advertising content set with the terminal device identifier and time slice checksum of the target object, assembling it into a network distribution data packet, and pushing the network distribution data packet to the terminal device via the communication network.

[0018] Preferably, the preset steady-state threshold and sensitivity constant vector are dynamically updated through the following steps: Step S801, collect historical distribution log data and click conversion feedback data of target objects according to a preset time sliding window; Step S802, calculate the variance of conversion rate fluctuation for different catering categories under different weather indices and traffic congestion indices, and recalibrate the sensitivity constant vector; Step S803, recalculate and set the preset steady-state threshold based on the quantile of the variance distribution of the overall feature dimension, and synchronize the updated preset steady-state threshold and the updated sensitivity constant vector to the distribution database.

[0019] Preferably, when obtaining the environmental fluctuation parameter in step S1, if a data timeout occurs during the reception of the environmental fluctuation parameter, a data loss protection step is triggered: Step S901, read the historical normal environmental average value of the current geographical area pre-stored on the server and fill it into the environmental fluctuation parameter as a replacement value; Step S902, continue to construct the three-dimensional input feature tensor based on the filled environmental fluctuation parameter to ensure the non-blocking continuous operation of the dynamic distribution method.

[0020] Compared with existing technologies, the dynamic distribution method for precise marketing in the catering industry proposed in this invention has the following advantages:

[0021] 1. In the dynamic distribution of precise marketing in the catering industry, the orthogonal feature decoupling network is used to process the relatively stable historical consumption preferences and the rapidly changing environmental information respectively. This can reduce the interference of short-term changes such as weather and traffic on the judgment of users' historical preferences. The micro-intent vector and the offset vector are kept linearly orthogonal, making the user intent information on which the ad matching is based clearer. This is conducive to improving the accuracy of catering ad matching results and maintaining the stability of the feature calculation process.

[0022] 2. The initial matching degree is dynamically compensated based on the offset vector and the sensitivity constant vector, so that the advertising distribution results can reflect the impact of environmental changes such as weather and traffic on different catering categories in a timely manner, and reduce the distribution lag caused by fixed historical data. When the dynamic distribution weight is negative, the invalid transmission of obviously mismatched advertisements can be reduced by truncating and setting it to zero and stopping the packetization and distribution of the corresponding advertising content, thereby reducing the waste of server computing resources and communication network bandwidth.

[0023] 3. By using high-concurrency adaptive degradation control, when concurrent requests exceed the server's processing capacity, dynamic compensation calculation is temporarily stopped, and the initial matching degree is used to continue to complete the ad sorting and distribution. This method can reduce the amount of calculation under high concurrency, avoid complex calculations from continuously occupying server resources, and maintain a low response latency while retaining basic ad matching capabilities, thereby reducing the risk of service blocking or system instability. Attached Figure Description

[0024] Figure 1 This is a flowchart of the dynamic distribution process for restaurant advertisements according to the present invention;

[0025] Figure 2 This is the dynamic distribution degradation control diagram for catering advertisements in this invention. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0027] A dynamic distribution method for precision marketing in the catering industry includes the following steps:

[0028] Step S1: Obtain the real-time spatiotemporal state data, historical consumption mapping data and environmental fluctuation parameters of the target object, and concatenate and construct a three-dimensional input feature tensor in the matrix space;

[0029] Step S2: Input the three-dimensional input feature tensor into the preset orthogonal feature decoupling network, strip away the static historical dimension whose variance change rate is lower than the preset steady-state threshold, extract the dynamic high-frequency dimension data whose variance change rate is higher than the preset steady-state threshold, generate the micro-intent vector, and at the same time extract the numerical sequence representing the magnitude of environmental factor anomalies, generate the offset vector that is linearly orthogonal to the micro-intent vector.

[0030] Step S3: Obtain the basic distribution benchmark value of each candidate restaurant advertisement in the candidate restaurant advertisement content library, calculate the cosine similarity between the micro intent vector and the feature identifier vector of each candidate restaurant advertisement, multiply the cosine similarity by the basic distribution benchmark value, and calculate the initial matching degree of each candidate restaurant advertisement.

[0031] Step S4: Based on the offset vector and the preset sensitivity constant vector, the dynamic compensation amount is added to the initial matching degree to generate the dynamic distribution weight under the current time slice. The sensitivity constant vector is determined by the mapping relationship between the catering category and the environmental parameters. When the dynamic distribution weight is less than 0, the dynamic distribution weight is forcibly truncated to 0, and the data packet of the corresponding candidate catering advertisement content is terminated.

[0032] Step S5: Sort the candidate catering advertisement content library in descending order according to the dynamic distribution weight, select the target advertisement content set, and encapsulate the target advertisement content set into a packet and transmit it to the target's terminal device via the communication network.

[0033] Preferably, the method further includes a high-concurrency adaptive degradation control step: Step S201, when the concurrent request frequency detected by the server cluster exceeds the preset throughput threshold, the dynamic compensation calculation in step S4 is automatically suspended; Step S202, the initial matching degree calculated in step S3 is assigned to the dynamic distribution weight, and the truncation and zeroing step when the dynamic distribution weight is less than 0 is skipped; Step S203, based on the assigned dynamic distribution weight, the descending sorting and distribution of the candidate catering advertisement content library is completed in a degraded manner to keep the latency response of the server cluster within a safe range of less than 10ms.

[0034] Preferably, step S1, which involves acquiring the real-time spatiotemporal state data, historical consumption mapping data, and environmental fluctuation parameters of the target object, and then assembling and constructing a three-dimensional input feature tensor in the matrix space, includes the following steps: Step S11, collecting the real-time geographic location coordinates, timestamps, and retrieval action sequences of the target object as real-time spatiotemporal state data; Step S12, extracting the category click frequency and historical conversion rate of the target object within a preset historical period as historical consumption mapping data; Step S13, collecting the real-time weather index and traffic congestion index of the area where the target object is currently located as environmental fluctuation parameters; Step S14, performing numerical normalization on the real-time spatiotemporal state data, historical consumption mapping data, and environmental fluctuation parameters, and assembling and constructing a fixed-size three-dimensional input feature tensor in the matrix space, wherein the dimension of the real-time spatiotemporal state data is 3, the dimension of the historical consumption mapping data is 4, and the dimension of the environmental fluctuation parameters is 2.

[0035] Preferably, the extraction of the micro-intent vector and offset vector by the orthogonal feature decoupling network in step S2 includes the following steps: Step S21, calculating the variance change rate of the three-dimensional input feature tensor in each dimension data channel; Step S22, comparing the variance change rate with a preset steady-state threshold, stripping the static historical dimensions with variance change rates lower than the preset steady-state threshold, and projecting the extracted dynamic high-frequency dimension data with variance change rates higher than the preset steady-state threshold in a low dimension to generate a micro-intent vector composed of one-dimensional values; Step S23, extracting the numerical sequences of weather variation amplitude and traffic variation amplitude representing the amplitude of environmental factor anomalies, generating an offset vector, and constraining the micro-intent vector and offset vector to satisfy a linear orthogonal relationship with a dot product of 0 in the decoupling space through orthogonal transformation.

[0036] Preferably, the calculation of the initial matching degree in step S3 includes the following steps: Step S31, read the basic distribution benchmark value of the candidate catering advertisement content from the advertising database. The basic distribution benchmark value represents the global historical average conversion rate of the candidate catering advertisement content; Step S32, extract the category attribute feature vector of the candidate catering advertisement content as the feature identifier vector; Step S33, calculate the cosine value of the vector angle between the micro intent vector and the feature identifier vector to obtain the cosine similarity; Step S34, perform a scalar multiplication operation between the cosine similarity and the basic distribution benchmark value to obtain the initial matching degree.

[0037] Preferably, step S4, which involves forcibly truncating and setting the dynamic distribution weight to 0 when it is less than 0, and terminating the packet encapsulation, includes the following steps: Step S41, determining the positive or negative polarity of the calculated dynamic distribution weight; Step S42, triggering a hard zeroing mechanism when the dynamic distribution weight is less than 0, assigning the dynamic distribution weight to 0; Step S43, removing the candidate catering advertisement content with a dynamic distribution weight of 0 from the candidate queue for distribution under the current time slice, and canceling the serialization of the network communication protocol packet corresponding to the candidate catering advertisement content.

[0038] Preferably, step S5, which involves selecting the target advertising content set and transmitting the target advertising content set via a communication network, includes the following steps: Step S51, rearranging the candidate restaurant advertising content library from largest to smallest according to dynamic distribution weights; Step S52, extracting the top 5 candidate restaurant advertising contents after the rearrangement to form the target advertising content set; Step S53, concatenating the data payload of the target advertising content set with the terminal device identifier and time slice checksum of the target object, assembling it into a network distribution data packet, and pushing the network distribution data packet to the terminal device via the communication network.

[0039] Preferably, the preset steady-state threshold and sensitivity constant vector are dynamically updated through the following steps: Step S801, collect historical distribution log data and click conversion feedback data of target objects according to a preset time sliding window; Step S802, calculate the variance of conversion rate fluctuation for different catering categories under different weather indices and traffic congestion indices, and recalibrate the sensitivity constant vector; Step S803, recalculate and set the preset steady-state threshold based on the quantile of the variance distribution of the overall feature dimension, and synchronize the updated preset steady-state threshold and the updated sensitivity constant vector to the distribution database.

[0040] Preferably, when obtaining the environmental fluctuation parameter in step S1, if a data timeout occurs during the reception of the environmental fluctuation parameter, a data loss protection step is triggered: Step S901, read the historical normal environmental average value of the current geographical area pre-stored on the server and fill it into the environmental fluctuation parameter as a replacement value; Step S902, continue to construct the three-dimensional input feature tensor based on the filled environmental fluctuation parameter to ensure the non-blocking continuous operation of the dynamic distribution method.

[0041] Example 1: When rainfall occurs in the commercial area where the target object is located, accompanied by widespread road congestion, the advertising distribution window has strong short-term timeliness and spatial relevance. After the advertising distribution server cluster receives the real-time geographical coordinates, timestamps and retrieval action sequences of the target object, if the environmental fluctuation parameters are directly mixed with the static historical dimensions, high-frequency environmental changes will interfere with the basic matching relationship, causing feature extraction delays. This prevents the fluctuation amplitude of environmental factors from being converted into compensation parameters for dynamic distribution weights in a timely manner, and increases the transmission of non-matching data and network bandwidth consumption.

[0042] To isolate environmental noise and perform dynamic compensation calculations within the orthogonal decoupling space, in this embodiment, the advertising distribution server cluster collects the real-time geographic location coordinates, timestamps, and retrieval action sequences of the target object as real-time spatiotemporal state data; extracts the category click frequency and historical conversion rate of the target object within a preset historical period as historical consumption mapping data; and collects the real-time weather index and traffic congestion index of the target object's current location as environmental fluctuation parameters. After the three sets of data are normalized, the three real-time spatiotemporal components are used as row indices, the four historical consumption components as column indices, and the two environmental components as channel indices. A correspondence is established according to the fixed order of rows, columns, and channels to form a three-dimensional input feature tensor with 3 rows, 4 columns, and 2 channels. Any index position in the tensor records the ordered combination of the corresponding real-time spatiotemporal component, historical consumption component, and environmental component, so that each component can be read according to the original index in subsequent decoupling operations.

[0043] The ad distribution server cluster inputs a 3D input feature tensor into a preset orthogonal feature decoupling network. This network includes an input layer, a channel variance calculation layer, a static and dynamic feature stripping module, a low-dimensional projection fully connected layer, and an environment anomaly coding branch. The channel variance calculation layer calculates the variance of each data channel in the current window and the previous window within a preset time sliding window, and uses the ratio of the difference between the variance of the current window and the variance of the previous window to the variance of the previous window as the variance change rate. The static and dynamic feature stripping module compares the variance change rate element-wise with a preset steady-state threshold τ = 0.15, stripping static historical dimensions with variance change rates lower than the preset steady-state threshold τ, and retaining variance change rates higher than the preset steady-state threshold. The dynamic, high-frequency dimensional data is used. The low-dimensional projection fully connected layer contains 128 hidden neurons and uses a linear activation function. The retained data is first unfolded according to a fixed data channel order, and then linear projection is performed according to the connection weights of the fully connected layer. The output is a micro-intent vector composed of 3 floating-point values. .

[0044] The environmental anomaly coding branch calculates the difference in weather index and traffic congestion index between adjacent time slices to obtain a numerical sequence characterizing the magnitude of environmental factor anomalies. In this embodiment, the two-dimensional numerical sequence composed of the magnitude of rainfall intensity anomaly and the magnitude of traffic flow decline is [0.50, -0.20]. After obtaining this two-dimensional numerical sequence through a two-node linear transformation, the environmental anomaly coding branch adds a third component to it. This third component is determined according to the condition that its dot product with the micro-intent vector is 0, generating an offset vector with the same dimension as the micro-intent vector. The orthogonal feature decoupling network applies orthogonal constraints to the output layer through the Schmitt orthogonalization operator, ensuring that the micro-intent vectors... With offset vector Maintaining linear orthogonality within the decoupling space, the first two components of the offset vector correspond to the weather variation amplitude and traffic variation amplitude, respectively, while the third component is used to satisfy the linear orthogonality relationship.

[0045] After feature decoupling is completed, the ad distribution server cluster reads the basic distribution benchmark value of hot pot restaurant category ad item A1 from the candidate restaurant ad content library in the ad database. This baseline distribution value represents the global historical average conversion rate of candidate restaurant advertising content. The server simultaneously reads the three category attribute feature values ​​of advertising item A1 and arranges them in the same component order as the micro intent vector to form a feature identifier vector. Calculate the micro-intention vector With feature identifier vector The cosine of the angle between the vectors is used to obtain the cosine similarity. Compare the cosine similarity with the baseline distribution value. Perform scalar multiplication to obtain the initial matching degree. .

[0046] The server reads the sensitivity constant vector of hot pot catering categories to weather and traffic from the sensitivity table. Based on the current rate of change of intent, a nonlinear adjustment factor α=1.20 is determined. The dynamic distribution weight W under the current time slice is calculated by adding dynamic compensation. To ensure that the dimensions of the sensitivity constant vector and the offset vector are consistent, a component with a value of 0 is added to the end of the sensitivity constant vector before performing the inner product. This component does not change the sensitivity values ​​corresponding to weather and traffic. The dynamic distribution weight is calculated according to the following formula: , where W is the dynamic distribution weight, which is a dimensionless probability factor; The initial matching degree is given by α, which is a non-linear adjustment factor. It is an offset vector in row vector form, whose first two components represent the magnitude of the change in environmental factors, and the third component is used to satisfy the linear orthogonality relationship; This is a constant vector representing the sensitivity of environmental fluctuation parameters to food and beverage categories. Sensitivity constant vector The transpose matrix, since the sensitivity value corresponding to the supplemented component is 0, the inner product of the offset vector and the sensitivity constant vector is still calculated as two environmental components, i.e., 0.50×0.30+(-0.20)×0.40=0.07; multiply this inner product by the nonlinear adjustment factor α, and then divide it by the initial matching degree. Adding them together, we get the dynamic distribution weight W = 0.447 + 1.20 × 0.07 = 0.531. The increase in this weight reflects the positive impact of rainfall on the potential conversion rate of hot pot catering categories.

[0047] The current intent change rate is determined by the ratio of the target object's search action frequency to the geographical location movement distance in the last 5 minutes. It is used to characterize the degree of change in consumption demand with time and location. When the intent change rate is between 0.20 and 0.50 per minute, the server obtains a non-linear adjustment factor of 1.20 from the mapping table. When the intent change rate is less than 0.20 per minute, the target object's consumption intent is in a stable state, and the non-linear adjustment factor is set to 0.80 to reduce the impact of high-frequency environmental noise on dynamic distribution weights. When the intent change rate is greater than 0.50 per minute, the target object is in a short-term decision-making state, and the non-linear adjustment factor is set to 1.50 to increase the dynamic compensation amount corresponding to sudden environmental changes.

[0048] For ad item A2 in the category of frozen desserts and cold drinks, the initial match degree is: The sensitivity constant vectors for weather and traffic are: Under rainfall and cooling conditions, the offset vector The first two environmental components are [-0.80, -0.60]. The third component is determined according to the condition that the dot product with the corresponding micro-intent vector is 0. The nonlinear adjustment factor is α=1.50. Substituting the above values ​​into the dynamic distribution weight calculation formula, we get 0.100+1.50×[(-0.80)×0.50+(-0.60)×0.50]=-0.950. Since the dynamic distribution weight is less than 0, the server triggers a hard zeroing mechanism, assigns the dynamic distribution weight of ad item A2 to 0, removes it from the current time slice's distribution candidate queue, and cancels the corresponding network communication protocol packet serialization.

[0049] When the concurrent request frequency detected by the server cluster reaches 12,000 requests per second, exceeding the preset throughput threshold of 10,000 requests per second, the server suspends the dynamic compensation calculation and directly adjusts the initial matching degree. The dynamic distribution weight W is assigned, skipping the truncation and zeroing step when the dynamic distribution weight is less than 0. The candidate restaurant advertisement content library is sorted and distributed in descending order according to the assigned dynamic distribution weight, keeping the response latency of the server cluster within 10ms. The initial matching degree is obtained by scalar multiplication of the vector cosine similarity in the range of 0.00 to 1.00 and the non-negative global historical average conversion rate. Therefore, the assigned dynamic distribution weight is greater than or equal to 0. Skipping the truncation and zeroing step can save the negative value judgment instruction in the high concurrency state and reduce the microprocessor CPU execution cycle. Since the candidate restaurant advertisement content is still sorted according to the initial matching degree, this process does not change the original filtering relationship and will not add the advertisement content corresponding to the negative weight to the distribution queue.

[0050] The server reorders candidate restaurant advertisements from largest to smallest according to dynamic distribution weights, extracts the top 5 candidate restaurant advertisements to form the target advertisement content set, and pushes this set to the target audience's terminal devices through the communication network. The data payload of the target advertisement content set corresponds to the environmental fluctuation parameters of the current time slice and the target audience's dynamic dining intentions, which can reduce the distribution delay caused by changes in spatiotemporal state, improve the advertisement matching degree within a short time window, and reduce the occupation of network bandwidth and computing resources by non-matching data. When traffic increases transiently, the multi-dimensional feature mapping architecture composed of orthogonal feature decoupling and dynamic parameter compensation continues to maintain stable data flow.

[0051] Example 2: The current distributed advertising distribution simulation cluster platform has 8 independent computing nodes, each with a memory throughput of 64GB per second and a network bandwidth capacity of 10Gbps. In this example, the dynamic distribution method in the precision marketing of catering is run on this platform. The data acquisition terminal accesses the catering placement historical log dataset containing physical node trajectories, and Gaussian white noise with a signal-to-noise ratio of 20dB is superimposed on the data stream to simulate high-frequency transient jitter in the network transmission layer. The sampling rate of the real-time environmental fluctuation parameter is set to 50,000 times per second so that the sampling resolution of the input data stream meets the calculation requirements of the variance change rate.

[0052] The nonlinear adjustment factor α is based on the offset vector The magnitude of environmental factor fluctuations is determined. When the magnitude of environmental factor fluctuations increases, the nonlinear adjustment factor α is increased to increase the dynamic compensation amount; when the environmental fluctuation parameters tend to stabilize, the nonlinear adjustment factor α is decreased to reduce the impact of high-frequency noise on the dynamic distribution weight W. Under the combined conditions of rainfall and local traffic congestion, the nonlinear adjustment factor α, the preset steady-state threshold τ, and the offset vector are adjusted respectively. With sensitivity constant vector The intermediate feature inner product value is set, and the dynamic distribution weight W, response latency, and distribution hit rate are recorded.

[0053] The concurrent request frequency of Experiment Group 1 was 5000 times per second, the nonlinear adjustment factor α was taken as the normal median value of 1.20, the preset steady-state threshold τ was taken as 0.15, the inner product of intermediate features was 0.07, after adding dynamic compensation to the initial matching degree, the dynamic distribution weight W was 0.531, the response latency was 3.2ms, and the distribution hit rate was 94.6%.

[0054] In Experiment 2, the concurrent request frequency was maintained at 5000 times per second, the preset steady-state threshold τ was 0.15, and the intermediate feature inner product value was 0.07. The nonlinear adjustment factor α was adjusted to the lower limit of the working range of 0.80. The dynamic distribution weight W obtained by dynamic compensation calculation was 0.503, the response latency was 3.1ms, and the distribution hit rate was 92.4%.

[0055] Experiment group three also set the concurrent request frequency to 5000 times per second and preset the steady-state threshold. The value was set to 0.15, the intermediate feature inner product value was set to 0.07, and the nonlinear adjustment factor α was adjusted to the upper limit of the working range of 1.50. After dynamic compensation, the dynamic distribution weight W was 0.552, the response delay was 3.4ms, and the distribution hit rate was 95.8%. The data from test groups one to three show that when the nonlinear adjustment factor α is within the set working range, the dynamic distribution weight changes with the compensation amplitude, and the distribution hit rate remains in the corresponding stable range.

[0056] Experimental group four was used to test the negative environmental compensation and hard zeroing mechanism. The concurrent request frequency of this group was 5000 times per second, the nonlinear adjustment factor α was 1.50, the preset steady-state threshold τ was 0.15, the intermediate feature inner product value was -0.70, and after the dynamic compensation calculation produced a negative result, the system truncated the dynamic distribution weight W to 0.000 and stopped packetizing the corresponding candidate catering advertisement content. The response latency of this group was 3.3ms, and the distribution hit rate was 91.2%.

[0057] Experiment group 5 was used to test high-concurrency adaptive degradation control. The concurrent request frequency in this group was 12,000 requests per second, exceeding the throughput threshold of 10,000 requests per second. The nonlinear adjustment factor α was 1.20, the preset steady-state threshold τ was 0.15, and the intermediate feature inner product value was 0.07. After the server cluster detected that the concurrent request frequency exceeded the throughput threshold, it suspended the dynamic compensation calculation and adjusted the initial matching degree according to the following formula. Assigning values ​​to dynamic distribution weights The assigned dynamic distribution weight W is 0.447. The system uses this weight to rearrange and distribute candidate restaurant advertisements, with a response latency of 8.6ms and a distribution hit rate of 89.1%.

[0058] In control group 1, the concurrent request frequency was maintained at 5000 times per second, the preset steady-state threshold τ was 0.15, and the intermediate feature inner product value was 0.07. The nonlinear adjustment factor α was reduced to 0.70. At this time, the dynamic distribution weight W was 0.496, the response latency was 3.1ms, and the distribution hit rate was 73.1%. This value is lower than the set working range, the dynamic compensation amount corresponding to environmental fluctuations is reduced, and short-term environmental changes are not fully reflected in the ranking results of candidate catering advertisement content.

[0059] Control group 2 set the concurrent request frequency to 5000 times per second, the preset steady-state threshold τ to 0.15, the intermediate feature inner product value to 0.07, and increased the nonlinear adjustment factor α to 1.60. The dynamic distribution weight W was 0.559, the response latency was 3.6ms, and the distribution hit rate was 68.5%. This value is higher than the set working range. High-frequency noise is amplified along with the dynamic compensation amount, causing the dynamic distribution weight to deviate from the actual matching relationship of the candidate catering advertisement content.

[0060] The control group removed the orthogonal feature decoupling network and directly fused the dynamic high-frequency dimension data with the static historical dimension. The concurrent request frequency of this group was 5000 times per second, the nonlinear adjustment factor α was 1.20, the preset steady-state threshold τ was 0.15, the inner product of intermediate features was 0.07, and the dynamic distribution weight W was 0.531. Due to the mutual interference between environmental fluctuation parameters and historical consumption mapping data in the same feature space, the response latency increased to 14.5ms, and the distribution hit rate dropped to 51.3%.

[0061] Control group four continued to perform dynamic compensation calculations when the concurrent request frequency was 12,000 times per second, without triggering high-concurrency adaptive degradation control. The nonlinear adjustment factor α of this group was 1.20, the preset steady-state threshold τ was 0.15, the intermediate feature inner product value was 0.07, and the dynamic distribution weight W was 0.531. After the server cluster continuously performed feature matrix calculations, the response latency reached 45.2ms, and the distribution hit rate was 93.2%. Compared with experimental group five, the high-concurrency adaptive degradation control reduced the amount of matrix calculations under high concurrency by suspending dynamic compensation calculations and directly assigning dynamic distribution weights, allowing ad reordering and distribution to continue running with lower response latency.

[0062] The above data corresponds to the construction of the three-dimensional input feature tensor and the micro-intent vector. With offset vector Orthogonal extraction based on sensitivity constant vector The process of dynamic compensation, negative weight truncation, and high-concurrency adaptive degradation control involves the environmental fluctuation parameter being orthogonally decoupled to form a correction amount for the dynamic distribution weight. Candidate catering advertisements are sorted according to the corrected weight. When the frequency of concurrent requests exceeds the throughput threshold, the server cluster switches to the degradation sorting driven by the initial matching degree, thereby reducing the response latency caused by continuous execution of dynamic compensation calculations.

[0063] Example 3: In this example, a preset steady-state threshold is used. The value range is 0.10 to 0.20, the operating range of the nonlinear adjustment factor α is 0.8 to 1.5, and the sensitivity constant vector... The values ​​of each dimension component range from 0.10 to 0.60. The input end of the distributed advertising distribution simulation cluster platform is connected to a historical log dataset of catering placements containing the physical trajectory of the target object. Gaussian white noise with a signal-to-noise ratio of 20dB is superimposed on the input data stream to test the impact of different parameter values ​​on the feature separation results of the orthogonal feature decoupling network and the computational latency of the server cluster.

[0064] When the preset steady-state threshold When set to 0.10, some transient channel environmental noise with low variance change rate is included in the dynamic high-frequency dimension data, and the micro-intent vector output by the orthogonal feature decoupling network is... Noise aliasing occurs, initial matching degree The calculation accuracy decreased, and the distribution hit rate was 82.1%, when the preset steady-state threshold was reached. When set to 0.20, some low-frequency consumer preference features with variance change rates near the boundary are stripped away, making the micro-intention vector... The distribution hit rate was 84.3%, deviating from the actual preference attributes of the target object.

[0065] When the preset steady-state threshold is τ=0.15 and the sensitivity constant vector is... At that time, the orthogonal feature decoupling network removes the static historical dimensions whose variance change rate is lower than a preset steady-state threshold τ, and extracts the dynamic high-frequency dimension data whose variance change rate is higher than the threshold, generating a micro-intent vector composed of one-dimensional numerical values. The environmental anomaly coding branch simultaneously extracts the magnitude of weather and traffic changes, resulting in a numerical sequence consisting of 0.50 and -0.20.

[0066] To make the numerical sequence of environmental factor anomalies correlate with the micro-intention vector Having the same feature dimensions, the orthogonal feature decoupling network appends a channel with a value of 0.00 to the end of the numerical sequence through a dimension expansion mapping operator, forming an orthogonal transformation input consisting of 0.50, -0.20, and 0.00. This input is then compared with the micro-intent vector. Element-wise multiplication is performed at corresponding positions. The product of the first dimension is 0.40, the product of the second dimension is -0.12, and the product of the third dimension is 0.00. Since the cumulative result of the products of each dimension is not zero, the network adjusts the values ​​of the first two dimensions using an orthogonal correction matrix, and uses the adjusted three-dimensional values ​​as the offset vector. Output, making the micro-intent vector With offset vector The sum of the product of the three dimensions of the vector equals 0, thus satisfying the linear orthogonality relation that the dot product is 0. This process makes the two vectors have the same dimension, allowing the dot product operation to be performed directly.

[0067] After feature decoupling is completed, the server cluster reads the basic distribution benchmark value of ad item A1 from the candidate restaurant ad content library. And utilize feature identifier vectors With micro-intention vector Calculate the cosine of the vector angle to obtain the initial matching degree. When the nonlinear adjustment factor α is set to 1.20, the dynamic parameter compensation operator is based on the offset vector. With sensitivity constant vector Adding the dynamic compensation amount, we get the dynamic distribution weight W=0.531.

[0068] In the above processing, the magnitude of rainfall intensity fluctuations and the magnitude of traffic congestion decreases are processed via offset vectors. The modified parameter is converted into a dynamic distribution weight W. The server cluster rearranges and distributes candidate catering advertisements according to this weight. The overall system calculation latency is kept within 3.2ms. The network bus bandwidth occupation of non-matching data is reduced accordingly. The feature rearrangement and distribution accuracy are kept within the preset operating specifications.

[0069] Example 4: This example combines Figures 1 to 2 This paper explains a dynamic distribution method in precision marketing for the catering industry, such as... Figure 1 As shown, step S1 obtains the real-time spatiotemporal state data, historical consumption mapping data, and environmental fluctuation parameters of the target object, and concatenates and constructs a three-dimensional input feature tensor in the matrix space. Step S2 inputs the three-dimensional input feature tensor into the orthogonal feature decoupling network, strips off the static historical dimension and extracts the dynamic high-frequency dimension data, and generates mutually orthogonal micro-intent vectors and offset vectors. Step S3 calculates the cosine similarity between the micro-intent vector and the feature identifier vector of each candidate catering advertisement content, and multiplies it with the basic distribution benchmark value to generate the initial matching degree. Step S4 adds a dynamic compensation amount to the initial matching degree based on the offset vector and the sensitivity constant vector to generate dynamic distribution weights. When the weight is less than 0, it is forcibly truncated to zero and the packet is terminated. Step S5 sorts the candidate catering advertisement content library in descending order according to the dynamic distribution weights, selects the target advertisement content set, and transmits its packet to the terminal device via the communication network.

[0070] like Figure 2As shown, when receiving environmental fluctuation parameters, a data timeout determination is performed. If a data timeout occurs, the historical average value of the normal environment is read and a three-dimensional input feature tensor is constructed. If a data timeout does not occur, the three-dimensional input feature tensor is directly constructed. After constructing the three-dimensional input feature tensor, the initial matching degree is calculated and generated. Then, a determination is made that the concurrent request frequency exceeds the threshold. If the concurrent request frequency exceeds the threshold, the dynamic compensation calculation is suspended, the truncation and zeroing step is skipped, and the delivery action is completed in a downgraded manner. If the concurrent request frequency exceeds the threshold, the dynamic compensation amount is added, and then a determination is made that the dynamic distribution weight is less than zero. If the dynamic distribution weight is less than zero, the packet is truncated and set to zero and the packet is terminated. If the dynamic distribution weight is not less than zero, the target advertising content set is transmitted.

[0071] Example 5: During the initial deployment of the distributed advertising distribution simulation cluster platform, 50,000 historical distribution logs and click-conversion feedback data accumulated over the past 30 days from the target commercial area are imported and categorized according to food and beverage categories such as hot pot and ice cream. The computing nodes calculate the variance of conversion rate fluctuations for each food and beverage category under different weather index levels and traffic congestion index levels. Then, the least squares method is used to fit the influence weights of each environmental fluctuation parameter on the conversion rate, and the sensitivity constant vector is calibrated accordingly. The sensitivity constant vector for hot pot catering is calibrated as [0.30, 0.40], and the sensitivity constant vector for ice cream and cold drinks is calibrated as [0.50, 0.50]. Within the data sampling sliding window, the computing nodes simultaneously extract the anomaly amplitude sequence of multi-source spatiotemporal environmental parameters, and determine the value range of the corresponding parameters through the convergence rule of the first derivative extreme point.

[0072] The computation node sorts the variance change rate of all feature channels in ascending order, calculates its 80th percentile, and sets 0.15 corresponding to this quantile as the preset steady-state threshold. The rate of change of variance is lower than the preset steady-state threshold. Feature channels are assigned to the static historical dimension, while feature channels with a variance change rate higher than this threshold are used to extract dynamic high-frequency dimension data. After initial calibration, the platform assigns sensitivity constant vectors to the hot pot catering category and the ice cream and cold beverage category. and preset steady-state threshold Write it into the distribution database as a parameter benchmark for subsequent dynamic compensation and feature stripping.

[0073] During the online operation of the ad distribution server cluster, the computing nodes read historical distribution log data and target object click conversion feedback data generated within the current period according to a preset time sliding window period. They then recalculate the variance of conversion rate fluctuations for different food and beverage categories under different weather and traffic congestion indices, and recalibrate the sensitivity constant vector accordingly. Simultaneously, the computing nodes recalculate the quantiles of the variance distribution for the overall feature dimensions; when the rate of change of variance deviates and exceeds the preset fluctuation tolerance, the system calls the reconstruction operator to adjust the preset steady-state threshold. The initial value of 0.15 is corrected to a new scalar value that matches the current variance distribution of the feature channels.

[0074] After recalibration is complete, the system will update the sensitivity constant vector. and preset steady-state threshold The data is synchronously written to the distribution database. In subsequent time slices, the orthogonal feature decoupling network reads the updated parameters and performs static historical dimension stripping and dynamic high-frequency dimension data extraction. This allows the parameters to be periodically adjusted according to changes in the target object retrieval sequence, weather index, and traffic congestion index, while maintaining the linear orthogonal relationship between the micro-intent vector and the offset vector.

[0075] Example 6: When the advertising distribution server cluster collects real-time spatiotemporal status data and environmental fluctuation parameters of the target object, network bus jitter or communication channel overload may cause the environmental fluctuation parameters to arrive late. The data acquisition module continuously records the request time and reception time of the environmental fluctuation parameters. When the time difference between the two exceeds the preset communication timeout threshold of 50ms, it is determined that the environmental fluctuation parameters have timed out and the data loss protection step is triggered.

[0076] In the data loss protection step, the processor addresses the pre-stored data in the local cache memory according to the regional identifier of the current geographic region, and reads the historical normal environmental average of the region. The historical normal environmental average includes the average weather index and the average traffic congestion index. The processor uses these two as substitute values ​​and fills them into the corresponding weather index dimension and traffic congestion index dimension of the environmental fluctuation parameter. After filling, the data acquisition module continues to perform numerical splicing according to the predetermined arrangement order of real-time spatiotemporal status data, historical consumption mapping data and environmental fluctuation parameter, maintaining a fixed size of 3 rows, 4 columns and 2 channels, to construct a three-dimensional input feature tensor, so that the subsequent orthogonal feature decoupling network can still continuously read complete input data in the event of environmental fluctuation parameter timeout.

[0077] The orthogonal feature decoupling network further strips features from the 3D input feature tensor after the replacement values ​​are added, generating micro-intent vectors and offset vectors. The ad distribution server cluster calculates dynamic distribution weights based on the offset vectors, sensitivity constant vectors, and initial matching degrees, and sorts the candidate catering ad content library from largest to smallest according to the dynamic distribution weights.

[0078] After sorting, the distribution control module extracts the top 5 candidate restaurant advertisements to form the target advertisement content set. The server then writes the data payload of the target advertisement content set, the terminal device identifier of the target object, and the current time slice checksum into the network distribution data packet in sequence. The current time slice checksum is generated by the timestamp of the current time slice according to the preset check field in the communication protocol, and is used to verify the time slice corresponding to the network distribution data packet after the terminal device receives the data.

[0079] The communication protocol stack assembles the network distribution data packets at the serialization layer and sends them to the target terminal device through the network interface card. The sorting, assembly and sending of the network distribution data packets are completed within a 10ms scheduling cycle. The terminal device receives the target advertising content set corresponding to the current time slice based on the current time slice check code, thereby maintaining the continuous operation of the advertising distribution process when environmental fluctuation parameters cause data timeouts.

[0080] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A dynamic distribution method for precision marketing in the catering industry, characterized in that, Includes the following steps: Step S1: Obtain the real-time spatiotemporal state data, historical consumption mapping data and environmental fluctuation parameters of the target object, and concatenate and construct a three-dimensional input feature tensor in the matrix space; Step S2: Input the three-dimensional input feature tensor into the preset orthogonal feature decoupling network, strip away the static historical dimension whose variance change rate is lower than the preset steady-state threshold, extract the dynamic high-frequency dimension data whose variance change rate is higher than the preset steady-state threshold, generate the micro-intent vector, and at the same time extract the numerical sequence representing the magnitude of environmental factor anomalies, generate the offset vector that is linearly orthogonal to the micro-intent vector. Step S3: Obtain the basic distribution benchmark value of each candidate restaurant advertisement in the candidate restaurant advertisement content library, calculate the cosine similarity between the micro intent vector and the feature identifier vector of each candidate restaurant advertisement, multiply the cosine similarity by the basic distribution benchmark value, and calculate the initial matching degree of each candidate restaurant advertisement. Step S4: Based on the offset vector and the preset sensitivity constant vector, the dynamic compensation amount is added to the initial matching degree to generate the dynamic distribution weight under the current time slice. The sensitivity constant vector is determined by the mapping relationship between the catering category and the environmental parameters. When the dynamic distribution weight is less than 0, the dynamic distribution weight is forcibly truncated to 0, and the data packet of the corresponding candidate catering advertisement content is terminated. Step S5: Sort the candidate catering advertisement content library in descending order according to the dynamic distribution weight, select the target advertisement content set, and encapsulate the target advertisement content set into a packet and transmit it to the target's terminal device via the communication network.

2. The dynamic distribution method in precision marketing for catering according to claim 1, characterized in that, The method also includes a high-concurrency adaptive degradation control step: Step S201, when the concurrent request frequency detected by the server cluster exceeds the preset throughput threshold, the dynamic compensation calculation in step S4 is automatically suspended; Step S202, the initial matching degree calculated in step S3 is assigned to the dynamic distribution weight, and the truncation and zeroing step when the dynamic distribution weight is less than 0 is skipped; Step S203, based on the assigned dynamic distribution weight, the descending sorting and distribution of the candidate catering advertisement content library is completed in a degraded manner to keep the latency response of the server cluster within a safe range of less than 10ms.

3. The dynamic distribution method in precision marketing for catering according to claim 1, characterized in that, Step S1 involves acquiring the real-time spatiotemporal state data, historical consumption mapping data, and environmental fluctuation parameters of the target object, and then assembling and constructing a three-dimensional input feature tensor in the matrix space. This includes the following steps: Step S11: Collect the real-time geographic location coordinates, timestamps, and retrieval action sequences of the target object as real-time spatiotemporal state data; Step S12: Extract the category click frequency and historical conversion rate of the target object within a preset historical period as historical consumption mapping data; Step S13: Collect the real-time weather index and traffic congestion index of the area where the target object is currently located as environmental fluctuation parameters; Step S14: Perform numerical normalization on the real-time spatiotemporal state data, historical consumption mapping data, and environmental fluctuation parameters, and assemble and construct a fixed-size three-dimensional input feature tensor in the matrix space, wherein the dimension of the real-time spatiotemporal state data is 3, the dimension of the historical consumption mapping data is 4, and the dimension of the environmental fluctuation parameters is 2.

4. The dynamic distribution method in precision marketing for catering according to claim 1, characterized in that, Step S2, the orthogonal feature decoupling network extracts the micro-intent vector and offset vector, including the following steps: Step S21, calculate the variance change rate of the three-dimensional input feature tensor in each dimension data channel; Step S22, compare the variance change rate with a preset steady-state threshold, remove the static historical dimensions with variance change rates lower than the preset steady-state threshold, and perform low-dimensional projection on the extracted dynamic high-frequency dimension data with variance change rates higher than the preset steady-state threshold to generate a micro-intent vector composed of one-dimensional values; Step S23, extract the numerical sequences of weather variation amplitude and traffic variation amplitude representing the amplitude of environmental factor anomalies, generate an offset vector, and constrain the micro-intent vector and offset vector to satisfy a linear orthogonal relationship with a dot product of 0 in the decoupling space through orthogonal transformation.

5. The dynamic distribution method in precision marketing for catering according to claim 1, characterized in that, Step S3 involves calculating the initial matching degree using the following steps: Step S31, read the basic distribution benchmark value of the candidate restaurant advertisement content from the advertising database. The basic distribution benchmark value represents the global historical average conversion rate of the candidate restaurant advertisement content; Step S32, extract the category attribute feature vector of the candidate restaurant advertisement content as the feature identifier vector; Step S33, calculate the cosine value of the angle between the micro intent vector and the feature identifier vector to obtain the cosine similarity; Step S34, perform a scalar multiplication operation between the cosine similarity and the basic distribution benchmark value to obtain the initial matching degree.

6. The dynamic distribution method in precision marketing for catering according to claim 1, characterized in that, Step S4, when the dynamic distribution weight is less than 0, forcibly truncates and sets it to 0 and terminates the packet encapsulation, includes the following steps: Step S41, perform boundary determination on the positive and negative polarities of the calculated dynamic distribution weight; Step S42, when the dynamic distribution weight is less than 0, trigger the hard zeroing mechanism and assign the dynamic distribution weight to 0; Step S43, remove the candidate catering advertisement content with a dynamic distribution weight of 0 from the candidate queue for distribution under the current time slice, and cancel the serialization of the network communication protocol packet corresponding to the candidate catering advertisement content.

7. The dynamic distribution method in precision marketing for catering according to claim 1, characterized in that, Step S5 involves selecting the target advertising content set and transmitting the target advertising content set via a communication network, which includes the following steps: Step S51, rearranging the candidate restaurant advertising content library from largest to smallest according to the dynamic distribution weight; Step S52, extracting the top 5 candidate restaurant advertising contents after the rearrangement to form the target advertising content set; Step S53, concatenating the data payload of the target advertising content set with the terminal device identifier and time slice check code of the target object, assembling it into a network distribution data packet, and pushing the network distribution data packet to the terminal device through the communication network.

8. The dynamic distribution method in precision marketing for catering according to claim 1, characterized in that, The preset steady-state threshold and sensitivity constant vector are dynamically updated through the following steps: Step S801, collect historical distribution log data and click conversion feedback data of target objects according to a preset time sliding window; Step S802, calculate the variance of conversion rate fluctuation for different catering categories under different weather indices and traffic congestion indices, and recalibrate the sensitivity constant vector; Step S803, recalculate and set the preset steady-state threshold based on the quantile of the variance distribution of the overall feature dimension, and synchronize the updated preset steady-state threshold and the updated sensitivity constant vector to the distribution database.

9. The dynamic distribution method in precision marketing for catering according to claim 1, characterized in that, If a data timeout occurs when receiving environmental fluctuation parameters in step S1, a data loss protection step is triggered: Step S901, read the historical normal environmental average value of the current geographic area pre-stored on the server and fill it into the environmental fluctuation parameters as a replacement value; Step S902, continue to construct the three-dimensional input feature tensor based on the filled environmental fluctuation parameters to ensure the non-blocking continuous operation of the dynamic distribution method.

Images