A smart vending cabinet product advertising recommendation management system

Abstract

The present invention relates to the technical field of vending machine advertising management, and specifically to an intelligent vending machine commodity advertising recommendation management system. By calculating the variance and entropy of the behavioral feature set and the environmental feature set, the complexity of the behavioral distribution and the complexity of the environment are described to dynamically adjust the dimensions of the projection weight matrix and the key. By optimizing the dimension of the key, the dimension of the model can be adjusted according to the complexity of the environment and behavioral data, the accuracy of the model in multimodal data fusion is improved, and the personalization of advertising recommendations is improved; by Gaussian process modeling click-through rate, using the Bayesian optimization algorithm, effective exploration is carried out in advertising recommendations, so that the vending machine can continuously adjust the recommendation strategy according to the actual effect of the advertisement, and combined with the multi-armed bandit algorithm, the optimal display strategy is found between known good performing advertisements and unknown potential effects advertisements, and the accuracy and efficiency of advertising recommendations are gradually improved, thereby realizing the optimization of recommended advertisements.

Description

Technical Field

[0001] The present invention relates to the technical field of vending machine advertising management, and in particular to an intelligent vending machine commodity advertising recommendation management system. Background Art

[0002] With the widespread application of smart vending machines, their advertising function has gradually attracted attention. Traditional advertising playback methods are mostly fixed modes, which cannot be accurately pushed according to users' real-time behavior and environmental changes, resulting in poor advertising effects:

[0003] For example, consumers’ purchasing preferences and acceptance of advertisements vary in different time periods and weather conditions. If the vending machine cannot capture this information, the pushed advertisement resources will be seriously wasted. In addition, the complexity of the environment around the vending machine, such as lighting, temperature and other factors, will also affect the display effect of the advertisement and the user’s viewing experience, and simultaneously affect the effect of the push of the vending machine advertisements.

[0004] In addition, there are a large number of push advertisements in the current vending machine database. How to combine the above factors with the actual status of the ads to be pushed in the current database to achieve accurate push of advertisements and reduce resource waste in order to find the most appropriate ad push display strategy in the complex and ever-changing advertising market is a problem that needs to be solved at present. Summary of the invention

[0005] In view of the above-mentioned shortcomings of the prior art, the present invention provides a smart vending machine product advertising recommendation management system, which can effectively solve the problem in the prior art that the advertising push display strategy cannot be improved.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0007] The present invention provides a smart vending cabinet commodity advertisement recommendation management system, which at least comprises:

[0008] The data collection and analysis module is used to collect behavioral data sets and environmental data sets and establish behavioral feature sets B and environmental feature sets C. The behavioral feature sets B and environmental feature sets C are spliced ​​to construct a multimodal input matrix Xt, and input projection is performed to calculate the attention weight matrix And get multimodal features ;

[0009] Dynamically adjust key dimensions based on environmental and behavioral characteristics , to achieve the attention weight matrix Optimization;

[0010] Advertisement recommendation analysis module, based on Determine the similarity between the current user and other users with cosine similarity ,in accordance with determining recommended ads;

[0011] The advertising recommendation management module models the click-through rate of advertisements through Gaussian processes and calculates the upper confidence bound of advertisements based on the predicted mean and standard deviation of advertisements. , to optimize the recommendation of advertisements, including:

[0012] The length scale, signal variance, and noise variance in the Gaussian process are used to construct a set of hyperparameters. The covariance matrix is ​​constructed and the log marginal likelihood function is calculated. The hyperparameters are updated using L-BFGS. The optimal value is determined by calculating the gradient of each hyperparameter to achieve optimization of the length scale, signal variance, and noise variance.

[0013] Furthermore, the behavior feature set B includes commodity category, purchase time, and purchase frequency; and the environment feature set C includes temperature, humidity, and light.

[0014] Furthermore, the multimodal feature The construction method is:

[0015] Execute X t Input projection of:

[0016]

[0017] Among them, Q t , K t 、V t Represents the feature matrices of query, key and value respectively, used to calculate self-attention, W Q , W K and W V The projection weight matrices representing query, key, and value respectively;

[0018] Calculate the attention weight matrix :

[0019]

[0020] Among them, softmax represents the normalization function, represents the attention weight distribution, represents the dimension of the key, represents the transposed inner product of the query feature matrix and the key feature matrix;

[0021] Multimodal features .

[0022] Furthermore, dynamic adjustment The method is:

[0023] Calculate the variance of the behavioral feature set B and entropy :

[0024]

[0025] in, Indicates behavioral characteristics The probability of occurrence, Represents the total number of elements in the behavior feature set B;

[0026] Calculate the variance of the environmental feature set C and the rank of the correlation matrix :

[0027]

[0028] Relevance matrix representing the set of environmental features;

[0029]

[0030] in, , They represent the corresponding weight coefficients, , They represent the coefficients used to adjust the impact of behavioral characteristics on key dimensions, , They represent the coefficients used to adjust the impact of environmental characteristics on the key dimensions.

[0031] in accordance with Recommended ads are determined by:

[0032] Based on the calculated similarity , select the K most similar users as similar users, and calculate the recommendation score of each advertisement for the current user based on the historical behaviors of similar users, then:

[0033]

[0034] in, Indicates the current user of the prediction About Advertising Interest score, represents a set of K other users similar to the current user, Indicates other users About Advertising Ratings;

[0035] Set a score threshold, and ads that exceed the score threshold will be identified and pushed.

[0036] Furthermore, the method for optimizing the recommendation of advertisements is to set an initial state for each advertisement to facilitate subsequent advertisement selection and reward update, which is:

[0037] Defining Ad Sets , N represents the number of advertisements, represents the i-th advertisement;

[0038] For each ad , confirm the ad Number of impressions N i , confirm the ad Number of clicks , get the click rate ;

[0039] The click-through rate of an ad is modeled through a Gaussian process, and we have:

[0040] Reward function Indicates advertisement The click rate return follows a Gaussian process:

[0041]

[0042] in, represents a Gaussian process, represents the mean function, represents the kernel function value;

[0043] A1. Perform Gaussian process prediction steps, including:

[0044] According to Gaussian process Predicted mean With standard deviation ,

[0045] ;

[0046] Where X represents the set of advertising data points, represents the kernel function matrix between all advertisements in the advertisement data point set X, represents the noise variance, I represents the identity matrix, Indicates advertisement Collection of advertising data points The kernel function value of all advertisements in , Represents all advertisements and advertisements in the advertisement data point set X Column vector of kernel function values;

[0047] A2. Calculate value:

[0048]

[0049] in, represents the upper confidence bound of the ith advertisement, represents the hyperparameter term;

[0050] A3. Update the click-through rate of ads :

[0051] Advertisement Get feedback after tth display , update ads Impressions and clicks for:

[0052] ;

[0053] Based on updated click-through rate Update the Gaussian process prediction step to get a new prediction mean With standard deviation , calculate the advertisement selected in the next round value;

[0054] A4. Update-based value, select the maximum The ads with the same value are optimized for the next round and steps A1-A3 are repeated to update Value, based on Value for recommended playback.

[0055] Furthermore, the kernel function value According to the following expression:

[0056]

[0057] in, represents the signal variance, Indicates the length scale.

[0058] Furthermore, the method for optimizing the length scale, signal variance, and noise variance is:

[0059] Initialize the hyperparameter set ;

[0060] Construct the covariance matrix K:

[0061]

[0062] in, represents the elements in the covariance matrix K, represents the noise term;

[0063] Compute the log marginal likelihood function :

[0064]

[0065] in, represents the inverse matrix, express The transpose of represents the target variable of the advertisement, represents the determinant of the covariance matrix, Indicates the number of advertisements;

[0066] Compute the gradient:

[0067] for Derivative, get the gradient :

[0068]

[0069] in, represents the weight of the posterior distribution, represents the trace sum of the matrix, express The transpose of

[0070] Determine the derivative form of the hyperparameters;

[0071] By updating each hyperparameter through the gradient formula, we have:

[0072]

[0073] in, Indicated in The hyperparameter vector at iteration , Indicated in The hyperparameter vector at iteration , represents the learning rate;

[0074] Update hyperparameters:

[0075]

[0076] in, Represents the approximate inverse matrix of the Hessian matrix at the current iteration;

[0077] After each iteration, check whether the convergence condition is met. If the norm of the gradient is less than the preset threshold, or if the change of the log-likelihood function is less than the set threshold, the optimization process ends and the optimized length scale is obtained. , signal variance , Noise Variance .

[0078] Furthermore, the derivative form of the hyperparameters is as follows:

[0079] Length scale :

[0080]

[0081] For the signal variance :

[0082]

[0083] For the noise variance :

[0084] .

[0085] Compared with the known prior art, the technical solution provided by the present invention has the following beneficial effects:

[0086] By calculating the variance and entropy of the behavioral feature set and the environmental feature set, the complexity of the behavioral distribution and the complexity of the environment are described, so as to dynamically adjust the dimensions of the projection weight matrix and the key, so that the matching between the query and the key is more relevant. By optimizing the dimension of the key, the dimension of the model can be adjusted according to the complexity of the environment and behavioral data, so as to improve the accuracy of the model in multimodal data fusion and enhance the personalization of advertising recommendations.

[0087] By modeling click-through rate through Gaussian process and using Bayesian optimization algorithm, effective exploration is carried out in advertising recommendation, so that the vending machine can continuously adjust the recommendation strategy according to the actual effect of the advertisement. In combination with the multi-armed bandit algorithm, the optimal display strategy is found between advertisements with known good performance and advertisements with unknown potential effects, gradually improving the accuracy and efficiency of advertising recommendations, thereby achieving the optimization of recommended advertisements. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0089] Figure 1 It is the overall module block diagram of the present invention. DETAILED DESCRIPTION

[0090] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0091] The present invention will be further described below in conjunction with the embodiments.

[0092] Example 1 (see Figure 1 ): A smart vending cabinet product advertising recommendation management system, comprising at least:

[0093] The data collection and analysis module uses sensors and cameras to collect the user's behavior data set and environmental data set when purchasing goods in the vending machine, including:

[0094] Collect the environmental characteristics of the vending machine, including temperature, humidity, light, etc., and establish the environmental characteristic set C;

[0095] Obtain the current user's purchase behavior characteristics from the sales counter records, including product category, purchase time, purchase frequency, etc., and establish a purchase behavior feature set B;

[0096] The behavior feature set B and the environment feature set C are concatenated to form a multimodal input matrix X t ;

[0097] Execute X t The input projection is:

[0098]

[0099] Among them, Q t , K t 、V t Represents the feature matrices of query, key and value respectively, used to calculate self-attention, W Q , W K and W V The projection weight matrices representing query, key, and value respectively;

[0100] Calculate the attention weight matrix :

[0101]

[0102] Among them, softmax represents the normalization function, which is used to map each row of the weight matrix to Range, indicating the attention weight distribution, Indicates the dimension of the key, used for scaling operations to avoid values ​​that are too large. represents the transposed inner product of the query feature matrix and the key feature matrix, generating an attention weight matrix representing the relevance between each query and all keys;

[0103] Therefore, the fused multimodal features are obtained The multimodal data fusion method can more comprehensively understand the user's preferences and needs in specific situations, thereby providing a more accurate basis for advertising recommendations. For example, when a user buys cold drinks in hot weather, the system can capture this behavioral feature and combine it with the ambient temperature data to accurately push cold drink-related advertisements, thereby increasing the attractiveness of the advertisements and the user's click-through rate.

[0104] Dynamically adjust based on environmental and behavioral characteristics , making the match between the query and the key more relevant, then we have,

[0105] Calculate the variance of the current user behavior feature set B and entropy , describing the behavior distribution characteristics:

[0106]

[0107] in, Indicates behavioral characteristics The probability of occurrence, Represents the total number of elements in the behavior feature set B;

[0108] Calculate the variance of the environmental feature set C and the rank of the correlation matrix , describing the complexity of the environment:

[0109]

[0110] The correlation matrix representing the set of environmental features describes the correlation between different environmental features;

[0111] Therefore, the dimension of the key is dynamically adjusted :

[0112]

[0113] in, , They represent the corresponding weight coefficients, , They represent the coefficients used to adjust the impact of behavioral characteristics on key dimensions, , They represent the coefficients used to moderate the effects of environmental characteristics on the key dimensions;

[0114] By optimizing , which can adjust the model's query, key, and value dimensions according to the complexity of behavioral and environmental characteristics, improve the model's accuracy in multimodal data fusion, and enhance the expressive power of the attention mechanism by performing feature analysis, dynamically adjusting key dimensions, and optimizing the projection weight matrix through this process.

[0115] It also includes an advertising recommendation analysis module based on the obtained multimodal features. , recommending advertisements for vending machines based on the behaviors of similar users, including:

[0116] Calculate the similarity between the current user and other users , then:

[0117]

[0118] in, and For current users and other users The characteristic vector of and They represent the corresponding Euclidean norms, respectively, and represent the size of the vector. The similarity of interests between users is measured by cosine similarity. The closer the similarity value is to 1, the more similar it is, and the closer the value is to 0, the lower the similarity is;

[0119] Then, based on the calculated user similarity, we select the K other users who are most similar to the user as similar users for recommendation. Then, we can calculate the recommendation score of each advertisement for the current user based on the historical behavior of similar users. Then, we have:

[0120]

[0121] in, Indicates the current user of the prediction About Advertising Interest score, represents a set of K other users similar to the current user, Indicates other users About Advertising Rating, The higher the score, the more likely the advertisement is to be accepted by the user. Therefore, all advertisements in the vending machine can be sorted, and a score threshold can be set. Advertisements exceeding the score threshold can be identified and pushed so that the vending machine can play them according to the actual situation.

[0122] The advertisement recommendation management module, based on the determined advertisement, combines the Bayesian optimization model with the multi-armed bandit algorithm to optimize the selection of vending machine advertisements, including the following steps:

[0123] Set the initial state for each ad to facilitate subsequent ad selection and return updates, then:

[0124] Defining Ad Sets , N represents the number of advertisements, represents the i-th advertisement;

[0125] For each ad , confirm the ad Impressions , confirm the ad Number of clicks C i , get the click rate (Click-through rate is a key indicator of advertising returns, indicating the probability of an ad being clicked after being displayed). as well as is 0;

[0126] The click-through rate of an ad is modeled through a Gaussian process, and we have:

[0127] Reward function Indicates advertisement The click rate return follows a Gaussian process:

[0128]

[0129] in, represents a Gaussian process, represents the mean function, describing the advertisement The expected value of the return, Represents the kernel function value, which measures the value of two ads and The similarity between them, where the kernel function value According to the following expression:

[0130]

[0131] in, represents the signal variance, represents the length scale, controlling the similarity between advertisements;

[0132] Further, A1, perform Gaussian process prediction step, that is, for advertisement , predicting the mean according to the Gaussian process With standard deviation ,

[0133] ;

[0134] Where X represents a set of advertising data points (describing the features and return data of the ads that have been displayed in the ad set A. Ad features are descriptive information of ads, such as the content, target audience, and delivery period of ads. Return data include click-through rate and conversion rate after ad display). represents the kernel function matrix between all advertisements in the advertisement data point set X, represents the noise variance, describing the observation noise of the data, and I represents the unit matrix, which is used to adjust the influence of the noise and ensure that the kernel matrix is ​​reversible after adding the noise variance. Indicates advertisement Collection of advertising data points The kernel function value of all advertisements in , Represents all advertisements and advertisements in the advertisement data point set X Column vector of kernel function values ​​describing the ad Similarity with all historical advertising data points;

[0135] Furthermore, A2, when selecting advertisements, we should not only pay attention to the predicted mean of the advertisements, but also consider the standard deviation of the advertising returns. Calculate its value:

[0136]

[0137] in, represents the upper confidence bound of the ith advertisement, describing the upper bound of the predicted revenue of the advertisement or action. Represents a hyperparameter item, which is used to control the balance between exploration and utilization. It will increase the intensity of exploration, making the system choose more uncertain ads for trial; reduce They will be more inclined to choose ads that are currently known to perform well;

[0138] Furthermore, after each ad is displayed, the number of ad impressions and clicks is updated based on the feedback (click or no click), and the Bayesian optimization model is updated using this data, including:

[0139] A3. Update the click-through rate of ads (Report data):

[0140] Hypothetical Advertisement Get feedback after tth display ,and 1 means click, 0 means no click, that is, update the ad Impressions and clicks for:

[0141] ;

[0142] Thus, the updated click rate is obtained again , you can use the updated click rate Update the Gaussian process prediction step to get a new prediction mean With standard deviation , from which the advertisement selected in the next round can be calculated Value, based on the updated value, select the value with the largest The ads with the highest click-through rate will be optimized for the next round, and steps A1-A3 will be repeated continuously. Update the Gaussian process prediction step and update value, select the optimal ad (maximum Value) is displayed to ensure a balance between exploration and utilization in advertising recommendation. Therefore, the Bayesian optimization model is combined with the multi-armed bandit algorithm to optimize the selection of vending machine advertisements. The advertisement click-through rate is modeled through the Gaussian process. The predicted mean and standard deviation are comprehensively considered when selecting advertisements, and the upper confidence limit of the advertisement is calculated to balance exploration and utilization. This enables the vending machine to find the best balance between constantly trying new advertisements and using advertisements with known good performance, and gradually improve the efficiency, accuracy and optimization effect of advertisement selection, so as to find the most suitable advertisement display strategy in the complex and ever-changing advertising market.

[0143] Furthermore, for the Gaussian process , , Solving, we have:

[0144] Initialize the hyperparameter set , defines the covariance structure of the Gaussian process;

[0145] Construct the covariance matrix K:

[0146]

[0147] in, Represents the elements in the covariance matrix K, describing and The covariance value between represents the noise term;

[0148] Compute the log marginal likelihood function :

[0149]

[0150] in, represents the inverse matrix, express The transpose of represents the target variable of the advertisement (such as the click-through rate of the advertisement), represents the determinant of the covariance matrix, represents the number of advertisements (equal to N);

[0151] Compute the gradient:

[0152] for About Hyperparameter Sets Derivative, get the gradient :

[0153]

[0154] in, represents the weight of the posterior distribution, represents the trace sum of the matrix, express The transpose of , that is, the update direction of the hyperparameters is provided by the gradient;

[0155] For each hyperparameter (length scale , signal variance, , Noise Variance ), then we have:

[0156] Length scale :

[0157]

[0158] For the signal variance :

[0159]

[0160] For the noise variance :

[0161]

[0162] By updating each hyperparameter through the gradient formula, we have:

[0163]

[0164] in, Indicated in The hyperparameter vector at iteration 1 (i.e., the set containing the hyperparameters), Indicated in The hyperparameter vector at iteration , represents the learning rate;

[0165] BFGS uses the following update formula to update the hyperparameters:

[0166]

[0167] in, Represents the approximate inverse matrix of the Hessian matrix at the current iteration. L-BFGS updates the hyperparameter set through the gradient and the approximate inverse matrix of the Hessian matrix. After each iteration, check whether the convergence condition is met. If the norm of the gradient is small enough (less than the preset threshold), it is considered to have converged, or if the change in the log-likelihood function is less than the set threshold, it means that it has converged, thereby maximizing the hyperparameter set of the log-likelihood function. , to achieve the length scale , signal variance , Noise Variance , so as to achieve more accurate recommendation and selection of vending machine advertisements.

[0168] A computer-readable storage medium stores a computer program, wherein the computer program implements any one of the above-mentioned systems when executed by a processor.

[0169] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart vending machine product advertising recommendation management system, characterized in that: include: The data collection and analysis module is used to collect behavioral data sets and environmental data sets and establish behavioral feature sets B and environmental feature sets C. The behavioral feature sets B and environmental feature sets C are spliced ​​to construct a multimodal input matrix Xt, and input projection is performed to calculate the attention weight matrix And get multimodal features ; Dynamically adjust key dimensions based on environmental and behavioral characteristics , to achieve the attention weight matrix Optimization; Dynamic Adjustment The method is: Calculate the variance of the behavioral feature set B and entropy : in, Indicates behavioral characteristics The probability of occurrence, Represents the total number of elements in the behavior feature set B; Calculate the variance of the environmental feature set C and the rank of the correlation matrix : Relevance matrix representing the set of environmental features; in, , They represent the corresponding weight coefficients, , They represent the coefficients used to adjust the impact of behavioral characteristics on key dimensions, , They represent the coefficients used to moderate the effects of environmental characteristics on the key dimensions; Advertisement recommendation analysis module, based on Determine the similarity between the current user and other users with cosine similarity ,in accordance with determining recommended ads; The advertising recommendation management module models the click-through rate of advertisements through Gaussian processes and calculates the upper confidence bound of advertisements based on the predicted mean and standard deviation of advertisements. , to optimize the recommendation of advertisements, including: The length scale, signal variance, and noise variance in the Gaussian process are used to construct a set of hyperparameters. The covariance matrix is ​​constructed and the log marginal likelihood function is calculated. The hyperparameters are updated using L-BFGS. The optimal value is determined by calculating the gradient of each hyperparameter to achieve optimization of the length scale, signal variance, and noise variance.

2. According to claim 1, the intelligent vending machine product advertisement recommendation management system is characterized in that: The behavior feature set B includes commodity category, purchase time, and purchase frequency; the environment feature set C includes temperature, humidity, and light.

3. According to claim 1, the intelligent vending machine product advertisement recommendation management system is characterized in that: The multimodal features The construction method is: Execute X t Input projection of: Among them, Q t , K t 、V t Represents the feature matrices of query, key and value respectively, used to calculate self-attention, W Q , W K and W V The projection weight matrices representing query, key, and value respectively; Calculate the attention weight matrix : Among them, softmax represents the normalization function, represents the attention weight distribution, represents the dimension of the key, represents the transposed inner product of the query feature matrix and the key feature matrix; Multimodal features .

4. According to claim 1, the intelligent vending machine product advertisement recommendation management system is characterized in that: The basis Recommended ads are determined by: Based on the calculated similarity , select the K most similar users as similar users, and calculate the recommendation score of each advertisement for the current user based on the historical behaviors of similar users, then: in, Indicates the current user of the prediction About Advertising Interest score, represents a set of K other users similar to the current user, Indicates other users About Advertising Ratings; Set a score threshold, and ads that exceed the score threshold will be identified and pushed.

5. The intelligent vending machine product advertisement recommendation management system according to claim 4, characterized in that: The method for optimizing the recommendation of advertisements is: setting an initial state for each advertisement to facilitate subsequent advertisement selection and reward update, then: Defining Ad Sets , N represents the number of advertisements, represents the i-th advertisement; For each ad , confirm the ad Impressions , confirm the ad Number of clicks , get the click rate ; The click-through rate of an ad is modeled through a Gaussian process, and we have: Reward function Indicates advertisement The click rate return follows a Gaussian process: in, represents a Gaussian process, represents the mean function, represents the kernel function value; A1. Perform Gaussian process prediction steps, including: According to Gaussian process Predicted mean With standard deviation , ； Where X represents the set of advertising data points, represents the kernel function matrix between all advertisements in the advertisement data point set X, represents the noise variance, I represents the identity matrix, Indicates advertisement Collection of advertising data points The kernel function value of all advertisements in , Represents all advertisements and advertisements in the advertisement data point set X Column vector of kernel function values; A2. Calculate value: in, represents the upper confidence bound of the ith advertisement, represents the hyperparameter term; A3. Update the click-through rate of ads : Advertisement Get feedback after tth display , update ads Impressions and clicks for: ； Based on updated click-through rate Update the Gaussian process prediction step to get a new prediction mean With standard deviation , calculate the advertisement selected in the next round value; A4. Update-based value, select the maximum The ads with the same value are optimized for the next round and steps A1-A3 are repeated to update Value, based on Value for recommended playback.

6. The intelligent vending machine product advertisement recommendation management system according to claim 5, characterized in that: The kernel function value According to the following expression: in, represents the signal variance, Indicates the length scale.

7. The intelligent vending machine product advertisement recommendation management system according to claim 6, characterized in that: The method for optimizing the length scale, signal variance, and noise variance is: Initialize the hyperparameter set ; Construct the covariance matrix K: in, represents the elements in the covariance matrix K, represents the noise term; Compute the log marginal likelihood function : in, represents the inverse matrix, express The transpose of represents the target variable of the advertisement, represents the determinant of the covariance matrix, Indicates the number of advertisements; Compute the gradient: for Derivative, get the gradient : in, represents the weight of the posterior distribution, represents the trace sum of the matrix, express The transpose of Determine the derivative form of the hyperparameters; By updating each hyperparameter through the gradient formula, we have: in, Indicated in The hyperparameter vector at iteration , Indicated in The hyperparameter vector at iteration , represents the learning rate; Update hyperparameters: in, Represents the approximate inverse matrix of the Hessian matrix at the current iteration; After each iteration, check whether the convergence condition is met. If the norm of the gradient is less than the preset threshold, or if the change of the log-likelihood function is less than the set threshold, the optimization process ends and the optimized length scale is obtained. , signal variance , Noise Variance .

8. The intelligent vending machine product advertisement recommendation management system according to claim 7, characterized in that: The derivative form of the hyperparameter is as follows: Length scale : For the signal variance : For the noise variance : 。 9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the system according to any one of claims 1 to 8 is implemented.

Images