User behavior analysis method and system for AI capability fusion platform

Abstract

The invention discloses a user behavior analysis method and system for an AI capability fusion platform, and belongs to the technical field of big data analysis. Historical data related to user behaviors is identified and collected, and the collected data is cleaned and standardized; defining target variables and features, and constructing a sample matrix; calculating correlation between each feature vector and a target variable in the sample matrix and behavior features of each sample sequence, and generating a behavior sample matrix; constructing a user prediction model, performing deep mining and analysis on the user behavior data, and predicting the purchase intention of the user; behavior data of the user is obtained in real time, and a product recommendation table is generated; and collecting feedback data of the user on the product recommendation table in real time, calculating the conversion rate of the product recommendation table and the similarity between the product recommendation table and the actual purchase condition, and dynamically adjusting the user prediction model.

Description

Technical Field

[0001] The present invention relates to the field of big data analysis technology, and more specifically to a user behavior analysis method and system for an AI capability fusion platform. Background Art

[0002] Traditional user behavior analysis methods are usually based on simple statistical and classification techniques, such as frequency statistics, user classification, etc. Although these methods can reveal some basic characteristics of user behavior to a certain extent, their limitations are becoming more and more obvious in the face of the complexity and diversity of the big data era; traditional statistical and classification techniques can often only capture the surface phenomena of user behavior, and it is difficult to deeply explore the deep-level laws and patterns behind user behavior; at the same time, these methods are greatly limited in efficiency and accuracy when processing large-scale data, and it is difficult to meet the real-time and accuracy requirements of modern enterprises for user behavior analysis. To this end, it is necessary to use more advanced and efficient big data analysis technologies, make full use of the diversity and complexity of big data, deeply explore the potential value in user behavior data, and provide users with more accurate behavior predictions and personalized recommendations.

[0003] The existing Chinese patent with authorization announcement number CN114925273B discloses a user behavior prediction method and AI prediction analysis system based on big data analysis. First, a user behavior data set including multiple user historical operations is obtained; then the confidence of the generalized user key behavior corresponding to each user historical operation is obtained, and the user historical operation distribution knowledge graph corresponding to the target user historical operation of the candidate narrow time dimension analysis is analyzed in a narrow time dimension. The narrow user key behavior confidence of the target user historical operation is used to perform confidence screening on its generalized user key behavior confidence. Finally, the target user historical operation after confidence screening is used to predict the user behavior of the user to be processed, and the target user historical operation used to predict user behavior is determined in combination with different time dimensions.

[0004] Although the prior art effectively solves the problem of unreliable prediction results in complex user behavior prediction based only on historical user operations that are close in time, it does not consider deeper user behavior mining, such as the impact of new trends, new products, and changes in the surrounding cultural environment on user behavior. The dynamic changes in user behavior will lead to deviations in prediction results, resulting in poor personalized recommendation effects. Therefore, this application provides a user behavior analysis method and system for an AI capability fusion platform. Summary of the invention

[0005] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a user behavior analysis method and system for an AI capability fusion platform, to build a user prediction model, to conduct in-depth mining and analysis of user behavior, to predict user purchase intentions, to generate personalized product recommendation tables, and to dynamically adjust the user prediction model based on user feedback data to ensure that the product recommendation table always keeps up with the user's latest needs and preference changes.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The user behavior analysis method used for the AI ​​capability fusion platform includes:

[0008] Step S1: Identify and collect historical data related to user behavior, and clean and standardize the collected historical data;

[0009] Step S2: define the target variable and features, and construct a sample matrix; calculate the correlation between each feature vector in the sample matrix and the target variable, as well as the behavior features of each sample sequence, and generate a behavior sample matrix;

[0010] Step S3: Build a user prediction model, analyze user behavior data, and predict the user's purchase intention; and obtain user behavior data in real time to generate a product recommendation table;

[0011] Step S4: Collect user feedback data on the product recommendation form in real time, calculate the conversion rate of the product recommendation form and the similarity between the product recommendation form and the actual purchase situation, and dynamically adjust the user prediction model.

[0012] Specifically, the step S2 includes:

[0013] S2.1: Define purchase decision as the target variable, the factors affecting the target variable as features, and construct a sample matrix (Ψ, Φ); where Ψ is the feature matrix, and Ψ=(Ψ1,…,Ψ a ), a is the number of feature types, Φ is the target variable vector, Ψ a is the ath eigenvector that affects the target variable;

[0014] S2.2: Calculate the eigenvector Ψ in the eigenmatrix x The correlation between the target variable vector and the target variable vector is generated, and the correlation matrix R = (r1, r2, ..., r a ), where Ψ x is the eigenvector of the xth feature in the feature matrix, x = 1, 2, ..., a, r a is the correlation between the xth feature and the target variable vector;

[0015] S2.3: Set the correlation threshold to rth , filter out the features related to the target variable; if |r x |≥r th , define the eigenvector Ψ x is the relevant eigenvector; if |r x | <r th , define the eigenvector Ψ x is an irrelevant eigenvector;

[0016] S2.4: Get relevant feature vectors and generate relevant sample matrix in, is the correlation feature matrix.

[0017] Specifically, the step S2 further includes:

[0018] S2.5: Collect the user's behavior records, set the time window to T, and calculate the behavior frequency and average duration of each behavior type of the user within the time window;

[0019] S2.6: Add the calculated behavior frequency and average duration as new features to the relevant sample matrix to generate a behavior sample matrix in, is the behavioral feature matrix.

[0020] Specifically, step S3 includes:

[0021] S3.1: Use a multi-layer perceptron (MLP) model to build a user prediction model, which includes an input layer, a hidden layer, and an output layer;

[0022] S3.2: Initializing parameters of the user prediction model;

[0023] S3.3: The behavior sample matrix Dividing the user prediction model into a training set and a test set, and using the training set to train the user prediction model to optimize the parameters of the user prediction model;

[0024] S3.4: Validating the user prediction model using the test set;

[0025] S3.5: Acquire the user's current behavior data in real time, use the user prediction model to predict the user's behavior data, and obtain the user's predicted purchase decision;

[0026] S3.6: Obtain the products with purchase intention output by the user prediction model, sort the products in descending order based on purchase probability, and generate a product recommendation table.

[0027] Specifically, the S3.1 also includes:

[0028] S3.11: Set the number of hidden layers to L, and the lth layer contains k l neurons, using an activation function to process the behavior feature matrix to generate a behavior hidden matrix;

[0029] S3.12: Process the behavior hidden matrix using the Sigmoid activation function to calculate the probability of the user's purchase decision;

[0030] S3.13: Set a binary classifier after the output layer and set the threshold to 0.5; if Y out ≥0.5, then let φ pred =1, it is predicted that the user has the intention to buy; if Y out <0.5, then let φ pred =0, it is predicted that the user has no purchase intention; where Y out is the probability of user purchase decision, φ pred Predicted value for user purchase decision.

[0031] Specifically, the S3.3 also includes:

[0032] S3.31: Input the behavior feature matrix in the training set into the user prediction model to obtain the predicted value φ of the user's purchase decision pred ;

[0033] S3.32: define cross entropy loss as the loss function of the user prediction model, and calculate the difference between the predicted value and the actual value;

[0034] S3.33: Calculate the gradient of the loss function with respect to the user prediction model parameters using a back propagation algorithm;

[0035] S3.34: Update the parameters of the user prediction model using stochastic gradient descent method;

[0036] S3.35: Repeat S3.32 to S3.34 until the preset number of iterations is reached and the loss function value is less than the preset loss threshold, and the training ends.

[0037] Specifically, step S4 includes:

[0038] S4.1: setting the feedback interval to t, collecting the user's real-time feedback data on the product recommendation table after the t time period of generating the product recommendation table, and preprocessing the real-time feedback data;

[0039] S4.2: Calculate the conversion rate P of users to the product recommendation table CR , the expression is as follows:

[0040]

[0041] In the formula, F buy F is the number of products that users actually purchase after clicking on the recommended product. click The total number of recommended products clicked by users.

[0042] Specifically, the step S4 further includes:

[0043] S4.3: For the product recommendation table and the products actually purchased by the user, extract the key features of each product and generate the recommended product vector H = (H1, H2, ..., H m ) and the actual product vector E=(E1,E2,…,E n ), where m is the number of recommended products in the product recommendation table, H m is the product vector of the mth recommended product in the product recommendation table, n is the number of products actually purchased, E n is the product vector of the nth product among the products actually purchased;

[0044] S4.4: Calculate the similarity between each product vector in the product recommendation table and each product vector actually purchased, and generate a similarity matrix S;

[0045] S4.5: Calculate the average similarity of the similarity matrix S

[0046] S4.6: Set the user's conversion threshold to P th The average similarity threshold is Determine whether the product recommendation table is consistent with the user's actual purchase situation; if P CR ≥P th and The product recommendation table is consistent with the user's actual purchase situation; if P CR <P th or The product recommendation table does not conform to the actual purchase situation of the user, and returns to S2.1 to update the sample matrix to update the parameters in the user prediction model.

[0047] User behavior analysis system for AI capability fusion platform, including: data collection module, feature extraction module and prediction module;

[0048] The data collection module is used to collect historical data related to user behavior and pre-process the data;

[0049] The feature extraction module is used to define the target variable and the features that affect the target variable, and to construct a sample matrix (Ψ, Φ); the sample matrix is ​​processed to generate a behavior sample matrix

[0050] The prediction module is used to build a user prediction model, predict the user's purchase intention, and generate a product recommendation table; collect user feedback data on the product recommendation table in real time, analyze the feedback data, and dynamically adjust the user prediction model.

[0051] Specifically, the feature extraction module includes a similarity unit and a behavior unit;

[0052] The similarity unit is used to calculate the correlation between each feature vector in the sample matrix and the target variable, screen out the features related to the target variable, and generate a related sample matrix;

[0053] The behavior unit is used to collect behavior records, set the time window to T, calculate the behavior frequency and average duration of each behavior type of the user within the time window, and generate a behavior sample matrix.

[0054] The prediction module includes a recommendation unit and a feedback update unit;

[0055] The recommendation unit is used to build a user prediction model, obtain the user's current behavior data in real time, calculate the user's predicted purchase decision, and generate a product recommendation table;

[0056] The feedback updating unit collects user feedback data on the product recommendation table in real time, calculates the conversion rate and average similarity, determines whether the product recommendation table is consistent with the user's actual purchase situation, updates the sample matrix, and updates the parameters in the user prediction model.

[0057] Beneficial effects of the present invention:

[0058] 1. By defining the target variable and screening the features that are highly correlated with the target variable, a behavior sample matrix is ​​constructed. The matrix only contains the features that are closely related to the target variable, which can more effectively capture the user's behavior pattern. At the same time, by removing the features with low correlation with the target variable, the amount of information that the model needs to process is reduced, the redundancy of the data is reduced, and the computational complexity of the model training process is significantly reduced, thereby reducing the risk of overfitting, improving the prediction accuracy of the model, and making the model more robust and reliable in practical applications.

[0059] 2. In order to achieve a more personalized user experience and precise marketing strategies, we obtain user behavior data in real time, build a user prediction model, conduct in-depth mining and analysis of user behavior, and capture users' potential needs and preferences in real time, so as to more accurately predict users' purchasing intentions and generate personalized product recommendation tables. Each user can get product recommendations tailored to their personal preferences and current needs, which not only meets the unique needs of different users, but also greatly improves users' shopping experience and satisfaction. At the same time, personalized recommendation strategies also improve users' purchase conversion rates, increase platform sales and user stickiness, and bring direct benefit growth and long-term competitive advantages to the business.

[0060] 3. Collect user feedback on the product recommendation form in real time, directly measure the effectiveness of the model by calculating the conversion rate and the similarity between the product recommendation form and the actual purchase situation, and dynamically adjust the model based on the feedback, optimize feature selection and weight distribution, so as to continuously improve the accuracy and relevance of the recommendation, ensure that the product recommendation form always keeps up with the latest needs and preference changes of users, and continuously optimize the recommendation effect; this not only improves user satisfaction and platform interactivity, but also promotes a virtuous cycle between users and the platform, laying a solid foundation for the long-term development of the business. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A schematic diagram of a user behavior analysis method for an AI capability fusion platform;

[0062] Figure 2 A feature extraction flow chart for the user behavior analysis method used in the AI ​​capability fusion platform;

[0063] Figure 3 A personalized recommendation flow chart for the user behavior analysis method used in the AI ​​capability fusion platform;

[0064] Figure 4 Real-time update of flowcharts for user behavior analysis methods used in AI capability fusion platforms;

[0065] Figure 5 This is a structural diagram of the user behavior analysis system used for the AI ​​capability fusion platform. DETAILED DESCRIPTION

[0066] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. The embodiments of the present invention and the technical features in the embodiments may be combined with each other unless there is a conflict.

[0067] Example 1

[0068] refer to Figures 1 to 4 As shown, this embodiment introduces a user behavior analysis method for an AI capability fusion platform, including the following steps:

[0069] Step S1: Identify and collect historical data related to user behavior, including historical user operations, social media data, geographic location data, and browsing history data, such as browsing, clicking, searching, and purchasing history. These multi-dimensional data are obtained through multiple channels, such as log files, API interfaces, and third-party data providers, to ensure the comprehensiveness and representativeness of the data in order to have a more comprehensive understanding of user behavior patterns and preferences; clean the collected data to remove useless data, such as missing values, duplicate values, and outliers. At the same time, standardize the data, such as normalizing the data to eliminate the dimensional differences between different features, and standardizing the data to improve the convergence speed and performance of the algorithm, so that the data is suitable for subsequent algorithm training and modeling, and improve the accuracy and reliability of data analysis;

[0070] Step S2: Based on the preprocessed data, define the target variable and the features that affect the target variable, and construct a sample matrix; calculate the correlation between each feature vector in the sample matrix and the target variable, screen out the features that are highly correlated with the target variable, collect user behavior records, calculate the behavior features of each sample sequence, capture more user behavior patterns, and generate a behavior sample matrix;

[0071] Step S3: Based on the behavior sample matrix and target variables, a user prediction model is constructed to conduct in-depth mining and analysis of user behavior data, identify users' potential needs and preferences, and predict users' purchase intentions for different types of products or services; obtain user behavior data in real time, use the user prediction model to obtain users' predicted purchase decisions, and generate a product recommendation table;

[0072] Step S4: Collect user feedback data on the product recommendation form in real time, and calculate the conversion rate of the product recommendation form and the similarity between the product recommendation form and the actual purchase situation based on the feedback data, evaluate the consistency between the product recommendation form and the actual purchase situation, and dynamically adjust the user prediction model to update the user's product recommendation form to improve the platform's interactivity and user satisfaction.

[0073] Specifically, the specific steps of step S2 include:

[0074] S2.1: Based on business needs and data characteristics, define the user's purchase decision on the product as the target variable, and define 0 for no purchase and 1 for purchase. Define various factors that affect the target variable as features. These factors are other factors in the historical data except the user's purchase decision, and construct a sample matrix Among them, the sample matrix is ​​a data set containing various features and target variables, Ψ is the feature matrix, a is the number of feature types, Ψ a is the ath eigenvector that affects the target variable, Φ is the target variable vector, b is the number of samples, (ψ1(b),…,ψ a (b),φ(b)) is the b-th group of sample sequences in the sample matrix,ψ a (b) is the ath eigenvalue in the bth group of sample sequences, φ(b) is the target variable value in the bth group of sample sequences;

[0075] S2.2: Using the Pearson correlation coefficient, calculate the eigenvector Ψ in the sample matrix x The correlation between the target variable vector and the target variable vector is generated, and the correlation matrix R = (r1, r2, ..., r a ), where each element in the correlation matrix represents the correlation between a feature and the target variable; the expression is as follows:

[0076]

[0077] In the formula, Ψ x is the eigenvector of the xth feature in the feature matrix, r x is the correlation between the xth feature and the target variable vector, x = 1, 2, ..., a, is the eigenvector Ψ x The mean of is the mean of the target variable vector Φ,

[0078] S2.3: Set the correlation threshold to r th , traverse the correlation matrix R, filter out the features that are highly correlated with the target variable, and reduce the data dimension; if |r x |≥r th , define the eigenvector Ψ x is the relevant eigenvector; if |r x | <r th , define the eigenvector Ψ x is an irrelevant eigenvector;

[0079] S2.4: Get the relevant eigenvectors in the sample matrix (Ψ, Φ) and generate the relevant sample matrix in, is the relevant feature matrix, a th is the number of feature types in the relevant feature matrix, a th ≤a, is the ath in the correlation feature matrix th related feature vectors;

[0080] S2.5: Since the related sample matrix It is directly extracted from the historical data of user behavior, and only records the data or behavior at a single time point, which cannot directly reflect some potential laws or patterns in the data. Perform calculations to capture more user behavior patterns, thereby enhancing the model's ability to predict user purchase decisions; collect user behavior records at the corresponding time points of each sample sequence in the relevant sample matrix, such as behavior records, behavior types, and behavior duration at each time point, set the time window to T, such as the past week or month, and for each sample sequence, calculate the user's behavior frequency and average duration of each behavior type within the time window. The expression is as follows:

[0081]

[0082] Where η c is the behavior frequency of the cth behavior type in the time window, c = 1, 2, ..., M, M is the total number of behavior types, N c is the number of occurrences of the cth behavior type in the time window, D c is the average duration of the cth behavior type in the time window, d z is the duration of the zth behavior in the cth behavior type, z=1,2,…,N c ;

[0083] S2.6: Add the calculated behavior frequency and average duration as new features to the relevant sample matrix to generate a behavior sample matrix in, is the behavior feature matrix, a c is the number of feature types in the behavior feature matrix, and a c >a th , is the ath in the behavior feature matrix c A behavioral feature vector.

[0084] Specifically, the specific steps of step S3 include:

[0085] S3.1: Use the multi-layer perceptron (MLP) model to build a user prediction model. The model includes an input layer, a hidden layer, and an output layer. The input layer is used to input the behavior sample matrix into the user prediction model, and the number of nodes in the input layer should match the number of features in the behavior feature matrix. The hidden layer is used to extract the deep-level features of user behavior. The output layer contains a neuron, using the Sigmoid activation function, and the output value is a probability value between 0 and 1, indicating the user's purchase possibility of different types of products or services. A binary classifier is set after the output layer to convert the probability value into a purchase decision prediction of 0 or 1;

[0086] S3.2: Initialize the parameters of the user prediction model (including the weight matrix and bias matrix of each layer in the user prediction model) to small random numbers;

[0087] S3.3: Behavior Sample Matrix Divide into a training set and a test set, the division ratio is α:(1-α), and use the training set to train the user prediction model to optimize the parameters of the user prediction model; wherein, in this embodiment, α=0.8;

[0088] S3.4: Use the test set to verify the user prediction model; if the verification fails, expand the training set and continue to train the model; if the verification passes, the user prediction model is successfully constructed;

[0089] S3.5: Obtain the user's current behavior data in real time, pre-process the data, use the user prediction model to predict the user's behavior data, and obtain the user's predicted purchase decision; wherein the predicted purchase decision includes purchase intention and no purchase intention;

[0090] S3.6: Obtain the products with purchase intention output by the user prediction model, sort the products in descending order based on the purchase probability, and generate a product recommendation table.

[0091] Specifically, the specific steps of S3.1 also include:

[0092] S3.11: Set the number of hidden layers to L, and the lth layer contains k l neurons, the hidden layer receives the behavior feature matrix from the input layer or the hidden matrix of the previous layer, processes the behavior feature matrix using the activation function, and generates a behavior hidden matrix. The expression is as follows:

[0093] H l =γ(W l ·H l-1 +B l )

[0094] Where l = 1, 2, ..., L, H l is the behavior hidden matrix of the lth hidden layer, W l is the weight matrix from layer l-1 to layer l, B l is the bias matrix of the lth layer, γ(·) is the ReLU function, which introduces nonlinearity and improves the expressiveness of the model; for the first hidden layer, its input is the output of the input layer, and the output is H1=f(W1·X+B1), where X is the output matrix of the input layer, and for subsequent hidden layers, its input is the output of the previous layer;

[0095] S3.12: Use the Sigmoid activation function to process the behavior hidden matrix and calculate the probability of the user's purchase decision. The expression is as follows:

[0096] Y out =σ(W out ·H L +B out )

[0097] Where Y out is the probability of user purchase decision, σ(·) is the Sigmoid activation function, H L is the output of the last hidden layer, W out , B out They are the weight matrix and bias matrix from the last hidden layer to the output layer respectively;

[0098] S3.13: Set a binary classifier after the output layer and set the threshold to 0.5; if Y out ≥0.5, then let φ pred =1, it is predicted that the user has the intention to buy; if Y out <0.5, then let φ pred =0, it is predicted that the user has no purchase intention; where φ pred Predicted value for user purchase decision.

[0099] Specifically, the specific steps of S3.3 also include:

[0100] S3.31: Input the behavioral feature matrix in the training set into the user prediction model, and obtain the predicted value φ of the user's purchase decision by layer-by-layer calculation pred , the target variable vector in the training set is taken as the actual value φ of the user's purchase decision;

[0101] S3.32: Define cross entropy loss as the loss function of the user prediction model, and calculate the difference f between the predicted value and the actual value loss , the expression is as follows:

[0102]

[0103] In the formula, G is the number of samples in the training set, φ(g) is the g-th element in the actual value, and φ pred (g) is the g-th element in the predicted value, g = 1, 2, ..., G;

[0104] S3.33: Use the back propagation algorithm to calculate the gradient of the loss function to the user prediction model parameters, calculate the gradient of the loss function to the output layer output, and calculate the gradient backward layer by layer based on the chain rule, and calculate the gradient of the loss function to the weight matrix and bias matrix in each layer. The expression is as follows:

[0105]

[0106] In the formula, is the gradient of the loss function to the output layer output, is the gradient of the output layer weight matrix, σ' is the derivative of the activation function σ(·), is the gradient of the output layer bias matrix, is the gradient of the weight matrix of the lth hidden layer, γ' is the derivative of the activation function γ(·), is the gradient of the bias matrix of the lth hidden layer;

[0107] S3.34: Update the parameters of the user prediction model using stochastic gradient descent to minimize the loss function, as shown below:

[0108]

[0109] Where W old is the weight matrix before update, W new is the updated weight matrix, B old is the bias matrix before updating, B new is the updated bias matrix, λ is the learning rate, is the gradient of the corresponding parameter;

[0110] S3.35: Repeat S3.32 to S3.34 until the preset number of iterations is reached and the loss function value is less than the preset loss threshold, and the training ends.

[0111] Specifically, the specific steps of step S4 include:

[0112] S4.1: Set the feedback interval to t. After the product recommendation table is generated for a period of t, collect the user's real-time feedback data on the product recommendation table, such as the number of clicks, browsing time, adding to shopping cart, purchase times, and actual purchases of the user during the period of t when the product recommendation table is generated. These data intuitively reflect the user's satisfaction and acceptance of the product recommendation table. Preprocess the real-time feedback data, including cleaning, denoising, and formatting, to provide a high-quality data basis for subsequent analysis.

[0113] S4.2: Calculate the conversion rate R of users to the product recommendation table CR , the conversion rate is used to measure the degree to which the recommended products in the product recommendation table meet the needs of users, reflecting the proportion of users who actually complete the purchase after clicking on the recommended products; the expression is as follows:

[0114]

[0115] In the formula, F buyThe number of products that users actually purchased after clicking on the recommended products. click The total number of recommended products clicked by users, indicating how many products in the product recommendation table were clicked by users;

[0116] S4.3: For the product recommendation table and the products actually purchased by the user, extract the key features of each product, such as price, brand, and category, and convert the features of each product into a feature vector with the same dimension to generate a recommended product vector H = (H1, H2, ..., H m ) and the actual product vector E=(E1,E2,…,E n ), where m is the number of recommended products in the product recommendation table, n is the number of products actually purchased, and H i is the product vector of the i-th recommended product in the product recommendation table, i=1,2,…,m, E j is the product vector of the jth product among the actually purchased products, j = 1, 2, …, n, and the dimension of the product vector of each product is k;

[0117] S4.4: Calculate the product vector H in the product recommendation table i Vector E of the product actually purchased j The similarity of i,j , the similarity reflects the similarity or consistency between the recommended products in the product recommendation table and the products actually purchased by the user, and generates a similarity matrix S, which is expressed as follows:

[0118]

[0119] In the formula, (·) T is the transpose operation of the vector, and ||·|| is the modulus operation of the vector;

[0120] S4.5: Calculate the average similarity of the similarity matrix S The average similarity is used to measure the consistency between the product recommendation table and the actual purchase situation. The expression is as follows:

[0121]

[0122] S4.6: Set the user's conversion threshold to P th The average similarity threshold is Determine whether the product recommendation table is consistent with the user's actual purchase situation; if P CR ≥P th and The product recommendation table is consistent with the user's actual purchase situation; if P CR <P th or The product recommendation table does not match the user's actual purchase situation, and returns to S2.1 to update the sample matrix to update the parameters in the user prediction model so that the model can more accurately reflect the user's preferences.

[0123] Example 2

[0124] See also Figure 5 , another embodiment provided by the present invention: a user behavior analysis system for an AI capability fusion platform, comprising: a data acquisition module, a feature extraction module and a prediction module;

[0125] The data collection module is used to obtain historical data related to user behavior from multiple channels, such as log files, API interfaces and third-party data providers, including user historical operations, social media data, geographic location data and browsing history data; and pre-process the data and save the pre-processed data into the user behavior database;

[0126] The feature extraction module is used to define the target variable and the features that affect the target variable, and to construct a sample matrix (Ψ, Φ); the sample matrix is ​​processed to screen out the features that have a significant impact on the prediction results, and the behavioral features in the sample matrix are calculated to generate a behavioral sample matrix. Among them, Ψ is the feature matrix, Φ is the target variable vector, is the behavioral feature matrix;

[0127] The prediction module is used to build a user prediction model, conduct in-depth mining and analysis of user behavior data, predict users' purchase intentions for different types of products or services, and generate a product recommendation form; collect user feedback data on the product recommendation form in real time, calculate the conversion rate of the product recommendation form, and the similarity between the product recommendation form and the actual purchase situation, evaluate the consistency between the product recommendation form and the actual purchase situation, and dynamically adjust the user prediction model.

[0128] Specifically, the feature extraction module includes a similarity unit and a behavior unit;

[0129] The similarity unit calculates the correlation between each eigenvector in the sample matrix and the target variable through correlation analysis methods, such as the Pearson correlation coefficient, generates a correlation matrix R, and sets the correlation threshold r th , traverse the correlation matrix R, compare the elements in the correlation matrix R with the correlation threshold, screen out the features that are highly correlated with the target variable, and generate the correlation sample matrix Reduce the data dimension; among them, is the correlation feature matrix;

[0130] The behavior unit is used to collect the relevant sample matrix The behavior record of each sample sequence at the corresponding time point in the time window is set as T, and the behavior frequency η of each behavior type of the user in the time window is calculated. c and the average duration D c , add the calculated behavior frequency and average duration as new features to the relevant sample matrix to generate the behavior sample matrix Among them, η c is the behavior frequency of the cth behavior type in the time window, c = 1, 2, ..., M, M is the total number of behavior types, D c is the average duration of the cth behavior type in the time window.

[0131] Specifically, the prediction module includes a recommendation unit and a feedback update unit;

[0132] The recommendation unit is used to build a user prediction model and transform the behavior sample matrix Divide into training set and test set, use the training set to train the user prediction model, optimize the parameters of the user prediction model, and use the test set to verify the user prediction model; obtain the user's current behavior data in real time, use the user prediction model to predict the behavior data, obtain the user's predicted purchase decision, and sort the products in descending order based on the purchase probability to generate a product recommendation table;

[0133] The feedback update unit collects the user's feedback data on the product recommendation form in real time through the set feedback interval t, and calculates the user's conversion rate P for the product recommendation form. CR , the conversion rate is used to measure the degree to which the recommended products in the product recommendation table meet the needs of users, reflecting the proportion of users who actually complete the purchase after clicking on the recommended products; extract the key features of each product in the product recommendation table and the actual purchase of users, and generate the recommended product vector H = (H1, H2, ..., H m ) and the actual product vector E=(E1,E2,…,E n ), calculate the average similarity between the product recommendation table and the user's actual purchase Based on the conversion rate and average similarity, determine whether the product recommendation table is consistent with the user's actual purchase situation, update the sample matrix, and update the parameters in the user prediction model to make the model more accurately reflect user preferences; where m is the number of recommended products in the product recommendation table, and n is the number of actually purchased products.

[0134] In summary, the present invention defines the target variable and the features that affect the target variable, constructs a sample matrix, calculates the similarity between the features and the target variable, screens out the features related to the target variable, calculates the behavioral features, and generates a behavioral sample matrix; by constructing a user prediction model, a product recommendation table is generated, and user feedback data on the product recommendation table is collected in real time, the conversion rate of the product recommendation table and the similarity between the product recommendation table and the actual purchase situation are calculated, and the user prediction model is dynamically adjusted.

[0135] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A user behavior analysis method for an AI capability fusion platform, characterized in that: include: Step S1: Identify and collect historical data related to user behavior, and clean and standardize the collected historical data; Step S2: define target variables and features, and construct a sample matrix; Calculating the correlation between each feature vector in the sample matrix and the target variable, as well as the behavior characteristics of each sample sequence, to generate a behavior sample matrix; Step S3: Build a user prediction model, analyze user behavior data, and predict user purchase intention; And obtain user behavior data in real time to generate product recommendation tables; Step S4: Collect user feedback data on the product recommendation form in real time, calculate the conversion rate of the product recommendation form and the similarity between the product recommendation form and the actual purchase situation, and dynamically adjust the user prediction model.

2. The user behavior analysis method for the AI ​​capability fusion platform according to claim 1, characterized in that: The step S2 comprises: S2.1: Define purchase decision as the target variable, the factors affecting the target variable as features, and construct a sample matrix (Ψ, Φ); where Ψ is the feature matrix, and Ψ = (Ψ1, ..., Ψ a ), a is the number of feature types, Φ is the target variable vector, Ψ A is the ath eigenvector that affects the target variable; S2.2: Calculate the eigenvector Ψ in the eigenmatrix x The correlation between the target variable vector and the target variable vector is generated, and the correlation matrix R = (r1, r2, ..., r a ), where Ψ x is the eigenvector of the xth feature in the feature matrix, x = 1, 2, ..., a, r a is the correlation between the xth feature and the target variable vector; S2.3: Set the correlation threshold to r th , filter out the features related to the target variable; if |r x |≥r th , define the eigenvector Ψ x is the relevant eigenvector; if |r x |<r th , define the eigenvector Ψ x is an irrelevant eigenvector; S2.4: Get relevant feature vectors and generate relevant sample matrix in, is the correlation feature matrix.

3. The user behavior analysis method for the AI ​​capability fusion platform according to claim 2 is characterized in that: The step S2 further comprises: S2.5: Collect the user's behavior records, set the time window to T, and calculate the behavior frequency and average duration of each behavior type of the user within the time window; S2.6: Add the calculated behavior frequency and average duration as new features to the relevant sample matrix to generate a behavior sample matrix in, is the behavioral feature matrix.

4. The user behavior analysis method for the AI ​​capability fusion platform according to claim 3 is characterized in that: The step S3 comprises: S3.1: Use a multi-layer perceptron model to build a user prediction model, which includes an input layer, a hidden layer, and an output layer; S3.2: Initializing parameters of the user prediction model; S3.3: The behavior sample matrix Dividing the user prediction model into a training set and a test set, and using the training set to train the user prediction model to optimize the parameters of the user prediction model; S3.4: Validating the user prediction model using the test set; S3.5: Acquire the user's current behavior data in real time, use the user prediction model to predict the user's behavior data, and obtain the user's predicted purchase decision; S3.6: Obtain the products with purchase intention output by the user prediction model, sort the products in descending order based on purchase probability, and generate a product recommendation table.

5. The user behavior analysis method for the AI ​​capability fusion platform according to claim 4 is characterized in that: The S3.1 also includes: S3.11: Set the number of hidden layers to L, and the lth layer contains k l neurons, using an activation function to process the behavior feature matrix to generate a behavior hidden matrix; S3.12: Process the behavior hidden matrix using the Sigmoid activation function to calculate the probability of the user's purchase decision; S3.13: Set a binary classifier after the output layer and set the threshold to 0.5; if Y out ≥0.5, then φ pred =1, it is predicted that the user has the intention to buy; if Y out <0.5, then φ pred =0, it is predicted that the user has no purchase intention; where Y out is the probability of user purchase decision, φ pred Predicted value for user purchase decision.

6. The user behavior analysis method for the AI ​​capability fusion platform according to claim 5 is characterized in that: The S3.3 also includes: S3.31: Input the behavior feature matrix in the training set into the user prediction model to obtain the predicted value φ of the user's purchase decision pred ; S3.32: define cross entropy loss as the loss function of the user prediction model, and calculate the difference between the predicted value and the actual value; S3.33: Calculate the gradient of the loss function with respect to the user prediction model parameters using a back propagation algorithm; S3.34: Update the parameters of the user prediction model using stochastic gradient descent method; S3.35: Repeat S3.32 to S3.34 until the preset number of iterations is reached and the loss function value is less than the preset loss threshold, and the training ends.

7. The user behavior analysis method for the AI ​​capability fusion platform according to claim 6 is characterized in that: The step S4 comprises: S4.1: setting the feedback interval to t, collecting the user's real-time feedback data on the product recommendation table after the t time period of generating the product recommendation table, and preprocessing the real-time feedback data; S4.2: Calculate the conversion rate P of users to the product recommendation table CR , the expression is as follows: In the formula, F buy F is the number of products that users actually purchase after clicking on the recommended product. click The total number of recommended products clicked by users.

8. The user behavior analysis method for the AI ​​capability fusion platform according to claim 7 is characterized in that: The step S4 further comprises: S4.3: For the product recommendation table and the products actually purchased by the user, extract the key features of each product and generate a recommended product vector H = (H1, H2, ..., H m ) and the actual product vector E = (E1, E2, ..., E n ), where m is the number of recommended products in the product recommendation table, H m is the product vector of the mth recommended product in the product recommendation table, n is the number of products actually purchased, E n is the product vector of the nth product among the products actually purchased; S4.4: Calculate the similarity between each product vector in the product recommendation table and each product vector actually purchased, and generate a similarity matrix S; S4.5: Calculate the average similarity of the similarity matrix S S4.6: Set the user's conversion threshold to P th The average similarity threshold is Determine whether the product recommendation table is consistent with the user's actual purchase situation; if P CR ≥P th and The product recommendation table is consistent with the user's actual purchase situation; if P CR <P th or The product recommendation table does not conform to the actual purchase situation of the user, and returns to S2.1 to update the sample matrix to update the parameters in the user prediction model.

9. A user behavior analysis system for an AI capability fusion platform, which is used to implement a user behavior analysis method for an AI capability fusion platform as described in any one of claims 1 to 8, characterized in that: include: Data acquisition module, feature extraction module and prediction module; The data collection module is used to collect historical data related to user behavior and pre-process the data; The feature extraction module is used to define the target variable and the features that affect the target variable, and to construct a sample matrix (Ψ, Φ); the sample matrix is ​​processed to generate a behavior sample matrix The prediction module is used to build a user prediction model, predict the user's purchase intention, and generate a product recommendation table; collect user feedback data on the product recommendation table in real time, analyze the feedback data, and dynamically adjust the user prediction model.

10. The user behavior analysis system for the AI ​​capability fusion platform according to claim 9, characterized in that: The feature extraction module includes a similarity unit and a behavior unit; The similarity unit is used to calculate the correlation between each feature vector in the sample matrix and the target variable, screen out the features related to the target variable, and generate a related sample matrix; The behavior unit is used to collect behavior records, set the time window to T, calculate the behavior frequency and average duration of each behavior type of the user within the time window, and generate a behavior sample matrix. The prediction module includes a recommendation unit and a feedback update unit; The recommendation unit is used to build a user prediction model, obtain the user's current behavior data in real time, calculate the user's predicted purchase decision, and generate a product recommendation table; The feedback updating unit collects user feedback data on the product recommendation table in real time, calculates the conversion rate and average similarity, determines whether the product recommendation table is consistent with the user's actual purchase situation, updates the sample matrix, and updates the parameters in the user prediction model.

Images