A machine learning-based user behavior profiling content generation method

Abstract

This invention relates to the field of behavioral profile generation technology, specifically disclosing a method for generating user behavior profile content based on machine learning, including the following steps: Step S1: Collect historical behavior logs of multiple users, count the total interaction frequency of a single user, distinguish active and inactive users based on thresholds, and construct frequency vectors for each user's behavior items; Step S2: For a single user, calculate the angle between its frequency vector and the vectors of all active users, select the K active users with the smallest angle, aggregate their vectors and concatenate them with the original vector of the target user to generate an enhanced behavioral feature vector; Step S3: Utilize two sets of linear mapping matrices to operate on the enhanced features respectively to obtain frequency and interest representation vectors, calculate the mean cosine similarity between the two, iteratively optimize the classification based on this value, determine the final interest representation vector, and output the user behavior profile text accordingly.

Description

Technical Field

[0001] This invention relates to the field of behavioral profile generation technology, and specifically to a method for generating user behavior profile content based on machine learning. Background Technology

[0002] In internet service and user operation scenarios, building accurate user behavior profiles is the core prerequisite for achieving personalized services and precision marketing.

[0003] However, in existing technologies, active user data is large and has rich behavioral patterns, while inactive user data is sparse and lacks behavioral information, resulting in a serious imbalance in sample distribution. User behavior profiles generated based on traditional methods inevitably bias towards high-frequency active groups and are difficult to comprehensively and objectively reflect the true behavioral characteristics of all users.

[0004] Traditional user profiling relies heavily on conventional machine learning models. These models are trained on a large sample base and, in cases of sample imbalance, prioritize learning the behavioral patterns of active users, resulting in weak capabilities for mining the behavioral characteristics of inactive users and significant biases in the profiling results. This bias not only prevents operational decisions from covering the inactive user group and missing opportunities to tap into potential value, but also renders personalized recommendations and targeted outreach strategies ineffective due to the one-sidedness of the profiling, thus reducing service efficiency and user experience. Summary of the Invention

[0005] The purpose of this invention is to provide a method for generating user behavior profiles based on machine learning, thereby solving the following technical problems.

[0006] The objective of this invention can be achieved through the following technical solutions: A method for generating user behavior profiles based on machine learning includes the following steps: Step S1: Obtain historical behavior logs of several users, count the total interaction frequency of each user, and divide all users into active users and inactive users according to the preset interaction frequency threshold; and generate the behavior item frequency vector of each user. Step S2: For any user, sequentially obtain the angle between the user's behavior item frequency vector and the behavior item frequency vector of each active user, select the K active users with the smallest angle, sum the behavior item frequency vectors of the K active users, and then concatenate them with the user's behavior item frequency vector to obtain the user's enhanced behavior feature vector. Step S3: Multiply the user's enhanced behavioral feature vector by the first linear mapping matrix and the second linear mapping matrix respectively to obtain the frequency representation vector and the interest representation vector, and obtain the absolute value of the cosine similarity between the user's frequency representation vector and the interest representation vector to obtain the average similarity value; perform iterative classification on the user's frequency representation vector and interest representation based on the total similarity value to obtain the final current interest representation vector; generate the user's behavioral profile text based on the current interest representation vector.

[0007] As a further aspect of the present invention: the historical behavior log includes each behavior item of the user and the interaction frequency of each behavior item; the total interaction frequency is the sum of the interaction frequencies of each behavior item of the user, and the interaction frequency threshold is the median of the total interaction frequencies of each user.

[0008] As a further aspect of the present invention: the process of generating the frequency vector of user behavior items includes: Each behavior item is numbered, and a two-dimensional coordinate system is established with the behavior item number as the x-axis and the interaction frequency as the y-axis. The numbers of each user's behavior items and the corresponding interaction frequencies are converted into coordinate points in the two-dimensional coordinate system. Based on each discrete coordinate point, a regression line segment is obtained using the least squares method. The endpoint of the regression line segment closest to the origin of the two-dimensional coordinate system is designated as the starting point, and the other endpoint is designated as the ending point. The behavior item frequency vector is obtained by taking the direction from the starting point to the ending point as the direction and the length of the regression line segment as the modulus.

[0009] As a further aspect of the present invention: the process of summing the frequency vectors of the behavior items of K active users includes: Input the included angles of the K active users into the softmax function to obtain normalized weight coefficients. Multiply the frequency vectors of the behavior items of the K active users by the corresponding normalized weight coefficients to obtain K weighted vectors. Add the K weighted vectors point by point to obtain the summed vector.

[0010] As a further embodiment of the present invention: K∈[5,20], and K is an integer.

[0011] As a further aspect of the present invention: the first linear mapping matrix is ​​a real number matrix with the number of rows being a first preset dimension and the number of columns being the dimension of the enhanced behavior feature vector; the second linear mapping matrix is ​​a real number matrix with the number of rows being a preset second preset dimension and the number of columns being the dimension of the enhanced behavior feature vector; the first preset dimension and the second preset dimension are both preset positive integers.

[0012] As a further aspect of the present invention: the process of iteratively classifying the user's frequency representation vector and interest representation based on the total similarity value includes: The frequency representation vector is input into a first classifier to obtain a first predicted category, and the interest representation vector is input into a second classifier to obtain a second preset category. A first classification error and a second classification error are obtained respectively. Then, parameter updates are performed: the parameters of the second linear mapping matrix and the second classifier are fixed, and the parameters of the first linear mapping matrix and the first classifier are updated in the direction of reducing the sum of the first classification error and the total similarity value; the parameters of the first linear mapping matrix, the first classifier, and the second classifier are fixed, and the parameters of the second linear mapping matrix are updated in the direction of increasing the second classification error and decreasing the total similarity value; the parameters of the first linear mapping matrix, the first classifier, and the second linear mapping matrix are fixed, and the parameters of the second classifier are updated in the direction of reducing the second classification error. Repeatedly acquire frequency representation vectors and interest representation vectors, acquire new first classification error and second classification error, and continue to perform parameter updates until the classification accuracy of the second classifier for the interest representation vector falls into the preset random guessing range, then stop parameter updates and extract the current interest representation vector.

[0013] As a further aspect of the present invention: the process of generating user behavior profile text based on the template of the current interest representation vector includes: A preset profile text template library is provided, which includes feature vectors of each template. The similarity between the current interest representation vector and the feature vectors of each template in the profile text template library is obtained, and the template with the highest similarity is selected as the corresponding template of the current interest representation vector. The current interest representation vector is multiplied by the behavior item projection matrix to obtain the preference value of each behavior item. Several behavior items with the highest preference values ​​are selected and filled into the corresponding template to generate the user's behavior profile text.

[0014] The beneficial effects of this invention are: This invention divides users into groups based on total interaction frequency, calculates the cosine similarity of each user's behavior frequency vector with that of each active user, selects the K most similar active users, performs weighted fusion, and concatenates it with their own vector to form an enhanced behavior feature vector. This step transfers the rich interaction information of active users to inactive users, making up for the feature loss caused by the sparse behavior of the latter, so that inactive users can also obtain sufficient feature expression, thus alleviating the profile bias caused by data imbalance from the source.

[0015] Secondly, this invention further separates the frequency representation vector and the interest representation vector by multiplying the enhanced behavioral feature vector with the first and second linear mapping matrices respectively, and forces the two representations to tend to be orthogonal in the vector space by minimizing the sum of the absolute values ​​of their cosine similarity. This step decouples the interaction frequency attribute from the content interest attribute, avoiding the pollution of the interest feature by the frequency information. During the alternating parameter update process, the first classifier uses frequency representation vectors to predict active and inactive categories, maintaining the distinguishability of frequency information. The second classifier predicts categories based on interest representation vectors and updates the second mapping matrix by fixing the first mapping matrix and the classifier, increasing the second classification error and decreasing the total similarity value, forming adversarial training. This gradually removes frequency traces from the interest representation vectors until the classification accuracy converges to the level of random guessing. At this point, the extracted interest representation vectors have been stripped of interaction frequency bias, retaining only content attributes that reflect the user's true preferences. Based on the similarity matching between this pure interest representation vector and the profile text template library, as well as the calculation of behavioral item preference scores, the generated user behavior profile text can accurately describe the user's intrinsic interests without bias due to the user's activity level, significantly improving the objectivity and fairness of the profile. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a schematic diagram illustrating the steps of a user behavior profile content generation method based on machine learning according to the present invention; Figure 2 This is a schematic diagram of the structure of a user behavior profile content generation method based on machine learning according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, this invention is a method for generating user behavior profiles based on machine learning, comprising the following steps: Step S1: Obtain historical behavior logs of several users, count the total interaction frequency of each user, and divide all users into active users and inactive users according to the preset interaction frequency threshold; and generate the behavior item frequency vector of each user. In a preferred embodiment of the present invention, the historical behavior log includes each behavior item of the user and the interaction frequency of each behavior item; the total interaction frequency is the sum of the interaction frequencies of each behavior item of the user, and the interaction frequency threshold is the median of the total interaction frequency of each user. In a preferred embodiment of the present invention, the process of generating the frequency vector of user behavior items includes: Each behavior item is numbered, and a two-dimensional coordinate system is established with the behavior item number as the x-axis and the interaction frequency as the y-axis. The numbers of each user's behavior items and the corresponding interaction frequencies are converted into coordinate points in the two-dimensional coordinate system. Based on each discrete coordinate point, a regression line segment is obtained using the least squares method. The endpoint of the regression line segment closest to the origin of the two-dimensional coordinate system is designated as the starting point, and the other endpoint is designated as the ending point. The behavior item frequency vector is obtained with the direction from the starting point to the ending point as the direction and the length of the regression line segment as the modulus. It should be noted that the historical behavior logs are sourced from the application's backend database. Each log entry contains a user identifier, a behavior item identifier, and a timestamp of the current interaction. The logs are grouped and aggregated by user identifier, and the number of interactions of each user with each behavior item is accumulated to obtain a user-behavior item interaction frequency table. For each user, the number of interactions on all behavior items is summed to obtain the user's total interaction frequency. The total interaction frequencies of all users are sorted from smallest to largest, and the median is taken as a preset interaction frequency threshold. Users whose total interaction frequency is greater than or equal to the interaction frequency threshold are classified as active users, and the remaining users are classified as inactive users. When generating the behavior item frequency vector, for each user, the number of interactions for each behavior item is extracted from the user-behavior item interaction frequency table. The interaction counts of all behavior items are organized into a fixed-length vector according to the preset behavior item arrangement order. The fixed-length vector is the behavior item frequency vector for that user. When a user has no interaction record for a certain behavior item, the value of the element at the corresponding position in the fixed-length vector is set to zero. The dimension of the fixed-length vector is consistent with the total number of all behavior items, and the value on each dimension is the interaction frequency of that user on that behavior item. Step S2: For any user, sequentially obtain the angle between the user's behavior item frequency vector and the behavior item frequency vector of each active user, select the K active users with the smallest angle, sum the behavior item frequency vectors of the K active users, and then concatenate them with the user's behavior item frequency vector to obtain the user's enhanced behavior feature vector. In a preferred embodiment of the present invention, the process of summing the frequency vectors of the behavior items of K active users includes: Input the included angles of K active users into the softmax function to obtain normalized weight coefficients. Multiply the frequency vectors of the behavior items of K active users by the corresponding normalized weight coefficients to obtain K weighted vectors. Add the K weighted vectors point by point to obtain the summed vector. In a preferred embodiment of the present invention, K∈[5, 20], and K is an integer; Specifically, by multiplying the frequency vectors of the behavior items of K active users by their corresponding normalized weight coefficients and then summing them, we get a vector with the same dimension as the frequency vector of a single behavior item. In a preferred embodiment of the present invention, the process of concatenating the user's behavior frequency vector includes concatenating the user's behavior frequency vector with the vector obtained by addition, generating an enhanced behavior feature vector whose dimension is the sum of the dimensions of the two vectors. Specifically: Assuming the frequency vector of a user's own behavior items is A with dimension D, and the vector obtained by adding them together is B with dimension D; then, the vectors A and B are joined end to end to form a new vector with dimension 2D, which is the enhanced behavior feature vector. The first D coordinate points of the enhanced behavior feature vector come from vector A, and the last D coordinate points come from vector B. It should be noted that the frequency vectors of all user behavior items obtained in step S1 are normalized using the L2 norm so that the magnitude of each vector is 1. For the user to be processed, the normalized frequency vector of the user and the normalized frequency vector of each active user are calculated as a vector dot product. The resulting dot product value is the cosine of the angle between the two vectors. The cosine values ​​are sorted from largest to smallest. Since the larger the cosine value, the smaller the angle, the K active users with the largest cosine values ​​are selected as the set of nearest active users, where K is a pre-set positive integer. Take the original behavior frequency vectors of the selected K active users before normalization, and sum these K vectors element-wise along the same dimension. The summed result is a single vector with the same dimension as the original behavior frequency vector. Then, concatenate the user's original behavior frequency vector with the summed vector, that is, take the elements of the user's behavior frequency vector as the first segment and the elements of the nearest neighbor's summed vector as the second segment, and merge them to generate a new vector with a length twice the dimension of the original vector. This new vector is the user's enhanced behavior feature vector. Perform the above operation on all users one by one to obtain the enhanced behavior feature vector of all users. Step S3: Multiply the user's enhanced behavior feature vector by the first linear mapping matrix and the second linear mapping matrix respectively to obtain the frequency representation vector and the interest representation vector, and obtain the absolute value of the cosine similarity between the user's frequency representation vector and the interest representation vector to obtain the average similarity value; perform iterative classification on the user's frequency representation vector and interest representation vector according to the total similarity value to obtain the final current interest representation vector; generate the user's behavior profile text according to the current interest representation vector; In a preferred embodiment of the present invention, the first linear mapping matrix is ​​a real number matrix with the number of rows being a first preset dimension and the number of columns being the dimension of the enhanced behavior feature vector; the second linear mapping matrix is ​​a real number matrix with the number of rows being a preset second preset dimension and the number of columns being the dimension of the enhanced behavior feature vector; the first preset dimension and the second preset dimension are both preset positive integers; In a preferred embodiment of the present invention, the process of iteratively classifying the user's frequency representation vector and interest representation based on the total similarity value includes: The frequency representation vector is input into a first classifier to obtain a first predicted category, and the interest representation vector is input into a second classifier to obtain a second preset category. A first classification error and a second classification error are obtained respectively. Then, parameter updates are performed: the parameters of the second linear mapping matrix and the second classifier are fixed, and the parameters of the first linear mapping matrix and the first classifier are updated in the direction of reducing the sum of the first classification error and the total similarity value; the parameters of the first linear mapping matrix, the first classifier, and the second classifier are fixed, and the parameters of the second linear mapping matrix are updated in the direction of increasing the second classification error and decreasing the total similarity value; the parameters of the first linear mapping matrix, the first classifier, and the second linear mapping matrix are fixed, and the parameters of the second classifier are updated in the direction of reducing the second classification error. Repeatedly acquire frequency representation vectors and interest representation vectors, acquire new first classification error and second classification error, and continue to perform parameter updates until the classification accuracy of the second classifier for the interest representation vector falls into the preset random guessing range, then stop parameter updates and extract the current interest representation vector. The random guessing interval is the interval where the classification accuracy falls within [45%, 55%]; It is worth noting that the first classifier and the second classifier each contain a fully connected layer and a softmax layer; the number of input nodes of the fully connected layer is the same as the dimension of the corresponding frequency representation vector or interest representation vector, and the number of output nodes is 2; In a preferred embodiment of the present invention, the process of generating user behavior profile text based on the template of the current interest representation vector includes: A preset profile text template library is provided, which includes feature vectors of each template. The similarity between the current interest representation vector and the feature vectors of each template in the profile text template library is obtained, and the template with the highest similarity is selected as the corresponding template of the current interest representation vector. The current interest representation vector is multiplied by the behavior item projection matrix to obtain the preference value of each behavior item. Several behavior items with the highest preference values ​​are selected and filled into the corresponding template to generate the user's behavior profile text. Understandably, two real-valued matrices are pre-constructed, namely a first linear mapping matrix and a second linear mapping matrix. The number of rows in the first linear mapping matrix is ​​a first preset dimension, and the number of columns is the dimension of the enhanced behavior feature vector. The number of rows in the second linear mapping matrix is ​​a second preset dimension, and the number of columns is the same as that in the first linear mapping matrix. The first and second preset dimensions are pre-defined positive integers, which can take the same or different values. The initial values ​​of each element in the two matrices are generated through random sampling. The enhanced behavior feature vectors of each user obtained in step S2 are used as column vectors, and then multiplied by the first linear mapping matrix and the second linear mapping matrix on the left to obtain the frequency representation vector and interest representation vector of the user. The dimension of the frequency representation vector is equal to the first preset dimension, and the dimension of the interest representation vector is equal to the second preset dimension. For any user, calculate the cosine similarity between its frequency representation vector and interest representation vector, and take the absolute value of the cosine similarity. Sum the absolute values ​​of the above for all users and divide by the total number of users to obtain the average similarity value. Further, the iterative classification process begins. The frequency representation vector is input into the first classifier to obtain the first predicted class probability; the interest representation vector is input into the second classifier to obtain the second predicted class probability; based on the user's actual activity or inactivity label, the first and second classification errors are calculated respectively; in each iteration, the parameters are updated alternately in the following order: First, fix the parameters of the second linear mapping matrix and the second classifier, and update the parameters of the first linear mapping matrix and the first classifier in the direction of reducing the sum of the first classification error and the average similarity value; then, fix the parameters of the first linear mapping matrix, the first classifier, and the second classifier, and update the parameters of the second linear mapping matrix in the direction of increasing the second classification error and decreasing the average similarity value; finally, fix the parameters of the first linear mapping matrix, the first classifier, and the second linear mapping matrix, and update the parameters of the second classifier in the direction of reducing the second classification error. Repeat the above iterative process, and calculate the classification accuracy of the second classifier for all users' interest representation vectors every preset number of rounds. Stop iterating when the classification accuracy falls into the preset random guessing interval for multiple consecutive rounds. At this time, extract the interest representation vector of each user as the final current interest representation vector. After obtaining the current interest representation vector, the cosine similarity between the vector and the template feature vectors of each template in the pre-built profile text template library is calculated one by one, and the template with the highest similarity is selected. The current interest representation vector is then multiplied by the preset behavior item projection matrix to obtain the preference score of each behavior item. The preset number of behavior items with the highest scores are selected as interest tags and filled into the template placeholders to generate the user's behavior profile text.

[0020] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A method for generating user behavior profile content based on machine learning, characterized in that, Includes the following steps: Step S1: Obtain historical behavior logs of several users, count the total interaction frequency of each user, and divide all users into active users and inactive users according to the preset interaction frequency threshold; and generate the behavior item frequency vector of each user. Step S2: For any user, sequentially obtain the angle between the user's behavior item frequency vector and the behavior item frequency vector of each active user, select the K active users with the smallest angle, sum the behavior item frequency vectors of the K active users, and then concatenate them with the user's behavior item frequency vector to obtain the user's enhanced behavior feature vector. Step S3: Multiply the user's enhanced behavior feature vector with the first linear mapping matrix and the second linear mapping matrix respectively to obtain the frequency representation vector and the interest representation vector, and obtain the absolute value of the cosine similarity between the user's frequency representation vector and the interest representation vector to obtain the average similarity value. Based on the total similarity value, the user's frequency representation vector and interest representation are iteratively classified to obtain the final current interest representation vector; based on the current interest representation vector, the user's behavioral profile text is generated.

2. The method for generating user behavior profile content based on machine learning according to claim 1, characterized in that, In step S1, the historical behavior log includes each behavior item of the user and the interaction frequency of each behavior item; the total interaction frequency is the sum of the interaction frequencies of each behavior item of the user, and the interaction frequency threshold is the median of the total interaction frequency of each user.

3. The user behavior profile content generation method based on machine learning according to claim 1, characterized in that, In step S1, the process of generating the user's behavior item frequency vector includes: Each behavior item is numbered, and a two-dimensional coordinate system is established with the behavior item number as the x-axis and the interaction frequency as the y-axis. The numbers of each user's behavior items and the corresponding interaction frequencies are converted into coordinate points in the two-dimensional coordinate system. Based on each discrete coordinate point, a regression line segment is obtained using the least squares method. The endpoint of the regression line segment closest to the origin of the two-dimensional coordinate system is designated as the starting point, and the other endpoint is designated as the ending point. The behavior item frequency vector is obtained by taking the direction from the starting point to the ending point as the direction and the length of the regression line segment as the modulus.

4. The method for generating user behavior profile content based on machine learning according to claim 1, characterized in that, In step S2, the process of summing the frequency vectors of the behavior items of K active users includes: Input the included angles of the K active users into the softmax function to obtain normalized weight coefficients. Multiply the frequency vectors of the behavior items of the K active users by the corresponding normalized weight coefficients to obtain K weighted vectors. Add the K weighted vectors point by point to obtain the summed vector.

5. The user behavior profile content generation method based on machine learning according to claim 1, characterized in that, In step S2, K∈[5,20], and K is an integer.

6. The method for generating user behavior profile content based on machine learning according to claim 1, characterized in that, In step S3, the first linear mapping matrix is ​​a real matrix with the number of rows being a first preset dimension and the number of columns being the dimension of the enhanced behavior feature vector; the second linear mapping matrix is ​​a real matrix with the number of rows being a preset second preset dimension and the number of columns being the dimension of the enhanced behavior feature vector; the first preset dimension and the second preset dimension are both preset positive integers.

7. The method for generating user behavior profile content based on machine learning according to claim 1, characterized in that, In step S3, the process of iteratively classifying the user's frequency representation vector and interest representation based on the total similarity value includes: The frequency representation vector is input into a first classifier to obtain a first predicted category, and the interest representation vector is input into a second classifier to obtain a second preset category. A first classification error and a second classification error are obtained respectively. Then, parameter updates are performed: the parameters of the second linear mapping matrix and the second classifier are fixed, and the parameters of the first linear mapping matrix and the first classifier are updated in the direction of reducing the sum of the first classification error and the total similarity value; the parameters of the first linear mapping matrix, the first classifier, and the second classifier are fixed, and the parameters of the second linear mapping matrix are updated in the direction of increasing the second classification error and decreasing the total similarity value; the parameters of the first linear mapping matrix, the first classifier, and the second linear mapping matrix are fixed, and the parameters of the second classifier are updated in the direction of reducing the second classification error. Repeatedly acquire frequency representation vectors and interest representation vectors, acquire new first classification error and second classification error, and continue to perform parameter updates until the classification accuracy of the second classifier for the interest representation vector falls into the preset random guessing range, then stop parameter updates and extract the current interest representation vector.

8. The method for generating user behavior profile content based on machine learning according to claim 1, characterized in that, In step S3, the process of generating user behavior profile text based on the template of the current interest representation vector includes: A preset profile text template library is provided, which includes feature vectors of each template. The similarity between the current interest representation vector and the feature vectors of each template in the profile text template library is obtained, and the template with the highest similarity is selected as the corresponding template of the current interest representation vector. The current interest representation vector is multiplied by the behavior item projection matrix to obtain the preference value of each behavior item. Several behavior items with the highest preference values ​​are selected and filled into the corresponding template to generate the user's behavior profile text.

Images