Digital user behavior big data mining method and system

Abstract

The present application provides a method and system for mining user behavior big data for digitization, which uses multiple user behavior log learning samples to debug the neural network. The second behavior feature mining neural network can automatically identify the behavior of the behavior log, thereby improving the recognition accuracy. In addition, based on the first behavior log feature of the first behavior log and the second sample user behavior feature of the second behavior log included in each tuple in the mining training database, the learning sample estimated commonality metric between the first sample user behavior feature and the second sample user behavior feature corresponding to each tuple is obtained, so as to debug the second behavior feature mining neural network based on the learning sample estimated commonality metric and the learning sample commonality metric corresponding to multiple tuples, and obtain the target behavior feature mining neural network. Reduce the focus on high-frequency behaviors and focus on other behavior data items in the behavior log. The mined behavior log features more accurately reflect the semantics of the behavior log.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method and system for mining digital user behavior big data. Background Art

[0002] With the advent of the digital age, user behavior data has shown explosive growth. These data contain rich behavioral information of users on various digital platforms, such as browsing history, search keywords, purchase behavior, etc., which are of great significance for enterprises to understand user needs, optimize service quality and formulate marketing strategies. However, how to effectively mine and analyze these massive user behavior data to extract valuable information has always been a technical challenge faced by the industry. Traditional user behavior analysis methods mainly rely on manually set rules and statistical analysis. This method can still meet the needs when processing small-scale data, but it often seems to be unable to cope with large-scale and high-dimensional user behavior data. On the one hand, it is difficult for manually set rules to cover all possible user behavior patterns, resulting in deviations in the analysis results; on the other hand, traditional statistical analysis methods are inefficient when processing high-dimensional data and it is difficult to respond to market changes in real time. In recent years, with the rapid development of machine learning and deep learning technologies, their application in the field of user behavior analysis has gradually received attention. By training machine learning models, key features can be automatically extracted from user behavior data and user behavior trends can be predicted. However, the diversity and complexity of user behavior data require that the model can accurately identify multiple aspects of user behavior and give corresponding confidence levels in order to analyze user behavior patterns more finely. Secondly, there are often correlations and similarities between user behaviors, which are important for understanding the overall user behavior patterns, but existing methods often ignore this point. Summary of the invention

[0003] In view of this, the present application at least provides a method, device, system and storage medium for digital user behavior big data mining. The technical solution of the present application is implemented as follows:

[0004] In a first aspect, an embodiment of the present application provides a method for mining user behavior big data for digitalization, the method comprising: obtaining a training database, the training database comprising a plurality of user behavior log learning samples and a priori labels of learning samples for each user behavior log learning sample, a plurality of user behavior log learning sample tuples and a learning sample commonality measurement value corresponding to each user behavior log learning sample tuple, the learning sample commonality measurement value corresponding to the user behavior log learning sample tuple being a commonality measurement value between a first user behavior log learning sample and a second user behavior log learning sample included in the user behavior log learning sample tuple; mining sample user behavior features corresponding to each user behavior log learning sample, and performing confidence distribution mapping on each sample user behavior feature to obtain confidences of each user behavior log learning sample for a plurality of candidate classification labels; determining a learning sample estimated label for each user behavior log learning sample based on the confidences of the plurality of candidate classification labels, and The learning sample estimated labels and learning sample prior labels corresponding to the user behavior log learning samples are used to debug the first behavior feature mining neural network to obtain the second behavior feature mining neural network; the first sample user behavior features of the first user behavior log learning sample and the second sample user behavior features of the second user behavior log learning sample included in each user behavior log learning sample binary are respectively mined to obtain the learning sample estimated commonality metric value between the first sample user behavior feature and the second sample user behavior feature corresponding to each user behavior log learning sample binary, and the second behavior feature mining neural network is debugged according to the learning sample estimated commonality metric values ​​and the learning sample commonality metric values ​​corresponding to the multiple user behavior log learning sample binary groups to obtain the target behavior feature mining neural network; the target behavior log of the target user is received, the behavior log features of the target behavior log are determined according to the target behavior feature mining neural network, and the behavior log profile matching the target user is determined according to the behavior log features of the target behavior log.

[0005] In a second aspect, the present application provides a user behavior big data mining device for digitalization, including: a training data acquisition module, used to acquire a training database, wherein the training database includes multiple user behavior log learning samples and a priori labels of learning samples for each user behavior log learning sample, multiple user behavior log learning sample tuples and learning sample commonality measurement values ​​corresponding to each user behavior log learning sample tuple, wherein the learning sample commonality measurement value corresponding to the user behavior log learning sample tuple is a commonality measurement value between a first user behavior log learning sample and a second user behavior log learning sample included in the user behavior log learning sample tuple; a behavior feature mining module, used to mine sample user behavior features corresponding to each user behavior log learning sample, and to perform confidence distribution mapping on each sample user behavior feature to obtain the confidence of each user behavior log learning sample for multiple candidate classification labels; a first network debugging module, used to determine the learning sample estimated label of each user behavior log learning sample based on the confidence of the multiple candidate classification labels, and to The learning sample estimated labels and learning sample prior labels corresponding to multiple user behavior log learning samples are used to debug the first behavior feature mining neural network to obtain the second behavior feature mining neural network; the second network debugging module is used to respectively mine the first sample user behavior feature of the first user behavior log learning sample and the second sample user behavior feature of the second user behavior log learning sample included in each user behavior log learning sample binary, obtain the learning sample estimated commonality measurement value between the first sample user behavior feature and the second sample user behavior feature corresponding to each user behavior log learning sample binary, and debug the second behavior feature mining neural network according to the learning sample estimated commonality measurement value and the learning sample commonality measurement value corresponding to the multiple user behavior log learning sample binary to obtain the target behavior feature mining neural network; the behavior profile determination module is used to receive the target behavior log of the target user, determine the behavior log feature of the target behavior log according to the target behavior feature mining neural network, and determine the behavior log profile matching the target user according to the behavior log feature of the target behavior log.

[0006] According to a third aspect, a computer system includes a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor implements the steps of the above method when executing the program.

[0007] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which implements the steps in the above-described method when executed by a processor.

[0008] The beneficial effects of the present application include at least: in the embodiment of the present application, by mining the sample user behavior features corresponding to each user behavior log learning sample in the training database, a confidence distribution mapping is performed on each sample user behavior feature to obtain the confidence of each user behavior log learning sample for multiple candidate classification labels, so as to determine the learning sample estimated label of each user behavior log learning sample, so that the first behavior feature mining neural network can be debugged based on the learning sample estimated labels and learning sample prior labels corresponding to multiple user behavior log learning samples to obtain the second behavior feature mining neural network. Using multiple user behavior log learning samples to debug the neural network, the second behavior feature mining neural network can automatically perform behavior recognition on the behavior log, thereby improving recognition accuracy and reliability. In addition, based on the first behavior log feature of the first behavior log and the second sample user behavior feature of the second behavior log included in each user behavior log learning sample binary in the mining training database, the learning sample estimated commonality measurement value between the first sample user behavior feature and the second sample user behavior feature corresponding to each user behavior log learning sample binary is obtained, so that the second behavior feature mining neural network can be debugged based on the learning sample estimated commonality measurement value and the learning sample commonality measurement value corresponding to multiple user behavior log learning sample binary groups to obtain the target behavior feature mining neural network. By debugging the neural network using the user behavior log learning sample binary group and the learning sample commonality measurement value, the target behavior feature mining neural network can learn the semantic commonality of the user behavior log learning samples in the user behavior log learning sample binary group. Then, when the behavior log is subjected to feature mining based on the target behavior feature mining neural network, the focus on high-frequency behavior can be reduced, and the focus on other behavior data items in the behavior log can be focused, so that the mined behavior log features can more accurately reflect the semantics of the behavior log.

[0009] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present application and are used together with the specification to illustrate the technical solution of the present application.

[0011] Figure 1 A schematic diagram of the implementation process of a digital user behavior big data mining method provided in an embodiment of the present application.

[0012] Figure 2 A schematic diagram of the composition structure of a digital user behavior big data mining device provided in an embodiment of the present application.

[0013] Figure 3A hardware entity diagram of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions of the present application are further elaborated in detail below in conjunction with the drawings and embodiments. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0015] The embodiment of the present application provides a method for mining user behavior big data for digitization, which can be executed by a processor of a computer system, wherein the computer system can refer to a device with data processing capabilities such as a server, a laptop, a tablet computer, and a desktop computer.

[0016] Figure 1 A schematic diagram of the implementation process of a digital user behavior big data mining method provided in the embodiment of the present application is as follows: Figure 1 As shown, the method includes:

[0017] Step S100: Obtain a training database, wherein the training database includes multiple user behavior log learning samples and a learning sample prior label of each user behavior log learning sample, multiple user behavior log learning sample tuples, and a learning sample commonality measurement value corresponding to each user behavior log learning sample tuple, wherein the learning sample commonality measurement value corresponding to the user behavior log learning sample tuple is a commonality measurement value between a first user behavior log learning sample and a second user behavior log learning sample included in the user behavior log learning sample tuple.

[0018] In the embodiment of the present application, the computer system constructs a training database, which contains a large number of user behavior log learning examples, which are essentially behavioral data records generated by users on digital platforms (such as e-commerce websites, social media platforms, etc.). For example, in an e-commerce website, user behavior log learning examples may include users' browsing history, search keywords, purchase records, click behaviors, etc., and each record describes in detail the specific behavior of the user at a specific time point.

[0019] In addition to the user behavior log learning samples themselves, the training database also contains the learning sample prior tags for each sample. These prior tags are pre-defined labels that are used to identify the category to which the user behavior log learning sample belongs or certain characteristics it has. For example, for user purchase behavior on an e-commerce website, prior tags may include "high purchase intention", "medium purchase intention", "low purchase intention", etc. These tags are determined based on factors such as the user's purchase history and browsing behavior, and are used to guide the learning process of the machine learning model.

[0020] In addition, the training database also contains multiple user behavior log learning sample tuples, which are composed of two related user behavior log learning samples. These tuples are selected based on some association rules or similarity measures, aiming to capture the intrinsic connections and regularities between user behaviors. For example, the user behavior log learning samples in two tuples may be browsing and purchasing records of the same type of goods, or they may be behavioral data from the same user at different time points.

[0021] For each user behavior log learning sample pair, the training database also records the corresponding learning sample commonality metric, that is, the commonality metric between the two samples (that is, the similarity score). This commonality metric is calculated by a specific similarity algorithm and reflects the degree of similarity between the two user behavior log learning samples in terms of content, pattern, or semantics. For example, if two user behavior log learning samples are browsing records about digital products, and the browsed product types, price ranges, and other features are highly consistent, then the commonality metric between them will be high.

[0022] Step S200: mining the sample user behavior features corresponding to each user behavior log learning sample, and performing confidence distribution mapping on each sample user behavior feature to obtain the confidence of each user behavior log learning sample being respectively labeled as a plurality of candidate classifications.

[0023] In the embodiment of the present application, the main task of step S200 is to mine the sample user behavior features corresponding to each user behavior log learning sample, and perform confidence distribution mapping on these features to determine the degree of association between the user behavior log learning sample and multiple candidate classification labels.

[0024] Specifically, the computer system uses various feature extraction techniques to mine sample user behavior features from user behavior log learning samples. These features may include the user's behavior type (such as browsing, searching, purchasing, etc.), the timestamp of the behavior, the associated product or content category, the user's interaction depth (such as page dwell time, number of clicks, etc.), etc. In order to more accurately capture the user's behavior pattern, the feature extraction process can use deep learning models, such as recurrent neural networks (RNN) or long short-term memory networks (LSTM), to process serialized user behavior data. Once the sample user behavior features are extracted, the computer system will then perform confidence distribution mapping, that is, multi-classification of these features. Multi-classification is a special classification task in which each sample may belong to multiple categories at the same time. In the scenario of user behavior big data mining, multi-label classification is particularly useful because user behavior often involves multiple aspects or points of interest.

[0025] In order to perform confidence distribution mapping, that is, to classify multiple labels, the computer system can use classic machine learning algorithms such as logistic regression, decision tree, random forest, etc., which are referred to as confidence distribution mapping algorithms in the embodiments of the present application, or use more advanced deep learning models, such as convolutional neural networks (CNN) or Transformer models. These models will assign multiple candidate classification labels to each user behavior log learning sample based on the extracted sample user behavior characteristics, and assign a confidence level to each label. The confidence level reflects the model's degree of certainty that the sample belongs to a certain category, and the value range is usually between 0 and 1. The larger the value, the higher the degree of certainty.

[0026] Taking the user behavior mining of e-commerce platforms as an example, suppose a user behavior log learning sample contains the user's browsing, searching and purchasing records on the platform. Through feature extraction, the computer system can obtain the following features: browsing mobile phone products, searching for Bluetooth headsets, purchasing smart watches, etc. Next, the confidence distribution mapping algorithm will assign multiple candidate classification tags to the sample based on these features, such as "interested in electronic products", "pursuing high-quality life", etc., and assign a confidence level to each tag. In this way, the computer system can fully and accurately understand the user's behavior patterns and interest preferences, providing strong support for subsequent user portrait construction and personalized recommendations.

[0027] Step S300: Determine the learning sample estimation label of each user behavior log learning sample based on the confidence of the multiple candidate classification labels, and debug the first behavior feature mining neural network according to the learning sample estimation labels and learning sample prior labels corresponding to the multiple user behavior log learning samples to obtain the second behavior feature mining neural network.

[0028] In step S300, the computer system determines the learning sample estimation mark of each user behavior log learning sample based on the confidence of multiple candidate classification marks, and then uses these estimation marks and known learning sample prior marks to debug the first behavior feature mining neural network, thereby obtaining the optimized second behavior feature mining neural network. Specifically, the computer system first performs multi-label classification on each user behavior log learning sample to generate multiple candidate classification marks and corresponding confidences. These candidate classification marks represent the categories or features to which the user behavior may belong, and the confidence reflects the certainty of the classification model on the classification result. For example, in the user behavior mining of the e-commerce platform, a user behavior log learning sample may be classified into multiple candidate classification marks such as "interested in electronic products" (confidence 0.8), "concerned about fashion trends" (confidence 0.6). Next, the computer system determines one or more learning sample estimation marks for each user behavior log learning sample based on the confidence of these candidate classification marks, combined with preset thresholds or other strategies. These estimation marks are the model's preliminary judgment of the user behavior characteristics, which will be used for subsequent model debugging and optimization. After determining the estimated labels of the learning samples, the computer system will use these estimated labels and the known prior labels of the learning samples to debug the first behavior feature mining neural network. Prior labels are the behavioral characteristics or classification results of users in the real world, which are usually obtained through manual annotation or historical data. During the debugging process, the computer system compares the difference between the estimated labels and the prior labels, calculates the loss function (such as the cross entropy loss function), and then uses the back propagation algorithm to update the weights and bias terms of the neural network to minimize the loss function value. Taking the convolutional neural network (CNN) in deep learning as an example, the network can extract high-level features in the user behavior log learning samples through multi-layer convolution and pooling operations, and output the confidence of multiple candidate classification labels through the fully connected layer. During the debugging process, CNN will adjust its convolution kernel and weight parameters according to the difference between the estimated labels and the prior labels to improve the recognition accuracy of user behavior features. After multiple iterations and optimizations, the first behavior feature mining neural network will gradually converge to a better state, at which time it will be saved as the second behavior feature mining neural network. This optimized neural network will have stronger generalization ability and higher accuracy, and will be able to more accurately mine and analyze user behavior characteristics, providing companies with more precise marketing and personalized service support.

[0029] Step S400: respectively mine the first sample user behavior feature of the first user behavior log learning sample and the second sample user behavior feature of the second user behavior log learning sample included in each user behavior log learning sample binary, obtain the learning sample estimated commonality metric value between the first sample user behavior feature and the second sample user behavior feature corresponding to each user behavior log learning sample binary, and debug the second behavior feature mining neural network according to the learning sample estimated commonality metric values ​​and the learning sample commonality metric values ​​corresponding to the multiple user behavior log learning sample binary groups to obtain the target behavior feature mining neural network.

[0030] In step S400, the computer system mines the first sample user behavior features of the first user behavior log learning sample and the second sample user behavior features of the second user behavior log learning sample in each user behavior log learning sample binary. These features may include the user's browsing habits, purchase history, search keywords, etc., which together constitute a comprehensive description of the user's behavior. For example, in the scenario of an e-commerce platform, a user behavior log learning sample binary may contain the behavior data of two users, one of whom often browses and purchases electronic products, while the other user shows a strong interest in fashion clothing. Next, the computer system calculates the learning sample estimated commonality metric between the first sample user behavior features and the second sample user behavior features corresponding to each user behavior log learning sample binary. This commonality metric reflects the degree of similarity between the two user behavior log learning samples in terms of features. Technically, this is usually achieved by calculating the cosine similarity, Euclidean distance or other metric methods between feature vectors. Continuing with the above example, if two users are both searching for the keyword “smart watch” and their browsing and purchasing records also show a common interest in electronic products, then the estimated commonality measure between these two user behavior log learning examples will be high.

[0031] Then, the computer system uses the estimated commonality metric values ​​of the learning samples and the known commonality metric values ​​of the learning samples corresponding to the binary groups of the learning samples of multiple user behavior logs to debug the second behavior feature mining neural network. The known commonality metric values ​​of the learning samples are calculated based on the actual user behavior data, and they represent the true similarity between user behaviors. During the debugging process, the computer system compares the difference between the estimated commonality metric values ​​and the known commonality metric values, and adjusts the parameters of the neural network accordingly to reduce the difference. This adjustment process is usually implemented through the back propagation algorithm and the gradient descent method, aiming to enable the neural network to better capture the similarity between user behaviors. Finally, after multiple iterations and optimizations, the computer system obtains a target behavior feature mining neural network. This neural network can not only accurately mine the characteristics of user behaviors, but also effectively identify the similarities between user behaviors, providing enterprises with more in-depth user behavior analysis and more accurate personalized services. In the scenario of e-commerce platforms, the target behavior feature mining neural network can help enterprises identify user groups with similar purchasing habits or interests, so as to recommend more accurate products and services to them.

[0032] Step S500: receiving a target behavior log of a target user, determining a behavior log feature of the target behavior log based on a neural network mining method according to the target behavior feature, and determining a behavior log profile that matches the target user based on the behavior log feature of the target behavior log.

[0033] In step S500, the computer system receives the target behavior log of the target user, and determines the specific features of these behavior logs based on the trained target behavior feature mining neural network, and then constructs a behavior log portrait that matches the target user. Specifically, when the computer system receives the target behavior log of the target user, these logs may contain various interaction records of the user on the digital platform, such as browsing records, search keywords, click behaviors, purchase records, etc. These behavior logs are an important data basis for understanding user interest preferences and consumption habits. Next, the computer system uses the target behavior feature mining neural network to conduct an in-depth analysis of the target behavior log. This neural network is trained through a large number of user behavior log learning examples in the previous steps, and it has learned how to accurately extract user behavior features and identify similarities between user behaviors. Therefore, when the target behavior log of the target user is input, the neural network can quickly and accurately extract the key behavior features therein. These behavior features may include user interests, consumption tendencies, active time periods, etc. For example, if the target user frequently browses electronic products such as smart watches and Bluetooth headsets on the e-commerce platform and searches for related keywords many times, the neural network can extract the key behavior feature of "interested in electronic products".

[0034] After determining the behavioral characteristics of the target behavior log, the computer system can build a behavior log profile that matches the target user based on these characteristics. The behavior log profile is a comprehensive description of the user's behavior characteristics, which can help companies understand the user's behavior patterns, interest preferences, and consumption habits more comprehensively. In this process, the computer system can match the extracted behavior characteristics with the preset user profile template, or directly generate a new user profile based on the behavior characteristics.

[0035] Continuing with the above example, if the target user's behavioral characteristics show a strong interest in electronic products, the computer system can generate a user profile containing labels such as "electronic product enthusiast" and "high spending power". This profile can not only be used for precision marketing and personalized recommendations, but also provide strong support for companies to formulate market strategies and optimize product design.

[0036] Through step S500, the computer system can achieve in-depth analysis and accurate profiling of target user behavior, providing enterprises with more effective user behavior insights and marketing support.

[0037] As an implementation mode, the sample user behavior feature corresponding to each user behavior log learning sample includes a representation array corresponding to each sample behavior data item in each user behavior log learning sample. Based on this, in step S200, confidence distribution mapping is performed on each sample user behavior feature to obtain the confidence of each user behavior log learning sample as a plurality of candidate classification labels, which may specifically include:

[0038] Step S210: integrating the representation information of the representation array corresponding to each sample behavior data item in each user behavior log learning sample to obtain the integrated behavior log feature of each user behavior log learning sample.

[0039] In an embodiment of the present application, step S210 integrates the representation array corresponding to each sample behavior data item in each user behavior log learning sample to obtain the integrated behavior log features of each user behavior log learning sample. First of all, it is necessary to clarify the concepts of user behavior log learning samples and user behavior data items. User behavior log learning samples usually refer to a series of behavior records generated by users on digital platforms, such as browsing, searching, purchasing, etc. Each sample behavior data item is a specific component of these behavior records. For example, a user browses a product page, and this browsing behavior is a sample behavior data item.

[0040] In the data processing process, in order to better understand and analyze user behavior, each sample behavior data item is usually feature extracted to obtain a corresponding characterization array. These characterization arrays contain key information describing the sample behavior data item, such as the category of goods browsed, browsing time, number of clicks, etc. These feature information are usually expressed in numerical form to facilitate subsequent calculations and analysis. Next, the role of step S210 is to integrate the characterization arrays corresponding to all sample behavior data items in each user behavior log learning sample. The purpose of integration is to merge the feature information of multiple sample behavior data items into a unified integrated behavior log feature in order to more comprehensively reflect the user's overall behavior pattern. Specifically, the integration process can be implemented in a variety of ways. A common method is to splice or superimpose the characterization arrays corresponding to all sample behavior data items to form a larger feature vector. For example, suppose a user behavior log learning sample contains three sample behavior data items: browsing the electronic product page, searching for Bluetooth headset keywords, and purchasing a smart watch. The representation arrays corresponding to these three sample behavior data items may be [1,0,0] (indicating browsing the electronic product page), [0,1,0] (indicating searching for the Bluetooth headset keyword), and [1,0,1] (indicating purchasing a smart watch). During the integration process, the computer system can splice these three representation arrays into a new feature vector [1,0,0,0,1,0,1,0,1], which represents the integrated behavior log features of the user behavior log learning sample.

[0041] The integrated behavior log features not only contain the specific information of user behavior, but also reflect the diversity and complexity of user behavior. By integrating the behavior log features, the computer system can more comprehensively understand the user's interests, preferences, consumption habits, etc., and provide strong support for the subsequent user portrait construction and personalized recommendations. In the scenario of e-commerce platforms, integrating behavior log features can help companies identify users' potential needs, push more accurate product recommendations, and improve user satisfaction and purchase conversion rates.

[0042] Step S220: estimating the confidence of the representation arrays corresponding to the plurality of candidate classification labels respectively for the integrated behavior log features.

[0043] In an embodiment of the present application, step S220 performs multi-label classification on the integrated behavior log features and estimates the confidence corresponding to each classification mark. First, after processing in step S210, the computer system has obtained the integrated behavior log features of each user behavior log learning sample. This integrated feature is a high-dimensional vector that contains various behavioral information generated by the user on the digital platform, such as browsing history, search keywords, purchase behavior, etc. This feature vector comprehensively reflects the user's interest preferences, consumption habits and other behavioral patterns. Next, in step S220, the computer system will use the confidence distribution mapping algorithm to classify the integrated behavior log features. Multi-label classification is a special classification task in which each sample may belong to multiple categories or have multiple features at the same time. In the scenario of user behavior big data mining, multi-label classification is particularly important because user behavior usually involves multiple aspects or points of interest.

[0044] The confidence distribution mapping algorithm can be a model based on traditional machine learning algorithms (such as logistic regression, support vector machine, decision tree, etc.) or a model based on deep learning (such as convolutional neural network, recurrent neural network, Transformer, etc.). These models have been trained with a large amount of data and have learned to extract key information from the integrated behavior log features and determine the multiple categories or features that a user may belong to.

[0045] In a specific implementation, the computer system inputs the integrated behavior log features into the confidence distribution mapping algorithm, and the model outputs a probability distribution vector, each element of which corresponds to the confidence of a candidate classification tag. The confidence indicates the degree of certainty of the model that the user belongs to the classification tag, and the value range is usually between 0 and 1, and the larger the value, the higher the degree of certainty. Taking the user behavior mining of the e-commerce platform as an example, it is assumed that the integrated behavior log features indicate that the user has shown a certain interest in both electronic products and fashion clothing. The confidence distribution mapping algorithm can output the following probability distribution vector: [0.8, 0.6], where the first element corresponds to the confidence of the classification tag "interested in electronic products", and the second element corresponds to the confidence of the classification tag "interested in fashion clothing". This probability distribution vector reflects the model's judgment result on the user's interest preference. Through the processing of step S220, the computer system can accurately estimate the confidence of the user behavior log learning sample belonging to multiple candidate classification tags, which provides an important basis for subsequent model debugging and optimization. At the same time, these confidence information can also be used in application scenarios such as building user portraits and making personalized recommendations to improve user experience and satisfaction.

[0046] The step S300, determining the learning sample estimation mark of each user behavior log learning sample according to the confidence of the multiple candidate classification marks, and debugging the first behavior feature mining neural network according to the learning sample estimation marks and the learning sample prior marks corresponding to the multiple user behavior log learning samples to obtain the second behavior feature mining neural network, may specifically include:

[0047] Step S310: determining one or more candidate classification labels from the multiple candidate classification labels according to the confidences of the multiple candidate classification labels, and determining them as learning sample estimated labels for each user behavior log learning sample.

[0048] In an embodiment of the present application, step S310 determines the estimated learning sample label of each user behavior log learning sample based on the confidence of multiple candidate classification labels. The purpose of this step is to select the classification label that best matches the actual user behavior from the multiple possible classifications output by the model for subsequent model debugging and optimization. Specifically, in the preceding step of step S300, the confidence distribution mapping algorithm has classified the integrated user behavior log features and generated corresponding confidence for each candidate classification label. These confidences reflect the model's degree of certainty that the user behavior belongs to a certain category, and the value range is usually between 0 and 1, and the larger the value, the higher the degree of certainty.

[0049] In step S310, the computer system determines the estimated learning sample label of each user behavior log learning sample based on these confidences. This process usually involves setting a confidence threshold, and only classification labels with confidence exceeding the threshold will be selected as estimated labels for user behavior log learning samples. For example, the threshold can be set to 0.7, and only when the confidence of a classification label is greater than or equal to 0.7, the label will be selected as the estimated label.

[0050] For example, suppose there is a user behavior log learning example, and the confidence distribution mapping algorithm outputs three candidate classification tags and their corresponding confidences, namely "interested in electronic products" (confidence 0.8), "interested in fashion apparel" (confidence 0.6), and "interested in travel" (confidence 0.4). If the confidence threshold is set to 0.7, then only the classification tag "interested in electronic products" will be selected as the estimated tag because it is the only tag with a confidence exceeding the threshold. This estimated tag will be used as the feature label of the user behavior log learning example for subsequent model debugging and optimization.

[0051] Through the processing of step S310, the computer system can accurately select the estimated label that best matches the user's actual behavior from multiple candidate classification labels, providing a reliable foundation for subsequent model optimization and application.

[0052] Step S320: Obtain a label commonality measure between a priori label of the learning sample of each user behavior log learning sample and an estimated label of each learning sample, and determine a first loss function of the first behavior feature mining neural network based on the label commonality measure between a priori label of the learning sample of each user behavior log learning sample and an estimated label of each learning sample.

[0053] In step S320, the computer system determines the first loss function of the first behavior feature mining neural network by calculating the label commonality measure between the learning sample prior label and the estimated label of each user behavior log learning sample. The purpose of this step is to quantify the difference between the model prediction result and the actual result, so as to provide guidance for the optimization of the model. The learning sample prior label is usually a pre-defined label for each user behavior log learning sample based on historical data or expert knowledge, which represents the real category or feature of the user behavior. The estimated label is the prediction result obtained after classifying the integrated user behavior log features through the confidence distribution mapping algorithm. In step S320, the computer system obtains the prior label and the estimated label of each user behavior log learning sample. Then, it calculates the commonality measure between the two labels to quantify the similarity or difference between them. The specific calculation method of the commonality measure depends on the specific business scenario and requirements. Common measurement methods include cosine similarity, Jaccard similarity, etc.

[0054] Taking cosine similarity as an example, assuming that both the prior tags and the estimated tags can be represented as high-dimensional vectors (for example, by one-hot encoding or embedding vector representation), the cosine similarity between them can be calculated by the following formula:

[0055]

[0056] Where a represents the prior label vector, b represents the estimated label vector, a·b represents the dot product of two vectors, |||a||| and |||b||| represent the modulus lengths of the two vectors respectively. The value range of cosine similarity is between -1 and 1, and the larger the value, the more similar the two vectors are.

[0057] After calculating the commonality metrics between the priori tags and the estimated tags of each user behavior log learning sample, the computer system determines the first loss function of the first behavior feature mining neural network based on these metrics. The loss function is a function that measures the difference between the model's predicted results and the actual results. Its goal is to make the predicted results as close to the actual results as possible.

[0058] A common loss function is the cross-entropy loss function, which is suitable for multi-label classification tasks. Assume that there are N user behavior log learning examples, for each example i, its prior label vector is The estimated label vector is Then the cross entropy loss function can be expressed as:

[0059]

[0060] It should be noted that since this is a multi-label classification task, and are all vectors, and each element of the vector corresponds to a classification label. In actual calculations, the cross entropy loss is usually calculated for each classification label separately, and then the loss values ​​of all classification labels are summed to obtain the total loss value.

[0061] By calculating the loss function, the computer system can quantify the difference between the model's predicted results and the actual results, thereby guiding the subsequent model optimization process. During the model training process, the computer system continuously adjusts the weights and bias terms of the neural network through the back-propagation algorithm and the gradient descent method to minimize the loss function value, thereby improving the model's prediction accuracy.

[0062] Step S330: Debugging the first behavior feature mining neural network based on the first loss function to obtain the second behavior feature mining neural network.

[0063] In an embodiment of the present application, the computer system debugs the first behavior feature mining neural network based on the first loss function calculated in step S320 to optimize the parameters of the model, thereby improving the prediction accuracy of the model, and ultimately obtaining a second behavior feature mining neural network with better performance.

[0064] Specifically, the first loss function quantifies the difference between the model's prediction results and the actual results (prior marks) of the user behavior log learning examples. This difference reflects the performance of the model in the current state and is the goal of model optimization. By minimizing the loss function value, the model's prediction results can be made closer to the actual situation, thereby improving the accuracy of the model. During the debugging process, the computer system uses the backpropagation algorithm and the gradient descent method. First, the gradient of the loss function with respect to the model parameters is calculated. This gradient represents the direction and rate at which the loss function changes with the parameters. Then, the parameters of the model are updated according to the gradient information so that the loss function value decreases in the opposite direction of the gradient. This process will be iterated multiple times, and each iteration will gradually reduce the loss function value and gradually improve the performance of the model.

[0065] Taking neural networks in deep learning as an example, the first behavioral feature mining neural network may be a multi-layer perceptron (MLP), convolutional neural network (CNN) or recurrent neural network (RNN). During the debugging process, the computer system adjusts parameters such as weights and biases in these neural networks. These parameters determine how the neural network processes input data and produces output predictions. By continuously optimizing these parameters, the prediction results of the neural network can be made more accurate.

[0066] As an implementation manner, in step S200, mining the sample user behavior features corresponding to each user behavior log learning sample may specifically include:

[0067] Step S201: performing a data item splitting operation on each user behavior log learning sample to obtain a behavior data item corresponding to each user behavior log learning sample.

[0068] In an embodiment of the present application, step S201 splits the complex user behavior log learning sample into more fine-grained behavior data items. The purpose of this step is to extract specific details in the user behavior so as to perform more refined feature extraction and model training later. Specifically, in step S201, the computer system performs a data item splitting operation on each user behavior log learning sample. User behavior log learning samples usually contain a series of behavior records of users on digital platforms, which may cover a variety of user behaviors such as browsing, searching, clicking, and purchasing. Each behavior record contains rich information, such as the time when the behavior occurred, the type of behavior, the object of the behavior (such as goods, pages, etc.), etc.

[0069] The purpose of the data item splitting operation is to split these complex user behavior log learning examples into separate behavior data items. Each behavior data item represents a specific user behavior and contains all relevant information about the behavior. Through splitting, the computer system can convert a user behavior log learning example into a data set consisting of multiple behavior data items.

[0070] Taking the e-commerce platform as an example, assume that a user behavior log learning sample contains all the shopping behaviors of the user in one day. This learning sample may include a series of behaviors such as the user browsing multiple product pages, searching for specific keywords, adding certain products to the shopping cart, and finally purchasing some of the products. In step S201, the computer system splits this learning sample into multiple separate behavior data items, each of which corresponds to a specific behavior. For example, one data item may indicate that the user browsed an electronic product page, including browsing time, browsing page URL, browsing duration and other information; another data item may indicate that the user searched for the keyword "smart watch", including search time, search keywords and other information.

[0071] Through the data item splitting operation, the computer system can transform complex user behavior log learning examples into a structured set of behavior data items. These behavior data items will serve as input data for subsequent feature extraction and model training, providing support for in-depth mining of user behavior features.

[0072] Step S202: Determine a target behavior data item in the behavior data items corresponding to each user behavior log learning sample, perform a mask operation on the target behavior data item corresponding to each user behavior log learning sample to obtain a masked behavior data item corresponding to each user behavior log learning sample, and determine a masked behavior log learning sample corresponding to each user behavior log learning sample based on the behavior data items other than the target behavior data item in each user behavior log learning sample and the masked behavior data items corresponding to each user behavior log learning sample.

[0073] In the embodiment of the present application, step S202 selects specific target items from the behavior data items corresponding to the user behavior log learning sample, and performs mask operations on these target items to simulate data missing or uncertainty. The purpose of this step is to train the machine learning model to accurately predict the user's behavior characteristics when some information is missing.

[0074] Specifically, in step S202, the computer system analyzes the behavior data items in each user behavior log learning sample. These behavior data items contain various behavior records of users on the digital platform, such as browsing products, searching keywords, clicking on advertisements, etc. Each behavior data item represents an aspect of user behavior and contains rich information, such as the timestamp of the behavior, the type of behavior, the object of the behavior, etc.

[0075] Next, the computer system determines the target behavior data item from these behavior data items. The selection of the target behavior data item can be based on a variety of strategies, such as random selection, selection based on the importance or frequency of the behavior, etc. The selected target item will be used as the object of the mask operation to simulate the missing or uncertainty in the data.

[0076] Once the target behavior data items are determined, the computer system will perform mask operations on these target items. Mask operations usually replace the value of the target item with a specific mask value (such as a null value, 0, or a specific placeholder) to mask its original information. Through mask operations, the computer system can generate masked behavior data items that are identical in appearance to the target behavior data items, but their actual values ​​have been replaced.

[0077] After obtaining the masked behavior data items, the computer system combines the other behavior data items except the target item in each user behavior log learning sample and the generated masked behavior data items to determine the masked behavior log learning sample. The masked behavior log learning sample is obtained by masking some behavior data items based on the original user behavior log learning sample. These masked samples are used to train the machine learning model to make predictions when some information is missing.

[0078] Taking the user behavior mining of e-commerce platforms as an example, assume that a user behavior log learning sample contains behavior data items of a user browsing multiple product pages. In step S202, the computer system can select the browsing behavior of one of the product pages as the target behavior data item and perform a mask operation on it. In this way, a masked behavior log learning sample is obtained, which contains the original data item of the user browsing other product pages and a masked browsing behavior item. Through this masked sample, the machine learning model can learn how to predict the user's interest preference or purchase intention based on other information in the absence of browsing information of a certain product page.

[0079] Through the processing of step S202, the computer system can generate masked behavior log learning samples containing masked behavior data items. These masked samples provide valuable data resources for the training of machine learning models, enabling the model to maintain high prediction accuracy when some information is missing, thereby improving the robustness and practicality of user behavior big data mining.

[0080] Step S203: mining sample user behavior features corresponding to each masked behavior log learning example, wherein the sample user behavior features corresponding to each masked behavior log learning example include a representation array corresponding to the masked behavior data item.

[0081] In the embodiment of the present application, step S203 mines sample user behavior features from the masked user behavior log learning samples. These features not only reflect the user's behavior pattern, but also contain relevant information of the masked behavior data items, providing an important data basis for subsequent model training and optimization.

[0082] Specifically, in step S203, the computer system performs feature extraction for each masked behavior log learning sample. The masked behavior log learning sample is obtained by selecting and masking some behavior data items based on the original user behavior log learning sample. The masking operation simulates the situation of missing or uncertain data, making the features extracted from the masked sample more generalized and practical. In the feature extraction process, the computer system uses various feature extraction algorithms or deep learning models to extract sample user behavior features from the masked behavior log learning sample. These features may include the user's behavior type, the timestamp of the behavior, the frequency of the behavior, the associated goods or content categories, etc. In particular, for the masked behavior data items, the computer system generates a corresponding representation array. This representation array contains key information of the masked behavior data items, such as possible behavior types, potential associated objects, etc. Although this information is not completely accurate, it provides the model with reasonable assumptions and prediction basis about the masked data items.

[0083] Taking the user behavior mining of e-commerce platforms as an example, suppose a masked behavior log learning sample contains behavioral data items of a user browsing multiple product pages, but the browsing behavior of one of the product pages is masked. In step S203, the computer system extracts sample user behavior features from this masked sample, such as the category of other product pages browsed by the user, browsing time, etc. At the same time, for the masked browsing behavior, the computer system generates a representation array, which may contain information such as the potential category of the product page (such as electronic products, household items, etc.), possible browsing time range, etc. These features will be used as input for subsequent machine learning models to help the model predict the specific content of the masked behavior data items.

[0084] Through the processing of step S203, the computer system can extract rich sample user behavior features from the masked behavior log learning samples and generate a representation array corresponding to the masked behavior data items. These features not only contain specific information about user behavior, but also simulate situations where data is missing or uncertain, providing valuable data resources for subsequent model training and optimization. By using these features for model training, the computer system can train a more robust and accurate user behavior prediction model, providing enterprises with more accurate user portraits and personalized recommendation services.

[0085] In the step S300, debugging the first behavior feature mining neural network according to the learning sample estimation marks and the learning sample prior marks corresponding to the multiple user behavior log learning samples to obtain the second behavior feature mining neural network may specifically include:

[0086] Step S301: perform behavior data item estimation on the representation array corresponding to the masked behavior data item corresponding to each user behavior log learning sample, obtain the confidence that the masked behavior data item corresponding to each user behavior log learning sample is a plurality of control behavior data items, and determine the sample estimated behavior data item of the masked behavior data item corresponding to each user behavior log learning sample based on the confidence of the plurality of control behavior data items.

[0087] In the embodiment of the present application, step S301 estimates the behavior data item for the masked behavior data item. The purpose of this step is to use the trained machine learning model to estimate the possible true value of the masked behavior data item according to the representation array corresponding to the masked behavior data item, and give the confidence level of each estimated value. Through this step, the computer system can simulate and predict the user's behavior in the case of missing or uncertain data, thereby enhancing the robustness and generalization ability of the model.

[0088] Specifically, in step S301, the computer system obtains the masked behavior data items corresponding to each user behavior log learning sample and the representation arrays corresponding to these masked items. These representation arrays are generated in step S203 and contain key information of the masked behavior data items, such as possible behavior types, potential associated objects, etc. Then, the computer system inputs the representation arrays corresponding to the masked behavior data items into the trained machine learning model. This model is usually a deep learning model, such as a recurrent neural network (RNN), a long short-term memory network (LSTM), or a Transformer, which have powerful sequence modeling and feature learning capabilities, and can extract useful information from the representation array and make predictions.

[0089] After receiving the representation array, the masked behavior data items are estimated based on its internal parameters and the knowledge learned. The result of the estimation is a set of control behavior data items and their corresponding confidences. These control behavior data items represent the possible true values ​​of the masked behavior data items, while the confidence measures the degree of certainty of the model for each estimated value.

[0090] Taking the user behavior mining of the e-commerce platform as an example, suppose that a user behavior log learning sample contains the behavior data items of the user browsing multiple product pages, but the browsing behavior of one of the product pages is masked. In step S301, the computer system obtains the representation array corresponding to the masked browsing behavior and inputs it into the trained deep learning model. The model will estimate the possible true value of the masked browsing behavior based on the information in the representation array, such as the user may have browsed a specific electronic product page. At the same time, the model will also give the confidence of this estimated value, such as 0.8, indicating that the model is 80% sure that this estimated value is correct. Through the processing of step S301, the computer system can estimate the behavior data items of the masked behavior data items and obtain multiple possible true values ​​and their confidence. These estimation results and confidence information will be used as inputs for subsequent steps to calculate the loss function and debug the model, thereby improving the prediction accuracy and robustness of the model.

[0091] Step S302: Obtain the sample masked behavior data items in each user behavior log learning sample, and obtain the second loss function for the first behavior feature mining neural network based on the sample estimated behavior data items corresponding to each user behavior log learning sample and the sample masked behavior data items in each user behavior log learning sample.

[0092] In the embodiment of the present application, step S302 calculates the second loss function to quantify the error of the model in predicting the mask behavior data item. The purpose of this step is to guide the optimization direction of the model through the loss function and improve the prediction accuracy of the model in the case of missing or uncertain data.

[0093] Specifically, in step S302, the computer system obtains sample mask behavior data items in each user behavior log learning sample. These mask behavior data items are generated by masking operations in step S202 and are used to simulate data missing or uncertain situations.

[0094] Next, the computer system calculates a second loss function based on the sample estimated behavior data item obtained in step S301 and the current sample mask behavior data item. This loss function measures the difference between the model prediction result and the actual mask behavior data item, providing guidance for model optimization.

[0095] When calculating the second loss function, a certain metric can be used to measure the inconsistency between the predicted results and the actual results. A common method is to use the cross-entropy loss function (Cross-Entropy Loss), especially in multi-classification or probabilistic prediction scenarios. However, since this involves the estimation of behavioral data items, a more specific metric may be required. Taking the mean squared error (MSE) as an example, assuming that the masked behavior data items and the estimated behavior data items are both numerical (such as browsing time, purchase amount, etc.), the second loss function can be expressed by the following formula:

[0096]

[0097] Where N is the number of user behavior log learning examples, yi is the true value of the masked behavior data item in the i-th example (known during the training process), is the estimated value of the mask behavior data item in the ith example (obtained in step S301). This formula calculates the square of the difference between the estimated value and the actual value in each example, and takes the average value as the value of the loss function.

[0098] For example, suppose that in a user behavior mining scenario of an e-commerce platform, there is a masked behavior data item representing the user's browsing time (in seconds). The true value is 120 seconds, but the model estimates the browsing time to be 110 seconds. Then for this example, the partial contribution of the loss function is (120-110)2=100. If there are multiple such examples, the loss function is the average of the losses of all examples.

[0099] By calculating the second loss function, the computer system can quantify the model's error in predicting the mask behavior data item, thus providing clear guidance for subsequent model debugging and optimization. This step is a key step in improving the model's prediction accuracy in the case of missing or uncertain data.

[0100] Step S303: Based on the second loss function, the learning sample estimation labels corresponding to the multiple user behavior log learning samples and the learning sample prior labels of the multiple user behavior log learning samples, the first behavior feature mining neural network is debugged to obtain the second behavior feature mining neural network.

[0101] In the embodiment of the present application, step S303 is the core step in the model training and optimization process. In this step, the computer system debugs and optimizes the first behavior feature mining neural network based on the second loss function, the learning sample estimation labels corresponding to the multiple user behavior log learning samples, and the learning sample prior labels, so as to obtain the second behavior feature mining neural network with better performance.

[0102] Specifically, the second loss function is calculated in step S302, which measures the error of the model in predicting the masked behavior data item. The learning sample estimated label is obtained by model estimation in step S301, which represents the classification or prediction result of the model for the user behavior data item. The learning sample prior label is a pre-defined label for each user behavior log learning sample based on historical data or expert knowledge, which represents the real category or feature of the user behavior.

[0103] In step S303, the computer system uses the second loss function, the learning sample estimated label and the learning sample prior label as optimization targets. By minimizing the second loss function, the model can be made more accurate in predicting the masked behavior data item; and by making the learning sample estimated label as close as possible to the learning sample prior label, the accuracy of the model in classifying or predicting user behavior can be improved.

[0104] Next, the computer system uses the back propagation algorithm and the gradient descent method to debug and optimize the first behavioral feature mining neural network. The back propagation algorithm determines the direction and rate at which the loss function changes with the parameters by calculating the gradient of the loss function with respect to the model parameters. The gradient descent method updates the model parameters based on this gradient information, so that the loss function value decreases in the opposite direction of the gradient, thereby gradually optimizing the performance of the model.

[0105] During the debugging process, the computer system iteratively performs forward propagation (calculating estimated labels and loss function values) and back propagation (updating model parameters) operations until the preset stop conditions are met (such as loss function value convergence, number of iterations reaching the upper limit, etc.). After multiple iterations of optimization, the performance of the first behavioral feature mining neural network will gradually improve, and finally the second behavioral feature mining neural network will be obtained.

[0106] For example, suppose there is a user behavior log learning sample, which contains the user's browsing, searching and purchasing behaviors. In step S303, the computer system calculates the second loss function based on the masked behavior data item of this sample, and adjusts the parameters of the model based on the difference between the purchase intention estimated by the model (the learning sample estimated mark) and the actual purchase behavior (the learning sample prior mark). Through iterative optimization, the model will be able to more accurately predict the user's purchase intention in the absence of data or uncertainty, thereby providing enterprises with more accurate personalized recommendation services.

[0107] Step S303 debugs and optimizes the first behavior feature mining neural network by combining the second loss function, the learning sample estimation mark and the learning sample prior mark, so that the model is more accurate and robust in predicting user behavior. This process is an indispensable part of digital user behavior big data mining.

[0108] As an implementation method, the sample user behavior features corresponding to each user behavior log learning sample include a representation array corresponding to each sample behavior data item in each user behavior log learning sample, and the representation array corresponding to each sample behavior data item includes a preset behavior data item representation array subjected to a mask operation.

[0109] Based on this, the step S301 performs behavior data item estimation on the representation array corresponding to the mask behavior data item corresponding to each user behavior log learning sample, obtains the confidence that the mask behavior data item corresponding to each user behavior log learning sample is a plurality of control behavior data items, and determines the sample estimated behavior data item of the mask behavior data item corresponding to each user behavior log learning sample according to the confidence of the plurality of control behavior data items, which may specifically include:

[0110] Step S3011: Obtain a representation array corresponding to a sample behavior data item that has a contextual relationship with the preset behavior data item representation array corresponding to each user behavior log learning sample, and determine the context behavior data item representation array of the preset behavior data item representation array corresponding to each user behavior log learning sample.

[0111] In step S3011, the main task of the computer system is to identify and obtain the representation array corresponding to the sample behavior data item that has a contextual relationship with the preset behavior data item representation array in each user behavior log learning sample. These contextual relationship data are crucial for predicting the true value of the masked behavior data item.

[0112] Specifically, in step S3011, the computer system analyzes each user behavior log learning example, especially those examples that contain a representation array of preset masked behavior data items. Preset masked behavior data items refer to those behavior data items that are masked in the actual data for some reason (such as missing data or intentionally hidden to protect user privacy). These masked items usually represent important information in the user behavior log, but due to incomplete data, they cannot be directly used for subsequent analysis and prediction.

[0113] In order to make up for this data loss, the computer system needs to find sample behavior data items that have a contextual relationship with the preset mask behavior data items. These contextual relationship data usually include behavior data items that are adjacent to the preset mask behavior data items in time series, or behavior data items that are logically associated with the preset mask behavior data items. These contextual relationship data can provide additional information about the preset mask behavior data items, which helps the model predict its true value more accurately.

[0114] Taking user behavior mining of e-commerce platforms as an example, assume that a user behavior log learning sample contains behavior data items of users browsing multiple product pages. Among them, a browsing behavior data item about a smart watch is preset as a mask item (that is, its true value is hidden). In step S3011, the computer system searches for other behavior data items that have a contextual relationship with the smart watch browsing behavior data item. These contextual relationship data may include other product browsing behaviors (such as mobile phones, headphones, etc.) before or after the smart watch browsing behavior, or other browsing behaviors that belong to the same product category as the smart watch. The representation array corresponding to these contextual relationship data will be determined as the contextual behavior data item representation array of the smart watch browsing behavior data item.

[0115] By obtaining the representation array of these contextual behavior data items, the computer system can provide rich information support for the subsequent estimation of masked behavior data items. These contextual relationship data not only help the model to more accurately predict the true value of the masked behavior data items, but also improve the accuracy and reliability of the entire user behavior big data mining.

[0116] Step S3012: Integrate the context behavior data item representation array of the preset behavior data item representation array corresponding to each user behavior log learning sample and the preset behavior data item representation array corresponding to each user behavior log learning sample to obtain an integrated behavior data item representation array corresponding to each user behavior log learning sample.

[0117] In step S3012, the computer system integrates the context behavior data item representation array of the preset behavior data item representation array in each user behavior log learning sample with the preset behavior data item representation array to obtain an integrated behavior data item representation array. This integration process helps to extract more comprehensive user behavior features and provide a richer information basis for subsequent masked behavior data item estimation.

[0118] Specifically, in step S3011, the computer system has obtained the representation arrays corresponding to the sample behavior data items that have a contextual relationship with the preset behavior data item representation array, namely, the context behavior data item representation array. These context behavior data item representation arrays contain behavior data features that are adjacent to or logically associated with the preset behavior data items in time series.

[0119] In step S3012, the computer system integrates the contextual behavior data item representation arrays and the preset behavior data item representation arrays. The integration method includes, for example, splicing, superimposing or feature fusion of multiple representation arrays. These operations are intended to combine multiple related representation arrays into a more comprehensive and rich integrated behavior data item representation array.

[0120] Taking the user behavior mining of e-commerce platforms as an example, suppose a user behavior log learning sample contains behavior data items of a user browsing multiple product pages, and one browsing behavior data item about a smart watch is preset as a mask item. In step S3011, the computer system has obtained the context behavior data item representation array of the smart watch browsing behavior data item, which may contain the behavior data features of the mobile phone, headphones and other products that the user has browsed before.

[0121] In step S3012, the computer system integrates these contextual behavior data item representation arrays with the preset representation array of the smart watch browsing behavior. The integrated representation array will contain some features of the smart watch browsing behavior itself (although masked), as well as the features of other related product browsing behaviors. Such an integrated behavior data item representation array not only contains direct information about the smart watch browsing behavior, but also incorporates the contextual information related to it, providing a more comprehensive and rich data basis for the subsequent estimation of masked behavior data items. Through the integration operation of step S3012, the computer system can construct a more complete and accurate representation of user behavior features, providing strong support for subsequent machine learning model training and masked behavior data item estimation. This integration method not only improves the accuracy and effectiveness of user behavior big data mining, but also enhances the model's ability to handle complex user behaviors.

[0122] Step S3013: Estimate the confidence of the integrated behavior data item representation array corresponding to each user behavior log learning sample as a plurality of control behavior data items, and determine the control behavior data item with the maximum confidence based on the plurality of confidences, and determine it as the sample estimated behavior data item of the mask behavior data item corresponding to each user behavior log learning sample.

[0123] In the embodiment of the present application, step S3013 performs behavior data item estimation on the integrated behavior data item representation array. In this step, the computer system uses the trained machine learning model to analyze the integrated representation array, estimates the possible true values ​​of the masked behavior data items, and quantifies the reliability of these estimated values ​​through confidence.

[0124] First, in step S3012, the computer system has completed the integration of the preset behavior data item representation array and its context behavior data item representation array, and obtained the integrated behavior data item representation array. This integrated array contains the direct information and related context information of the masked behavior data item, providing a rich data basis for subsequent estimation. Then, in step S3013, the computer system inputs the integrated behavior data item representation array into the trained machine learning model. This model is usually a deep learning model, such as a recurrent neural network (RNN), a long short-term memory network (LSTM) or a Transformer, which have powerful sequence modeling and feature learning capabilities. The model will analyze and process the integrated representation array based on its internal parameters and learned knowledge.

[0125] When analyzing the integrated representation array, the model estimates multiple possible control behavior data items for the masked behavior data item and assigns a confidence level to each control item. These control behavior data items represent the possible true values ​​of the masked behavior data item, and the confidence level measures the degree of certainty of the model for each estimated value. The confidence level is usually calculated based on the model's analysis of the integrated representation array and the internal probability distribution.

[0126] Taking the user behavior mining of e-commerce platforms as an example, assume that the smartwatch browsing behavior in a user behavior log learning sample is masked. In step S3013, the computer system inputs the integrated behavior data item representation array (including other product browsing behavior features before and after the smartwatch browsing behavior) into the trained model. The model will analyze this integrated representation array, estimate multiple smartwatch models that the user may browse (as control behavior data items), and assign a confidence level to each model. For example, the model may estimate that the confidence level that the user browsed model A smartwatch is 0.8, and the confidence level that the user browsed model B smartwatch is 0.6.

[0127] Finally, the computer system selects the control item with the highest confidence as the sample estimated behavior data item based on the confidence of multiple control behavior data items. In this example, since the confidence of model A smartwatch is the highest (0.8), the computer system determines it as the sample estimated behavior data item of the masked smartwatch browsing behavior. This estimation result will be used as input for subsequent steps to calculate the loss function and debug the model, thereby improving the prediction accuracy of the model in the case of missing or uncertain data.

[0128] Through the processing of step S3013, the computer system can make full use of the integrated behavioral data item representation array and the trained machine learning model to accurately estimate the masked behavioral data items and quantify the reliability of the estimated value through confidence measurement. This process not only improves the accuracy of user behavior big data mining, but also provides strong support for subsequent model optimization and personalized recommendation.

[0129] As an implementation manner, in the step S400, debugging the second behavior feature mining neural network according to the learning sample estimated commonality metric values ​​and the learning sample commonality metric values ​​corresponding to the multiple user behavior log learning sample binary groups to obtain the target behavior feature mining neural network may specifically include:

[0130] Step S410: Determine a third loss function between the estimated commonality metric value of the learning sample corresponding to each user behavior log learning sample binary and the learning sample commonality metric value corresponding to each user behavior log learning sample binary; determine the third loss function of the multiple user behavior log learning sample binary groups based on the third loss function between the estimated commonality metric value of the learning sample corresponding to each user behavior log learning sample binary and the learning sample commonality metric value corresponding to each user behavior log learning sample binary.

[0131] In the embodiment of the present application, step S410 calculates the difference between the estimated commonality metric value of the learning sample and the actual commonality metric value of the learning sample binary of the user behavior log, and determines the loss function accordingly. The purpose of this step is to guide the adjustment of the model parameters by quantifying the difference between the model prediction results and the actual results, thereby improving the model's prediction accuracy for user behavior similarity.

[0132] A user behavior log learning sample tuple is composed of two user behavior log learning samples, which represent two related behaviors of users on digital platforms. The learning sample estimated commonality metric is the prediction result of the model based on the current parameters of the similarity of the behaviors of the two samples in the tuple, while the learning sample commonality metric is the real similarity between the two samples calculated based on the actual user behavior data.

[0133] In step S410, the computer system calculates the difference between the estimated commonality metric value of the learning sample and the actual commonality metric value of the learning sample for each user behavior log learning sample binary. This difference can be quantified by a loss function, the goal of which is to minimize the difference between the predicted result and the actual result.

[0134] Taking Mean Squared Error (MSE) as an example of a loss function, the following third loss function () can be defined for each user behavior log learning sample binary:

[0135]

[0136] Among them, Losspairi represents the loss function value of the i-th user behavior log learning sample binary, ypairi is the actual learning sample commonality metric value of the binary, is the model's estimated commonality measure for the bigram.

[0137] Next, in order to obtain the total loss function of all the binary pairs of user behavior log learning examples, the computer system accumulates the loss function values ​​of all binary pairs and takes the average. The formula of the total loss function can be expressed as:

[0138]

[0139] Where N is the total number of binary pairs of user behavior log learning examples. This total loss function will be used as the optimization target for subsequent model debugging to guide the update of model parameters.

[0140] For example, suppose that in the user behavior mining scenario of an e-commerce platform, there are two user behavior log learning example tuples. The actual commonality metric value of the first tuple is 0.9 (indicating that the behaviors of the two users are very similar), while the commonality metric value estimated by the model is 0.8. The actual commonality metric value of the second tuple is 0.5 (indicating that the behaviors of the two users have a certain degree of similarity), while the commonality metric value estimated by the model is 0.6. According to the definition of the MSE loss function, the loss function values ​​of the two tuples can be calculated, and they can be added and averaged to obtain the total loss function value, which is used to guide further optimization of the model.

[0141] Step S420: Debugging the second behavior feature mining neural network according to the third loss function of the plurality of user behavior log learning sample pairs to obtain a target behavior feature mining neural network.

[0142] In the embodiment of the present application, step S420 debugs the second behavior feature mining neural network according to the third loss function of the binary group of multiple user behavior log learning samples, thereby obtaining a target behavior feature mining neural network with better performance. This step is the key to improving the model's ability to predict user behavior similarity.

[0143] In step S420, the computer system calculates a third loss function for multiple user behavior log learning example pairs, which measures the difference between the user behavior similarity predicted by the model (estimated commonality metric value) and the actual user behavior similarity (actual commonality metric value). This loss function provides a clear direction and metric for model optimization.

[0144] Next, the computer system uses the back propagation algorithm and the gradient descent method to debug the second behavioral feature mining neural network. The back propagation algorithm determines the direction and rate at which the loss function changes with the parameters by calculating the gradient of the loss function with respect to the model parameters. The gradient descent method updates the model parameters based on this gradient information, so that the loss function value decreases in the opposite direction of the gradient, thereby gradually optimizing the performance of the model.

[0145] During the debugging process, the computer system iteratively performs forward propagation (calculating the estimated commonality metric value and loss function value) and back propagation (updating model parameters). In the forward propagation stage, the computer system uses the user behavior log learning sample as input data, performs forward calculations through the second behavior feature mining neural network, obtains the estimated commonality metric value of each user behavior log learning sample binary, and calculates the loss function value accordingly. In the back propagation stage, the computer system updates the model parameters according to the gradient information of the loss function, so that the model's prediction results are closer to the actual results.

[0146] After multiple iterations of optimization, the performance of the second behavior feature mining neural network will gradually improve, and the target behavior feature mining neural network will eventually be obtained. This optimized neural network can better capture the similarities between user behaviors and provide more accurate data support for subsequent user portrait construction, personalized recommendation and other applications.

[0147] For example, suppose there is a set of user behavior log learning sample tuples, each of which contains the behavior data of two users. By calculating the estimated commonality metric and the actual commonality metric of each tuple and determining the third loss function accordingly, the back propagation algorithm and the gradient descent method can be used to debug the second behavior feature mining neural network. After multiple iterations of optimization, the neural network will be able to more accurately predict the similarities between user behaviors, such as identifying user groups with similar shopping preferences or browsing behaviors. This will help e-commerce platforms provide users with more accurate personalized recommendation services and improve user experience and satisfaction.

[0148] As an embodiment, the method further includes:

[0149] Step S10: Acquire multiple control behavior logs, and mine control behavior log features of each of the multiple control behavior logs using a neural network based on the target behavior feature mining method.

[0150] In the embodiment of the present application, step S10 aims to extract features of multiple control behavior logs through a target behavior feature mining neural network to provide reference data for subsequent target user behavior analysis.

[0151] First, the computer system needs to obtain multiple control behavior logs. These control behavior logs usually come from historical user behavior data, which can be selected by random sampling or screened based on specific conditions (such as similar user groups, similar behavior patterns, etc.). The purpose of obtaining these control behavior logs is to build a rich reference data set for subsequent target user behavior analysis.

[0152] Once multiple control behavior logs are obtained, the computer system will then mine the control behavior log features of each control behavior log based on the trained target behavior feature mining neural network. The target behavior feature mining neural network is trained with a large amount of user behavior data in the previous step, and it has learned how to accurately extract the key features of user behavior.

[0153] In the process of mining the characteristics of the control behavior log, the computer system uses each control behavior log as input data and performs forward propagation through the target behavior feature mining neural network. The neural network will extract and transform the input data based on its internal parameters and learned knowledge, and finally output the control behavior log characteristics of each control behavior log. These features are usually represented in the form of high-dimensional vectors or arrays, which contain key information of the control behavior log, such as behavior type, behavior time, associated objects, etc.

[0154] Taking the e-commerce platform as an example, assume that 100 control behavior logs are obtained, which contain the browsing, searching, purchasing and other behavior data of different users on the platform. In step S10, the computer system inputs these control behavior logs into the target behavior feature mining neural network in sequence, and obtains the control behavior log features of each control behavior log through feature extraction and conversion of the neural network. For example, for one of the control behavior logs, the neural network can output a vector containing multiple features, such as [0.8, 0.2, 0.5], where the first element represents the user's interest in electronic products (0.8 means high interest), the second element represents the interest in fashion clothing (0.2 means low interest), and the third element may represent other features such as the user's purchase frequency or browsing time. Through the processing of step S10, the computer system can obtain the control behavior log features of multiple control behavior logs, which will serve as important reference data for the target user behavior analysis in subsequent steps. This process not only enriches the content of the reference data set, but also improves the accuracy and reliability of the target user behavior analysis.

[0155] Step S20: adding the control behavior log feature of each control behavior log to the characterization array set.

[0156] In the embodiment of the present application, step S20 is a key step in constructing and enriching the reference data set, which involves adding the control behavior log features of each control behavior log to the characterization array set.

[0157] In step S20, the computer system maintains a representation array set, which is used to store the control behavior log features of multiple control behavior logs. These features are extracted from the control behavior logs by the target behavior feature mining neural network mentioned in the previous step S10, and they represent the key information of different user behaviors. Then, for each control behavior log, the computer system adds its corresponding control behavior log features to the representation array set. This process is similar to adding new data items to a database or data set. Specifically, the computer system adds each control behavior log feature (usually represented in the form of a high-dimensional vector or array) as an independent data item to the representation array set to ensure that each feature can be uniquely identified and retrieved.

[0158] Taking the e-commerce platform as an example, it is assumed that the corresponding control behavior log features have been extracted from 100 control behavior logs through step S10. In step S20, the computer system adds these features to the representation array set one by one. Each feature may contain multiple dimensions, such as user interest preferences, purchase frequency, browsing time, etc., which together describe the user's behavior pattern on the platform. For example, a control behavior log feature may be represented as a vector [0.8, 0.2, 0.5], where 0.8 represents the user's interest in electronic products, 0.2 represents the interest in fashion apparel, and 0.5 may represent the characteristic value of some other behavior.

[0159] By adding the control behavior log features of each control behavior log to the representation array set, the computer system can build a reference data set containing rich user behavior information. This data set not only provides important reference data for subsequent target user behavior analysis, but also helps to discover the similarities and correlations between user behaviors, further improving the accuracy and practicality of user behavior analysis. At the same time, the construction of the representation array set also provides data support for subsequent model training and optimization, which helps to improve the performance of machine learning models in user behavior prediction and classification.

[0160] Based on this, in step S500, determining a behavior log profile that matches the target user according to the behavior log features of the target behavior log may specifically include:

[0161] Step S510: determining a commonality metric value between the behavior log feature of the target behavior log and the comparison behavior log features of a plurality of comparison behavior logs in the representation array set.

[0162] In the embodiment of the present application, step S510 is an important part of the target user behavior analysis, which involves determining the commonality measurement value between the behavior log features of the target behavior log and the reference behavior log features of multiple reference behavior logs in the characterization array set. The purpose of this step is to quantify the similarity between the target user behavior and the reference user behavior, and provide a basis for the subsequent behavior log profile construction.

[0163] Specifically, in step S510, the computer system obtains the behavior log features of the target behavior log. These features are extracted from the behavior log of the target user through the target behavior feature mining neural network, and they represent the key behavior information of the target user on the digital platform. At the same time, the computer system also accesses the representation array set, which stores the control behavior log features of multiple control behavior logs, which are extracted from the historical user behavior data through the same neural network.

[0164] Next, the computer system calculates the commonality metric between the behavior log feature of the target behavior log and the control behavior log feature of each control behavior log in the characterization array set. The commonality metric is used to quantify the similarity between two behavior log features, which can be calculated by a variety of algorithms or formulas, such as cosine similarity, Euclidean distance, etc. These algorithms or formulas can calculate the similarity between two vectors based on the specific values ​​of the feature vectors.

[0165] In practical applications, the computer system traverses each control behavior log feature in the characterization array set, and calculates the cosine similarity (or other commonality measurement value) between it and the behavior feature of the target behavior log. In this way, the computer system can obtain a commonality measurement value set, and each element in the set corresponds to the similarity between a control behavior log and a target behavior log.

[0166] Taking the user behavior analysis of an e-commerce platform as an example, suppose the behavior log feature vector A of the target user indicates that the user has a strong interest in electronic products, and a control behavior log feature vector B in the representation array set also shows an interest in electronic products. By calculating the cosine similarity of vector A and vector B, the computer system can obtain a high similarity value, indicating that the behavior of the target user is highly similar to that of the control user.

[0167] Through the processing of step S510, the computer system can quantify the similarity between the target user's behavior and the reference user's behavior, providing strong support for the subsequent construction of the behavior log profile. This process not only helps to deeply understand the behavior pattern of the target user, but also provides a data basis for applications such as personalized recommendation and market segmentation.

[0168] Step S520: Determine a behavior log profile that matches the target user based on a commonality measurement value between the behavior log feature of the target behavior log and the control behavior log features of the plurality of control behavior logs.

[0169] In the embodiment of the present application, step S520 is a key step in constructing a target user behavior profile. This step determines a behavior log profile that matches the target user based on the commonality measurement between the behavior log features of the target behavior log and the control behavior log features of multiple control behavior logs.

[0170] In step S520, the computer system reviews the set of commonality metrics calculated in step S510. This set contains similarity scores between the behavior log features of the target behavior log and each reference behavior log feature in the characterization array set. These scores quantify the similarity between the target user behavior and the reference user behavior, providing an important basis for determining the behavior profile of the target user.

[0171] Next, the computer system uses a strategy to determine the behavior log profile that best matches the target user. This strategy can be based on a variety of methods such as sorting of commonality metrics, threshold screening, or comprehensive evaluation. The following is an example strategy based on threshold screening and sorting:

[0172] Setting a threshold: The computer system sets a threshold for the commonality metric. This threshold is used to filter out control behavior logs that are sufficiently similar to the target user's behavior. For example, the threshold can be set to 0.7, which means that only control behavior logs with a commonality metric greater than or equal to 0.7 will be further considered.

[0173] Screening control behavior logs: The computer system traverses the common measurement value set and screens out control behavior logs whose common measurement values ​​exceed the threshold. These logs have a high similarity with the target user's behavior, so their characteristics are of great significance for building the target user's behavior profile.

[0174] Sorting and selection: For the selected control behavior logs, the computer system can sort them according to the commonality metric value and select one or more logs with the highest commonality metric value as representatives. The characteristics of these logs best reflect the behavior patterns of the target users, so they will be used to construct the behavior log portrait of the target users.

[0175] Constructing a behavior log profile: Finally, the computer system constructs a behavior log profile of the target user based on the characteristics of the selected control behavior log. This profile is a comprehensive description of the target user's behavior characteristics, including the user's interest preferences, behavior frequency, associated objects, and other aspects. By integrating the characteristics of multiple similar control behavior logs, the computer system can more accurately capture the behavior patterns of the target user, providing strong support for subsequent personalized recommendations, market segmentation, and other applications.

[0176] Taking the user behavior analysis of an e-commerce platform as an example, assume that the behavior log characteristics of the target user show that he has a strong interest in electronic products. In step S520, the computer system screens out control behavior logs whose commonality metric values ​​exceed the threshold, and these logs also show an interest in electronic products. Next, the computer system selects the most representative control behavior logs according to the commonality metric value sorting, and constructs a behavior log portrait of the target user based on the characteristics of these logs. This portrait may include information such as the user's preference for a specific category of electronic products, purchase frequency, browsing time, etc., to provide accurate personalized recommendations and marketing strategy support for the e-commerce platform.

[0177] As an implementation manner, the step S520, determining the behavior log profile matching the target user based on the commonality metric between the behavior log feature of the target behavior log and the control behavior log features of the multiple control behavior logs, may specifically include:

[0178] Step S521: determining b candidate behavior logs having the largest commonality metric values ​​with the target behavior log from the multiple control behavior logs, where b≥1.

[0179] In the embodiment of the present application, step S521 is a key step in constructing the target user behavior profile. In this step, the main task of the computer system is to determine b candidate behavior logs with the largest commonality metric value with the target behavior log from multiple control behavior logs, where b is a positive integer representing the number of selected candidate behavior logs.

[0180] First, the computer system reviews the set of commonality metrics calculated in step S510. This set contains the commonality metrics between the target behavior log and each control behavior log in the characterization array set, and these metrics quantify the similarity between the target user behavior and the reference user behavior. The commonality metrics are usually calculated by a specific algorithm or formula, such as cosine similarity, Euclidean distance, etc. These algorithms can calculate the similarity between two vectors based on the specific values ​​of the feature vectors. Then, the computer system sorts the control behavior logs according to the size of the commonality metrics. The larger the commonality metric value, the higher the similarity between the control behavior log and the target behavior log. Through sorting, the computer system can quickly identify the control behavior log that is most similar to the target user behavior. Then, the computer system selects the first b control behavior logs with the largest commonality metrics as candidate behavior logs. Here, b is a pre-set parameter that can be determined according to specific application scenarios and requirements. The b behavior logs with the largest commonality metrics are selected as candidates because these behavior logs are most similar to the behavior of the target user, so their characteristics have important reference value for constructing the behavior portrait of the target user.

[0181] Taking the e-commerce platform as an example, assume that the target user's behavior log shows that he has a strong interest in electronic products, especially in smartphones and smart watches. In step S521, the computer system traverses all control behavior logs, calculates the commonality metric between them and the target behavior log, and sorts them according to the metric value. Then, the computer system selects the first b control behavior logs with the largest commonality metric as candidate behavior logs. These candidate behavior logs also show an interest in electronic products, and may have a high similarity in specific commodity categories such as smartphones and smart watches. Through the processing of step S521, the computer system can screen out the b candidate behavior logs that are most similar to the target user's behavior, providing important data support for subsequent steps S522 and S523. This process not only helps to deeply understand the behavior patterns of target users, but also provides a strong data foundation for applications such as personalized recommendations and market segmentation.

[0182] Step S522: Obtain user basic attribute information, determine a target screening mechanism corresponding to the user basic attribute information based on the user basic attribute information, screen the b candidate behavior logs based on the target screening mechanism to obtain a candidate behavior logs, 1≤a≤b.

[0183] In the embodiment of the present application, step S522 further screens the candidate behavior logs according to the basic attribute information of the user to ensure that the behavior log finally selected better matches the actual attributes of the target user.

[0184] In step S521, the computer system has determined b candidate behavior logs with the largest commonality metric values ​​from multiple control behavior logs. However, although these b candidate behavior logs have a high similarity with the target user's behavior, they may not completely conform to the target user's actual attributes. Therefore, in step S522, the computer system needs to further filter these candidate behavior logs based on the user's basic attribute information.

[0185] First, the computer system obtains the basic attribute information of the target user. This basic attribute information may include the user's age, gender, region, occupation, income status, etc. This information can usually be obtained from the user's registration information, transaction records or third-party data sources. For example, in the scenario of an e-commerce platform, the basic attribute information of the target user may show that the user is a 30-year-old woman who lives in a first-tier city and works in the IT industry. Then, the computer system determines a target screening mechanism based on the basic attribute information of the target user. This screening mechanism can be defined according to specific application scenarios and requirements. Its core purpose is to filter out candidate behavior logs that do not match the target user's attributes based on the user's basic attributes. The specific implementation methods of the screening mechanism can be varied. For example, a rule-based filter can be used to set a series of screening conditions based on the user's basic attributes; a machine learning model (such as a classifier or regression model) can also be used to predict the user's behavior pattern and perform screening accordingly.

[0186] After determining the target screening mechanism, the computer system applies the mechanism to the b candidate behavior logs determined in step S521. By comparing the basic attributes of each candidate behavior log with the target screening mechanism, the computer system screens out the candidate behavior logs that best match the basic attributes of the target user. The number of candidate behavior logs after screening is recorded as a, where 1≤a≤b, indicating that the number of candidate behavior logs after screening may be less than or equal to the original number of candidate behavior logs b.

[0187] Continuing with the example of the e-commerce platform, assume that five candidate behavior logs with the highest commonality metrics are identified in step S521. In step S522, the computer system obtains that the target user is a 30-year-old female who lives in a first-tier city and works in the IT industry. Based on this information, the computer system sets up a screening mechanism to filter out candidate behavior logs from other age groups, genders, or regions, while retaining those behavior logs related to the IT industry. After screening, three candidate behavior logs that best match the basic attributes of the target user may be obtained (i.e., a=3).

[0188] Through the screening in step S522, the computer system can further ensure that the candidate behavior logs finally selected not only have a high similarity with the behavior of the target user, but also conform to the actual basic attributes of the target user. This helps to more accurately build the behavior portrait of the target user and provide strong support for subsequent personalized recommendations, market segmentation and other applications.

[0189] Step S523: Determine a behavior log profile that matches the target user from among the a candidate behavior logs.

[0190] In the embodiment of the present application, step S523 is the last key step to construct the target user behavior profile. In the previous steps, the computer system has screened out candidate behavior logs similar to the target user behavior through commonality metrics, and further accurately screened the candidate behavior logs based on the user's basic attribute information. The purpose of step S523 is to combine the characteristics of these screened candidate behavior logs to determine a behavior log profile that is highly matched with the target user.

[0191] In step S523, the computer system obtains a candidate behavior logs after being screened in step S522. These candidate behavior logs not only have a high similarity with the behavior of the target user, but also conform to the basic attribute information of the target user, so they can more accurately reflect the behavioral characteristics of the target user. Then, the computer system extracts the key features of each candidate behavior log. These features may include multiple aspects such as user interest preferences, behavior frequency, and associated objects. They are usually extracted from the original behavior log through the target behavior feature mining neural network and expressed in the form of a high-dimensional vector or array. For example, the feature vector of a candidate behavior log may contain the user's interest in different commodity categories such as electronic products and fashion apparel, as well as the user's purchase frequency, browsing time and other behavioral data. Then, the computer system conducts a comprehensive analysis and integration of the features of these candidate behavior logs. This process may involve a variety of data processing technologies such as weighted average of feature vectors, cluster analysis, and association rule mining. Through comprehensive analysis, the computer system can identify key information such as the main behavior patterns and interest preferences of the target user, and integrate this information into a complete behavior log portrait.

[0192] Taking the e-commerce platform as an example, assume that three candidate behavior logs are obtained after screening in step S522. The characteristics of these candidate behavior logs show that the target user has a high interest in both electronic products and fashion apparel, and has a high purchase frequency and often browses related product pages. In step S523, the computer system conducts a comprehensive analysis and integration of these characteristics to construct a behavior log portrait of the target user. This portrait may describe the target user as a young consumer who has a strong interest in electronic products and fashion apparel, often browses and purchases related products, and has a high spending power and willingness to buy.

[0193] Through the processing of step S523, the computer system can generate a behavior log portrait that is highly matched with the target user. This portrait not only fully reflects the behavioral characteristics of the target user, but also provides strong data support for subsequent personalized recommendations, market segmentation and other applications. For example, the e-commerce platform can recommend products that meet the interests and preferences of the target user based on the behavior portrait of the target user, thereby improving the user experience and satisfaction; at the same time, it can also perform market segmentation based on the user portrait and formulate more accurate marketing strategies.

[0194] Based on the foregoing embodiments, the embodiments of the present application provide a user behavior big data mining device for digitalization. The units included in the device and the modules included in each unit can be implemented by a processor in a computer device; of course, they can also be implemented by a specific logic circuit; in the implementation process, the processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP) or a field programmable gate array (FPGA), etc.

[0195] Figure 2 A schematic diagram of the structure of a digital user behavior big data mining device provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the digital user behavior big data mining device 200 includes:

[0196] The training data acquisition module 210 is used to acquire a training database, wherein the training database includes a plurality of user behavior log learning samples and a learning sample priori label of each user behavior log learning sample, a plurality of user behavior log learning sample tuples, and a learning sample commonality metric value corresponding to each user behavior log learning sample tuple, wherein the learning sample commonality metric value corresponding to the user behavior log learning sample tuple is a commonality metric value between a first user behavior log learning sample and a second user behavior log learning sample included in the user behavior log learning sample tuple;

[0197] The behavior feature mining module 220 is used to mine the sample user behavior features corresponding to each user behavior log learning sample, and to perform confidence distribution mapping on each sample user behavior feature to obtain the confidence of each user behavior log learning sample being respectively labeled as a plurality of candidate classifications;

[0198] A first network debugging module 230 is used to determine the learning sample estimation mark of each user behavior log learning sample according to the confidence of the multiple candidate classification marks, and debug the first behavior feature mining neural network according to the learning sample estimation marks and the learning sample prior marks corresponding to the multiple user behavior log learning samples to obtain a second behavior feature mining neural network;

[0199] The second network debugging module 240 is used to mine the first sample user behavior feature of the first user behavior log learning sample and the second sample user behavior feature of the second user behavior log learning sample included in each user behavior log learning sample binary, obtain the learning sample estimated commonality metric value between the first sample user behavior feature and the second sample user behavior feature corresponding to each user behavior log learning sample binary, and debug the second behavior feature mining neural network according to the learning sample estimated commonality metric values ​​and the learning sample commonality metric values ​​corresponding to the multiple user behavior log learning sample binary groups to obtain the target behavior feature mining neural network;

[0200] The behavior profile determination module 250 is used to receive the target behavior log of the target user, determine the behavior log characteristics of the target behavior log based on the target behavior characteristic mining neural network, and determine the behavior log profile matching the target user based on the behavior log characteristics of the target behavior log.

[0201] The description of the above device embodiment is similar to the description of the above method embodiment, and has similar beneficial effects as the method embodiment. In some embodiments, the functions or modules provided by the device provided in the embodiment of the present application can be used to execute the method described in the above method embodiment. For the technical details not disclosed in the device embodiment of the present application, please refer to the description of the method embodiment of the present application for understanding. It should be noted that in the embodiment of the present application, if the above-mentioned digital-oriented user behavior big data mining method is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present application is essentially or the part that contributes to the relevant technology can be embodied in the form of a software product, which is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read Only Memory, ROM), disk or optical disk and other media that can store program code. In this way, the embodiment of the present application is not limited to any specific hardware, software or firmware, or any combination of hardware, software and firmware.

[0202] An embodiment of the present application provides a computer system, including a memory and a processor, wherein the memory stores a computer program that can be executed on the processor, and when the processor executes the program, some or all of the steps in the above method are implemented.

[0203] The embodiment of the present application provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, some or all of the steps in the above method are implemented. The computer-readable storage medium can be transient or non-transient.

[0204] An embodiment of the present application provides a computer program, including a computer-readable code. When the computer-readable code is run in a computer device, a processor in the computer device executes some or all of the steps for implementing the above method.

[0205] The embodiment of the present application provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and when the computer program is read and executed by a computer, some or all of the steps in the above method are implemented. The computer program product can be implemented specifically by hardware, software or a combination thereof. In some embodiments, the computer program product is specifically embodied as a computer storage medium, and in other embodiments, the computer program product is specifically embodied as a software product, such as a software development kit (SDK) and the like.

[0206] It should be noted here that the description of the various embodiments above tends to emphasize the differences between the various embodiments, and the same or similar aspects can be referenced to each other. The description of the above device, storage medium, computer program and computer program product embodiments is similar to the description of the above method embodiment, and has similar beneficial effects as the method embodiment. For technical details not disclosed in the embodiments of the device, storage medium, computer program and computer program product of this application, please refer to the description of the method embodiment of this application for understanding.

[0207] Figure 3 A hardware entity diagram of a computer system provided in an embodiment of the present application is shown in FIG. Figure 3 As shown, the hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

[0208] The memory 1002 stores computer programs that can be run on the processor. The memory 1002 is configured to store instructions and applications executable by the processor 1001. It can also cache data to be processed or processed by the processor 1001 and various modules in the computer system 1000 (for example, image data, audio data, voice communication data, and video communication data). This can be achieved through flash memory (FLASH) or random access memory (Random Access Memory, RAM).

[0209] When the processor 1001 executes the program, the steps of any one of the above-mentioned methods for mining digital user behavior big data are implemented. The processor 1001 generally controls the overall operation of the computer system 1000.

[0210] An embodiment of the present application provides a computer storage medium, which stores one or more programs. The one or more programs can be executed by one or more processors to implement the steps of the digital user behavior big data mining method in any of the above embodiments.

[0211] It should be pointed out here that the description of the above storage medium and device embodiments is similar to the description of the above method embodiments, and has similar beneficial effects as the method embodiments. For technical details not disclosed in the storage medium and device embodiments of this application, please refer to the description of the method embodiments of this application for understanding. The above is only an implementation method of this application, but the protection scope of this application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered within the protection scope of this application.

Claims

1. A digital user behavior big data mining method, characterized in that: The method comprises: Acquire a training database, wherein the training database includes a plurality of user behavior log learning samples and a learning sample priori label of each user behavior log learning sample, a plurality of user behavior log learning sample tuples, and a learning sample commonality metric value corresponding to each user behavior log learning sample tuple, wherein the learning sample commonality metric value corresponding to the user behavior log learning sample tuple is a commonality metric value between a first user behavior log learning sample and a second user behavior log learning sample included in the user behavior log learning sample tuple; Mining the sample user behavior features corresponding to each user behavior log learning sample, and performing confidence distribution mapping on each sample user behavior feature to obtain the confidence of each user behavior log learning sample being respectively labeled as a plurality of candidate classifications; Determining a learning sample estimation label of each user behavior log learning sample according to the confidence of the multiple candidate classification labels, and debugging a first behavior feature mining neural network according to the learning sample estimation labels and learning sample prior labels corresponding to the multiple user behavior log learning samples to obtain a second behavior feature mining neural network; Respectively mine the first sample user behavior feature of the first user behavior log learning sample and the second sample user behavior feature of the second user behavior log learning sample included in each user behavior log learning sample binary, obtain the learning sample estimated commonality metric value between the first sample user behavior feature and the second sample user behavior feature corresponding to each user behavior log learning sample binary, and debug the second behavior feature mining neural network according to the learning sample estimated commonality metric values ​​and the learning sample commonality metric values ​​corresponding to the multiple user behavior log learning sample binary groups to obtain the target behavior feature mining neural network; Receive a target behavior log of a target user, determine a behavior log feature of the target behavior log by mining a neural network based on the target behavior feature, and determine a behavior log profile that matches the target user based on the behavior log feature of the target behavior log; The method of determining the learning sample estimation mark of each user behavior log learning sample according to the confidence of the multiple candidate classification marks, and debugging the first behavior feature mining neural network according to the learning sample estimation marks and the learning sample prior marks corresponding to the multiple user behavior log learning samples to obtain the second behavior feature mining neural network includes: Determine one or more candidate classification labels from the multiple candidate classification labels according to the confidences of the multiple candidate classification labels, and determine them as learning sample estimated labels for each user behavior log learning sample; Obtaining a label commonality metric value between a priori label of the learning sample of each user behavior log learning sample and a predicted label of each learning sample, and determining a first loss function of the first behavior feature mining neural network according to the label commonality metric value between a priori label of the learning sample of each user behavior log learning sample and a predicted label of each learning sample; Debugging the first behavior feature mining neural network based on the first loss function to obtain the second behavior feature mining neural network; The step of debugging the second behavior feature mining neural network based on the estimated commonality metric values ​​of the learning samples and the commonality metric values ​​of the learning samples corresponding to the plurality of user behavior log learning sample tuples to obtain the target behavior feature mining neural network includes: Determine a third loss function between the estimated commonality metric value of the learning sample corresponding to each user behavior log learning sample binary and the commonality metric value of the learning sample corresponding to each user behavior log learning sample binary, and determine the third loss function of the multiple user behavior log learning sample binary according to the third loss function between the estimated commonality metric value of the learning sample corresponding to each user behavior log learning sample binary and the commonality metric value of the learning sample corresponding to each user behavior log learning sample binary; The second behavior feature mining neural network is debugged according to the third loss function of the plurality of user behavior log learning sample pairs to obtain a target behavior feature mining neural network.

2. The method according to claim 1, characterized in that The sample user behavior feature corresponding to each user behavior log learning sample includes a representation array corresponding to each sample behavior data item in each user behavior log learning sample; the confidence distribution mapping is performed on each sample user behavior feature to obtain the confidence of each user behavior log learning sample as a plurality of candidate classification labels, including: Integrating the representation information of the representation array corresponding to each sample behavior data item in each user behavior log learning sample to obtain the integrated behavior log feature of each user behavior log learning sample; The integrated behavior log features are estimated as confidences of representation arrays corresponding to the multiple candidate classification labels.

3. The method according to claim 1 or 2, characterized in that: The mining of sample user behavior features corresponding to each user behavior log learning sample includes: Performing a data item splitting operation on each of the user behavior log learning samples to obtain a behavior data item corresponding to each of the user behavior log learning samples; Determine a target behavior data item in the behavior data items corresponding to each user behavior log learning sample, perform a masking operation on the target behavior data item corresponding to each user behavior log learning sample to obtain a masked behavior data item corresponding to each user behavior log learning sample, and determine a masked behavior log learning sample corresponding to each user behavior log learning sample based on the behavior data items other than the target behavior data item in each user behavior log learning sample and the masked behavior data item corresponding to each user behavior log learning sample; Mining a sample user behavior feature corresponding to each masked behavior log learning example, wherein the sample user behavior feature corresponding to each masked behavior log learning example includes a representation array corresponding to the masked behavior data item; The debugging of the first behavior feature mining neural network according to the learning sample estimation labels and the learning sample prior labels corresponding to the multiple user behavior log learning samples to obtain the second behavior feature mining neural network includes: Performing behavior data item estimation on the representation array corresponding to the masked behavior data item corresponding to each user behavior log learning sample, obtaining the confidence that the masked behavior data item corresponding to each user behavior log learning sample is a plurality of control behavior data items, and determining the sample estimated behavior data item of the masked behavior data item corresponding to each user behavior log learning sample according to the confidence of the plurality of control behavior data items; Obtaining a sample mask behavior data item in each user behavior log learning sample, and obtaining a second loss function for the first behavior feature mining neural network based on the sample estimated behavior data item corresponding to each user behavior log learning sample and the sample mask behavior data item in each user behavior log learning sample; Based on the second loss function, the learning sample estimation labels corresponding to the multiple user behavior log learning samples and the learning sample prior labels of the multiple user behavior log learning samples, the first behavior feature mining neural network is debugged to obtain the second behavior feature mining neural network.

4. The method according to claim 3, characterized in that The sample user behavior feature corresponding to each user behavior log learning sample includes a representation array corresponding to each sample behavior data item in each user behavior log learning sample, and the representation array corresponding to each sample behavior data item includes a masked preset behavior data item representation array; The step of estimating the behavior data items of the representation array corresponding to the masked behavior data item corresponding to each of the user behavior log learning samples, obtaining the confidence that the masked behavior data item corresponding to each of the user behavior log learning samples is a plurality of control behavior data items, and determining the sample estimated behavior data item of the masked behavior data item corresponding to each of the user behavior log learning samples according to the confidence of the plurality of control behavior data items, comprises: Obtaining a representation array corresponding to a sample behavior data item that has a contextual relationship with the preset behavior data item representation array corresponding to each user behavior log learning sample, and determining the context behavior data item representation array of the preset behavior data item representation array corresponding to each user behavior log learning sample; Performing representation array integration on the context behavior data item representation array of the preset behavior data item representation array corresponding to each user behavior log learning sample and the preset behavior data item representation array corresponding to each user behavior log learning sample to obtain an integrated behavior data item representation array corresponding to each user behavior log learning sample; Estimate the confidence of the integrated behavior data item representation array corresponding to each user behavior log learning sample as a plurality of control behavior data items, and determine the control behavior data item with the maximum confidence based on the plurality of confidences, and determine it as the sample estimated behavior data item of the mask behavior data item corresponding to each user behavior log learning sample.

5. The method according to claim 1, characterized in that The method further comprises: Acquire multiple control behavior logs, and mine control behavior log features of each of the multiple control behavior logs using a neural network according to the target behavior feature mining method; Adding the control behavior log feature of each control behavior log to the characterization array set; The determining, based on the behavior log features of the target behavior log, a behavior log portrait matching the target user includes: Determine a commonality metric value between a behavior log feature of the target behavior log and a comparison behavior log feature of a plurality of comparison behavior logs in the characterization array set; A behavior log profile matching the target user is determined based on a commonality metric value between the behavior log feature of the target behavior log and the comparison behavior log features of the plurality of comparison behavior logs.

6. The method according to claim 5, characterized in that The determining, based on the commonality measurement value between the behavior log feature of the target behavior log and the comparison behavior log features of the plurality of comparison behavior logs, a behavior log profile matching the target user includes: Determine b candidate behavior logs having the largest commonality metric value with the target behavior log from the multiple control behavior logs, where b≥1; Obtaining user basic attribute information, determining a target screening mechanism corresponding to the user basic attribute information according to the user basic attribute information, screening the b candidate behavior logs according to the target screening mechanism to obtain a candidate behavior logs, 1≤a≤b; Determine a behavior log profile that matches the target user among the a candidate behavior logs.

7. A digital user behavior big data mining device, characterized in that: include: A training data acquisition module is used to acquire a training database, wherein the training database includes a plurality of user behavior log learning samples and a learning sample priori label of each user behavior log learning sample, a plurality of user behavior log learning sample tuples, and a learning sample commonality metric value corresponding to each user behavior log learning sample tuple, wherein the learning sample commonality metric value corresponding to the user behavior log learning sample tuple is a commonality metric value between a first user behavior log learning sample and a second user behavior log learning sample included in the user behavior log learning sample tuple; A behavior feature mining module is used to mine the sample user behavior features corresponding to each user behavior log learning sample, and to perform confidence distribution mapping on each sample user behavior feature to obtain the confidence that each user behavior log learning sample is respectively labeled as a plurality of candidate classifications; A first network debugging module is used to determine the learning sample estimation mark of each user behavior log learning sample according to the confidence of the multiple candidate classification marks, and debug the first behavior feature mining neural network according to the learning sample estimation marks and the learning sample prior marks corresponding to the multiple user behavior log learning samples to obtain a second behavior feature mining neural network; Specifically, the first network debugging module determines one or more candidate classification labels from the multiple candidate classification labels according to the confidences of the multiple candidate classification labels, and determines them as the learning sample estimated labels of each user behavior log learning sample; Obtaining a label commonality measure between a priori label of each learning sample and an estimated label of each learning sample of each user behavior log learning sample, and determining a first loss function of the first behavior feature mining neural network based on the label commonality measure between a priori label of each learning sample of each user behavior log learning sample and an estimated label of each learning sample; debugging the first behavior feature mining neural network based on the first loss function to obtain the second behavior feature mining neural network; The second network debugging module is used to respectively mine the first sample user behavior feature of the first user behavior log learning sample and the second sample user behavior feature of the second user behavior log learning sample included in each user behavior log learning sample binary, obtain the learning sample estimated commonality metric value between the first sample user behavior feature and the second sample user behavior feature corresponding to each user behavior log learning sample binary, and debug the second behavior feature mining neural network based on the learning sample estimated commonality metric value and the learning sample commonality metric value corresponding to the multiple user behavior log learning sample binary groups to obtain the target behavior feature mining neural network; wherein, the learning sample estimated commonality metric value and the learning sample commonality metric value corresponding to the multiple user behavior log learning sample binary groups are debugged. The second behavior feature mining neural network obtains the target behavior feature mining neural network, including: determining a third loss function between the estimated commonality metric value of the learning sample corresponding to each user behavior log learning sample binary and the learning sample commonality metric value corresponding to each user behavior log learning sample binary, and determining the third loss function of the multiple user behavior log learning sample binary according to the third loss function between the estimated commonality metric value of the learning sample corresponding to each user behavior log learning sample binary and the learning sample commonality metric value corresponding to each user behavior log learning sample binary; debugging the second behavior feature mining neural network according to the third loss function of the multiple user behavior log learning sample binary to obtain the target behavior feature mining neural network; The behavior profile determination module is used to receive the target behavior log of the target user, determine the behavior log characteristics of the target behavior log by mining the neural network based on the target behavior characteristics, and determine the behavior log profile matching the target user based on the behavior log characteristics of the target behavior log.

8. A computer system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, wherein: When the processor executes the program, the steps in the method according to any one of claims 1 to 6 are implemented.

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 steps in the method according to any one of claims 1 to 6 are implemented.

Images