User interestingness generation method, system and equipment for content tag and storage medium

Abstract

The embodiment of the invention discloses a content tag user interestingness generation method, system and device and a storage medium. According to the technical scheme provided by the embodiment of the invention, the tag features of the user historical behavior data are calculated through weighting in different time periods, and the user interestingness is calculated by using the user interestingness estimation model based on the tag features of different behavior tag types; therefore, the situation that the calculation precision is low due to the fact that the user interestingness is simply calculated according to the number of user behaviors is avoided, the calculation precision and the individuation degree of the user interestingness are improved, the content recommendation effect of the recommendation system is improved, and the content watching experience of the user is improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of computer technology, and in particular to a method, system, device, and storage medium for generating user interest in content tags. Background Art

[0002] With the increasing use of digital information, recommendation systems based on digital content have become an indispensable key technology for various internet applications. Recommendation systems aim to accurately filter content of potential interest to users from vast amounts of information resources, significantly improving the efficiency and experience of user information acquisition. Accurately calculating user interest in content tags is a key component of achieving accurate recommendations in recommendation systems. Related methods for calculating user interest primarily use methods based on user behavior frequency. These methods measure user interest in a specific content tag by counting the number of clicks, views, and saves performed by users.

[0003] However, related solutions for calculating user interest in corresponding content tags based on the frequency of user behavior typically analyze and calculate based solely on the number of user behaviors (such as the number of clicks, views, and favorites on content tagged with a specific tag). In reality, the frequency with which content with the corresponding tag is distributed also directly affects the frequency and diversity of user behavior. For example, if content with a certain tag receives too many clicks, it is likely because the recommendation system has distributed too much content related to that tag. Clearly, simply analyzing and calculating user interest based on the number of user behaviors will result in a low level of accuracy, which can easily affect the accuracy and personalization of the recommendation system's content recommendations, and thus the user's experience. Summary of the Invention

[0004] The embodiments of the present application provide a method, system, device, and storage medium for generating user interest in content tags. This technical solution calculates user interest by weighting the tag features of user historical behavior data over different time periods, and uses a user interest estimation model based on the tag features of different behavior tag types to calculate user interest. This avoids the situation where simply calculating user interest based on the number of user behaviors leads to low calculation accuracy, thereby improving the calculation accuracy and personalization of user interest. Compared with related user interest calculation solutions, this application can accurately analyze user interest in content tags by combining diverse user features, meeting users' needs for accurate content recommendations, and solving the technical problem of low calculation accuracy of user interest in content tags.

[0005] In a first aspect, an embodiment of the present application provides a method for generating user interest in content tags, comprising: Obtain the specified historical behavior data of the target user corresponding to the target content tag, use the set behavior tag type as the filtering dimension, and filter out target behavior data of multiple different target time periods from the specified historical behavior data; Based on multiple target behavior data of the same behavior tag type, weighted calculation of behavior sub-features is performed respectively. The behavior tag feature is obtained by concatenating the behavior sub-features obtained based on the weighted calculation. The feature weight of each target behavior data is configured from low to high according to the distance between the corresponding target time period and the current time; The behavior tag features of the target user corresponding to different behavior tag types are input into the pre-built user interest estimation model, and the user interest of the target user corresponding to the target content tag is output.

[0006] In a second aspect, an embodiment of the present application provides a system for generating user interest in content tags, including: A data screening module is configured to obtain the specified historical behavior data of the target user corresponding to the target content tag, and use the set behavior tag type as a screening dimension to screen out target behavior data of multiple different target time periods from the specified historical behavior data; A feature concatenation module is configured to perform weighted calculations on behavior sub-features based on multiple target behavior data of the same behavior tag type, and to concatenate the behavior sub-features obtained through the weighted calculations to obtain behavior tag features. The feature weights of the weighted calculations on each target behavior data are configured from low to high according to the distance between the corresponding target time period and the current time; The model calculation module is configured to input the behavior label features of the target user corresponding to different behavior label types into a pre-built user interest estimation model, and output the user interest of the target user corresponding to the target content label. During model training, the user interest estimation model constructs label samples corresponding to the target content label based on the historical behavior data of different users, and annotates the interest probability of the label samples. The label samples are used as model input for model training, and the label samples contain sample features of multiple behavior label types.

[0007] In a third aspect, an embodiment of the present application provides a device for generating user interest in content tags, including: memory and one or more processors; The memory is configured to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for generating user interest in content tags as described in the first aspect.

[0008] In a fourth aspect, an embodiment of the present application provides a non-volatile computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by a computer processor, they are configured to execute the method for generating user interest in content tags as described in the first aspect.

[0009] In a fifth aspect, an embodiment of the present application provides a computer program product, which includes instructions. When the instructions are executed on a computer or processor, the computer or processor executes the method for generating user interest in content tags as described in the first aspect.

[0010] The embodiment of the present application obtains the designated historical behavior data of the target user corresponding to the target content tag, uses the set behavior tag type as the screening dimension, and filters out target behavior data of multiple different target time periods from the designated historical behavior data; performs weighted calculation of behavior sub-features based on multiple target behavior data of the same behavior tag type, and splices the behavior sub-features obtained based on the weighted calculation to obtain behavior tag features, and the feature weights of the weighted calculation of each target behavior data are configured from low to high according to the distance between the corresponding target time period and the current time; inputs the behavior tag features of the target user corresponding to different behavior tag types into a pre-built user interest estimation model, and outputs the user interest of the target user corresponding to the target content tag. By adopting the above technical means, the tag features of the user's historical behavior data are weightedly calculated in different time periods, and the user interest is calculated using the user interest estimation model based on the tag features of different behavior tag types, thereby avoiding the situation where the calculation accuracy of the user interest is low due to simply calculating the user interest based on the number of user behaviors, improving the calculation accuracy and personalization of the user interest, and thus improving the content recommendation effect of the recommendation system and the user's content viewing experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a flow chart of a method for generating user interest in content tags provided by an embodiment of the present application; Figure 2 This is a schematic diagram of the structure of the behavior tag feature in the embodiment of the present application; Figure 3 is a training flow chart of the user interest estimation model in an embodiment of the present application; Figure 4 is an application flow chart of the user interest estimation model in an embodiment of the present application; Figure 5 This is a flowchart of user interest calculation of the user interest estimation model in an embodiment of the present application; Figure 6 Schematic diagram of vector conversion of user interest in an embodiment of the present application; Figure 7 This is a schematic diagram of the structure of a system for generating user interest in content tags provided by an embodiment of the present application; Figure 8 This is a structural diagram of a device for generating user interest in content tags provided in an embodiment of the present application. DETAILED DESCRIPTION

[0012] To further clarify the objectives, technical solutions, and advantages of this application, specific embodiments of the present application are described in further detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are intended only to illustrate this application and are not intended to limit it. It should also be noted that, for ease of description, the drawings only illustrate portions relevant to this application, not all of them. Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts depict the various operations (or steps) as sequential processes, many of the operations can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process may terminate upon completion of its operations, but may also include additional steps not shown in the accompanying drawings. The process may correspond to a method, function, procedure, subroutine, subprogram, or the like.

[0013] The user interest generation method for content tags provided in this application aims to calculate the tag features of user historical behavior data by weighting them in different time periods, and calculate the user interest using a user interest estimation model based on the tag features of different behavior tag types, thereby avoiding the situation where the calculation accuracy is low due to simply calculating the user interest based on the number of user behaviors, and improving the calculation accuracy and personalization of the user interest.

[0014] In today's era of digital information explosion, recommendation systems have become an indispensable key technology for various internet applications. From simple recommendation methods based on content matching or collaborative filtering in the early days to today's intelligent recommendation systems that integrate multiple complex algorithms, recommendation systems aim to accurately filter content of potential interest to users from massive amounts of information resources, greatly improving the efficiency and experience of user information acquisition.

[0015] In recommendation systems, accurately calculating user interest in content tags is a key component of achieving precise recommendations. There are two main methods for calculating user interest in relevant content tags: First, methods based on user behavior frequency measure user interest in a tag by counting the number of clicks, views, and saves on content with a specific tag. For example, if a user frequently clicks on content tagged "technology," the system will assume the user has a high interest in that tag. This method is simple, straightforward, easy to implement, and highly interpretable, allowing users to intuitively understand the basis for interest calculation. However, it only considers the number of clicks, ignoring the direct impact of content delivery on the frequency and diversity of user behavior. For example, if a user clicks too many articles tagged "technology," this is likely due to the recommendation system delivering too much "technology"-related content to the user, but this method cannot effectively distinguish these.

[0016] The second approach is machine learning. This approach uses deep learning models, such as neural networks, to combine multiple user characteristics (including demographics, historical behavioral data, and social connections) to build a user interest model to predict user interest in content tags. This approach can mine complex relationships between data, effectively improving the accuracy of interest calculations and addressing complex patterns of interest fluctuations. However, it requires extremely high data volumes and computing resources, resulting in high model training costs and poor model interpretability. This poses certain limitations in practical applications, making it difficult for users to clearly understand the specific logic behind the model's predictions.

[0017] Clearly, the development of content recommendation systems (such as short videos) has exposed numerous issues that urgently need to be addressed in methods for calculating user interest. Simple methods, such as counting the frequency of user tagging behaviors, offer certain computational advantages and can complete initial data processing with low resource consumption. However, these methods only provide a superficial analysis of user behavior and fail to provide deep insights into underlying user interests and preferences. This directly results in recommendations that are neither accurate nor personalized enough to meet the growing and diverse needs of users. Deep model learning methods, while leveraging complex neural network architectures to deeply mine and analyze massive amounts of data, thus achieving relatively accurate recommendations, also have significant drawbacks. Firstly, the training and inference processes of deep models require significant computing resources, including high-performance server clusters and large-scale memory and storage devices, which significantly increases operational costs. Secondly, the complex internal operational logic of the models makes interpreting parameters and predictions extremely difficult. For both ordinary users and professional developers, it is difficult to intuitively understand the rationale behind the recommendations. This lack of interpretability has led to numerous inconveniences in practical applications.

[0018] It is precisely because related user interest calculation methods have difficulty achieving an effective balance between explainability and computational cost that they are gradually becoming unable to meet the ever-increasing accuracy and personalization requirements of recommendation systems. Based on this, the present application provides a method for generating user interest in content tags in an embodiment to address the technical issue of low accuracy in calculating user interest in content tags, thereby improving the calculation accuracy and personalization of user interest.

[0019] By comprehensively collecting a series of behavioral data on users' responses to content with corresponding content tags over a period of time (such as the past 180 days), taking short videos as an example, including but not limited to average viewing time, video completion rate, fast scrolling frequency, and the incidence of interactive behaviors such as likes, shares, and reports, we use pre-designed feature engineering methods to construct user behavioral tag features for this content tag. With the help of the analytical capabilities of the user interest estimation model, we can accurately calculate the user's interest preference for this content tag, that is, the user interest. The core idea is to use the user's historical behavioral data on a certain content tag to predict the user's click-through rate when the content with this tag is sent to the user again, and use this as the user's interest in the content tag. This provides users with more convincing and transparent interest calculation results, helping the recommendation system to serve users more accurately and efficiently.

[0020] Example: Figure 1 A flowchart of a method for generating user interest in content tags, provided in an embodiment of the present application, is provided. This method can be performed by a device for generating user interest in content tags. This device can be implemented via software and / or hardware, and can be comprised of two or more physical entities, or a single physical entity. Generally speaking, this device can be a computing device such as a computer or a recommendation system server.

[0021] The following description will be made by taking the content recommendation system device as an example to describe the method for generating user interest in executing content tags. Figure 1 The method for generating user interest in content tags specifically includes: S110: Obtain designated historical behavior data of a target user corresponding to a target content tag, use a set behavior tag type as a screening dimension, and screen out target behavior data of multiple different target time periods from the designated historical behavior data.

[0022] When calculating user interest, this application first collects historical user behavior data for content tagged with the content for which interest is currently being calculated. The content tag corresponding to that content is the target content tag, and the collected historical behavior data is defined as the designated historical behavior data. It is understood that different content may have corresponding content tags pre-assigned, such as music, movies, and food. To calculate user interest for a corresponding content tag, it is sufficient to simply obtain the user's historical behavior data for that content tag.

[0023] When users use applications (such as short videos and graphic apps), the recommendation system collects comprehensive historical user behavior data. For example, when users browse articles, videos, and other content, not only are clicks, favorites, and comments recorded, but the system also accurately monitors the duration of the user's browsing. For reading content, the system tracks how long users spend in different paragraphs to determine their level of attention to different content sections and quantify their reading depth. The system has real-time data integration capabilities, rapidly aggregating all types of user behavior data collected into a unified data center. This ensures the timeliness and integrity of the data, providing the latest and most comprehensive data support for subsequent interest calculations.

[0024] Specifically, taking the short video application as an example, this application tracks user behavior in real time by burying code when collecting historical behavior data of users. For example, record the length of time users browse videos and the completeness of watching videos to understand the degree of user attention to the video content; monitor users' operations of liking, commenting, forwarding, and collecting short videos to judge the user's preference and degree of interaction with the video content; track users' social behaviors such as following other creators and joining fan groups, and analyze users' social relationships and social circles of interest. On the other hand, the server side of the content recommendation system will record in detail the user's login time, usage time, video category pages visited, frequency of switching videos and other information. These log data can show the user's activity trajectory within the application and help analyze the user's usage habits and interest preferences.

[0025] Furthermore, based on the various user historical behavior data monitored by the code tracking, according to the target content tag (such as food) that needs to be calculated at present, the historical behavior data of the user for the target content tag in the past period of time (such as the past 180 days) is collected. These historical behavior data are then split into historical behavior data of different time periods, and the time period obtained by the split is defined as the target time period, and the historical behavior data obtained by filtering different target time periods is defined as the target behavior data. It can be understood that the historical behavior data of users in different time periods have different degrees of influence on the user's interest in the current time. The historical behavior data of the target time period that is closer to the current time has a higher degree of influence on the user's current interest in the target content tag, and the historical behavior data of the target time period that is farther away from the current time has a lower degree of influence on the user's current interest in the target content tag. Therefore, this application splits the specified historical behavior data into multiple target time periods to filter the target behavior data, so as to be used for the subsequent time-based collection of the user's behavioral characteristics for the target content tag, so as to achieve accurate collection of behavioral characteristics.

[0026] Optionally, target behavior data of multiple target time periods are filtered out from the specified historical behavior data, including: Long-term behavior data, short-term behavior data and real-time behavior data are filtered out from the specified historical behavior data. The long-term behavior data is the historical behavior data corresponding to the first specified time period in the specified historical behavior data. The short-term behavior data is the historical behavior data corresponding to the second specified time period in the specified historical behavior data. The real-time behavior data is the historical behavior data collected at the previous historical behavior time of the current time. The duration of the first specified time period, the second specified time period and the previous historical behavior time is shortened in turn, and the time difference between the first specified time period, the second specified time period and the previous historical behavior time and the current time is shortened in turn.

[0027] This application divides the specified historical behavior data of the target content tag into long-term behavior data, short-term behavior data, and real-time behavior data. The filtering time is the first specified period, the second specified period, and the previous historical behavior time of the current time. Taking the specified historical behavior data in the past 180 days as an example, the first specified period is the historical behavior data from the past 7th to the 180th day, and the second specified period is the historical behavior data from the past 1st to the 6th day.

[0028] Among them, long-term behavioral data mainly counts the proportion of users' views of content with target content tags in the past few months. In addition, by analyzing the proportion of users' views of content with different content tags, users' long-term and stable areas of interest can be determined. By recording the types and number of creators that users follow for a long time, we can understand the focus of users' interests. Short-term behavioral data mainly focuses on changes in users' recent preferences for viewing content with target content tags, such as a sudden increase in the number of views of content with target content tags, and analysis of users' recent interactive activities with content with target content tags. Real-time behavioral data accurately records the user's last historical behavioral data on the target content tag.

[0029] Based on the collected long-term, short-term, and real-time behavioral data, we store the long-term data in HDFS (Distributed File System), given its large volume and the need for long-term preservation for in-depth analysis. HDFS is highly fault-tolerant and suitable for storing massive amounts of data. It meets the reliability and high capacity requirements of long-term behavioral data storage, facilitating subsequent complex data analysis and mining to explore users' long-term and stable interest patterns.

[0030] To quickly acquire and process recent user behavior data and respond to dynamic changes in user interests in real time, short-term behavior data is stored in a Redis (Remote Dictionary Server) database. Redis is a high-performance, in-memory cache database with fast read and write speeds. It can quickly respond to data query requests, meeting the need for rapid processing of short-term behavior data and capturing recent changes in user interests.

[0031] Real-time behavior data is collected by the client to account for Redis data latency issues. The next time a user initiates a behavior request, the real-time behavior data (i.e., the data from the previous request) is transparently transmitted to the recommendation system server through the request body. This approach ensures that the recommendation system can obtain the most timely and accurate real-time user behavior data, providing the latest behavioral basis for accurate recommendations. In addition, the recommendation system can also collect basic user attributes, including age, gender, region, device model, etc. This basic information helps to initially categorize users and understand the differences in behavior among different user groups within the app.

[0032] Optionally, this application uses cloud computing instances as the core computing platform for its historical behavior data processing service. Its unique elastic computing architecture provides the system with powerful resource allocation capabilities. For example, in the actual operation of short video applications, the dynamic changes in user volume and the frequent development of business activities will lead to real-time fluctuations in data volume and computing task scale. For example, in the short period after the release of a popular short video, the influx of users will trigger a massive amount of viewing, liking, and other behavioral data. Leveraging the elastic scaling capabilities of cloud computing instances, the system can rapidly expand the number of CPU cores from the initial 8 to 32 cores and memory from 16GB to 64GB during runtime, ensuring smooth system operation under high load and effectively avoiding lags during task processing. Furthermore, the servers' high-speed 10G network interfaces greatly ensure efficient data transmission between the front-end and back-end, and within the server cluster, meeting the stringent requirements of short video loading in seconds, real-time data acquisition, and rapid model training. Compared to the demanding computing resource requirements of deep model learning methods, this application can run efficiently on relatively standard hardware configurations, significantly reducing computing costs.

[0033] The server side of the recommendation system optionally uses the Linux operating system. Its open-source nature allows developers to deeply customize the kernel. Fine-tuning key kernel parameters, such as the file system cache size and network connection timeout, significantly improves overall system performance. When processing large-scale short video data and model training tasks, system stability ensures continuous and reliable task execution, effectively preventing data loss or task interruption due to system failures and guaranteeing uninterrupted 24 / 7 service for short video applications.

[0034] The system's data processing framework leverages the powerful data processing capabilities of the Hadoop (distributed system infrastructure) ecosystem and utilizes the Hive data warehouse tool, allowing users to conveniently query and analyze structured user behavior data using a SQL-like (Structured Query Language) language. For example, Hive SQL statements can easily count key metrics such as the number of views and total number of likes for short videos with different content tags within a specific time period, greatly facilitating data management and aggregate calculations. The data processing framework splits large-scale data processing tasks into multiple subtasks, executing them in parallel on different nodes in the cluster, significantly improving computational efficiency. For example, calculating the completion rate of short videos, by distributing the task across multiple nodes for simultaneous processing, massive data computations can be completed in a short period of time. For tasks requiring high real-time performance, such as real-time statistics of user likes and comments, the in-memory computing feature can significantly increase data processing speed, making it ideal for processing real-time user behavior data.

[0035] S120. Perform weighted calculations on behavior sub-features based on multiple target behavior data of the same behavior tag type, and concatenate the behavior sub-features obtained based on the weighted calculation to obtain behavior tag features. The feature weights of the weighted calculations on each target behavior data are configured from low to high according to the distance between the corresponding target time period and the current time.

[0036] Furthermore, based on the multiple target behavior data collected above, in the feature engineering phase, the recommendation system performs weighted aggregation calculations on multi-period data under the same behavior tag type. By building a dynamic weight allocation mechanism, behavioral data in recent periods are given a higher weight coefficient, while the weight of behavioral data with a long history decays exponentially. For example, the weight of the most recent click behavior may be set to 0.8, while the weight of similar behavior 7 days ago is reduced to 0.2. By performing weighted calculations on sub-features of different target behavior data, the user's current true interest status can be highlighted. For each behavior tag type of a target content tag (such as clicks, likes, comments, etc.), the system performs weighted calculations on the sub-features of each target time period to generate corresponding behavior sub-features.

[0037] Subsequently, the sub-features of the same behavior type label dimension (corresponding to different time periods) are horizontally spliced together to form a multimodal behavior label feature vector that incorporates a time-attenuation effect. This process not only addresses the superficiality problem caused by simple frequency statistics, but also, through time-weighted calibration, reduces the interference of the recommendation system's delivery strategy on user behavior data, making the feature representation more closely aligned with the user's subjective interests.

[0038] Specifically, based on multiple target behavior data of the same behavior tag type, weighted calculation of behavior sub-features is performed separately, including: Obtain the pre-configured feature weights of each target behavior data of the same behavior tag type, calculate the behavior mean and behavior rate value of each target behavior data based on the feature weights, and use the calculated behavior mean and behavior rate value as the behavior sub-features of the corresponding target behavior data.

[0039] Taking the feature processing of long-term, short-term, and real-time behavioral data for different target time periods as an example, the system obtains the target user's long-term, short-term, and real-time behavioral data corresponding to the target content tag from different storage nodes. Based on the different target behavioral data, the system pre-defines a set of time-related weight coefficients. These weights decrease according to the distance between the target time period and the current time. For example, the behavioral weight of real-time behavioral data is set to 0.8, the weight of short-term behavioral data is set to 0.5, and the weight of long-term behavioral data is set to 0.3.

[0040] For each target time period's behavioral data (such as clicks, comments, and likes), the system first calculates the behavioral mean. The system then normalizes the data (e.g., by daily or hourly averages), multiplies the data by the corresponding time period's weight coefficient, and sums the results. For example, if a user clicks on a certain tag 10 times (weight 0.5) in short-term behavioral data and 30 times (weight 0.3) in long-term behavioral data, and there are clicks on the tag in real-time behavioral data, the weighted behavioral mean is (1×0.8+10×0.5+30×0.3) / (0.8+0.5+0.3). The behavioral mean in real-time behavioral data can be expressed as 1 or 0 to indicate whether a click occurred. This process not only smooths out differences in the magnitude of behaviors across time periods but also emphasizes the contribution of recent behaviors through weighted distribution.

[0041] To eliminate the impact of the recommendation system's frequency on the number of actions, the system further calculates an action rate (e.g., click-through rate = number of clicks / number of content exposures). For example, if a piece of content with a certain tag is recommended 100 times during the target time period and users click on it 10 times, the click-through rate is 10%. By multiplying and aggregating the action rate with the corresponding time period weight, the system can quantify the user's active interest in the content, rather than the passive result of receiving recommendations. For real-time behavioral data, since only one action is counted, the action rate can be expressed as 1 or 0 to indicate whether a click occurred.

[0042] Ultimately, each behavior tag type (such as click behavior) generates behavior sub-features for target behavior data across different target periods, with one sub-feature for each target period for each behavior tag type. These sub-features include the corresponding weighted behavior mean (reflecting the time-decay distribution of behavior intensity) and the weighted behavior rate (reflecting the time-decay distribution of behavior efficiency). These sub-features are then concatenated by behavior type dimension to form a multidimensional feature vector containing temporal dynamic information.

[0043] Optionally, the behavior label feature is obtained by concatenating the behavior sub-features obtained based on weighted calculation, including: Obtain the basic attribute characteristics of the target user, and concatenate the basic attribute characteristics, the various behavioral sub-features obtained by weighted calculation, and the corresponding behavioral label types to obtain the behavioral label characteristics.

[0044] By constructing a multi-dimensional feature vector that includes basic user attribute characteristics, long-term behavior, short-term behavior, and real-time behavior, it provides comprehensive and structured data support for subsequent user interest calculations, thereby achieving more accurate personalized recommendations.

[0045] Reference Figure 2Taking the "quick scrolling" behavior as an example, for the past 180 days of user historical behavior data, the mean and rate values for each target period are calculated, yielding corresponding feature information such as average viewing duration and skip rate. This process is weighted according to the set feature weights to obtain sub-features for each target period. Taking the current time node as t, the historical behavior data from days (t-180) to (t-7) is long-term behavior data, the historical behavior data from days (t-6) to (t-1) is short-term behavior data, and the most recent behavior data is real-time behavior data. These sub-features for different target periods are combined with the user's basic attributes and positive / negative sample label type information to obtain a label feature for a specific user behavior type corresponding to the target content label. Each behavior type is assigned a corresponding label feature. The positive / negative sample label type information can be annotated as needed, such as clicks and likes as positive samples and reports as negative samples.

[0046] S130. Input the behavior label features of the target user corresponding to different behavior label types into the pre-built user interest estimation model, and output the user interest of the target user corresponding to the target content label. During model training, the user interest estimation model constructs label samples corresponding to the target content label based on the historical behavior data of different users, and annotates the interest probability of the label samples. The label samples are used as model input for model training. The label samples contain sample features of multiple behavior label types.

[0047] The final user interest calculation stage of this application is implemented using a pre-trained user interest estimation model. This model uses a deep learning architecture and can introduce a transfer learning mechanism during the training phase: first, a label sample library is constructed based on a large number of users' historical behavior logs. Each sample contains multiple behavior type features (such as number of clicks, average browsing time, collection frequency, etc.) and corresponding target content labels. Interest probability labels are then generated through subsequent user behavior (such as whether clicks are generated, whether active searches for similar content are performed, etc.) or manual labeling. The model is trained using a multi-task learning paradigm, while optimizing auxiliary objectives such as click-through rate prediction and interest decay curve fitting to enhance feature generalization capabilities.

[0048] During the model inference phase, the model receives the concatenated behavior tag feature vectors and automatically learns the importance weights of each behavior type feature through an attention mechanism, ultimately outputting an interest probability value between 0 and 1. This end-to-end modeling approach overcomes the linear weighting limitations of traditional rule engines, capturing the nonlinear evolution of user interests and accurately calculating the user's interest in the target content tag, providing a precise data foundation for subsequent content recommendations.

[0049] Through the above technical solutions, we can build a comprehensive calculation system for user interest in content tags from data collection, calculation logic to interactive display. It has low computational cost, strong interpretability, and can fully consider the diversity and depth of user behavior, laying a solid foundation for the accurate recommendation and good user experience of the recommendation system. By comprehensively considering user behavior in different time periods, such as browsing time, interactive behavior, interest decay, etc., we can more accurately measure the true interest of users in content tags. In predicting user interest, we comprehensively refer to the user's long-term behavior, short-term behavior, and real-time behavior. Long-term behavior can reflect the user's stable interest preferences, short-term behavior can capture the user's recent interest changes, and real-time behavior can reflect the user's current immediate needs. By integrating behavioral data from these three dimensions, we can more accurately predict the user's interest in tags, thereby further improving the accuracy and adaptability of the recommendation system.

[0050] In summary, this application uses innovative computational methods to comprehensively integrate users' long-term, short-term, and real-time behaviors, fully considering the diversity and depth of user behavior. This effectively reduces computational costs while significantly enhancing the interpretability of interest calculations. This allows for a more accurate grasp of user interests, providing strong support for recommendation systems, significantly improving the accuracy and personalization of recommendation systems, meeting the growing demand for recommendation systems from users and the market, and promoting the further development of recommendation system technology.

[0051] Optionally, refer to Figure 3 ,The training process of the user interest estimation model includes: S1001. Collect historical behavior data corresponding to target content tags from different users. Use the behavior tag type as a screening dimension to filter the collected historical behavior data to determine subsamples corresponding to different target time periods. Perform weighted calculation and concatenation of subfeatures based on the subsamples to obtain multiple sample features corresponding to different behavior tag types. Use the multiple sample features of the same user as the tag sample. S1002: Label the corresponding interest probability for the labeled sample, and use the labeled sample as a model input to train a user interest estimation model; S1003: Calculate a loss function according to the model output, and adjust the model parameters of the user interest estimation model based on the loss function until the user interest estimation model converges.

[0052] When training the user interest estimation model, refer to the above-mentioned method of determining label features to obtain label samples for model training. Figure 4 As shown in FIG, by filtering long-term behavior data, short-term behavior data and real-time behavior data from the user's historical behavior data, and then obtaining label samples through data cleaning, feature organization and feature processing.

[0053] Specifically, during the feature organization process, taking the feature calculation of click counts as an example, since click counts are easily influenced by content delivery times, this application avoids directly using frequency data in subsequent calculations. Smoothing algorithms such as sliding averages are used to process the acquired data, eliminating noise and fluctuations, ensuring data stability and reliability, and laying a solid foundation for subsequent analysis. In addition to cleaning anomalous data, this application also cleans data where it is unclear whether it is positive or negative. For example, in sample classification, positive samples are clearly defined as user likes, shares, completed broadcasts, and views of 6 seconds or longer, with a view duration greater than 80% of the video's duration. Negative samples are user reports, views of 6 seconds or less, and less than 20% of the video's duration. Data in an intermediate state, where it is difficult to clearly distinguish between positive and negative samples according to established criteria, is removed, as it may interfere with subsequent model training and reduce model accuracy and reliability. This provides a high-quality data foundation for subsequent analysis and model training.

[0054] Furthermore, this application focuses on the calculation of sample means and rate values. For example, the weighted sliding average method is used to calculate the average duration per view, fully considering the timeliness of user behavior and giving higher weight to recent data to accurately reflect the user's recent behavior trends. For example, when calculating click-through rate (CTR) and completion rate, the original data is smoothed to eliminate fluctuations caused by accidental factors, so that these indicators can truly reflect the user's interaction with the tagged content from multiple dimensions. After completing the data cleaning, the sample classification standards are further clarified to build a clear and effective sample system. Positive samples and negative samples are strictly divided according to established standards to ensure the validity and analyzability of the data.

[0055] Then, based on the clear definition of positive and negative samples, historical behavior data for the corresponding period is extracted from the data storage node to construct sample labels. For strong samples such as likes, shares, and reports, corresponding weighting coefficients can be set according to business needs. For example, the weighting coefficient of likes is set to 1.5, the weighting coefficient of shares is set to 2, and the weighting coefficient of reports is set to 3. This is used to calculate user interest, allowing the model to pay more attention to these strong interactive behaviors, thereby more accurately capturing user interest preferences. For older historical behavior data, an exponential decay function is used to reasonably decay its weight by setting the optimal decay value, effectively reducing the impact of expired behavior on the model.

[0056] During training, a logistic regression model is used to predict user interest. Using optimization algorithms such as gradient descent, training is performed in parallel across multiple servers via a distributed training framework. The predicted user interest is compared against the pre-annotated interest probabilities of the labeled samples to calculate a loss function. Based on this loss function, the parameters of the logistic regression model are continuously adjusted to minimize the loss function on the training set, thereby improving the model's accuracy and interpretability. The pre-annotated interest probabilities of the labeled samples can be calculated based on actual behavior in historical data or manually set.

[0057] Compared with the complex training process of deep models, the training process of logistic regression models is simpler and more efficient, and the model parameters are highly interpretable. It can clearly demonstrate the impact of each feature on user interest, providing a more convincing decision-making basis for the recommendation system.

[0058] Optionally, during the feature splicing process of the user interest estimation model, the present application also evenly divides the value range of each sub-feature into 100 equal-frequency intervals, and uses a distributed computing framework to count the amount of data in each interval, and then accurately calculates the bucket threshold. Merging data points with the same value into the same bucket effectively reduces the data dimension, reduces data redundancy, and improves model training efficiency. For feature splicing and encoding, the present application organically splices the basic attribute features (such as gender, language, country, etc.) on the user side with behavioral features. Each feature is converted into an input format that the model can accept, such as encoding the gender feature as [0, 1] to represent male and female respectively, and encoding the regional feature as the corresponding one-hot vector, eliminating the influence of different user basic attributes on label preference, and providing high-quality input data for model training.

[0059] Optionally, when bucketing features, in addition to equal-frequency bucketing, you can also use equal-interval bucketing. Equal-interval bucketing divides the feature value range into fixed intervals, which is relatively simple to calculate and can improve computational efficiency.

[0060] Afterwards, based on the above-mentioned pre-trained user interest estimation model, the behavior label features collected from the user's historical behavior data are input into the user interest estimation model to output the user interest of the target user corresponding to the target content label.

[0061] Among them, reference Figure 5 , output the user interest of the target user corresponding to the target content label, including: S1301: Determine the sample type of each behavior tag feature according to the behavior tag type, and determine the interest weight of each behavior tag feature based on the sample type, where the sample type includes positive samples and negative samples corresponding to the user's interest; S1302: Calculate the interest score of the target user corresponding to the target content tag based on each behavior tag feature and the corresponding interest weight; S1303: Calculate the user interest level of the target user corresponding to the target content tag based on the interest score.

[0062] Referencing the weighting coefficients set for different behavior types during model training, we also set corresponding weighting coefficients (i.e., interest weights) for different behavior sample types (such as likes, shares, and reports) based on business needs. For example, the weighting coefficient for likes can be set to 1.5, the weighting coefficient for shares to 2, and the weighting coefficient for reports to 3.

[0063] Exemplarily, the interest score of the target user corresponding to the target content tag is calculated based on each behavior tag feature of the user and the corresponding interest weight.

[0064] Among them, interest score = ∑(positive sample weight) + ∑(negative sample weight) For example, if user A likes a video tagged "Technology" 3 times, it gets 3 × 1.5 = 4.5 points; shares it once, it gets 1 × 2 = 2 points; reports it once, it gets -3 points; and skips it twice, it gets 2 × (-1) = -2 points (assuming the weight is 1). The total interest score is 4.5 + 2 - 3 - 2 = 1.5.

[0065] Based on the actual setting of the behavior label features, the behavior label features that characterize a type of behavior are obtained through the behavior mean, behavior rate value, etc., and combined with the interest weight, the corresponding interest score can be obtained. This application does not impose fixed restrictions on the specific interest score calculation rules, so we will not elaborate on them here.

[0066] Based on the interest score calculated above, the user interest is calculated using a logistic regression model of a pre-trained user interest estimation model. Optionally, the logistic regression model is as follows:

[0067] Among them, w is the influence coefficient of the interest score, which is adaptively adjusted according to model training; x is the interest score. Optionally, x can also be a collection of interest scores and behavioral label features of different behavior types; b is the bias term that is adaptively adjusted according to model training.

[0068] Through the above model calculation method, the corresponding user interest level can be output, thereby achieving accurate calculation of the user interest level estimation model.

[0069] After outputting the user interest level of the target user corresponding to the target content tag, the following is also included: Obtain the target user's interest in different content tags, and rank the content recommendations for different content tags according to the user's interest.

[0070] After the model calculates the predicted user interest in the current target content tag based on the input features, the prediction result can be returned to the front-end application via a web service interface for personalized recommendations of short video content and other applications. Simultaneously, the system monitors new user behavior data in real time. Once new data is uploaded, it immediately transmits the data to the real-time data processing module via a message queue, updates the model's input features, recalculates the prediction value, and achieves real-time model optimization. For example, when a user clicks "like" while watching a short video, the system can update the model within 1 second and recommend short videos that better suit the user's interests. This real-time feedback and optimization mechanism can dynamically adjust the recommendation strategy based on the user's real-time behavior, further improving the accuracy and personalization of the recommendation system. Furthermore, because the logistic regression model used has low computational cost, it can quickly complete model updates and predictions, meeting real-time requirements and providing users with higher-quality and more efficient recommendation services.

[0071] At the same time, when sorting content, using the user interest in the content tags calculated by this application as an important feature can significantly improve the accuracy and rationality of the sorting. When recommending short videos to users, the sorting model will comprehensively consider factors such as the popularity, quality, and relevance of the user to the video, and the user's interest in the tag to which the video belongs is one of the key indicators for measuring relevance. For example, when recommending videos to users, if the popularity and quality of two videos are similar, but the user is more interested in the tag to which one of the videos belongs, then the sorting model will rank the video higher, ensuring that users can see the content they are more interested in first, thereby improving the click-through rate of recommendations and user retention rate.

[0072] By introducing a multi-channel recall strategy based on user tag interests and building a real-time feature fusion mechanism to dynamically inject user interest prediction values into the ranking model, the content matching efficiency of the short video recommendation system is significantly optimized.

[0073] Optionally, for the development and training of the deep learning model of the present application, mainstream frameworks such as TensorFlow or PyTorch can be used. These frameworks provide a wealth of neural network layers and optimization algorithms. When developing deep learning models for understanding short video content, these frameworks can be used to build complex convolutional neural networks, recurrent neural networks, etc. On this basis, the present application combines the logistic regression model, which can have the significant characteristics of being simple to use and computationally efficient. It can complete the training and prediction of the model in a short time, while meeting the requirements of model parameter interpretability, and facilitates in-depth analysis of the intrinsic relationship between user interest and various features. Compared with deep models, the training process of the logistic regression model is simpler and more efficient, without the need for complex parameter adjustment and a large amount of computing resources. While ensuring that the computing cost is reasonable, it provides users with more convincing and transparent interest calculation results. For model training optimization algorithms, in addition to gradient descent, the Adam adaptive learning rate algorithm can also be used. The Adam algorithm can automatically adjust the learning rate and converge faster during training. It is particularly suitable for processing large-scale data and high-dimensional parameter spaces.

[0074] Optionally, when calculating user interest, this application can also use deep neural networks (DNNs) based on deep learning methods to directly model user behavior data. For example, multi-layer perceptrons (MLPs) can process high-dimensional and complex data, automatically learn nonlinear relationships between features, and predict user interest directly from raw behavior data through end-to-end training.

[0075] Furthermore, when recommending content, this application also uses collaborative filtering to analyze similarities between users, identifying groups of users with similar interests. Content is then recommended based on the behavior of these similar users. For example, user-based collaborative filtering identifies other users with similar behavior patterns as the target user and recommends content of interest to these similar users. This approach does not require complex feature engineering and is relatively simple to implement.

[0076] In one embodiment, the present application also provides a variety of interest recall schemes to further improve the accuracy of content recommendations. Among them. In terms of long-term and short-term interest recall, through the long-term accumulation and analysis of user behavior data, combined with the time decay mechanism, the user's long-term stable interests and short-term sudden interests can be effectively distinguished. For long-term interests, the system will comprehensively consider the user's interest in various tagged videos over a long period of time in the past (such as more than 180 days). For those tags whose interest has always remained at a high level, the corresponding videos will be used as candidate sets for long-term interest recall. For example, if the user has maintained a high level of interest in videos tagged "travel" in the past few months, the system will continue to include newly released travel videos in the recommendation range.

[0077] For short-term interests, the system focuses on recent changes in user behavior (e.g., the past seven days). When a user frequently views and likes videos with a specific tag within a short period of time, such as the "World Cup" tag during a tournament, the system quickly detects this short-term interest spike and prioritizes relevant videos to meet the user's interests during that specific timeframe.

[0078] Based on the user interest calculated by this application, it is possible to accurately recall positive and negative feedback interests. For positive feedback interests, that is, videos with labels in which users show positive behaviors (likes, shares, completions, etc.), the system will increase the recall of videos with similar labels. For example, if a user frequently likes and shares videos with the label "pets are funny", the system will select more videos of this type during the recall phase, providing users with more high-quality content of the same type. For negative feedback interests, that is, content that users clearly express no interest in (reports, dislikes, etc.), the system will automatically filter out videos with related labels during recall to avoid pushing content they don't like to users, thereby improving the user experience. At the same time, negative feedback information will also be incorporated into the model's training process to further optimize the model's judgment of user interest preferences.

[0079] In terms of long-tail recall, this application can tap into users' potential interest in some niche, low-frequency tags through refined calculations of user interest. Even if the number of videos with certain tags is relatively small and the number of views is not high, as long as the user shows a certain degree of interest in them, the system will include these long-tail tag videos in the recall range. For example, for long-tail tags such as "hand-woven niche patterns", if a user has browsed or liked related videos, the system will recommend more such niche videos that meet their interests to the user based on the calculated interest level during recall, enriching the user's content selection and meeting the user's diverse interest needs.

[0080] In addition, considering the diversity of user interests, this application also implements multi-interest recall. Through a comprehensive analysis of user behavior data, the user's interest tags in different fields are identified, such as "food cooking", "fitness and sports", "film and television entertainment", etc. During the recall stage, the system will recall videos from multiple interest areas at the same time based on the user's interest in different tags, and provide users with diversified content recommendations. For example, for a user who likes both food and fitness, the system will recall food preparation tutorials and fitness training videos at the same time to meet the user's needs in multiple interest dimensions and improve the comprehensiveness and accuracy of the recommendations.

[0081] Optionally, this application also performs collaborative recall of users based on the user's interest in each tag. Figure 6To build a user interest profile system, we first collect user interest data for all content tags. These interest values are accurately output by the user interest estimation model as probabilities between 0 and 1. These values are then combined to form a high-dimensional vector. Next, we introduce the AutoEncoder to perform dimensionality reduction on the high-dimensional vector. With its unique encoder-decoder architecture, the AutoEncoder can keenly capture and retain the core information in the high-dimensional vector, condensing it into a low-dimensional vector representation of the user.

[0082] The resulting low-dimensional vector possesses two significant properties: first, it accurately captures user interests and preferences for each key tag; second, it intelligently integrates interest information for similar tags, significantly improving the generalization and accuracy of user interest characteristics, making it ideal for representing user tag interest profiles. Once a user representation (i.e., a low-dimensional vector User Embedding) is successfully obtained, it can quickly locate similar user groups for each user and fully recall the browsing history of these similar users, laying a solid foundation for subsequent in-depth analysis and personalized recommendations.

[0083] In the above, by obtaining the specified historical behavior data of the target user corresponding to the target content tag, the set behavior tag type is used as the filtering dimension to filter out target behavior data of multiple different target time periods from the specified historical behavior data; based on the multiple target behavior data of the same behavior tag type, weighted calculation of behavior sub-features is performed respectively, and the behavior tag features obtained based on the weighted calculation are spliced to obtain the behavior tag features, and the feature weights of the weighted calculation of each target behavior data are configured from low to high according to the distance between the corresponding target time period and the current time; the behavior tag features of the target user corresponding to different behavior tag types are input into the pre-built user interest estimation model, and the user interest of the target user corresponding to the target content tag is output. Using the above technical means, the tag features of the user's historical behavior data are weightedly calculated in different time periods, and the user interest is calculated using the user interest estimation model based on the tag features of different behavior tag types, thereby avoiding the situation where the user interest is simply calculated based on the number of user behaviors, which leads to low calculation accuracy, improving the calculation accuracy and personalization of the user interest, and thus improving the content recommendation effect of the recommendation system and the user's content viewing experience.

[0084] Based on the above embodiments, Figure 7 This is a schematic diagram of the structure of a content tag user interest generation system provided by this application. Figure 7 The user interest generation system for content tags provided in this embodiment specifically includes: The data screening module 21 is configured to obtain the specified historical behavior data of the target user corresponding to the target content tag, and use the set behavior tag type as the screening dimension to screen out target behavior data of multiple different target time periods from the specified historical behavior data; The feature concatenation module 22 is configured to perform weighted calculations on behavior sub-features based on multiple target behavior data of the same behavior tag type, and to concatenate the behavior sub-features obtained through the weighted calculations to obtain a behavior tag feature. The feature weights of the weighted calculations on each target behavior data are configured from low to high according to the distance between the corresponding target time period and the current time. The model calculation module 23 is configured to input the target user's behavior tag features corresponding to different behavior tag types into a pre-built user interest estimation model, and output the target user's user interest corresponding to the target content tag.

[0085] Specifically, target behavior data of multiple target time periods are filtered out from the specified historical behavior data, including: Long-term behavior data, short-term behavior data and real-time behavior data are filtered out from the specified historical behavior data. The long-term behavior data refers to the historical behavior data corresponding to the first specified time period in the specified historical behavior data. The short-term behavior data refers to the historical behavior data corresponding to the second specified time period in the specified historical behavior data. The second specified time period is shorter than the first specified time period and closer to the current time. The real-time behavior data refers to the historical behavior data closest to the current time in the specified historical behavior data.

[0086] Specifically, based on multiple target behavior data of the same behavior tag type, weighted calculation of behavior sub-features is performed separately, including: Obtain the pre-configured feature weights of each target behavior data of the same behavior tag type, calculate the behavior mean and behavior rate value of each target behavior data based on the feature weights, and use the calculated behavior mean and behavior rate value as the behavior sub-features of the corresponding target behavior data.

[0087] Specifically, the behavior label features are obtained by concatenating the behavior sub-features obtained based on weighted calculation, including: Obtain the basic attribute characteristics of the target user, and concatenate the basic attribute characteristics, the various behavioral sub-features obtained by weighted calculation, and the corresponding behavioral label types to obtain the behavioral label characteristics.

[0088] Specifically, the training process of the user interest estimation model includes: Collect historical behavior data corresponding to target content tags from different users, use the behavior tag type as a screening dimension to filter the collected historical behavior data to determine subsamples corresponding to different target time periods, perform weighted calculation and splicing of subfeatures based on the subsamples, and obtain multiple sample features corresponding to different behavior tag types. Use multiple sample features of the same user as label samples; Label the corresponding interest probability for the labeled samples and use the labeled samples as model input to train the user interest estimation model; The loss function is calculated according to the model output, and the model parameters of the user interest estimation model are adjusted based on the loss function until the user interest estimation model converges.

[0089] Specifically, outputting the user interest level of the target user corresponding to the target content tag includes: Determine the sample type of each behavior tag feature according to the behavior tag type, and determine the interest weight of each behavior tag feature based on the sample type. The sample type includes positive samples and negative samples corresponding to the user's interest; Calculate the interest score of the target user corresponding to the target content tag based on the characteristics of each behavior tag and the corresponding interest weight; Calculate the user interest level of the target user corresponding to the target content tag based on the interest score.

[0090] In addition, after outputting the user interest level of the target user corresponding to the target content tag, the following is also included: Obtain the target user's interest in different content tags, and rank the content recommendations for different content tags according to the user's interest.

[0091] In the above, by obtaining the specified historical behavior data of the target user corresponding to the target content tag, the set behavior tag type is used as the filtering dimension to filter out target behavior data of multiple different target time periods from the specified historical behavior data; based on the multiple target behavior data of the same behavior tag type, weighted calculation of behavior sub-features is performed respectively, and the behavior tag features obtained based on the weighted calculation are spliced to obtain the behavior tag features, and the feature weights of the weighted calculation of each target behavior data are configured from low to high according to the distance between the corresponding target time period and the current time; the behavior tag features of the target user corresponding to different behavior tag types are input into the pre-built user interest estimation model, and the user interest of the target user corresponding to the target content tag is output. Using the above technical means, the tag features of the user's historical behavior data are weightedly calculated in different time periods, and the user interest is calculated using the user interest estimation model based on the tag features of different behavior tag types, thereby avoiding the situation where the user interest is simply calculated based on the number of user behaviors, which leads to low calculation accuracy, improving the calculation accuracy and personalization of the user interest, and thus improving the content recommendation effect of the recommendation system and the user's content viewing experience.

[0092] The system for generating user interest in content tags provided in the embodiments of the present application can be configured to execute the method for generating user interest in content tags provided in the above embodiments, and has corresponding functions and beneficial effects.

[0093] Based on the above practical example, the present application embodiment also provides a device for generating user interest in content tags, referring to Figure 8 The device for generating user interest in content tags includes: a processor 31, a memory 32, a communication module 33, an input device 34, and an output device 35. The memory, as a computer-readable storage medium, can be configured to store software programs, computer-executable programs, and modules, such as program instructions / modules corresponding to the method for generating user interest in content tags described in any embodiment of the present application (for example, the various modules in the system for generating user interest in content tags). The communication module is configured to transmit data. The processor executes the software programs, instructions, and modules stored in the memory to execute various functional applications and data processing of the device, thereby implementing the above-mentioned method for generating user interest in content tags. The input device can be configured to receive input digital or character information and generate key signal input related to user settings and function control of the device. The output device can include a display device such as a display screen. The device for generating user interest in content tags provided above can be configured to execute the method for generating user interest in content tags provided in the above-mentioned embodiments, and has corresponding functions and beneficial effects.

[0094] Based on the above embodiments, embodiments of the present application further provide a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores computer-executable instructions. When executed by a computer processor, the computer-executable instructions are configured to execute a method for generating user interest in content tags. The storage medium can be any of various types of memory devices or storage devices. Of course, the computer-executable instructions of the non-volatile computer-readable storage medium provided in embodiments of the present application are not limited to the method for generating user interest in content tags described above, but can also execute related operations in the method for generating user interest in content tags provided in any embodiment of the present application.

[0095] Based on the above embodiments, the embodiments of the present application also provide a computer program product. The essence of the technical solution of the present application or the part that contributes to the existing technology or all or part of the technical solution can be embodied in the form of a software product. The computer program product is stored in a storage medium and includes a number of instructions for enabling a computer device, a mobile terminal or a processor therein to execute all or part of the steps of the method for generating user interest in content tags described in each embodiment of the present application.

Claims

1. A method for generating user interest in content tags, characterized in that: include: Obtaining designated historical behavior data of a target user corresponding to a target content tag, and using a set behavior tag type as a filtering dimension, filtering out target behavior data of multiple target time periods from the designated historical behavior data; Based on the multiple target behavior data of the same behavior tag type, weighted calculation of behavior sub-features is performed respectively, and the behavior sub-features obtained based on the weighted calculation are spliced to obtain the behavior tag feature, and the feature weight of the weighted calculation of each target behavior data is configured from low to high according to the distance between the corresponding target time period and the current time; The behavior tag features of the target user corresponding to different behavior tag types are input into a pre-built user interest estimation model, and the user interest of the target user corresponding to the target content tag is output.

2. The method for generating user interest in content tags according to claim 1, characterized in that: The step of filtering out target behavior data of multiple target time periods from the designated historical behavior data includes: Long-term behavior data, short-term behavior data and real-time behavior data are filtered out from the specified historical behavior data, the long-term behavior data is the historical behavior data corresponding to the first specified time period in the specified historical behavior data, the short-term behavior data is the historical behavior data corresponding to the second specified time period in the specified historical behavior data, and the real-time behavior data is the historical behavior data collected at the previous historical behavior time of the current time. The lengths of the first specified time period, the second specified time period and the previous historical behavior time are shortened successively, and the time difference between the first specified time period, the second specified time period and the previous historical behavior time and the current time are shortened successively.

3. The method for generating user interest in content tags according to claim 1, characterized in that: The weighted calculation of behavior sub-features of the plurality of target behavior data based on the same behavior tag type includes: Obtain pre-configured feature weights for each target behavior data of the same behavior tag type, and calculate the behavior mean and behavior rate value of each target behavior data based on the feature weights, so as to use the calculated behavior mean and behavior rate value as the corresponding behavior sub-features of the target behavior data.

4. The method for generating user interest in content tags according to claim 3, wherein: The behavior label feature is obtained by concatenating the behavior sub-features obtained based on the weighted calculation, including: The basic attribute features of the target user are obtained, and the basic attribute features, each of the behavioral sub-features obtained by weighted calculation, and the corresponding behavioral label type are combined to obtain a behavioral label feature.

5. The method for generating user interest in content tags according to any one of claims 1 to 4, characterized in that: The training process of the user interest estimation model includes: Collect historical behavior data corresponding to the target content tag from different users, use the behavior tag type as a screening dimension to filter the collected historical behavior data to determine subsamples corresponding to different target time periods, perform weighted calculation and splicing of subfeatures based on the subsamples to obtain multiple sample features corresponding to different behavior tag types, and use the multiple sample features of the same user as label samples; Labeling the label samples with corresponding interest probabilities, and using the label samples as model inputs to train the user interest estimation model; A loss function is calculated according to the model output, and model parameters of the user interest estimation model are adjusted based on the loss function until the user interest estimation model converges.

6. The method for generating user interest in content tags according to any one of claims 1 to 4, characterized in that: Outputting the user interest level of the target user corresponding to the target content tag includes: Determine a sample type of each behavior tag feature according to the behavior tag type, and determine an interest weight of each behavior tag feature based on the sample type, wherein the sample type includes positive samples and negative samples corresponding to user interest; Calculating the target user's interest score corresponding to the target content tag based on each of the behavior tag features and the corresponding interest weight; The user interest degree of the target user corresponding to the target content tag is calculated based on the interest score.

7. The method for generating user interest in content tags according to any one of claims 1 to 4, characterized in that: After outputting the user interest level of the target user corresponding to the target content tag, the method further includes: The user interest level of the target user corresponding to different content tags is obtained, and content recommendation ranking of the different content tags is performed according to the user interest level.

8. A system for generating user interest in content tags, characterized in that: include: a data screening module configured to obtain designated historical behavior data of a target user corresponding to a target content tag, and to screen target behavior data of multiple target time periods from the designated historical behavior data using a set behavior tag type as a screening dimension; a feature splicing module configured to perform weighted calculation of behavior sub-features based on multiple target behavior data of the same behavior tag type, and splice the behavior sub-features obtained based on the weighted calculation to obtain a behavior tag feature, wherein the feature weight of each target behavior data in the weighted calculation is configured from low to high according to the distance between the corresponding target time period and the current time; The model calculation module is configured to input the behavior label features of the target user corresponding to different behavior label types into a pre-built user interest estimation model, and output the user interest of the target user corresponding to the target content label. During model training, the user interest estimation model constructs label samples corresponding to the target content label based on the historical behavior data of different users, and annotates the interest probability of the label samples. The model training is performed using the label samples as model input, and the label samples contain sample features of multiple behavior label types.

9. A device for generating user interest in content tags, characterized in that: include: memory and one or more processors; The memory is configured to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for generating user interest in content tags according to any one of claims 1 to 7.

10. A non-volatile computer-readable storage medium, characterized in that: The non-volatile computer-readable storage medium stores computer-executable instructions, which, when executed by a computer processor, are configured to execute the method for generating user interest in content tags according to any one of claims 1 to 7.

11. A computer program product, characterized in that The computer program product includes instructions, and when the instructions are executed on a computer or a processor, the computer or the processor executes the method for generating user interest in content tags according to any one of claims 1 to 7.

Images