Model training and blog article quality evaluation method and device, equipment and storage medium

Abstract

The invention provides a model training and blog article quality evaluation method and device, equipment and a storage medium, and the method comprises the steps: obtaining a plurality of training samples based on an exposure log of a blog article recommendation system; preprocessing the blog article information of the sample blog article in each training sample and the user information of the corresponding sample blogger, and generating blog article side behavior characteristics and user side behavior characteristics corresponding to each training sample; performing correction processing on the blog article side behavior characteristics and the user side behavior characteristics to obtain blog article correction characteristics and user correction characteristics corresponding to each training sample; and in the process of performing iterative training on the initial blog article quality evaluation model by using the plurality of training samples, adjusting model parameters based on the blog article side behavior characteristics, the user side behavior characteristics, the blog article correction characteristics and the user correction characteristics corresponding to each training sample so as to obtain a trained blog article quality evaluation model. Through the method and the device, more accurate blog article recommendation can be realized.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method, device, equipment and storage medium for model training and blog post quality assessment. Background Art

[0002] At present, the recommendation system is one of the most effective ways to solve information overload. In the relationship flow distribution scenario, the recommendation system is generally divided into two stages, namely the recall stage and the sorting stage. In the recall stage, the blog posts that users may be interested in are fully mined through multi-link methods such as user profiling, collaborative filtering, graph networks, and approximate search, and the multi-path recalled blog posts are truncated and merged according to time sequence or snake sorting; in the refined sorting stage, the recommendation algorithm based on deep learning makes personalized recommendations for the recalled blog posts to promote users' consumption desire.

[0003] However, by fully considering users' multiple interests and interest changes, the multi-link recalled blog posts have serious dimensionality problems in the quality calculation results of the recalled blog posts in different interest areas, because different blog posts target different interest areas and face different user bases, which affects the accurate recommendation of blog posts. Summary of the invention

[0004] The present application provides a model training and blog post quality assessment method, device, equipment and storage medium to achieve more accurate blog post recommendations.

[0005] The technical solution of this application is implemented as follows:

[0006] The present application provides a method for training a blog post quality assessment model, the method comprising: obtaining a plurality of training samples based on an exposure log of a blog post recommendation system; wherein each training sample comprises blog post information of a sample blog post and user information of a corresponding sample blogger; a sample blogger is a user who publishes a sample blog post; preprocessing the blog post information of a sample blog post in each training sample and user information of a corresponding sample blogger, generating blog post-side behavior features and user-side behavior features corresponding to each training sample; wherein the blog post-side behavior features are used to characterize the interaction of a sample blog post within a preset first time period; and the user-side behavior features are used to characterize the interaction of a sample blog post published by a sample blogger within a preset second time period. The interaction of the blog posts that have been read; by correcting the blog post-side behavioral features and the user-side behavioral features, the corrected blog post features and the user-corrected features corresponding to each training sample are obtained; the confidence level of the corrected blog post features is higher than that of the blog post-side behavioral features, and the confidence level of the user-corrected features is higher than that of the user-side behavioral features; in the process of iteratively training the initial blog post quality assessment model using multiple training samples, the model parameters are adjusted based on the blog post-side behavioral features, the user-side behavioral features, the blog post correction features and the user correction features corresponding to each training sample until the blog post quality assessment model converges to obtain a trained blog post quality assessment model; the blog post quality assessment model is used to assess the quality of blog posts.

[0007] In some possible implementations, the above method also includes: after obtaining multiple training samples, preprocessing the blog post information of the sample blog post in each training sample and the user information of the corresponding sample blogger to generate blog post short-term features and user information features corresponding to each training sample; the blog post short-term features are used to characterize the interaction of the sample blog post within a preset third time period, the third time period is shorter than the first time period, and the user information features are used to characterize the user portrait of the sample blogger.

[0008] In some possible implementations, during the iterative training of an initial blog post quality assessment model using multiple training samples, model parameters are adjusted based on the blog post-side behavior features, user-side behavior features, blog post correction features, and user correction features corresponding to each training sample until the blog post quality assessment model converges, thereby obtaining a trained blog post quality assessment model, including: determining the labeling information corresponding to each sample blog post based on the blog post-side behavior features, user-side behavior features, blog post correction features, and user correction features corresponding to each training sample; the labeling information is used to label the sample blog post as a positive sample or a negative sample; inputting the blog post short-term features and user information features corresponding to each training sample into the blog post quality assessment model to obtain a quality assessment result output by the blog post quality assessment model; the quality assessment result includes a first quality score corresponding to the sample blog post, and a prediction result of whether the sample blog post is a positive sample or a negative sample; the first quality score is used to characterize the user's interest in the sample blog post; a comparison result is obtained by comparing the prediction result and labeling information corresponding to each sample blog post; based on the comparison result of each sample blog post, adjusting the parameters of the blog post quality assessment model until the blog post quality assessment model converges.

[0009] In some possible implementations, based on the blog post-side behavioral features, user-side behavioral features, blog post correction features, and user correction features corresponding to each training sample, the labeling information corresponding to each sample blog post is determined, including: linear fusion processing is performed on the blog post-side behavioral features, user-side behavioral features, blog post correction features, and user correction features corresponding to each training sample to obtain an initial quality score corresponding to the sample blog post in each training sample; the initial quality score is used to characterize the user's interest in the sample blog post without being processed by the blog post quality assessment model; when the initial quality score of the sample blog post meets a preset threshold, the sample blog post is marked as a positive sample; when the initial quality score of the sample blog post does not meet the preset threshold, the sample blog post is marked as a negative sample.

[0010] In some possible implementations, by correcting the blog post behavior features and the user behavior features, the corrected blog post features and the user corrected features corresponding to each training sample are obtained, including: performing credibility correction on the blog post behavior features according to the Wilson confidence interval to obtain the corrected blog post features corresponding to each training sample; performing credibility correction on the user behavior features according to the Wilson confidence interval to obtain the corrected user features corresponding to each training sample.

[0011] The present application provides a method for evaluating the quality of a blog post, the method comprising: obtaining a plurality of exposed historical blog posts and historical bloggers who published the historical blog posts based on an exposure log of a blog post recommendation system; inputting the blog post information of each historical blog post and the user information of its corresponding historical blogger into a trained blog post quality evaluation model to obtain a second quality score corresponding to each historical blog post output by the blog post quality evaluation model; the blog post quality evaluation model is trained using the training method of the blog post quality evaluation model as described above, and the second quality score is used to characterize the user's interest in the historical blog post; based on the second quality score, determining a target blog post from a plurality of exposed historical blog posts; and the target blog post is used for a rough ranking of blog posts or a fine ranking of blog posts.

[0012] In some possible implementations, based on the second quality score, a target blog post is determined from multiple exposed historical blog posts, including: calculating a third quality score corresponding to each historical blog post according to a preset attenuation mapping relationship, a preset time period, and the second quality score; the third quality score is used to characterize the user's interest in the historical blog post after removing the influence of time; the attenuation mapping relationship is a mapping relationship between the second quality score and the third quality score; the blog post information of each historical blog post and the third quality score corresponding to each historical blog post are associated and saved in a trained blog quality assessment model; based on the third quality score, a target blog post is determined from multiple exposed historical blog posts.

[0013] In some possible implementations, a target blog post is determined from multiple exposed historical blog posts based on the third quality score, including: screening out multiple recalled blog posts from multiple exposed historical blog posts through multi-link recall; obtaining the third quality score corresponding to each recalled blog post in a trained blog post quality assessment model; and selecting a target blog post from multiple recalled blog posts based on the third quality score corresponding to each recalled blog post.

[0014] The present application provides a training device for a blog post quality assessment model, the device comprising: a sample data acquisition module, for obtaining a plurality of training samples based on an exposure log of a blog post recommendation system; wherein each training sample comprises the blog post information of the sample blog post and the user information of its corresponding sample blogger, the sample blogger being the user who publishes the sample blog post; a sample data preprocessing module, for preprocessing the blog post information of the sample blog post in each training sample and the user information of its corresponding sample blogger, to generate blog post-side behavior features and user-side behavior features corresponding to each training sample; wherein the blog post-side behavior features are used to characterize the interaction of the sample blog post within a preset first time period; the user-side behavior features are used to characterize the interaction of the sample blogger within a preset second time period The interaction status of published blog posts; a sample data processing module, which is used to obtain the corrected blog post features and user corrected features corresponding to each training sample by correcting the blog post-side behavioral features and the user-side behavioral features; the confidence level of the blog post corrected features is higher than that of the blog post-side behavioral features, and the confidence level of the user corrected features is higher than that of the user-side behavioral features; a training module, which is used to adjust the model parameters based on the blog post-side behavioral features, user-side behavioral features, blog post corrected features and user corrected features corresponding to each training sample during the iterative training of the initial blog post quality assessment model using multiple training samples, until the blog post quality assessment model converges to obtain a trained blog post quality assessment model; the blog post quality assessment model is used to assess the quality of blog posts.

[0015] In some possible implementations, the sample data preprocessing module is also used to preprocess the blog post information of the sample blog post in each training sample and the user information of the corresponding sample blogger after obtaining multiple training samples, so as to generate blog post short-term features and user information features corresponding to each training sample; the blog post short-term features are used to characterize the interaction of the sample blog post within a preset third time period, the third time period being shorter than the first time period, and the user information features are used to characterize the user portrait of the sample blogger.

[0016] In some possible implementations, a training module is used to determine the labeling information corresponding to each sample blog post based on the blog post-side behavioral features, user-side behavioral features, blog post correction features, and user correction features corresponding to each training sample; the labeling information is used to label the sample blog post as a positive sample or a negative sample; the blog post short-term features and user information features corresponding to each training sample are input into a blog post quality assessment model to obtain a quality assessment result output by the blog post quality assessment model; the quality assessment result includes a first quality score corresponding to the sample blog post, and a prediction result of whether the sample blog post is a positive sample or a negative sample; the first quality score is used to characterize the user's interest in the sample blog post; a comparison result is obtained by comparing the prediction result and the labeling information corresponding to each sample blog post; based on the comparison result of each sample blog post, the blog post quality assessment model is adjusted until the blog post quality assessment model converges.

[0017] In some possible implementations, the training module is used to perform linear fusion processing on the blog post-side behavior features, user-side behavior features, blog post correction features and user correction features corresponding to each training sample to obtain an initial quality score corresponding to the sample blog post in each training sample; the initial quality score is used to characterize the user's interest in the sample blog post without being processed by the blog post quality assessment model; when the initial quality score of the sample blog post meets a preset threshold, the sample blog post is marked as a positive sample; when the initial quality score of the sample blog post does not meet the preset threshold, the sample blog post is marked as a negative sample.

[0018] In some possible implementations, the sample data processing module is used to perform credibility correction on the blog post behavior characteristics according to the Wilson confidence interval to obtain the blog post correction characteristics corresponding to each training sample; and perform credibility correction on the user side behavior characteristics according to the Wilson confidence interval to obtain the user correction characteristics corresponding to each training sample.

[0019] The present application provides a blog post quality assessment device, which includes: a data acquisition module, which is used to obtain multiple exposed historical blog posts and historical bloggers who posted the historical blog posts based on the exposure log of a blog post recommendation system; a data processing module, which is used to input the blog post information of each historical blog post and the user information of its corresponding historical blogger into a trained blog post quality assessment model, and obtain a second quality score corresponding to each historical blog post output by the blog post quality assessment model; the blog post quality assessment model is trained using a training method such as a blog post quality assessment model, and the second quality score is used to characterize the user's interest in the historical blog post; a screening module, which is used to determine a target blog post from multiple exposed historical blog posts based on the second quality score; the target blog post is used for rough blog post sorting or fine blog post sorting.

[0020] In some possible implementations, a screening module is used to calculate a third quality score corresponding to each historical blog post based on a preset attenuation mapping relationship, a preset time period, and a second quality score; the third quality score is used to characterize the user's interest in the historical blog post after removing the influence of time; the attenuation mapping relationship is a mapping relationship between the second quality score and the third quality score; the blog post information of each historical blog post and the third quality score corresponding to each historical blog post are associated and saved in a trained blog quality assessment model; based on the third quality score, a target blog post is determined from multiple exposed historical blog posts.

[0021] In some possible implementations, the screening module is used to screen out multiple recalled blog posts from multiple exposed historical blog posts through multi-link recall; obtain a third quality score corresponding to each recalled blog post in a trained blog post quality assessment model; and select a target blog post from multiple recalled blog posts based on the third quality score corresponding to each recalled blog post.

[0022] The present application provides an electronic device, comprising: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement the method provided by the present application.

[0023] The present application provides a computer storage medium storing executable instructions for implementing a blog post quality assessment model training method or a blog post quality assessment method provided in the present application when the executable instructions are executed by a processor.

[0024] The present application provides a computer program product, including a computer program or instructions, which are used to implement the training method of the blog post quality assessment model or the blog post quality assessment method provided in the present application when the computer program or instructions are executed by a processor.

[0025] This application has the following beneficial effects:

[0026] In the present application, by correcting the blog post side behavior features and the user side behavior features, we obtain the blog post corrected features and user corrected features corresponding to each training sample. The confidence level of the blog post corrected features is higher than that of the blog post side behavior features, and the confidence level of the user corrected features is higher than that of the user side behavior features. This can solve the problem of inconsistent dimensions between sample blog posts, achieve consistency in training samples, and provide a more accurate data processing basis for the blog post recommendation process, thereby achieving more accurate blog post recommendations.

[0027] Furthermore, according to the preset attenuation mapping relationship, the preset validity period and the second quality score, the third quality score corresponding to each historical blog post is calculated, and blog posts with strong timeliness can be corrected to more intuitively represent the quality of historical blog posts. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the architecture of the recommendation system provided in the embodiment of the present application;

[0029] Figure 2 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application;

[0030] Figure 3 It is an optional flow chart of the training method of the blog post quality assessment model provided in the embodiment of the present application;

[0031] Figure 4 It is an optional flow chart of the training method of the blog post quality assessment model and the blog post quality assessment method provided in the embodiment of the present application. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. 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 this application.

[0033] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it can be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0034] If similar descriptions of "first / second" appear in the application documents, the following explanation is added. In the following description, the terms "first / second / third" involved are merely to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first / second / third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0035] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0036] With the popularization of smart devices and the rapid development of Internet technology, modern society has fully entered the era of mobile Internet. As the Web 2.0 model becomes mainstream, Internet users are not only consumers of online content, but also producers and disseminators of online information. The amount of information on the Internet is growing explosively. For example, social platforms such as Weibo can post more than 100 million posts per day. When users are faced with a huge amount of information, actively recommending high-quality information that may suit their interests is an indispensable and even very important part of the entire operation.

[0037] At present, the recommendation system is one of the most effective ways to solve information overload. In the relationship flow distribution scenario, the recommendation system is generally divided into two stages, namely the recall stage and the sorting stage. In the recall stage, the blog posts that users may be interested in are fully mined through multi-link methods such as user profiling, collaborative filtering, graph networks, and approximate search, and the multi-path recalled blog posts are truncated and merged according to time sequence or snake sorting; in the refined sorting stage, the recommendation algorithm based on deep learning makes personalized recommendations for the recalled blog posts to promote users' consumption desire.

[0038] However, by fully considering the user's multiple interests and interest changes, the blog posts recalled through multiple links have serious dimensionality problems in the quality calculation results of the blog posts recalled through each link that measure the user's preference due to different strategies, which affects the accurate recommendation of blog posts.

[0039] In response to the above problems, an embodiment of the present application provides a method for training a blog post quality assessment model to achieve more accurate blog post recommendations.

[0040] The following describes an exemplary application of the electronic device provided by the embodiment of the present application. The electronic device provided by the embodiment of the present application can be implemented as various types of user terminals such as a laptop computer, a tablet computer, a desktop computer, a mobile device (for example, a mobile phone, a wearable smart watch, a dedicated messaging device), and can also be implemented as a server. The following describes an exemplary application when the electronic device is implemented as a server.

[0041] See also Figure 1 , Figure 1 It is a schematic diagram of the architecture of the recommendation system provided in an embodiment of the present application. Electronic devices (electronic device 40-1 and electronic device 40-2 are shown as examples) are connected to server 200 via network 300. Network 300 can be a wide area network or a local area network, or a combination of the two.

[0042] In some possible implementations, user A may publish material information A through terminal 40-1, and user B may publish material information B through terminal 40-2. Material information A and material information B are uploaded to server 200 through network 300. Server 200 may store material information A and material information B in database 500. In order to accurately push materials according to the interests of users, server 200 may push materials that have the highest degree of matching with the interests of users to users after sorting, and the materials pushed to users may be displayed on the graphical interface of electronic device 400 (graphical interface 41-1 and graphical interface 41-2 are shown as examples).

[0043] In some embodiments, the server 200 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, and big data and artificial intelligence platforms. The terminal may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart watch, etc., but is not limited thereto. The terminal and the server may be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiments of the present invention.

[0044] See also Figure 2 , Figure 2 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application, Figure 2 The electronic device 400 shown may be the above-mentioned terminal and / or server 200. The electronic device 400 includes: at least one processor 410, a memory 450, at least one network interface 420 and a user interface 430. The various components in the electronic device 400 are coupled together via a bus system 440. It is understood that the bus system 440 is used to achieve connection and communication between these components. In addition to the data bus, the bus system 440 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the bus system 440 is not shown in FIG. Figure 2 Various buses are labeled as bus system 440 .

[0045] Processor 410 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0046] The user interface 430 includes one or more output devices 431 that enable presentation of media content, including one or more speakers and / or one or more visual display screens. The user interface 430 also includes one or more input devices 432, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0047] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drives, optical disk drives, etc. The memory 450 may optionally include one or more storage devices that are physically remote from the processor 410.

[0048] The memory 450 includes a volatile memory or a non-volatile memory, and may also include both volatile and non-volatile memories. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 450 described in the embodiments of the present application is intended to include any suitable type of memory.

[0049] In some embodiments, memory 450 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplarily described below.

[0050] Operating system 451, including system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks;

[0051] A network communication module 452, for reaching other computing devices via one or more (wired or wireless) network interfaces 420, exemplary network interfaces 420 include: Bluetooth, wireless compatibility certification (WiFi), and universal serial bus (USB), etc.;

[0052] a presentation module 453 for enabling presentation of information via one or more output devices 431 (e.g., display screen, speaker, etc.) associated with the user interface 430 (e.g., a user interface for operating peripherals and displaying content and information);

[0053] The input processing module 454 is used to detect one or more user inputs or interactions from one of the one or more input devices 432 and translate the detected inputs or interactions.

[0054] In some embodiments, the training device for the blog post quality assessment model provided in the embodiments of the present application can be implemented in software. Figure 2 A training device 455 and a blog post quality assessment device 456 of a blog post quality assessment model stored in a memory 450 are shown, which may be software in the form of programs and plug-ins, etc. The training device 455 of the blog post quality assessment model includes the following software modules: a sample data acquisition module 4551, a sample data preprocessing module 4552, a sample data processing module 4553 and a training module 4554; the blog post quality assessment device 456 includes the following software modules: a data acquisition module 4561, a data processing module 4562 and a screening module 4563.

[0055] These modules are logical, so they can be arbitrarily combined or further divided according to the functions to be implemented. The functions of each module will be explained below.

[0056] In other embodiments, the training device of the blog post quality assessment model and the blog post quality assessment device provided in the embodiments of the present application can be implemented in hardware. As an example, the device provided in the embodiments of the present application can be a processor in the form of a hardware decoding processor, which is programmed to execute the training method of the blog post quality assessment model and the blog post quality assessment method provided in the embodiments of the present application. For example, the processor in the form of a hardware decoding processor can adopt one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs) or other electronic components.

[0057] In some embodiments, the recommendation system provided in the embodiments of the present application may be a blog post recommendation system. The following uses the blog post recommendation system as an example to illustrate the training method of the blog post quality assessment model provided in the embodiments of the present application.

[0058] The following will explain the training method of the blog post quality assessment model provided in the embodiment of the present application in combination with the exemplary application and implementation of the electronic device provided in the embodiment of the present application.

[0059] See also Figure 3 , Figure 3 This is an optional flow chart of the training method of the blog post quality assessment model provided in the embodiment of the present application. Figure 3 The steps shown illustrate the training method of the blog post quality assessment model.

[0060] It should be noted that, for the sake of convenience, the training device of the blog post quality assessment model can be referred to as the training device.

[0061] S301, based on the exposure log of the blog recommendation system, a plurality of training samples are obtained, wherein each training sample includes blog information of a sample blog and user information of a corresponding sample blogger; the sample blogger is a user who publishes the sample blog.

[0062] It is understandable that the training device can collect the blog information of the sample blog and the user information of the sample blogger who published the sample blog by accessing the exposure log of the blog recommendation system, and use it as a training sample. In the process of collection, in order to meet the data volume requirements of the training data, the blog information of multiple sample blogs can be collected according to actual needs, and accordingly, the user information of the sample bloggers corresponding to the multiple sample blogs is also collected. A sample blogger can publish only one sample blog or multiple sample blogs, and the embodiments of the present application do not specifically limit this.

[0063] In some embodiments, the acquired blog post information of the plurality of sample blog posts and the user information of the corresponding sample bloggers may be stored in a database of the blog post recommendation system.

[0064] S302, pre-processing the blog post information of each sample blog post in each training sample and the user information of the corresponding sample blogger, and generating blog post-side behavior features and user-side behavior features corresponding to each training sample. The blog post-side behavior features are used to characterize the interaction of the sample blog post in a preset first time period; the user-side behavior features are used to characterize the interaction of the blog posts published by the sample blogger in a preset second time period.

[0065] It is understandable that the training device can obtain the blog information of multiple sample blogs and the user information of the corresponding sample bloggers in the database of the blog recommendation system, and then perform data preprocessing on the above blog information and the above user information to obtain blog post side behavior characteristics and user side behavior characteristics. Among them, each sample blog corresponds to a blog post side behavior feature, and each sample blogger corresponds to a user side behavior feature; since each sample blogger corresponds to a sample blogger, accordingly, each sample blog also corresponds to a user side behavior feature.

[0066] In some embodiments, the blog post-side behavior characteristics are used to characterize the interaction of the sample blog posts within a preset first time period; the user-side behavior characteristics are used to characterize the interaction of the blog posts published by the sample blogger within a preset second time period.

[0067] In one example, taking the sample blog post as blog post A, the sample blogger as blogger A, the preset first time period of three days, and the preset second time period of one month as an example, the blog post-side behavior characteristics of blog post A represent the exposure, clicks, favorites, forwarding, comments, likes, and viewing time of blog post A in the past three days. The user-side behavior characteristics of blogger A who published blog post A represent the exposure, clicks, favorites, forwarding, comments, likes, and viewing time of historical blog posts published by blogger A in the past month. It should be noted that in the above example, the preset first time period of three days and the preset second time period of one month are only one example, and the appropriate time length can be selected according to actual needs. The embodiment of the present application does not specifically limit this.

[0068] In some possible implementations, after obtaining multiple training samples, the training method of the blog quality assessment model may further include: preprocessing the blog information of the sample blog in each training sample and the user information of the corresponding sample blogger, and generating blog short-term features and user information features corresponding to each training sample. The blog short-term features are used to characterize the interaction of the sample blog in a preset third time period, which is shorter than the first time period, and the user information features are used to characterize the user portrait of the sample blogger.

[0069] It is understandable that the training device can obtain the blog information of multiple sample blogs and the user information of the corresponding sample bloggers in the database of the blog recommendation system, and then perform data preprocessing on the above blog information and the above user information to obtain blog short-term features and user information features. Among them, each sample blog corresponds to a blog short-term feature and a user information feature.

[0070] In one example, taking the sample blog post as blog post B, the sample blogger as blogger B, and the preset third time period as one hour, the short-term blog post features of blog post B represent the interactive conditions such as exposure, clicks, favorites, forwarding, comments, likes, and viewing time of blog post B in the past hour. The user information features of blogger B who published blog post B represent the user portraits such as the number of fans, interest preferences, and active time periods of blogger B. It should be noted that the short-term blog post features target the short-term features of the sample blog post, and the blog post side behavior features target the long-term features of the sample blog post. Therefore, the third time period corresponding to the short-term blog post features is shorter than the first time period corresponding to the blog post side behavior features.

[0071] S303, by correcting the blog post side behavior features and the user side behavior features, obtain the blog post corrected features and the user corrected features corresponding to each training sample. The confidence level of the blog post corrected features is higher than that of the blog post side behavior features, and the confidence level of the user corrected features is higher than that of the user side behavior features.

[0072] It is understandable that in the blog recommendation system, the interest fields corresponding to sample blogs of different contents will also be different. For example, the interest field corresponding to sample blog A is "entertainment stars", the interest field corresponding to sample blog B is "finance", and the interest field corresponding to sample blog C is "pets". Compared with the interest field of "entertainment stars", "finance" and "pets" are more niche and target smaller audiences (i.e., other users in the blog recommendation system except sample bloggers). In this case, the interactive data such as exposure, clicks, collections, forwarding, comments, likes, and viewing time corresponding to "finance" and "pets" will be significantly smaller than the interactive data corresponding to "entertainment stars", which will in turn affect the blog recommendation system's recommendation of sample blogs B and sample blogs C. Therefore, it is necessary to correct the blog post-side behavioral features and user-side behavioral features to obtain blog post correction features and user correction features with a higher confidence level.

[0073] In some possible implementations, the above S303 may include: performing credibility correction on the blog post side behavior characteristics according to the Wilson confidence interval to obtain the blog post correction characteristics corresponding to each training sample; performing credibility correction on the user side behavior characteristics according to the Wilson confidence interval to obtain the user correction characteristics corresponding to each training sample.

[0074] It can be understood that by using the preset Wilson confidence interval formula, the credibility of the blog post side behavior characteristics and the user side behavior characteristics can be corrected to obtain the blog post corrected characteristics and the user corrected characteristics.

[0075] In an embodiment of the present application, the Wilson confidence interval is used to perform credibility correction on the blog post side behavioral characteristics and the user side behavioral characteristics, which can provide better confidence interval estimates for small sample data (i.e., sample blog posts and sample bloggers in niche interest areas), and effectively alleviate the problem of dilution of the characteristics of sample blog posts and sample bloggers in niche interest areas.

[0076] In some embodiments, the Wilson confidence interval is used to modify the blog post behavior feature to obtain the blog post modified feature expression (1) as follows:

[0077]

[0078] In the formula, is the post correction feature of the ith blog post (i.e., the post correction feature corresponding to the ith sample blog post), p i represents the behavior rate corresponding to the behavior feature of the i-th blog post, n i represents the exposure corresponding to the behavioral characteristics of the i-th blog post, and z is the confidence level statistic under the preset normal distribution; the behavioral rate includes: click rate, viewing rate, interaction rate, etc.

[0079] In some embodiments, z in expression (1) can be obtained by using a preset standard normal distribution table. For example, at a confidence level of 95%, the value of z is 1.96.

[0080] In some embodiments, the expression (2) for calculating the behavior rate in the i-th blog post side behavior feature is as follows:

[0081] p i =x i / n i (2)

[0082] In the formula, x i Indicates the number of behaviors corresponding to the behavioral features of the i-th blog post; the number of behaviors includes: number of clicks, number of views, number of interactions, etc.

[0083] In an embodiment of the present application, when a sample blog post has just been published and has not yet been exposed and interacted with, in order to evaluate the quality of the sample blog post, user-side behavioral features and user correction features can be used to evaluate the quality of the newly published sample blog post, so as to effectively alleviate the cold start problem caused by the lack of features on the blog post side (i.e., blog post-side behavioral features and blog post correction features).

[0084] S304, in the process of iteratively training the initial blog post quality assessment model using multiple training samples, adjusting the model parameters based on the blog post behavior feature, user behavior feature, blog post correction feature and user correction feature corresponding to each training sample until the blog post quality assessment model converges, thereby obtaining a trained blog post quality assessment model. The blog post quality assessment model is used to assess the quality of blog posts.

[0085] It can be understood that after determining the blog post side behavior characteristics, user side behavior characteristics, blog post correction characteristics and user correction characteristics, a blog post quality assessment model can be trained based on the blog post side behavior characteristics, user side behavior characteristics, blog post correction characteristics and user correction characteristics.

[0086] In some possible implementations, the above S304 may include: determining the labeling information corresponding to each sample blog post based on the blog post-side behavioral features, user-side behavioral features, blog post correction features and user correction features corresponding to each training sample; the labeling information is used to mark the sample blog post as a positive sample or a negative sample; inputting the blog post short-term features and user information features corresponding to each training sample into the blog post quality assessment model to obtain a quality assessment result output by the blog post quality assessment model; the quality assessment result includes a first quality score corresponding to the sample blog post, and a prediction result of whether the sample blog post is a positive sample or a negative sample; the first quality score is used to characterize the user's interest in the sample blog post; a comparison result is obtained by comparing the prediction result and the labeling information corresponding to each sample blog post; based on the comparison result of each sample blog post, the blog post quality assessment model is adjusted until the blog post quality assessment model converges.

[0087] It can be understood that each sample blog post can be scored based on the blog post behavior characteristics, user-side behavior characteristics, blog post correction characteristics and user correction characteristics, and then the sample blog posts can be divided into positive samples and negative samples according to the scoring results (i.e., the initial quality score); then, the blog post short-term characteristics and user information characteristics corresponding to each training sample are input into the blog post quality assessment model to obtain the quality assessment results output by the blog post quality model. The quality assessment results include the prediction results of the sample blog posts, and the comparison results are obtained by comparing the prediction results and the marking information. Based on the comparison results, the parameters of the blog post quality assessment model are adjusted until a trained blog post quality assessment model is obtained.

[0088] In some possible implementations, the above-mentioned determination of labeling information corresponding to each sample blog post based on the blog post-side behavior features, user-side behavior features, blog post correction features and user correction features corresponding to each training sample may include: performing linear fusion processing on the blog post-side behavior features, user-side behavior features, blog post correction features and user correction features corresponding to each training sample to obtain an initial quality score corresponding to the sample blog post in each training sample; wherein the initial quality score is used to characterize the user's interest in the sample blog post without being processed by the blog post quality assessment model; when the initial quality score of the sample blog post meets a preset threshold, the sample blog post is marked as a positive sample; when the initial quality score of the sample blog post does not meet the preset threshold, the sample blog post is marked as a negative sample.

[0089] It can be understood that by linear fusion processing, assigning corresponding weights to each feature and then performing sum calculation, each sample blog post can be scored to obtain a scoring result (i.e., an initial quality score); the initial quality scores are sorted, and multiple sample blog posts are divided into positive samples and negative samples according to the sorting results.

[0090] In some embodiments, the linear fusion process of the blog post behavior feature, the user behavior feature, the blog post correction feature and the user correction feature is performed as follows:

[0091] S i =αM wi +βM hi +γA hi +δA wi (3)

[0092] In the formula, S i represents the initial quality score of the i-th sample blog post, M wi represents the revised feature of the i-th blog post, M hi represents the behavior characteristics of the i-th blog post, A hi represents the behavior characteristics of the i-th user side, A wi represents the user correction feature, α is the linear fusion weight coefficient corresponding to the blog post correction feature, β is the linear fusion weight coefficient corresponding to the blog post side behavior feature, γ is the linear fusion weight coefficient corresponding to the user side behavior feature, and δ is the linear fusion weight coefficient corresponding to the user correction feature.

[0093] In some embodiments, α, β, γ and δ are preset linear fusion weight coefficients with a value interval of [0.1]. The values ​​of the linear fusion weight coefficients can be adjusted according to actual needs. For example, α, β, γ and δ can be 0.2, 0.5, 0.2, and 0.1 respectively. The embodiments of the present application do not specifically limit the values ​​of the linear fusion weight coefficients.

[0094] In some embodiments, the above-mentioned marking the sample blog post as a positive sample when the initial quality score of the sample blog post meets a preset threshold; marking the sample blog post as a negative sample when the initial quality score of the sample blog post does not meet the preset threshold, can include: sorting multiple initial quality scores according to equidistant bucketing and percentile demarcation; marking the sample blog post corresponding to the initial quality score greater than the 80th percentile as a positive sample; marking the sample blog post corresponding to the initial quality score less than the 80th percentile as a negative sample.

[0095] In some embodiments, sample markings may also include difficult positive samples and difficult negative samples; then, the above-mentioned marking of the sample blog post as a positive sample when the initial quality score of the sample blog post meets the preset threshold; marking the sample blog post as a negative sample when the initial quality score of the sample blog post does not meet the preset threshold may include: sorting multiple initial quality scores according to the method of equidistant bucketing and percentile demarcation; marking the sample blog posts corresponding to the initial quality scores greater than the 80th percentile as positive samples; randomly selecting 15% of the sample blog posts between the 70th and 80th percentiles and marking them as difficult positive samples, and marking the remaining 85% of the sample blog posts between the 70th and 80th percentiles as difficult negative samples; marking the sample blog posts corresponding to the initial quality scores less than the 70th percentile as negative samples.

[0096] In the embodiment of the present application, by constructing difficult positive samples and difficult negative samples, the learning difficulty of the microblog quality assessment model can be effectively increased, overfitting can be prevented, and the generalization ability of the microblog quality assessment model can be greatly improved.

[0097] In some embodiments, after determining the positive samples and negative samples, the short-term features of each sample blog post and the user information features corresponding to the blog post can be input into the blog post quality assessment model to obtain the quality assessment result output by the blog post quality assessment model, and the quality assessment result includes the first quality score corresponding to the sample blog post and the prediction result of whether the sample blog post is a positive sample or a negative sample; the first quality score is used to characterize the user's interest in the sample blog post. After obtaining the quality assessment result, it is determined whether the prediction result of the sample blog post is the same as the tag information of the sample blog post to obtain a comparison result; based on the comparison result, the parameters of the blog post quality assessment model are adjusted until the blog post quality assessment model converges.

[0098] In an embodiment of the present application, positive samples and negative samples are determined based on blog post behavior features, user side behavior features, blog post correction features, and user correction features. In the process of using blog post short-term features and user information features as training data for a blog post quality assessment model, since blog post short-term features and user information features do not participate in the division of positive samples and negative samples, the problem of feature leakage in model training will not occur.

[0099] In some embodiments, the blog post quality assessment model can select a recommendation model, such as the DeepFM (deep factorization machine) model; the DeepFM model not only retains the strong memory ability of the traditional model, strengthens the combination of some features of the shallow network, but also integrates the advantages of the generalization ability of the deep learning model. In addition, the structure of the DeepFM model is not complicated, and it is relatively easy to implement training and online in engineering, which is suitable for wide application in recommendation scenarios.

[0100] In some embodiments, the blog post quality assessment model uses the DeepFM model as an example; the DeepFM model includes an FM layer, a Deep layer, a fully connected layer, and an activation function layer. During the training process, the output of the FM layer and the output of the Deep layer of the DeepFM model can be spliced ​​and input into the fully connected layer and the activation function layer, wherein the activation function used in the activation function layer is a sigmoid activation function; and the activation function layer outputs the prediction result of the model.

[0101] In the embodiment of the present application, the DeepFM model is selected as an example of a blog post quality assessment model; the output of the FM layer and the output of the Deep layer are spliced ​​into the fully connected layer and the activation function layer, which can, to a certain extent, alleviate the problem of the magnitude explosion caused by the addition method, causing most of the output values ​​(i.e., logis) to fall into the sigmoid saturation area.

[0102] In some embodiments, the training method of the blog post quality assessment model may include: improving the loss function of the blog post quality assessment model by weighting the loss of positive samples.

[0103] In one example, the loss function expression (4) of the improved blog post quality assessment model is as follows:

[0104]

[0105] In the formula, p is the probability value of the sample blog post predicted by the blog post quality assessment model as a positive sample. p can also represent the first quality score of the sample blog post. ω is the preset positive sample weighting coefficient. a is the preset positive and negative sample importance factor. b is the preset difficult sample weighting factor. logis is the output value of the activation function layer. The value range of a is [0,1], the preferred value is 0.9615, and the preferred value of b is 2.

[0106] The present application provides a blog post quality assessment method. After obtaining a trained blog post quality assessment model, the above-mentioned blog post quality assessment method may include: based on the exposure log of the blog post recommendation system, obtaining multiple exposed historical blog posts and historical bloggers who published the historical blog posts; inputting the blog post information of each historical blog post and the user information of its corresponding historical blogger into the trained blog post quality assessment model to obtain a second quality score corresponding to each historical blog post output by the blog post quality assessment model; the blog post quality assessment model is trained using the training method of the above-mentioned blog post quality assessment model, and the second quality score is used to characterize the user's interest in the historical blog post; based on the second quality score, determining a target blog post from multiple exposed historical blog posts; the target blog post is used for rough blog post ranking or fine blog post ranking.

[0107] It is understandable that after obtaining the trained blog post quality assessment model, the trained blog post quality assessment model can be put into online use, that is, the blog post information of the historical blog post and the user information of the historical blog post corresponding to the historical blog post are input into the trained blog post quality assessment model to obtain the output second quality score. Based on the second quality score of each historical blog post, multiple historical blog posts are sorted, and then the historical blog posts corresponding to the second quality scores that meet the preset threshold are screened out as target blog posts, and the target blog posts can enter the next stage of the blog post recommendation process, that is, the rough blog post ranking or the fine blog post ranking.

[0108] In some possible implementations, the above-mentioned determining a target blog post from a plurality of exposed historical blog posts based on the second quality score includes: calculating a third quality score corresponding to each historical blog post according to a preset attenuation mapping relationship, a preset time period, and the second quality score; the third quality score is used to characterize the user's interest in the historical blog post after removing the influence of time; the attenuation mapping relationship is a mapping relationship between the second quality score and the third quality score; associating the blog post information of each historical blog post and the third quality score corresponding to each historical blog post and saving them in a trained blog quality assessment model; and determining a target blog post from a plurality of exposed historical blog posts based on the third quality score.

[0109] It is understandable that during online use, considering the timeliness of blog posts, the quality scores of historical blog posts (i.e., the second quality scores) output by the trained blog post quality assessment model can be subjected to time decay processing to remove the impact of the timeliness of the blog posts on the recommendation.

[0110] In one example, the exposed historical blog posts include blog post D, blog post E, and blog post F, and the historical blogger who published blog post D is blogger D, the historical blogger who published blog post E is blogger E, and the historical blogger who published blog post F is blogger F. In the exposure log of the blog post recommendation system, the blog post information of blog post D, blog post E, and blog post F and the user information of blogger D, blogger E, and blogger F are obtained. The blog post information of each historical blog post and the user information of the historical blogger corresponding to the historical blog post are input into the trained blog post quality assessment model for quality assessment, and the second quality scores corresponding to blog post D, blog post E, and blog post F are obtained respectively. Through the preset attenuation mapping process, the second quality score of each historical blog post is subjected to time-dependent attenuation processing, and the third quality scores corresponding to blog post D, blog post E, and blog post F are obtained respectively. Then, the blog post information and the third quality score of each historical blog post are associated and saved in the trained blog post quality assessment model and database; then, based on the third quality score of each historical blog post, multiple historical blog posts are sorted to screen out target blog posts.

[0111] In some embodiments, the expression (5) of the above attenuation mapping relationship is as follows:

[0112]

[0113] In the formula, c is the decay rate, the time decay interval is [f, d], f is the time when the historical blog post was published, d is the time when the popularity of the historical blog post disappears, m is the preset blog post validity period, l is the decay change value, g is the time when the historical blog post has been published, p is the quality score of the historical blog post output by the trained blog post quality assessment model (i.e., the second quality score), S d is the third mass fraction. The time decay interval is preferably [0.1, 1].

[0114] In the embodiment of the present application, the third quality score obtained through time decay processing can remove the influence of the timeliness of the blog post.

[0115] In some embodiments, every time a newly exposed historical blog post appears, the newly exposed historical blog post is input into the trained blog post quality assessment model for quality assessment, that is, the third quality score of the historical blog post is calculated. Each historical blog post will be exposed multiple times, and the third quality score of the historical blog post can be calculated once each time it is exposed, and the latest third quality score replaces the third quality score stored last time in the blog post quality assessment model, that is, the third quality score of the historical blog post is updated once each time the historical blog post is exposed.

[0116] In some embodiments, the latest third quality score of the historical blog post is obtained through real-time updating; then, based on the third quality score, personalized strategic support can be formulated for the historical blog post in vertical fields (such as beauty, sports and fitness, etc.) according to the level of the historical blogger and the type of the historical blog post, so as to enhance the exposure and user coverage of the historical blog post in the relationship flow.

[0117] In some possible implementations, the above-mentioned determining a target blog post from multiple exposed historical blog posts based on the third quality score may include: screening out multiple recalled blog posts from multiple exposed historical blog posts through multi-link recall; obtaining the third quality score corresponding to each recalled blog post in a trained blog post quality assessment model; and selecting a target blog post from multiple recalled blog posts based on the third quality score corresponding to each recalled blog post.

[0118] It can be understood that in the process of using the trained blog quality assessment model online, N recalled blog posts that meet the preset requirements can be preliminarily screened out from the M historical blog posts in the blog recommendation system by multi-link recall, where N < M. Then, the third quality score corresponding to each recalled blog post is obtained in the database of the trained blog quality assessment model. Finally, based on the N third quality scores, the N recalled blog posts are sorted, and the first n recalled blog posts corresponding to the third quality scores that meet the preset threshold are selected as target blog posts. The n target blog posts will enter the rough sorting stage or fine sorting stage of the blog recommendation link for further screening.

[0119] Figure 4 This is an optional flow chart of the training method of the blog post quality assessment model and the blog post quality assessment method provided in the embodiment of the present application. Figure 4 The steps shown illustrate the training method of the blog post quality assessment model.

[0120] Below, we take the blog post quality assessment model as the DeepFM model as an example.

[0121] S401, the training device obtains blog post information of multiple sample blog posts and user information of their corresponding sample bloggers from the exposure log of the blog post recommendation system, and stores the blog post information and user information in a database of the blog post recommendation system.

[0122] S402, the training device obtains the blog post information of the sample blog post and the user information of the sample blogger in the database, and performs data preprocessing on the above blog post information and user information to obtain blog post side behavior characteristics, user side behavior characteristics, blog post short-term characteristics and user information characteristics; then, the blog post side behavior characteristics and user side behavior characteristics are corrected to obtain blog post corrected characteristics and user corrected characteristics corresponding to each sample blog post.

[0123] S403, the training device performs linear fusion processing on the blog post side behavior features, user side behavior features, blog post correction features and user correction features to obtain an initial quality score corresponding to each sample blog post; and marks multiple sample blog posts as positive samples and negative samples based on the initial quality scores.

[0124] S404, the training device iteratively trains the blog post quality assessment model based on the short-term features of the blog post, the user information features, and the labels of the positive samples or negative samples corresponding to each sample blog post to obtain a trained blog post quality assessment model; then, the blog post quality assessment device uses the trained blog post quality assessment model to perform quality assessment on each exposed historical blog post to obtain a third quality score corresponding to each historical blog post.

[0125] S405, the blog post quality assessment device obtains a recalled blog post from multiple historical blog posts by a multi-link recall method, and obtains a third quality score of the recalled blog post from a database of a trained blog post quality assessment model.

[0126] S406, the blog post quality assessment device sorts the recalled blog posts based on the third quality scores of the recalled blog posts to obtain target blog posts, and sends the target blog posts to the refined sorting stage.

[0127] At this point, the recommendation system has completed the training method of the above-mentioned blog post quality assessment model and the blog post quality assessment method.

[0128] In an embodiment of the present application, by correcting the blog post side behavior features and the user side behavior features, the corrected blog post features and the user corrected features corresponding to each training sample are obtained. The confidence level of the blog post corrected features is higher than that of the blog post side behavior features, and the confidence level of the user corrected features is higher than that of the user side behavior features. This can solve the problem of inconsistent dimensions between sample blog posts, achieve consistency in training samples, and provide a more accurate data processing basis for the blog post recommendation process, thereby achieving more accurate blog post recommendations.

[0129] Furthermore, according to the preset attenuation mapping relationship, the preset validity period and the second quality score, the third quality score corresponding to each historical blog post is calculated, and blog posts with strong timeliness can be corrected to more intuitively represent the quality of historical blog posts.

[0130] The following is an exemplary structure of the training device 455 of the blog post quality evaluation model provided in the embodiment of the present application implemented as a software module. In some embodiments, for example Figure 2As shown, the software modules in the training device 455 of the blog post quality assessment model stored in the memory 450 may include: a sample data acquisition module 4551, which is used to obtain multiple training samples based on the exposure log of the blog post recommendation system; wherein each training sample includes the blog post information of the sample blog post and the user information of the corresponding sample blogger, and the sample blogger is the user who publishes the sample blog post; a sample data preprocessing module 4552, which is used to preprocess the blog post information of the sample blog post in each training sample and the user information of the corresponding sample blogger, and generate blog post-side behavior characteristics and user-side behavior characteristics corresponding to each training sample; wherein the blog post-side behavior characteristics are used to characterize the interaction of the sample blog post within a preset first time period; the user-side behavior characteristics are used to characterize the interaction of the sample blogger within a preset The interaction of the blog posts published in the second time period; the sample data processing module 4553 is used to obtain the blog post correction features and user correction features corresponding to each training sample by correcting the blog post side behavior features and the user side behavior features; the confidence level of the blog post correction features is higher than that of the blog post side behavior features, and the confidence level of the user correction features is higher than that of the user side behavior features; the training module 4554 is used to adjust the model parameters based on the blog post side behavior features, user side behavior features, blog post correction features and user correction features corresponding to each training sample in the process of iteratively training the initial blog post quality assessment model using multiple training samples, until the blog post quality assessment model converges to obtain a trained blog post quality assessment model; the blog post quality assessment model is used to perform quality assessment on blog posts.

[0131] In some possible implementations, the sample data preprocessing module 4552 is also used to preprocess the blog post information of the sample blog post in each training sample and the user information of the corresponding sample blogger after obtaining multiple training samples, so as to generate blog post short-term features and user information features corresponding to each training sample; the blog post short-term features are used to characterize the interaction of the sample blog post within a preset third time period, where the third time period is shorter than the first time period, and the user information features are used to characterize the user portrait of the sample blogger.

[0132] In some possible implementations, training module 4554 is used to determine the labeling information corresponding to each sample blog post based on the blog post-side behavioral features, user-side behavioral features, blog post correction features and user correction features corresponding to each training sample; the labeling information is used to mark the sample blog post as a positive sample or a negative sample; the blog post short-term features and user information features corresponding to each training sample are input into the blog post quality assessment model to obtain a quality assessment result output by the blog post quality assessment model; the quality assessment result includes a first quality score corresponding to the sample blog post and a prediction result of whether the sample blog post is a positive sample or a negative sample; the first quality score is used to characterize the user's interest in the sample blog post; a comparison result is obtained by comparing the prediction result and labeling information corresponding to each sample blog post; based on the comparison result of each sample blog post, the blog post quality assessment model is adjusted until the blog post quality assessment model converges.

[0133] In some possible implementations, the training module 4554 is used to perform linear fusion processing on the blog post-side behavior features, user-side behavior features, blog post correction features and user correction features corresponding to each training sample to obtain an initial quality score corresponding to the sample blog post in each training sample; the initial quality score is used to characterize the user's interest in the sample blog post without being processed by the blog post quality assessment model; when the initial quality score of the sample blog post meets a preset threshold, the sample blog post is marked as a positive sample; when the initial quality score of the sample blog post does not meet the preset threshold, the sample blog post is marked as a negative sample.

[0134] In some possible implementations, the sample data processing module 4553 is used to perform credibility correction on the blog post side behavioral characteristics according to the Wilson confidence interval to obtain the blog post correction characteristics corresponding to each training sample; and perform credibility correction on the user side behavioral characteristics according to the Wilson confidence interval to obtain the user correction characteristics corresponding to each training sample.

[0135] The present application provides a blog post quality assessment device 456, which includes: a data acquisition module 4561, which is used to obtain multiple exposed historical blog posts and historical bloggers who posted historical blog posts based on the exposure log of the blog post recommendation system; a data processing module 4562, which is used to input the blog post information of each historical blog post and the user information of its corresponding historical blogger into a trained blog post quality assessment model, and obtain a second quality score corresponding to each historical blog post output by the blog post quality assessment model; the blog post quality assessment model is trained using the training method of the blog post quality assessment model as described above, and the second quality score is used to characterize the user's interest in the historical blog post; a screening module 4563, which is used to determine a target blog post from multiple exposed historical blog posts based on the second quality score; the target blog post is used for rough blog post sorting or fine blog post sorting.

[0136] In some possible implementations, screening module 4563 is used to calculate a third quality score corresponding to each historical blog post based on a preset attenuation mapping relationship, a preset time period, and a second quality score; the third quality score is used to characterize the user's interest in the historical blog post after removing the influence of time; the attenuation mapping relationship is a mapping relationship between the second quality score and the third quality score; the blog post information of each historical blog post and the third quality score corresponding to each historical blog post are associated and saved in a trained blog quality assessment model; based on the third quality score, a target blog post is determined from multiple exposed historical blog posts.

[0137] In some possible implementations, the screening module 4563 is used to screen out multiple recalled blog posts from multiple exposed historical blog posts through multi-link recall; obtain the third quality score corresponding to each recalled blog post in the trained blog post quality assessment model; and select the target blog post from the multiple recalled blog posts based on the third quality score corresponding to each recalled blog post.

[0138] The embodiment of the present application provides a computer program product or a computer program, which includes computer instructions, and the computer instructions are stored in a computer storage medium. The processor of the computer device reads the computer instructions from the computer storage medium, and the processor executes the computer instructions, so that the computer device executes the training method of the blog post quality assessment model or the blog post quality assessment method described in the embodiment of the present application.

[0139] The present application embodiment provides a computer storage medium storing executable instructions, wherein the executable instructions are stored. When the executable instructions are executed by a processor, the processor will be caused to execute the training method of the blog post quality assessment model or the blog post quality assessment method provided in the present application embodiment, for example Figure 3 The training method of the blog post quality assessment model and the blog post quality assessment method are shown.

[0140] In some embodiments, the computer storage medium may be a memory such as FRAM, ROM, PROM, EPRO M, EEPROM, flash memory, magnetic surface storage, optical disk, or CD-ROM; or various devices including one or any combination of the above memories.

[0141] In some embodiments, executable instructions may be in the form of a program, software, software module, script or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine or other unit suitable for use in a computing environment.

[0142] As an example, executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a hypertext markup language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files storing one or more modules, subroutines, or code portions).

[0143] By way of example, executable instructions may be deployed to be executed on one computing device, or on multiple computing devices located at one site, or on multiple computing devices distributed across multiple sites and interconnected by a communication network.

[0144] The above are only embodiments of the present application and are not intended to limit the protection scope of the present application. Any modifications, equivalent substitutions and improvements made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. A method for training a blog post quality assessment model, characterized in that: The method comprises: Based on the exposure log of the blog recommendation system, multiple training samples are obtained; wherein each training sample includes blog post information of a sample blog post and user information of a corresponding sample blogger, and the sample blogger is a user who publishes the sample blog post; Preprocess the blog post information of each sample blog post in each training sample and the user information of the corresponding sample blogger to generate blog post-side behavior features and user-side behavior features corresponding to each training sample; wherein the blog post-side behavior features are used to characterize the interaction of the sample blog post in a preset first time period; and the user-side behavior features are used to characterize the interaction of the blog posts published by the sample blogger in a preset second time period; By performing correction processing on the blog post side behavior feature and the user side behavior feature, a blog post correction feature and a user correction feature corresponding to each training sample are obtained; the confidence level of the blog post correction feature is higher than that of the blog post side behavior feature, and the confidence level of the user correction feature is higher than that of the user side behavior feature; In the process of iteratively training the initial blog post quality assessment model using the multiple training samples, the model parameters are adjusted based on the blog post side behavior characteristics, user side behavior characteristics, blog post correction characteristics and user correction characteristics corresponding to each training sample until the blog post quality assessment model converges to obtain a trained blog post quality assessment model; the blog post quality assessment model is used to perform quality assessment on blogs.

2. The method according to claim 1, characterized in that The method further comprises: After obtaining multiple training samples, the blog post information of the sample blog posts in each training sample and the user information of the corresponding sample bloggers are preprocessed to generate blog post short-term features and user information features corresponding to each training sample; the blog post short-term features are used to characterize the interaction of the sample blog posts within a preset third time period, and the third time period is shorter than the first time period; the user information features are used to characterize the user portrait of the sample blogger.

3. The method according to claim 2, characterized in that In the process of iteratively training the initial blog post quality assessment model using the multiple training samples, adjusting the model parameters based on the blog post behavior feature, user behavior feature, blog post correction feature and user correction feature corresponding to each training sample until the blog post quality assessment model converges, thereby obtaining a trained blog post quality assessment model, including: Based on the blog post-side behavior features, user-side behavior features, blog post correction features, and user correction features corresponding to each training sample, determining the labeling information corresponding to each sample blog post; the labeling information is used to label the sample blog post as a positive sample or a negative sample; Inputting the blog post short-term features and user information features corresponding to each training sample into the blog post quality assessment model to obtain a quality assessment result output by the blog post quality assessment model; the quality assessment result includes a first quality score corresponding to the sample blog post and a prediction result of whether the sample blog post is a positive sample or a negative sample; the first quality score is used to characterize the user's interest in the sample blog post; By comparing the prediction results and label information corresponding to each sample blog post, the comparison results are obtained; Based on the comparison result of each sample blog post, the blog post quality assessment model is adjusted until the blog post quality assessment model converges.

4. The method according to claim 3, characterized in that The determining of the tag information corresponding to each sample blog post based on the blog post-side behavior feature, user-side behavior feature, blog post correction feature and user correction feature corresponding to each training sample includes: Linear fusion processing is performed on the blog post behavior features, user behavior features, blog post correction features and user correction features corresponding to each training sample to obtain an initial quality score corresponding to the sample blog post in each training sample; the initial quality score is used to represent the user's interest in the sample blog post without being processed by the blog post quality assessment model; When the initial quality score of the sample blog post meets a preset threshold, marking the sample blog post as a positive sample; When the initial quality score of the sample blog post does not meet the preset threshold, the sample blog post is marked as a negative sample.

5. The method according to claim 1, characterized in that: The correcting process of the blog post behavior feature and the user behavior feature to obtain the blog post corrected feature and the user corrected feature corresponding to each training sample includes: According to the Wilson confidence interval, the credibility of the blog post behavior feature is corrected to obtain the blog post correction feature corresponding to each training sample; According to the Wilson confidence interval, the user-side behavior feature is corrected for credibility to obtain a user-corrected feature corresponding to each training sample.

6. A method for evaluating the quality of a blog post, characterized in that: The method comprises: Based on the exposure log of the blog recommendation system, multiple exposed historical blog posts and historical bloggers who published the historical blog posts are obtained; Inputting the blog post information of each historical blog post and the user information of its corresponding historical blogger into the trained blog post quality assessment model to obtain a second quality score corresponding to each historical blog post output by the blog post quality assessment model; the blog post quality assessment model is trained by the method according to any one of claims 1 to 5, and the second quality score is used to represent the user's interest in the historical blog post; Based on the second quality score, a target blog post is determined from the multiple exposed historical blog posts; the target blog post is used for rough blog post ranking or fine blog post ranking.

7. The method according to claim 6, characterized in that The step of determining a target blog post from the plurality of exposed historical blog posts based on the second quality score includes: Calculate a third quality score corresponding to each historical blog post according to a preset attenuation mapping relationship, a preset time limit, and the second quality score; the third quality score is used to characterize the user's interest in the historical blog post after removing the time limit effect; the attenuation mapping relationship is a mapping relationship between the second quality score and the third quality score; Associating the blog post information of each historical blog post and the third quality score corresponding to each historical blog post and saving them to the trained blog post quality assessment model; Based on the third quality score, the target blog post is determined from the plurality of exposed historical blog posts.

8. The method according to claim 7, characterized in that The step of determining the target blog post from the plurality of exposed historical blog posts based on the third quality score includes: Through multi-link recall, multiple recalled blog posts are selected from the multiple exposed historical blog posts; In the trained blog post quality assessment model, obtaining a third quality score corresponding to each recalled blog post; The target blog post is selected from the multiple recalled blog posts based on the third quality score corresponding to each recalled blog post.

9. A blog post quality assessment device, characterized in that: The device comprises: A data collection module, for obtaining a plurality of exposed historical blog posts and historical bloggers who published the historical blog posts based on the exposure log of the blog post recommendation system; A data processing module, for inputting the blog post information of each historical blog post and the user information of the corresponding historical blogger into a trained blog post quality assessment model, and obtaining a second quality score corresponding to each historical blog post output by the blog post quality assessment model; the blog post quality assessment model is trained by the method according to any one of claims 1 to 5, and the second quality score is used to represent the user's interest in the historical blog post; A screening module is used to determine a target blog post from the plurality of exposed historical blog posts based on the second quality score; the target blog post is used for rough blog post ranking or fine blog post ranking.

10. An electronic device, characterized in that: The electronic device comprises: A memory for storing executable instructions; A processor, configured to implement the method of any one of claims 1 to 5 or the method of any one of claims 6 to 8 when executing the executable instructions or computer programs stored in the memory.

Images