Request processing method, training method and apparatus for browsing information retrieval model

Abstract

This disclosure relates to a request processing method, a training method, apparatus, computer device, and storage medium for a browsing information retrieval model, belonging to the field of computer technology. The method includes: responding to a target account's access request for interactive information about content items; acquiring account characteristics of the target account, global information characteristics of the interactive information, and triggering scenario characteristics of the access request; acquiring target browsing information of the target account based on the account characteristics, the global information characteristics, and the triggering scenario characteristics; and, if the target browsing information meets target conditions, sending a recommendation request for the target account to a recommendation model. This disclosure can reduce the number of requests to the recommendation model during the ranking stage under conditions of limited computing resources.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to a request processing method, a training method for a browsing information retrieval model, an apparatus, a computer device, and a storage medium. Background Technology

[0002] With the development of computer technology and the diversification of terminal functions, users can watch short videos anytime and anywhere on their terminals. When browsing video content, users usually also have a need to view interactive information.

[0003] Taking interactive information as video comments as an example, when a terminal displays video comments in the comment section, it can make personalized recommendations for different users. When the server recommends video comments to the terminal, it usually uses some complex deep neural networks as recommendation models to rank the candidate video comments. This ranking process consumes a lot of computing resources. Therefore, how to reduce the number of requests to the recommendation model used for ranking under the condition of limited computing resources has become an urgent problem to be solved. Summary of the Invention

[0004] This disclosure provides a request processing method, a training method for a browsing information retrieval model, an apparatus, a computer device, and a storage medium to reduce the number of requests to the recommendation model during the ranking stage under conditions of limited computing resources. The technical solution of this disclosure is as follows:

[0005] According to one aspect of the embodiments of this disclosure, a request processing method is provided, including:

[0006] In response to a target account's request to access interactive information of a content item, the account characteristics of the target account, the global information characteristics of the interactive information, and the triggering scenario characteristics of the access request are obtained. The triggering scenario characteristics are used to characterize the features of the interface that triggers the access request.

[0007] Based on the account characteristics, the global information characteristics, and the triggering scenario characteristics, the target browsing information of the target account is obtained, and the target browsing information is used to characterize the expected browsing behavior of the target account towards the content item's interactive information;

[0008] If the target browsing information meets the target conditions, a recommendation request for the target account is sent to the recommendation model. The recommendation request is used to request target interaction information that matches the target account from the interaction information of the content item.

[0009] In one possible implementation, obtaining the target browsing information of the target account based on the account characteristics, the global information characteristics, and the triggering scenario characteristics includes:

[0010] The account features, global information features, and trigger scenario features are fused to obtain fused features;

[0011] The fused features are input into the browsing information acquisition model, and the browsing information acquisition model obtains the browsing category with the highest matching degree with the target account;

[0012] The target browsing information is determined based on the browsing category.

[0013] In one possible implementation, the browsing information acquisition model is a composite tree model integrating multiple tree models, each tree model being used to determine the contribution of a type of feature component to the browsing situation category using a class of feature components.

[0014] The step of inputting the fused features into the browsing information acquisition model, and obtaining the browsing category with the highest matching degree with the target account through the browsing information acquisition model, includes:

[0015] The fusion feature is input into each tree model in the browsing information acquisition model, and the fusion feature is split through the tree model to obtain the feature component in the fusion feature corresponding to the tree model.

[0016] Based on the feature threshold of each node in the tree model, the leaf node in which the feature component falls in the tree model is determined, and the leaf node score corresponding to the leaf node is determined as the contribution of the feature component to the browsing category.

[0017] Based on the multiple contribution values ​​output by the multiple tree models, the browsing category with the highest matching degree with the target account is determined.

[0018] In one possible implementation, the method further includes:

[0019] If the target browsing information meets the target conditions, the target interaction information indicated by the recommendation model is returned to the target account;

[0020] If the target browsing information does not meet the target conditions, the preset interactive information of the content item is returned to the target account.

[0021] In one possible implementation, the account features include at least one of the attribute features of the target account or the behavioral features of the target account. The attribute features are used to characterize the account attributes of the target account, and the behavioral features are used to characterize the behavior of the target account towards content items within a preset time period.

[0022] In one possible implementation, the global information feature includes global statistical information on at least one business metric for each interactive information of the content item.

[0023] According to another aspect of the embodiments of this disclosure, a method for training a browsing information acquisition model is provided, comprising:

[0024] For any content item, identify the sample accounts that have initiated historical access requests for the interaction information of the content item;

[0025] The account characteristics of the sample account, the global information characteristics of the interaction information, the triggering scenario characteristics of the historical access request, and the historical browsing information of the sample account are obtained. The triggering scenario characteristics are used to characterize the features of the interface that triggers the historical access request.

[0026] The account features, global information features, and trigger scenario features are input into the initial information acquisition model to obtain the target browsing information of the sample account;

[0027] Based on the historical browsing information and target browsing information of each sample account for each content item, the initial information acquisition model is iteratively trained to obtain the browsing information acquisition model. The browsing information acquisition model is used to acquire the target browsing information of the account's expected interaction with the content item.

[0028] In one possible implementation, the step of iteratively training the initial information acquisition model based on the historical browsing information and target browsing information of each sample account for each content item to obtain the browsing information acquisition model includes:

[0029] Based on the historical browsing information and target browsing information of each sample account for each content item, the loss function value is obtained;

[0030] When the loss function value does not meet the stopping iteration condition, the initial information acquisition model is iteratively trained.

[0031] When the loss function value meets the stopping iteration condition, the training of the initial information acquisition model is stopped, and the browsing information acquisition model is obtained.

[0032] In one possible implementation, the sample accounts that have initiated historical access requests to the content item's interaction information include:

[0033] A / B testing is performed on multiple accounts that have initiated the historical access requests; wherein, in the A / B testing, the multiple accounts are divided into experimental group accounts and control group accounts, the target interaction information matching the experimental group accounts is returned to the experimental group accounts, and the preset interaction information of the content items is returned to the control group accounts.

[0034] Based on the browsing history of the control group accounts, the sample accounts were selected from the experimental group accounts.

[0035] In one possible implementation, the sample accounts include positive sample accounts and negative sample accounts, wherein the historical browsing time of the target interaction information by the positive sample accounts is greater than the historical browsing time of the target interaction information by the negative sample accounts.

[0036] In one possible implementation, the step of selecting the sample accounts from the experimental group accounts based on the historical browsing information of the control group accounts includes:

[0037] Based on the historical browsing information of the control group account, multiple browsing categories are determined. The browsing category is determined based on the range of historical browsing duration of the control group account for the preset interactive information. The browsing category includes a first browsing category and a second browsing category, and the average browsing duration corresponding to the first browsing category is greater than the average browsing duration corresponding to the second browsing category.

[0038] For each experimental group account, if the historical browsing duration of the experimental group account for the target interaction information belongs to the first browsing category, the experimental group account is determined as a positive sample account;

[0039] If the historical browsing duration of the experimental group account for the target interaction information belongs to the second browsing category, the experimental group account will be identified as a negative sample account.

[0040] In one possible implementation, the first browsing category is located in the first target position in the sorting of the plurality of browsing categories according to the average browsing time from largest to smallest, and the second browsing category is located in the second target position in the sorting.

[0041] For each experimental group account, the method further includes:

[0042] From multiple browsing categories, determine the target browsing category to which the historical browsing duration of the target interaction information of the experimental group account belongs;

[0043] If the target browsing category is ranked first in the sorting, it is determined that the historical browsing duration of the experimental group account for the target interaction information belongs to the first browsing category;

[0044] If the target browsing category is located in the second-to-last target position in the sorting, it is determined that the historical browsing duration of the experimental group account for the target interaction information belongs to the second browsing category.

[0045] According to another aspect of the embodiments of this disclosure, a request processing apparatus is provided, comprising:

[0046] The first acquisition unit is configured to execute an access request in response to the target account’s interaction with content items, and acquire the account characteristics of the target account, the global information characteristics of the interaction information, and the triggering scenario characteristics of the access request, wherein the triggering scenario characteristics are used to characterize the features of the interface that triggers the access request.

[0047] The second acquisition unit is configured to acquire target browsing information of the target account based on the account characteristics, the global information characteristics, and the triggering scenario characteristics. The target browsing information is used to characterize the expected browsing behavior of the target account towards the content item's interactive information.

[0048] The recommendation unit is configured to send a recommendation request for the target account to the recommendation model when the target browsing information meets the target conditions. The recommendation request is used to request target interaction information that matches the target account from the interaction information of the content item.

[0049] In one possible implementation, the second acquisition unit includes:

[0050] The fusion subunit is configured to fuse the account features, the global information features, and the trigger scenario features to obtain fused features;

[0051] The acquisition subunit is configured to execute the input of the fused features into the browsing information acquisition model, and obtain the browsing category with the highest matching degree with the target account through the browsing information acquisition model;

[0052] The determination subunit is configured to perform the determination of the target browsing information based on the browsing category.

[0053] In one possible implementation, the browsing information acquisition model is a composite tree model integrating multiple tree models, each tree model being used to determine the contribution of a type of feature component to the browsing situation category using a class of feature components.

[0054] The acquisition subunit is configured to execute:

[0055] The fusion feature is input into each tree model in the browsing information acquisition model, and the fusion feature is split through the tree model to obtain the feature component in the fusion feature corresponding to the tree model.

[0056] Based on the feature threshold of each node in the tree model, the leaf node in which the feature component falls in the tree model is determined, and the leaf node score corresponding to the leaf node is determined as the contribution of the feature component to the browsing category.

[0057] Based on the multiple contribution values ​​output by the multiple tree models, the browsing category with the highest matching degree with the target account is determined.

[0058] In one possible implementation, the device further includes:

[0059] The return module is configured to return the target interaction information indicated by the recommendation model to the target account if the target browsing information meets the target conditions.

[0060] The return module is also configured to return preset interactive information of the content item to the target account if the target browsing information does not meet the target conditions.

[0061] In one possible implementation, the account features include at least one of the attribute features of the target account or the behavioral features of the target account, wherein the attribute features are used to characterize the account attributes of the target account, and the behavioral features are used to characterize the behavior of the target account towards content items within a preset time period.

[0062] In one possible implementation, the global information feature includes global statistical information on at least one business metric for each interactive information of the content item.

[0063] According to another aspect of the embodiments of this disclosure, a training apparatus for a browsing information acquisition model is provided, comprising:

[0064] The determining unit is configured to perform the following actions: for any content item, determine the sample account that has initiated historical access requests for the interaction information of the content item;

[0065] The acquisition unit is configured to acquire the account characteristics of the sample account, the global information characteristics of the interaction information, the triggering scenario characteristics of the historical access request, and the historical browsing information of the sample account. The triggering scenario characteristics are used to characterize the features of the interface that triggers the historical access request.

[0066] The output unit is configured to execute an initial information acquisition model by inputting the account features, the global information features, and the triggering scenario features into the sample account;

[0067] The training unit is configured to iteratively train the initial information acquisition model based on the historical browsing information and target browsing information of each sample account for each content item, thereby obtaining a browsing information acquisition model. The browsing information acquisition model is used to acquire the target browsing information of the account's expected interaction with the content item.

[0068] In one possible implementation, the training unit is configured to perform:

[0069] Based on the historical browsing information and target browsing information of each sample account for each content item, the loss function value is obtained;

[0070] When the loss function value does not meet the stopping iteration condition, the initial information acquisition model is iteratively trained.

[0071] When the loss function value meets the stopping iteration condition, the training of the initial information acquisition model is stopped, and the browsing information acquisition model is obtained.

[0072] In one possible implementation, the determining unit includes:

[0073] The testing subunit is configured to perform A / B testing on multiple accounts that have initiated the historical access requests; wherein, in the A / B testing, the multiple accounts are divided into experimental group accounts and control group accounts, the target interaction information matching the experimental group accounts is returned to the experimental group accounts, and the preset interaction information of the content items is returned to the control group accounts.

[0074] The filtering subunit is configured to perform filtering to obtain the sample accounts from the experimental group accounts based on the historical browsing information of the control group accounts.

[0075] In one possible implementation, the sample accounts include positive sample accounts and negative sample accounts, wherein the historical browsing time of the target interaction information by the positive sample accounts is greater than the historical browsing time of the target interaction information by the negative sample accounts.

[0076] In one possible implementation, the filtering subunit includes:

[0077] The first determining sub-unit is configured to perform the determination of multiple browsing situation categories based on the historical browsing information of the control group account. The browsing situation categories are determined based on the value range of the historical browsing duration of the control group account for the preset interactive information. The browsing situation categories include a first browsing category and a second browsing category, and the average browsing duration corresponding to the first browsing category is greater than the average browsing duration corresponding to the second browsing category.

[0078] The second determining sub-unit is configured to determine the experimental group account as a positive sample account if the historical browsing duration of the experimental group account for the target interaction information belongs to the first browsing category.

[0079] The second determining sub-unit is further configured to determine the experimental group account as a negative sample account if the historical browsing duration of the experimental group account on the target interaction information belongs to the second browsing category.

[0080] In one possible implementation, the first browsing category is located in the first target position in the sorting of the plurality of browsing categories according to the average browsing time from largest to smallest, and the second browsing category is located in the second target position in the sorting.

[0081] For each experimental group account, the second determined sub-unit is also configured to execute:

[0082] From multiple browsing categories, determine the target browsing category to which the historical browsing duration of the target interaction information of the experimental group account belongs;

[0083] If the target browsing category is ranked first in the sorting, it is determined that the historical browsing duration of the experimental group account for the target interaction information belongs to the first browsing category;

[0084] If the target browsing category is located in the second-to-last target position in the sorting, it is determined that the historical browsing duration of the experimental group account for the target interaction information belongs to the second browsing category.

[0085] According to another aspect of the embodiments of this disclosure, a computer device is provided, comprising:

[0086] One or more processors;

[0087] One or more memories for storing the one or more processor-executable instructions;

[0088] The one or more processors are configured to execute the request processing method or the training method of the browsing information acquisition model in any of the possible implementations of the above aspects.

[0089] According to another aspect of the present disclosure, a computer-readable storage medium is provided, which, when at least one instruction in the computer-readable storage medium is executed by one or more processors of a computer device, enables the computer device to perform a request processing method or a training method for a browsing information retrieval model in any possible implementation of the above aspect.

[0090] According to another aspect of the present disclosure, a computer program product is provided, including one or more instructions that can be executed by one or more processors of a computer device, enabling the computer device to perform a request processing method or a training method for a browsing information retrieval model in any possible implementation of the above aspect.

[0091] The technical solutions provided by the embodiments of this disclosure have at least the following beneficial effects:

[0092] By analyzing the access requests for interactive information issued by each account, and utilizing account characteristics, global information features of interactive information, and triggering scenario features of access requests, the overall dwell time of the current account on content items—that is, the target browsing information—can be quickly and accurately predicted. Only when the target browsing information meets the target conditions is a request for personalized recommendations sent to the recommendation model initiated; otherwise, there is no need to send such a request to the recommendation model. This significantly reduces the number of requests to the recommendation model under limited resources, thereby optimizing server resource allocation.

[0093] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0094] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0095] Figure 1 This is a schematic diagram illustrating the implementation environment of a request processing method according to an exemplary embodiment;

[0096] Figure 2 This is a flowchart illustrating a request processing method according to an exemplary embodiment;

[0097] Figure 3 This is an interactive flowchart illustrating a request processing method according to an exemplary embodiment;

[0098] Figure 4 This is a flowchart of a training method for a browsing information acquisition model according to an embodiment of this disclosure;

[0099] Figure 5 This is a logical structure block diagram of an information recommendation device according to an exemplary embodiment;

[0100] Figure 6 This illustration shows a structural block diagram of a terminal provided in an exemplary embodiment of the present disclosure;

[0101] Figure 7This is a schematic diagram of the structure of a server provided in an embodiment of this disclosure. Detailed Implementation

[0102] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0103] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0104] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in this disclosure are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the account characteristics of the target account, the interactive information of the content items, and the interface that triggers the access request involved in this application were all obtained with full authorization.

[0105] In some embodiments, the meaning of A and / or B includes three cases: A and B, and A and B.

[0106] Figure 1 This is a schematic diagram illustrating an implementation environment of a request processing method according to an exemplary embodiment. See also: Figure 1 The implementation environment includes terminal 101 and server 102.

[0107] Terminal 101 has an application installed and running that supports content items, i.e., a client application. This application includes, but is not limited to, short video applications, live streaming applications, social applications, food ordering applications, payment applications, ride-hailing applications, etc.

[0108] Terminal 101 is directly or indirectly connected to server 102 via wired or wireless communication. This embodiment of the present disclosure does not limit the connection method.

[0109] Server 102 is used to provide background services for the aforementioned applications. Server 102 includes at least one of a single server, multiple servers, a cloud computing platform, or a virtualization center. Optionally, server 102 undertakes the primary computing work, and terminal 101 undertakes the secondary computing work; or, server 102 undertakes the secondary computing work, and terminal 101 undertakes the primary computing work; or, server 102 and terminal 101 collaborate on computing using a distributed computing architecture.

[0110] In some embodiments, server 102 is 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 communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0111] In some embodiments, terminal 101 includes, but is not limited to: smartphones, tablets, laptops, desktop computers, smart voice interaction devices, smart home appliances, in-vehicle terminals, smart speakers, smartwatches, aircraft, MP3 (Moving Picture Experts Group Audio Layer III) players, MP4 (Moving Picture Experts Group Audio Layer IV) players, e-book readers, etc., but is not limited to these.

[0112] Those skilled in the art will understand that the number of terminals 101 described above can be more or less. For example, there may be only one terminal 101, or there may be dozens or hundreds of terminals 101, or even more. This disclosure does not limit the number or type of terminals 101.

[0113] Figure 2 This is a flowchart illustrating a request processing method according to an exemplary embodiment, see [link to flowchart]. Figure 2 The request processing method is executed by the computer device. The following explanation uses the computer device as a server as an example.

[0114] In step 201, in response to the target account's access request for interactive information of content items, the server obtains the account characteristics of the target account, the global information characteristics of the interactive information, and the triggering scenario characteristics of the access request. The triggering scenario characteristics are used to characterize the features of the interface that triggers the access request.

[0115] In step 202, the server obtains the target browsing information of the target account based on the account characteristics, the global information characteristics, and the triggering scenario characteristics. The target browsing information is used to characterize the expected browsing behavior of the target account towards the content item.

[0116] In step 203, if the target browsing information meets the target conditions, the server sends a recommendation request to the recommendation model for the target account. The recommendation request is used to request the acquisition of target interaction information that matches the target account from the interaction information of the content item.

[0117] The method provided in this disclosure, by taking into account the access requests for interactive information issued by each account, utilizes account characteristics, global information characteristics of interactive information, and triggering scenario characteristics of access requests to quickly and accurately predict the overall dwell time performance of the current account on the content item's interactive information, i.e., the target browsing information. Only when the target browsing information meets the target conditions is a request for personalized recommendation sent to the recommendation model initiated; otherwise, there is no need to send the above request to the recommendation model. Under the premise of limited resources, this method can greatly reduce the number of requests to the recommendation model, thereby optimizing the resource configuration of the server.

[0118] In one possible implementation, obtaining the target browsing information of the target account based on the account characteristics, the global information characteristics, and the triggering scenario characteristics includes:

[0119] The account features, global information features, and trigger scenario features are fused together to obtain the fused features;

[0120] The fused feature is input into the browsing information acquisition model, which then obtains the browsing category that best matches the target account.

[0121] The target browsing information is determined based on this browsing category.

[0122] In one possible implementation, the browsing information acquisition model is a composite tree model that integrates multiple tree models, each tree model being used to determine the contribution of a feature component to the browsing situation category using a class of feature components.

[0123] The fused feature is input into the browsing information acquisition model, which then obtains the browsing categories that best match the target account, including:

[0124] The fused feature is input into each tree model in the browsing information acquisition model. The fused feature is then split through the tree model to obtain the feature components in the fused feature that correspond to the tree model.

[0125] Based on the feature threshold of each node in the tree model, the leaf node in which the feature component falls in the tree model is determined, and the leaf node score corresponding to the leaf node is determined as the contribution of the feature component to the browsing category.

[0126] Based on multiple contribution values ​​output by the tree model, the browsing category with the highest matching degree with the target account is determined.

[0127] In one possible implementation, the method further includes:

[0128] If the browsing information of the target matches the target condition, return the target interaction information indicated by the recommendation model to the target account;

[0129] If the browsing information of the target does not meet the target conditions, return the preset interactive information of the content item to the target account.

[0130] In one possible implementation, the account features include at least one of the attribute features or the behavioral features of the target account. The attribute features are used to characterize the account attributes of the target account, and the behavioral features are used to characterize the behavior of the target account towards content items within a preset time period.

[0131] In one possible implementation, the global information feature includes global statistical information on the various interactive information of the content item in at least one business metric.

[0132] All of the above-mentioned optional technical solutions can be combined in any way to form optional embodiments of this disclosure, and will not be described in detail here.

[0133] Figure 3 This is an interactive flowchart illustrating a request processing method according to an exemplary embodiment, such as... Figure 3 As shown, the request processing method is executed by a computer device. The following description uses a computer device as a server as an example. This embodiment includes the following steps.

[0134] In step 301, the server responds to the target account's access request for interactive information of the content item by obtaining the account characteristics of the target account, the global information characteristics of the interactive information, and the triggering scenario characteristics of the access request.

[0135] The target account refers to any account logged in on an application that supports the content item.

[0136] Content items refer to multimedia resources used to present or carry specific content. Multimedia resources include, but are not limited to, video resources, audio resources, image resources, text resources, and web page resources. Among them, video resources include short videos, mini videos, edited videos, and composite videos; audio resources include songs, radio dramas, and radio broadcasts; image resources include animated images and image news; and text resources include articles, blog posts, and news. This disclosure does not specifically limit the types of multimedia resources.

[0137] Interactive information refers to information generated by the interaction between an account and a content item. For example, interactive information refers to various comments sent by each account to the content item, or interactive information also includes forwarding information, sharing information, bullet screen information, etc. sent by each account to the content item. This embodiment of the disclosure does not specifically limit interactive information.

[0138] An access request refers to a request to access the aforementioned interactive information. Typically, when a target account browses content items, to conserve computing resources, it will collapse the interactive information and provide an entry point for expanding the interactive information. This allows the terminal to only retrieve the content items when displaying them, without needing to retrieve the interactive information. Only when a user triggers the expansion entry is an access request sent to the server to retrieve the interactive information. Optionally, the terminal may retrieve both the content items and the interactive information simultaneously when displaying them. This embodiment of the disclosure does not specifically limit when the terminal sends the access request.

[0139] Account characteristics refer to the inherent attributes of a target account. Optionally, account characteristics include at least one of the target account's attribute characteristics or behavioral characteristics. The attribute characteristics are used to characterize the target account's account attributes and can reflect the basic attributes and interests of the user corresponding to the target account. The behavioral characteristics are used to characterize the behavioral data generated by the target account on content items during the target time period and can reflect whether the user corresponding to the target account has browsed interactive information of other content items or posted interactive information on other content items in the short term.

[0140] It should be noted that the aforementioned attribute features and behavioral features are all extracted after the user has authorized or fully authorized the attribute data and behavioral data. Optionally, taking the content item as a short video and the interactive information as the comment information of the short video as an example, the aforementioned attribute data includes: the age, gender, geographical location, occupation, etc. of the user corresponding to the target account, and the aforementioned behavioral data includes: the recent behavior count of the user corresponding to the target account in the comment section, wherein the behavior count includes: the number of recent likes, the number of times recently viewed the personal homepage of the comment author (referring to the account that posted the comment), the number of times recently followed the comment author, and the number of times recently expanded sub-comments (referring to secondary comments that further reply to the comment information). After collecting the attribute data and behavioral data that have been fully authorized and individually agreed to by the user, the server extracts attribute features from the attribute data and behavioral features from the behavioral data. Optionally, embedding features are used when extracting features, or one-hot features are used when extracting features. This embodiment of the disclosure does not specifically limit the feature type. In this context, "recent" refers to the last 3 days, the last 7 days, or the last 1 month, etc. This disclosure does not specifically limit these terms.

[0141] Global information features refer to the features obtained by extracting features from various interactive information of the same content item, taking each content item as a unit. Optionally, global information features include at least one of the content features or statistical features of the interactive information. The content features refer to the features of the interactive information itself. For example, if the interactive information is a comment, the content features refer to the embedding features of a text comment or the image features of an image comment. The statistical features refer to the statistical features of each interactive information of the content item. These statistical features are used to characterize the statistical information of each interactive information on at least one business indicator. This business indicator refers to the business indicator that needs to be considered when recommending interactive information in a personalized way. Pre-configured by technical personnel, for example, since each content item usually has a large amount of interactive information, this interactive information can form an information pool. The overall information pool can be statistically analyzed for each business indicator. Indicatively, the statistical information includes: the counting features of all candidate comments in the comment section, such as the size of the comment candidate pool, the total number of likes for all candidate comments, the total number of expanded sub-comments, the total number of exposures, the total number of hot comments and witty comments, etc. By extracting features from the statistical information, the statistical features can be obtained. Optionally, embedding features or one-hot features can be used when extracting features. This disclosure does not specifically limit the feature type.

[0142] Triggering scenario features refer to the scenario features that trigger the access request. In other words, the triggering scenario features are used to characterize the features of the interface that triggers the access request. The triggering scenario features are regarded as the context features of the access request, which can indicate which page the access request comes from, such as from the main feeds interface, or from the small video overlay, or from the short video overlay, etc. This disclosure does not specifically limit the type of interface. By extracting features from the interface identifier of the interface that triggers the access request, the triggering scenario features (i.e., context features) of the access request can be obtained. Optionally, embedding features or one-hot features can be used when extracting features. This disclosure does not specifically limit the feature type.

[0143] In some embodiments, the server maintains account characteristics for each account, stores the account characteristics of each account in the platform into an account characteristic database, and updates the account characteristics of the account in the account characteristic database when the attribute data or behavior data of the account changes.

[0144] In some embodiments, the server maintains global information features for each content item, stores the global information features of each content item in the global information feature library, and updates the global information features of the content item in the global information feature library when new interaction information for the content item is collected, as the statistical information changes.

[0145] In some embodiments, the server pre-configures a unique interface identifier for each interface that can support access to interactive information, and extracts trigger scene features for each interface identifier and stores them in the trigger scene feature library.

[0146] In some embodiments, the terminal logs into the target account in an application, retrieves the content item from the server based on the user's viewing action, and displays the content item and an expansion entry for interactive information in the application. In response to the user's triggering action on the expansion entry, the terminal sends an access request to the server for the interactive information. Optionally, the access request includes at least the account identifier of the target account, the content item identifier, and the interface identifier of the interface that triggered the access request. Optionally, the terminal retrieves the content item from the server while simultaneously sending the access request to the server to retrieve the interactive information based on the user's viewing action.

[0147] In some embodiments, when a server receives a request from any terminal, it parses the header field of the request. If the header field carries a request identifier for an access request for interactive information, it determines that the request is an access request for interactive information. Then, it parses the data field of the request to obtain the account identifier of the target account that issued the access request, the content item identifier of the content item associated with the requested interactive information, and the interface identifier of the interface that triggered the access request.

[0148] In some embodiments, after obtaining the account identifier, content item identifier, and interface identifier, the server uses the account identifier as an index to search for the target account's account features in the account feature database, uses the content item identifier as an index to search for the global information features of the content item's interaction information in the global information feature database, and uses the interface identifier as an index to search for the trigger scenario features of the access request in the trigger scenario feature database. Because the server maintains its own feature database for each feature, it can quickly find the corresponding feature after receiving an access request without needing to extract features in real time, thus speeding up the response to access requests and reducing request latency.

[0149] In some embodiments, after the server obtains the account identifier, content item identifier, and interface identifier, it uses the account identifier as an index to search for the attribute data and behavior data of the target account, thereby extracting attribute features and behavior features. The attribute features and behavior features are then identified as account features. Next, using the content item identifier as an index, the server searches for the statistical features of each interactive information of the content item, and the statistical features are identified as global information features. Then, using the interface identifier as an index, the server extracts the embedding feature of the interface identifier as the trigger scene feature. This ensures that the account features, global information features, and trigger scene features are all extracted based on the latest data. This embodiment does not specifically limit whether the features are pre-stored in the feature library.

[0150] In step 302, the server fuses the account feature, the global information feature, and the trigger scenario feature to obtain the fused feature.

[0151] In some embodiments, the server concatenates the account feature, the global information feature, and the trigger scenario feature to obtain the fused feature.

[0152] In some embodiments, in addition to splicing, the fusion method for fusing the account feature, the global information feature and the triggering scenario feature also includes: element-wise addition, element-wise multiplication, bilinear merging, etc. The embodiments of this disclosure do not specifically limit the fusion method.

[0153] In step 303, the server inputs the fused feature into the browsing information acquisition model, and uses the browsing information acquisition model to obtain the browsing category with the highest matching degree to the target account.

[0154] This browsing information acquisition model is used to obtain the browsing category with the highest degree of matching for each account, based on the account characteristics, global information characteristics, and trigger scenario characteristics of that account. Indicatively, the browsing category refers to the browsing duration range. Therefore, the browsing category with the highest degree of matching for an account refers to the browsing duration range in which the account's expected interaction with the content item falls. In other words, this browsing information acquisition model is used to estimate the overall dwell time of an account in the comment section of the current content item in a personalized and accurate manner.

[0155] Indicatively, when the browsing information retrieval model is a classification model, each optional browsing situation category is also called a browsing situation category of the classification model. The browsing information retrieval model obtains the degree of matching (e.g., matching probability) between the target account and each browsing situation category, and determines the browsing situation category with the highest degree of matching with the target account from multiple optional browsing situation categories. At this time, the browsing information retrieval model can be divided into two cases: a binary classification model and a multi-classification model, which will be discussed separately.

[0156] In some embodiments, the browsing information acquisition model is a binary classification model. This model is used to predict whether an account meets preset browsing conditions. Preset browsing conditions refer to browsing conditions that require personalized recommendations to be enabled. For example, preset browsing conditions mean that the account's browsing time for interacting with content items exceeds a browsing time threshold. The binary classification model then estimates whether the account's browsing time for interacting with content items exceeds the browsing time threshold. The binary classification model involves two categories of browsing conditions: one is a "Yes (YES, Y)" label, representing that the input account is expected to meet the preset browsing conditions, for example, that the account's browsing time for interacting with content items is expected to exceed the browsing time threshold; the other is a "No (NO, N)" label, representing that the input account is expected not to meet the preset browsing conditions, for example, that the account's browsing time for interacting with content items is expected to be less than or equal to the browsing time threshold. In other words, it is expected that the user will quickly leave the interaction area (such as the comment section). The browsing time threshold is any value greater than 0.

[0157] Optionally, the browsing information acquisition model is a multi-classification model. This multi-classification model is used to estimate which browsing situation an account specifically matches among multiple alternative browsing situations. For example, the multi-classification model is used to determine the browsing situation category with the highest matching degree to the input account. The browsing situation categories involved in the multi-classification model can be set by technical personnel. For example, the browsing situation categories include: longer browsing time, medium browsing time, and shorter browsing time. Alternatively, the browsing situation categories include: browsing time greater than 10 minutes, browsing time less than or equal to 10 minutes but greater than 3 minutes, and browsing time less than or equal to 3 minutes, etc. Therefore, compared with the binary classification model, it can perform fine quantification of browsing situation categories. This embodiment of the disclosure does not specifically limit the setting of browsing situation categories.

[0158] Optionally, in a server cluster architecture, the training server uses a sample dataset to train the initial information retrieval model. After training, a browsing information retrieval model is obtained. Then, the request server pulls the model parameter set of the trained browsing information retrieval model from the training server. After receiving an access request from a target account and obtaining the account features, global information features, and trigger scenario features of the target account, the request server inputs the fused feature obtained by fusing the account features, global information features, and trigger scenario features into the trained browsing information retrieval model. The browsing information retrieval model processes the fused feature and then outputs the matching probability of the target account with each browsing category. For example, when the browsing information retrieval model is a binary classification model, it outputs the matching probability of each of the two categories. When the browsing information retrieval model is a multi-class classification model, it outputs the matching probability of each of the multiple categories. Then, from the various matching probabilities output by the browsing information retrieval model, the browsing category with the highest matching probability is selected as the browsing category with the highest degree of matching with the target account.

[0159] In some embodiments, the model architecture of the browsing information acquisition model includes tree models such as GB (Gradient Boosting), GBDT (Gradient Boosting Decision Tree), and XGBoost (eXtremeGradient Boosting). This disclosure does not specifically limit the model structure of the browsing information acquisition model.

[0160] Taking the XGBoost tree model as an example to illustrate the browsing information acquisition model, the XGBoost tree model is a composite tree model integrating multiple tree models. Each tree model is used to determine the contribution of a feature component to the browsing category using a class of feature components. Optionally, the tree model is a CART (Classification and Regression Tree), or it is a linear classifier (Gblinear). This embodiment does not specifically limit the structure of the tree model. In other words, the tree model is equivalent to a weak learner, which corresponds to a constraint. This constraint is used to characterize the rules to be followed when determining the browsing category with the highest matching degree with the account. Assuming the interaction information is video comment information, the constraint includes, but is not limited to: the cumulative number of comments posted by the account in the comment section of other videos in the recent period is greater than a count threshold; the total number of candidate comments in the comment section of the current video is greater than a total comment threshold; the total number of popular and insightful comments in the comment section of the current video is greater than a popular comment threshold, etc. This embodiment does not specifically limit the constraint.

[0161] Schematic, the server inputs the fused feature into each tree model in the browsing information acquisition model. For each tree model, the server performs feature splitting on the fused feature to obtain the feature component corresponding to that tree model. Then, based on the feature threshold of each node in the tree model, the server determines the leaf node in which the feature component falls within the tree model. The leaf node score corresponding to this leaf node is determined as the contribution of the feature component to the browsing category. In other words, the process of obtaining the leaf node score is equivalent to dividing the feature component into leaf nodes of the decision tree in which the tree model resides. Optionally, the decision tree in which each tree model resides is a binary tree, so the tree model will be split into two parts during feature splitting: a left subtree (i.e., the left leaf node) and a right subtree (i.e., the right leaf node). Next, a score is assigned based on the constraints of the tree model, and the corresponding leaf node score is output. Schematic, during the scoring process, a positive value is assigned to the leaf node score for leaf nodes that meet the corresponding constraints, and a negative value is assigned to the leaf node score for leaf nodes that do not meet the corresponding constraints.

[0162] Since the XGBoost tree model is a composite tree model integrating multiple tree models, and each tree model outputs a leaf node score (i.e., a contribution), the multiple tree models output multiple leaf node scores (i.e., multiple contribution values). Based on these multiple contribution values ​​output by the multiple tree models, the browsing category with the highest matching degree to the target account can be determined. Illustratively, when the XGBoost tree model is a binary classification model, it only needs to output the matching probability for the "Yes" tag. The server weights the contribution values ​​output by each of the multiple tree models to obtain the matching probability for the "Yes" tag. Subtracting the matching probability of the "Yes" tag from 1 gives the matching probability of the "No" tag. In other words, if the matching probability of the "Yes" tag is α, then the matching probability of the "No" tag is (1-α).

[0163] In some embodiments, after obtaining the matching probability of the target account with each browsing category, the server obtains the browsing category corresponding to the highest matching probability among all matching probabilities, and takes the browsing category with the highest matching probability as the browsing category with the highest degree of matching with the target account.

[0164] In some embodiments, after obtaining the matching probability of the target account with each browsing category, the server sorts the matching probabilities in descending order and selects the browsing category corresponding to the matching probability at the top of the sort as the browsing category with the highest matching degree with the target account.

[0165] To illustrate, for the binary classification model, after calculating the matching probability of the "Yes" label and the matching probability of the "No" label respectively, the browsing category with the highest matching probability is taken as the browsing category with the highest matching degree with the target account.

[0166] In step 304, the server determines the target browsing information of the target account based on the browsing category. The target browsing information is used to characterize the expected browsing behavior of the target account towards the content item.

[0167] In some embodiments, since the browsing situation category may be a qualitative category (such as a binary classification model, estimating whether the browsing situation meets the preset criteria or does not meet the preset criteria), or a quantitative category (such as a multi-class classification model, estimating whether the browsing time is greater than 10 minutes or less than or equal to 3 minutes), at least one of the browsing situation category or the description tag information corresponding to the browsing situation category can be directly obtained as the target browsing information. Alternatively, the target browsing information can be determined based on at least one of the browsing situation category or the description tag information corresponding to the browsing situation category. For example, a binary target browsing information can be determined based on at least one of the browsing situation category or the description tag information corresponding to the browsing situation category. The binary target browsing information means that the value is either 1 or 0. Or, a Boolean target browsing information can be determined. The Boolean target browsing information means that the value is either True or False. This disclosure does not specifically limit the method of determining the target browsing information.

[0168] In some embodiments, when the browsing information acquisition model is a binary classification model, if the browsing situation category with the highest matching probability is the "Yes" label, then the target browsing information associated with the "Yes" label is determined to be 1, which means that the target account is expected to browse the interactive information in accordance with the preset browsing situation; if the browsing situation category with the highest matching probability is the "No" label, then the target browsing information associated with the "No" label is determined to be 0, which means that the target account is expected to browse the interactive information in accordance with the preset browsing situation. This makes it easy to binarize the target browsing information, which is convenient for subsequent determination of whether to forward personalized recommendation requests to the recommendation model.

[0169] In some embodiments, when the browsing information acquisition model is a multi-classification model, it is assumed that three browsing categories are provided: browsing duration greater than 10 minutes, browsing duration less than or equal to 10 minutes but greater than 3 minutes, and browsing duration less than or equal to 3 minutes. Taking a browsing duration threshold of 3 minutes as an example, if the browsing category with the highest matching probability is the label "browsing duration greater than 10 minutes" or the label "browsing duration less than or equal to 10 minutes but greater than 3 minutes", then the associated target browsing information is determined to be 1, which means that the target account expects to browse the interactive information for a duration greater than the browsing duration threshold. In other words, it reflects that the target account's expected browsing behavior for the interactive information conforms to the preset browsing behavior. If the browsing category with the highest matching probability is "browsing duration less than or equal to 3 minutes", then the associated target browsing information is determined to be 0, which means that the target account expects to browse the interactive information for a duration less than or equal to the browsing duration threshold. In other words, it reflects that the target account's expected browsing behavior for the interactive information does not conform to the preset browsing behavior. This also makes it easy to binarize the target browsing information, which is convenient for subsequent determination of whether to forward personalized recommendation requests to the recommendation model.

[0170] In the above process, for each account's access request for interactive information, after the server-side has loaded the model parameter set of the trained browsing information acquisition model, it can use the browsing information acquisition model to process the account features, global information features, and trigger scenario features to obtain the target browsing information of the target account, so as to predict the overall dwell performance of this request in real time. In other words, the server-side obtains the target browsing information of the target account based on the account features, global information features, and trigger scenario features. Since the target browsing information can characterize whether the target account meets the target conditions, it can further determine whether the current access request needs to be forwarded to the recommendation model for personalized recommendation of interactive information. This is equivalent to filtering the access request before the recommendation model used for personalized recommendation, thereby greatly reducing the number of requests to the recommendation model under the premise of limited computing resources.

[0171] In step 305, if the target browsing information meets the target conditions, the server sends a recommendation request to the recommendation model for the target account. The recommendation request is used to request the acquisition of target interaction information that matches the target account from the interaction information of the content item.

[0172] Here, the target condition refers to the condition for enabling personalized recommendations of interactive information for the current access request. For example, the target condition is that the target browsing information is 1, or the target browsing information is that the target account's expected browsing behavior of the interactive information conforms to preset browsing behavior. This embodiment of the disclosure does not specifically limit the content of the target condition.

[0173] Optionally, when the target browsing information obtained in step 304 above is 1, it is determined that the target browsing information meets the target conditions. At this time, the server can make personalized recommendations for interactive information of the target account. That is, the server sends a recommendation request for the target account to the recommendation model, or the server forwards the access request to the recommendation model (e.g., a deep neural network) used for personalized recommendations.

[0174] In step 306, in response to the recommendation request, the server filters the interaction information of the content item through the recommendation model to obtain the target interaction information that matches the target account, and returns the target interaction information indicated by the recommendation model to the target account.

[0175] In some embodiments, upon receiving a recommendation request, a recommendation model selects the top K (K≥1) matching target interaction messages from each interaction message of the content item for the target account, and returns K target interaction messages to the target account so that the target account can display K target interaction messages, thereby achieving personalized interaction information recommendations for the target account. Here, target interaction information refers to interaction information that matches the target account.

[0176] Indicatively, the aforementioned recommendation model is an XTR (X-Through-Rate) model. For example, the XTR model includes at least one of the following: CTR (Click Through Rate) model, LTR (Like Through Rate) model, FTR (Follow Through Rate) model, etc. This disclosure does not specifically limit the recommendation model.

[0177] In some embodiments, the personalized recommendation process includes two stages: a recall stage (i.e., coarse ranking stage) and a ranking stage (i.e., fine ranking stage). The recall stage is used to quickly filter out a small portion of interactive information that the target account may be interested in from the full candidate pool consisting of massive interactive information of the content item for any access request. The interactive information filtered out in the recall stage is called the first interactive information. Optionally, multiple different methods are used for multi-way recall in the recall stage. Since the first interactive information filtered out by multi-way recall has different measurement standards in each way, it is difficult to compare them directly. Therefore, the ranking stage is used to uniformly score and rank each first interactive information recalled by multi-way recall in the recall stage. Optionally, the ranking stage integrates rich features for each first interactive information and uses a complex deep learning model to accurately predict the probability of various interactive behaviors between the target account and each first interactive information. Taking into account the above factors, personalized recommendations of interactive information are finally made for the target account. The aforementioned recommendation model can be used in the ranking stage to uniformly score and rank each first interaction information, while in the recall stage, some strategy-oriented recall methods or supervised coarse-ranking models can be used for multi-way recall.

[0178] In some embodiments, during the ranking phase, the server needs to use a recommendation model to predict the probability that the target account will interact with each first interactive message. For example, when multiple video comments are retrieved from the first interactive message, it is necessary to predict the target account's like rate, probability of expanding sub-comments, probability of entering the comment author's profile, probability of following the comment author, probability of replying to the comment author, etc., for each video comment. Then, all the above probabilities are combined to score the current video comment (e.g., by weighted summation). Since a large number of features are used when estimating various probabilities, including the account features of the comment reader (target account), the account features of the comment author (account that posted the comment), the comment's own features, and the comment context features, the server needs to use a recommendation model to predict the probability of the target account interacting with each video comment. Furthermore, the ranking process involves complex recommendation models (such as deep neural networks), which consume significant computational resources. Additionally, the volume of interactive information accessed by content items on the platform is enormous, while server-side computational resources are limited. Given these limited resources, the method provided in this embodiment can filter out requests whose target browsing information does not meet the target conditions. For these requests, preset interactive information is sent via step 307 without personalized recommendations. Only requests whose target browsing information meets the target conditions will be subject to precise personalized recommendations through the recall and ranking stages. This significantly reduces the number of requests to the recommendation model and maximizes the estimated benefits of the XTR model.

[0179] In step 307, if the target browsing information does not meet the target conditions, the server returns the preset interactive information of the content item to the target account.

[0180] Optionally, when the target browsing information obtained in step 304 above is 0, it is determined that the target browsing information does not meet the target conditions. At this time, the server only needs to return the preset interaction information of the content item visited this time to the target account. For example, return the N (N≥1) interaction information with the highest number of likes, or return the N interaction information with the most sub-comments, etc. The present disclosure embodiment does not specifically limit the selection rules of the preset interaction information for each content item.

[0181] If the target browsing information does not meet the target conditions, it means that the target account's expected browsing behavior for the interactive information does not meet the preset browsing behavior. Therefore, the server does not need to make personalized recommendations for interactive information for the target account, nor does it need to send the current access request or recommendation request to the recommendation model, thus avoiding the waste of computing resources.

[0182] All of the above-mentioned optional technical solutions can be combined in any way to form optional embodiments of this disclosure, and will not be described in detail here.

[0183] The method provided in this disclosure, by taking into account the access requests for interactive information issued by each account, utilizes account characteristics, global information characteristics of interactive information, and triggering scenario characteristics of access requests to quickly and accurately predict the overall dwell time performance of the current account on the content item's interactive information, i.e., the target browsing information. Only when the target browsing information meets the target conditions is a request for personalized recommendation sent to the recommendation model initiated; otherwise, there is no need to send the above request to the recommendation model. Under the premise of limited resources, this method can greatly reduce the number of requests to the recommendation model, thereby optimizing the resource configuration of the server.

[0184] Furthermore, when using a tree model as a browsing information acquisition model to predict the target browsing information, the tree model itself has the characteristics of simple structure, lightweight, and personalization. It can rate-limit requests forwarded to the recommendation model. That is, by restricting short-stay (i.e., not meeting the target conditions) access requests in advance through the tree model, these access requests that do not meet the target conditions will not be forwarded to the recommendation model, which can save server-side computing resources. At the same time, it retains long-stay (i.e., meeting the target conditions) access requests. These access requests that meet the target conditions will still be forwarded to the recommendation model, which can maximize the benefits of each account's stay time in the interaction area. In other words, it realizes personalized rate limiting for access requests, alleviates the pressure of server-side computing resource shortages, and can be directly deployed on the server receiving the requests. Its overall deployment cost is also relatively low, and there is no need to deploy it on a new machine.

[0185] Furthermore, since the prediction process of tree models can be parallelized, is fast, and has a small memory footprint, it can also be executed locally on the server receiving the request, such as a gateway server or a proxy server, without needing to request Remote Procedure Call (RPC) or gRPC services. This also reduces the time consumption and minimizes storage and communication overhead.

[0186] The above embodiments detailed how to use the browsing information acquisition model for online evaluation. In this disclosure, the training process of the browsing information acquisition model will be described in detail. Figure 4 This is a flowchart illustrating a training method for a browsing information acquisition model according to an embodiment of this disclosure. This training process is executed by a computer device, and the following description uses a server as an example. It should be noted that in a server cluster architecture, the device used to train the browsing information acquisition model and the device used to apply the browsing information acquisition model can be the same or different; this is not limited here.

[0187] In step 401, the server performs A / B testing on multiple accounts that have initiated historical access requests to the interactive information of the content items. In the A / B test, the multiple accounts are divided into experimental group accounts and control group accounts. The server returns the target interactive information that matches the experimental group accounts and returns the preset interactive information of the content items to the control group accounts.

[0188] A / B testing, also known as A / B experiment, is a method used for testing new versions or features of internet products. It typically involves developing two (or more) different solutions for the same optimization goal. For example, optimization goals may include modifying the interface layout, modifying the appearance of controls, launching new features, or modifying the interaction flow. During the same testing period, different solutions are randomly deployed to various users' clients, allowing each client to randomly use any one of the solutions. Behavioral data from the accounts logged into each client is collected. Finally, significance testing is used to analyze and evaluate the results, and the solution with the best performance is adopted in the official version.

[0189] During the A / B testing process, two schemes were designed. One was the experimental group, in which personalized recommendations for interactive information were enabled for each content item browsed by each experimental group account. That is, for access requests for interactive information of content items initiated by experimental group accounts, the recommendation model returned target interactive information matching their own accounts. The other was the control group, in which personalized recommendations for interactive information were disabled for each content item browsed by each control group account. That is, for access requests for interactive information of content items initiated by control group accounts, the preset interactive information of the content items was returned directly without accessing the recommendation model.

[0190] When the server receives an access request for interactive information of a content item from any client, it randomly assigns the account logged into that client to either an experimental group account or a control group account. If the account is assigned to the experimental group, the access request is forwarded to the recommendation model or a recommendation request is sent to the recommendation model. Through multi-path recall and precise ranking, personalized recommendations for interactive information are achieved for the experimental group account. Please refer to the description in steps 305-306 above for the recommendation method, which will not be repeated here. If the account is assigned to the control group, there is no need to forward the access request to the recommendation model or send a recommendation request to the recommendation model. It is only necessary to return the preset interactive information of the corresponding content item to the client. Please refer to the description in step 307 above for the method of selecting the preset interactive information, which will not be repeated here.

[0191] In step 402, the server determines multiple browsing categories based on the historical browsing information of the control group account. These browsing categories are determined based on the range of historical browsing durations of the control group account for preset interactive information. The browsing categories include a first browsing category and a second browsing category.

[0192] The average browsing time corresponding to the first browsing category is greater than the average browsing time corresponding to the second browsing category. The first and second browsing categories will be introduced in steps 403-404 below, and will not be repeated here.

[0193] In some embodiments, since the server returns preset interaction information associated with content items to each control group account in the A / B test, that is, the server does not make personalized recommendations for interaction information for each control group account, after the client logged in by the control group account receives the preset interaction information returned by the server, the client obtains the historical browsing time of the control group account for the returned preset interaction information, and then reports the historical browsing time to the server. The server stores the content item identifier of the content item, the historical browsing time reported by the client and the account identifier of the control group account accordingly. Finally, the server collects the historical browsing time reported by each control group account in the control group each time, and the collection of these historical browsing times constitutes the historical browsing information of the control group account.

[0194] In some embodiments, browsing scenarios are provided as browsing duration intervals that each historical browsing duration may fall into. In this case, the minimum and maximum values ​​of each historical browsing duration are determined, and the interval from the minimum to the maximum value is the value interval of the historical browsing duration. Then, the multiple browsing duration intervals obtained by dividing the value interval are used as multiple browsing scenarios.

[0195] In some embodiments, the server obtains the number of experimental buckets set by the technician. For example, the technician inputs 10 experimental buckets. Alternatively, if the technician does not set a number of experimental buckets, the number of experimental buckets is set to a default value, such as 10. Of course, the number of experimental buckets can also be any integer greater than or equal to 1, such as 8, 12, 20, etc. This disclosure does not specifically limit this. The number of experimental buckets represents how many experimental buckets need to be divided into for the historical browsing time of each control group account. That is, the number of experimental buckets represents the total number of the above-mentioned browsing situation categories (i.e., browsing time intervals).

[0196] In some embodiments, after obtaining the number of experimental buckets, the above value range is divided into multiple browsing duration intervals at equal intervals, such that the total number of browsing duration intervals is equal to the number of experimental buckets. For example, assuming the number of experimental buckets is 10, the above value range is divided into 10 browsing duration intervals at equal intervals.

[0197] In some embodiments, after obtaining the number of experimental buckets, the above-mentioned value interval is divided into multiple browsing duration intervals at non-equidistant intervals, such that the total number of browsing duration intervals is equal to the number of experimental buckets. For example, assuming the number of experimental buckets is 10, the above-mentioned value intervals are divided into 10 browsing duration intervals at non-equidistant intervals. This disclosure does not specifically limit whether the browsing duration intervals are divided at equidistant intervals.

[0198] In step 403, for each experimental group account, if the server determines the experimental group account as a positive sample account if the historical browsing time of the experimental group account to the returned target interaction information belongs to the first browsing category.

[0199] The first browsing category refers to the browsing category that ranks first in the sorting of the multiple browsing categories obtained in step 402, according to the average browsing time from largest to smallest. In other words, the server sorts the multiple browsing categories obtained in step 402 according to the average browsing time (e.g., the median of the browsing time interval) from largest to smallest, and determines the browsing category that ranks first in the first N1 positions as the first browsing category, where N1 is the number of digits in the first target position, and N1 is an integer greater than or equal to 1, for example, N1 = 3.

[0200] In some embodiments, during A / B testing, the server returns target interaction information matching the experimental group account to each experimental group account. That is, the server makes personalized recommendations for interaction information for each experimental group account. After receiving the target interaction information matching its own account returned by the server, the client logged in by the experimental group account obtains the historical browsing time of the experimental group account for the returned target interaction information, and then reports the historical browsing time to the server. The server stores the content item identifier of the content item, the historical browsing time reported by the client, and the account identifier of the experimental group account accordingly. Finally, the server collects the historical browsing time reported by each experimental group account in the experimental group each time. The collection of these historical browsing times constitutes the historical browsing information of the experimental group account.

[0201] For each historical browsing duration in the historical browsing information of the experimental group account, it is necessary to determine which browsing category this historical browsing duration belongs to, as obtained in step 402 above. Furthermore, if the current historical browsing duration belongs to the first browsing category, then the corresponding experimental group account will be obtained as a positive sample account in the browsing information acquisition model training process. If the current historical browsing duration belongs to the second browsing category, then the corresponding experimental group account will be obtained as a negative sample account in the browsing information acquisition model training process. If the current historical browsing duration belongs to neither the first nor the second browsing category, then the corresponding experimental group account will not participate in this training process and will be filtered out as a low-quality sample.

[0202] In some embodiments, for each experimental group account, from the multiple browsing categories obtained in step 402 above, the target browsing category to which the historical browsing time of the experimental group account for the target interaction information belongs is determined. For example, when the browsing category is provided as a browsing time range, the target browsing category refers to the browsing time range into which the historical browsing time of the experimental group account for the target interaction information falls. If the target browsing category is in the first target position in the sorting of the multiple browsing categories according to the average browsing time from largest to smallest, in other words, if the target browsing category is in the first N1 positions in the above sorting, it is determined that the historical browsing time of the current experimental group account for the target interaction information belongs to the first browsing category. At this time, the experimental group account is determined as a positive sample account. Obviously, since the positive sample account falls into the first browsing category ranked in the first N1 positions, it will necessarily have a high degree of matching with the returned target interaction information.

[0203] In step 404, for each experimental group account, if the server determines the experimental group account as a negative sample account if the historical browsing time of the experimental group account to the returned target interaction information belongs to the second browsing category.

[0204] The second browsing category refers to the browsing category that is the second-to-last target position in the sorting of the multiple browsing categories obtained in step 402, according to the average browsing time from largest to smallest. In other words, the server sorts the multiple browsing categories obtained in step 402 according to the average browsing time (e.g., the median of the browsing time interval) from largest to smallest, and determines the browsing category that is the last N2 position (i.e., the last N2 position) as the second browsing category. Here, N2 is the number of digits in the second target position, and N2 is an integer greater than or equal to 1, for example, N2 = 3.

[0205] To illustrate, assuming N1 = N2 = 3, after obtaining 10 browsing categories, the 10 browsing categories are sorted in descending order of average browsing time. The top 3 browsing categories (ranked 1, 2, and 3) are taken as the first browsing category, and the bottom 3 browsing categories (ranked 10, 9, and 8) are taken as the second browsing category.

[0206] In some embodiments, for each experimental group account, from the multiple browsing categories obtained in step 402 above, the target browsing category to which the historical browsing time of the experimental group account for the target interaction information belongs is determined. For example, when the browsing category is provided as a browsing time range, the target browsing category refers to the browsing time range into which the historical browsing time of the experimental group account for the target interaction information falls. If the target browsing category is the second-to-last target category in the sorting of the multiple browsing categories according to the average browsing time from largest to smallest, in other words, if the target browsing category is the last N2 positions in the above sorting, it is determined that the current experimental group account's historical browsing time for the target interaction information belongs to the second browsing category. In this case, the experimental group account is identified as a negative sample account. Obviously, since the negative sample account falls into the second browsing category ranked last N2, it means that the user quickly swiped past it, and therefore has a low degree of matching with the returned target interaction information.

[0207] In the above process, the experiment was essentially binned according to the historical browsing duration of the control group accounts, resulting in multiple browsing categories that met the required number of bins. Each browsing category can be considered as a range of historical browsing duration values ​​corresponding to an experimental bin. Next, for each experimental group's data, based on the experimental bin (i.e., the target browsing category) to which the historical browsing duration of the experimental group's accounts fell, it was determined whether the experimental group's accounts were positive sample accounts, negative sample accounts, or low-quality samples that needed to be filtered.

[0208] In some embodiments, each positive sample account is added to the positive sample set, and each negative sample account is added to the negative sample set to facilitate the storage of positive and negative samples. Since the average browsing time of the first browsing category to which each positive sample account belongs is greater than the average browsing time of the second browsing category to which each negative sample account belongs, the historical browsing time of the positive sample account on the target interaction information is greater than the historical browsing time of the negative sample account on the target interaction information. The positive sample accounts and negative sample accounts together constitute the sample accounts for this training.

[0209] In an exemplary scenario, taking a video as the content item, video comments as the interactive information, and the duration of viewing the interactive information as the dwell time in the comment section as an example, the A / B testing system sets up an experimental (exp) group and a control (base) group. The base group does not request the recommendation XTR model at all, while the exp group requests the recommendation XTR model. In the base group, the distribution of the dwell time of all comment section readers (i.e., users corresponding to the control group accounts) in the comment section of a certain video within a time slice (e.g., 1 hour) is statistically analyzed, and these dwell times are divided into 10 experimental buckets at equal intervals (i.e., 10 browsing categories are obtained). In the exp group, the dwell time of each comment section reader (i.e., users corresponding to the experimental group accounts) in the comment section of the same video within the same time slice is obtained, and then the dwell time of each experimental group account is assigned to the experimental buckets divided by the base group, thus determining the target browsing category. If any experimental group account stays in the comment section of the exp group for the duration that matches the first 3 experimental buckets of the base group (belonging to the first browsing category), it is marked as a positive sample account. If it matches the last 3 experimental buckets of the base group (belonging to the second browsing category), it is marked as a negative sample account.

[0210] In steps 402-404 above, the server selects sample accounts from the experimental group accounts based on the historical browsing information of the control group accounts. These sample accounts are experimental group accounts that have initiated historical access requests for any content item's interaction information in the A / B test. The sample accounts include positive sample accounts and negative sample accounts. This selection process sets a first browsing category and a second browsing category for browsing situations. The first browsing category can be used to select the positive sample set composed of each positive sample account, and the second browsing category can be used to select the negative sample set composed of each negative sample account, without the need for manual labeling of positive and negative samples, thus realizing the automated construction of positive and negative sample sets.

[0211] In step 405, for each sample account that has initiated a historical access request to the content item's interaction information, the server obtains the sample account's account characteristics, the global information characteristics of the interaction information, the triggering scenario characteristics of the historical access request, and the sample account's historical browsing information.

[0212] The trigger scenario feature is used to characterize the features of the interface that triggers the historical access request.

[0213] The sample account is an experimental group account that has initiated historical access requests for any content item in the A / B test and has passed the screening in steps 402-404 above. The sample account includes positive sample accounts and negative sample accounts.

[0214] In some embodiments, for each positive sample account and each negative sample account, account features, global information features, trigger scenario features and historical browsing information are obtained. The acquisition methods of the above account features, global information features and trigger scenario features are referred to step 301 in the above embodiments. The acquisition can be achieved by simply replacing the target account with the sample account, and will not be described again here.

[0215] The following will explain how the historical browsing information of the sample account is obtained. This historical browsing information is used to characterize the historical browsing of the target interaction information of the requested content item by the sample account. For example, the historical browsing information is the historical browsing duration itself, or the historical browsing information is the reference browsing category to which the historical browsing duration belongs, or the historical browsing information is the descriptive tag information corresponding to the reference browsing category to which the historical browsing duration belongs. For example, when the browsing information acquisition model is a binary classification model, it involves two types of tags: the first type of tag ("Yes") and the second type of tag ("No"). The first duration interval corresponds to the first type of tag ("Yes"), and the second duration interval corresponds to the second type of tag ("No").

[0216] In some embodiments, taking historical browsing information as a reference browsing category as an example, after obtaining positive sample accounts and negative sample accounts, it is necessary to set their respective reference browsing categories. The reference browsing category is the actual browsing category to which the sample account belongs. This reference browsing category can be regarded as historical browsing information determined based on the target browsing category to which the actual historical browsing duration belongs.

[0217] In some embodiments, when the browsing information acquisition model is a binary classification model, two types of labels are involved: a first type label (“Yes”) and a second type label (“No”). The first type label, “Yes”, indicates that the account’s browsing of interactive information conforms to preset browsing conditions, such as the account’s browsing time of interactive information is greater than a time threshold. The second type label, “No”, indicates that the account’s browsing of interactive information does not conform to preset browsing conditions, such as the account’s browsing time of interactive information is less than or equal to the time threshold.

[0218] To illustrate, the first type of label, "Yes," is configured as the reference browsing category for each positive sample account. For example, setting the reference browsing category Label for a positive sample account to 1 represents configuring it as the first type of label, indicating that the corresponding access request is a request that needs to be given priority, and that personalized recommendations using the XTR recommendation model will yield higher benefits for this access request. On the other hand, the second type of label, "No," is configured as the reference browsing category for each negative sample account. For example, setting the reference browsing category Label for a negative sample account to 0 represents configuring it as the second type of label, indicating that the corresponding access request is not a priority request and that personalized recommendations do not need to be enabled.

[0219] In step 406, for each sample account, the server inputs the account features, the global information features, and the triggering scenario features into the initial information acquisition model to obtain the target browsing information of the sample account.

[0220] In some embodiments, the account features, global information features, and trigger scenario features of each sample account are fused to obtain sample fusion features. The sample fusion features are then input into the initial information acquisition model. Using a method similar to steps 303-304 above, the initial information acquisition model obtains the browsing category with the highest matching degree for the sample account. Based on the browsing category, the target browsing information of the sample account is determined. For example, the target browsing information is the browsing category with the highest matching degree for the sample account output by the initial information acquisition model.

[0221] In step 407, the server iteratively trains the initial information acquisition model based on the historical browsing information and target browsing information of each sample account for each content item, and obtains the browsing information acquisition model. This browsing information acquisition model is used to acquire the target browsing information of the account's expected interaction with the content item.

[0222] In some embodiments, the server obtains a loss function value based on the historical browsing information and target browsing information of each sample account for each content item. Schematic, historical browsing information refers to the reference browsing category of the sample account, and target browsing information refers to the predicted browsing category of the sample account (i.e., the browsing category output by the initial information acquisition model that best matches the sample account). When the reference browsing category and the predicted browsing category of the sample account are consistent, it means that the prediction has obtained an accurate result; when the reference browsing category and the predicted browsing category of the sample account are inconsistent, it means that the prediction has obtained an incorrect result. Based on the reference browsing category and the predicted browsing category of each sample account, a loss function value is constructed to characterize the degree of difference between the actual result and the predicted result.

[0223] In some embodiments, when obtaining the loss function value based on the predicted browsing category and reference browsing category of each sample account, taking the XGBoost tree model as an example as the initial information acquisition model, the XGBoost tree model performs a second-order Taylor expansion of the loss function during training, using both first and second derivatives to accelerate the gradient descent process. Furthermore, the XGBoost tree model introduces a regularization term to control model complexity, adds handling for missing values, and supports data sampling, effectively preventing overfitting and improving the model's generalization ability. In addition, the XGBoost tree model significantly optimizes algorithm efficiency, resulting in fast training speed. Therefore, choosing the XGBoost tree model as the model structure for the browsing information acquisition model allows for accurate prediction of the overall dwell time of any account in the current video comment section with relatively low training cost. In one example, when the XGBoost tree model is a binary classification model, the binary logistic regression loss function, i.e., cross-entropy, is used. In other words, the loss function value is calculated based on the cross-entropy of the predicted browsing category and reference browsing category of each sample account.

[0224] After obtaining the loss function value for this iteration, it is determined whether the loss function value meets the stopping iteration condition. Optionally, the stopping iteration condition includes at least one of the following: the loss function value is less than the loss threshold; or, the number of iterations is greater than the iteration threshold. The loss threshold is any value greater than or equal to 0 and less than or equal to 1, for example, a loss threshold of 0.2. The iteration threshold is an integer greater than or equal to 1, for example, an iteration threshold of 1000.

[0225] In some embodiments, when the loss function value does not meet the stopping iteration condition, the server iteratively trains the initial information acquisition model; when the loss function value meets the stopping iteration condition, the server stops training the initial information acquisition model, thus obtaining the final browsing information acquisition model that needs to be put online and used to limit the rate of requests sent to the recommendation model.

[0226] In some embodiments, at each iteration, it is necessary to adjust the model parameters of the initial information acquisition model. Optionally, cross-validation is used to search for the optimal values ​​of key parameters of the XGBoost tree model in a grid-like manner. For example, these optimal values ​​include, but are not limited to: the maximum depth of the decision tree (max_depth), the minimum weights of all observations of a node (min_child_weight), the sample sampling rate (subsample), and the learning rate (learning_rate).

[0227] After the initial information acquisition model is trained offline, a set of model parameters for the browsing information acquisition model can be obtained. This browsing information acquisition model can be used in the previous embodiment for online prediction. Optionally, the set of model parameters is stored in JSON (JavaScript Object Notation). The set of model parameters includes, but is not limited to, all nodes on each decision tree. Each node stores the node's depth, the features used for splitting, the threshold of the splitting features, whether the node has leaf nodes, and information about the leaf nodes.

[0228] All of the above-mentioned optional technical solutions can be combined in any way to form optional embodiments of this disclosure, and will not be described in detail here.

[0229] In this embodiment of the disclosure, by training the initial information acquisition model offline, the resulting browsing information acquisition model can accurately obtain the browsing information of any account, such as the browsing category with the highest matching degree to the account. The information obtained by the browsing information acquisition model can be used to perform personalized rate limiting on requests sent to the recommendation model, that is, retaining personalized recommendations for requests whose browsing duration meets the target conditions, but restricting personalized recommendations for requests whose browsing duration does not meet the target conditions. This can greatly optimize the resource configuration of the server under the condition of limited computing resources.

[0230] Figure 5 This is a logical structure block diagram of a request processing apparatus according to an exemplary embodiment. (Refer to...) Figure 5 The device includes a first acquisition unit 501, a second acquisition unit 502, and a recommendation unit 503.

[0231] The first acquisition unit 501 is configured to execute an access request in response to the target account’s interaction with content items, and acquire the account characteristics of the target account, the global information characteristics of the interaction information, and the triggering scenario characteristics of the access request. The triggering scenario characteristics are used to characterize the features of the interface that triggers the access request.

[0232] The second acquisition unit 502 is configured to acquire target browsing information of the target account based on the account characteristics, the global information characteristics and the triggering scenario characteristics. The target browsing information is used to characterize the expected browsing behavior of the target account for the content item.

[0233] Recommendation unit 503 is configured to send a recommendation request to the recommendation model when the target browsing information meets the target conditions. The recommendation request is used to request target interaction information that matches the target account from the interaction information of the content item.

[0234] The apparatus provided in this embodiment of the present disclosure, by taking into account the access requests for interactive information issued by each account, and by utilizing account characteristics, global information characteristics of interactive information, and triggering scenario characteristics of access requests, can quickly and accurately predict the overall dwell time of the current account on the interactive information of the content item, i.e. the target browsing information. Only when the target browsing information meets the target conditions is a request for personalized recommendation sent to the recommendation model initiated; otherwise, there is no need to send the above request to the recommendation model. Under the premise of limited resources, the number of requests to the recommendation model can be greatly reduced, thereby optimizing the resource configuration of the server.

[0235] In one possible implementation, based on Figure 5 The device comprises, wherein the second acquisition unit 502 includes:

[0236] The fusion subunit is configured to perform the fusion of the account feature, the global information feature, and the trigger scenario feature to obtain a fused feature.

[0237] The acquisition subunit is configured to execute the input of the fused feature into the browsing information acquisition model, and obtain the browsing category with the highest matching degree to the target account through the browsing information acquisition model;

[0238] The determination subunit is configured to perform the determination of the target browsing information based on the browsing category.

[0239] In one possible implementation, the browsing information acquisition model is a composite tree model that integrates multiple tree models, each tree model being used to determine the contribution of a feature component to the browsing situation category using a class of feature components.

[0240] This acquisition subunit is configured to execute:

[0241] The fused feature is input into each tree model in the browsing information acquisition model. The fused feature is then split through the tree model to obtain the feature components in the fused feature that correspond to the tree model.

[0242] Based on the feature threshold of each node in the tree model, the leaf node in which the feature component falls in the tree model is determined, and the leaf node score corresponding to the leaf node is determined as the contribution of the feature component to the browsing category.

[0243] Based on multiple contribution values ​​output by the tree model, the browsing category with the highest matching degree with the target account is determined.

[0244] In one possible implementation, based on Figure 5 The device comprises:

[0245] The return module is configured to return the target interaction information indicated by the recommendation model to the target account if the target browsing information meets the target conditions.

[0246] The return module is also configured to return preset interactive information of the content item to the target account if the target browsing information does not meet the target conditions.

[0247] In one possible implementation, the account features include at least one of the attribute features or the behavioral features of the target account. The attribute features are used to characterize the account attributes of the target account, and the behavioral features are used to characterize the behavior of the target account towards content items within a preset time period.

[0248] In one possible implementation, the global information feature includes global statistical information on the various interactive information of the content item in at least one business metric.

[0249] All of the above-mentioned optional technical solutions can be combined in any way to form optional embodiments of this disclosure, and will not be described in detail here.

[0250] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments relating to the request processing method, and will not be elaborated upon here.

[0251] Figure 6 This is a logical structure block diagram of a training device for a browsing information acquisition model, according to an exemplary embodiment. (Refer to...) Figure 6 The device includes a determination unit 601, an acquisition unit 602, an output unit 603, and a training unit 604.

[0252] The determining unit 601 is configured to perform the following actions: for any content item, determine the sample account that has initiated historical access requests for the interaction information of that content item.

[0253] The acquisition unit 602 is configured to acquire the account characteristics of the sample account, the global information characteristics of the interaction information, the triggering scenario characteristics of the historical access request, and the historical browsing information of the sample account. The triggering scenario characteristics are used to characterize the features of the interface that triggers the historical access request.

[0254] The output unit 603 is configured to execute the input of the account features, the global information features and the triggering scenario features into the initial information acquisition model to obtain the target browsing information of the sample account.

[0255] Training unit 604 is configured to iteratively train the initial information acquisition model based on the historical browsing information and target browsing information of each sample account for each content item, thereby obtaining a browsing information acquisition model. This browsing information acquisition model is used to acquire the target browsing information of the account's expected interaction with the content item.

[0256] The apparatus provided in this disclosure, through offline training of an initial information acquisition model, can accurately obtain the browsing information of any account for any given account after training. For example, it can obtain the browsing category with the highest matching degree to the account. The information obtained by the browsing information acquisition model can be used to perform personalized rate limiting on requests sent to the recommendation model. That is, personalized recommendations are retained for requests whose browsing duration meets the target conditions, but personalized recommendations are restricted for requests whose browsing duration does not meet the target conditions. This can greatly optimize the resource configuration of the server when computing resources are limited.

[0257] In one possible implementation, the training unit 604 is configured to perform:

[0258] Based on the historical browsing information and target browsing information of each sample account for each content item, the loss function value is obtained;

[0259] If the loss function value does not meet the stopping iteration condition, the initial information acquisition model is iteratively trained.

[0260] When the loss function value meets the stopping iteration condition, the training of the initial information acquisition model is stopped, and the browsing information acquisition model is obtained.

[0261] In one possible implementation, based on Figure 6 The device comprises, wherein the determining unit 601 includes:

[0262] The test subunit is configured to perform A / B testing on multiple accounts that initiated the historical access request. In the A / B test, the multiple accounts are divided into experimental group accounts and control group accounts. The target interaction information matching the experimental group account is returned to the experimental group account, and the preset interaction information of the content item is returned to the control group account.

[0263] The filtering subunit is configured to perform a process of filtering the experimental group accounts based on the historical browsing information of the control group accounts to obtain the sample accounts.

[0264] In one possible implementation, the sample account includes a positive sample account and a negative sample account, wherein the historical browsing time of the target interaction information by the positive sample account is greater than the historical browsing time of the target interaction information by the negative sample account.

[0265] In one possible implementation, based on Figure 6 The device comprises a screening subunit including:

[0266] The first determining sub-unit is configured to perform the determination of multiple browsing categories based on the historical browsing information of the control group account. The browsing category is determined based on the range of historical browsing duration of the control group account for the preset interactive information. The browsing category includes a first browsing category and a second browsing category, and the average browsing duration corresponding to the first browsing category is greater than the average browsing duration corresponding to the second browsing category.

[0267] The second determination sub-unit is configured to perform the following action: for each experimental group account, if the historical browsing duration of the experimental group account for the target interaction information belongs to the first browsing category, the experimental group account is determined as a positive sample account.

[0268] The second determining sub-unit is also configured to determine the experimental group account as a negative sample account if the historical browsing duration of the experimental group account's interaction information with the target belongs to the second browsing category.

[0269] In one possible implementation, the first browsing category is located in the first target position in the sorting of the multiple browsing categories in descending order of average browsing time, and the second browsing category is located in the second target position in the sorting.

[0270] For each experimental group account, this second determined sub-unit is also configured to execute:

[0271] From multiple browsing categories, determine the target browsing category to which the historical browsing duration of the target interaction information of the experimental group account belongs;

[0272] If the target browsing category is ranked first in the sorting, it is determined that the historical browsing duration of the experimental group account's interaction information with the target belongs to the first browsing category;

[0273] If the target browsing category is the second target in the sorting, it is determined that the historical browsing time of the experimental group account for the target's interaction information belongs to the second browsing category.

[0274] All of the above-mentioned optional technical solutions can be combined in any way to form optional embodiments of this disclosure, and will not be described in detail here.

[0275] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operations has been described in detail in the embodiments concerning the training method of the browsing information acquisition model, and will not be elaborated upon here.

[0276] Figure 7This is a schematic diagram of the structure of a computer device provided in an embodiment of this disclosure. Taking server 700 as an example, the server 700 can vary considerably due to different configurations or performance. It may include one or more Central Processing Units (CPUs) 701 and one or more memories 702. The memory 702 stores at least one line of program code, which is loaded and executed by the processor 701 to implement the request processing method or the training method for the browsing information acquisition model provided in the above embodiments. Of course, the server 700 may also have wired or wireless network interfaces, a keyboard, and input / output interfaces for input and output. The server 700 may also include other components for implementing device functions, which will not be elaborated here.

[0277] In an exemplary embodiment, a computer-readable storage medium including at least one instruction is also provided, such as a memory including at least one instruction, which can be executed by a processor in a computer device to complete the request processing method or the training method for a browsing information retrieval model in the above embodiments. Optionally, the computer-readable storage medium may be a non-transitory computer-readable storage medium, such as ROM (Read-Only Memory), RAM (Random-Access Memory), CD-ROM (Compact Disc Read-Only Memory), magnetic tape, floppy disk, and optical data storage device, etc.

[0278] In an exemplary embodiment, a computer program product is also provided, including one or more instructions that can be executed by a processor of a computer device to complete the request processing method or the training method for the browsing information acquisition model provided in the above embodiments.

[0279] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0280] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A request processing method characterized by comprising: The method comprises: in response to an access request of a target account for interaction information of a content item, obtaining account features of the target account, global information features of the interaction information, and trigger scene features of the access request, the trigger scene features being used to represent features of an interface triggering the access request; fusing the account features, the global information features, and the trigger scene features to obtain fused features; inputting the fused features into each tree model in a browsing information acquisition model, performing feature splitting on the fused features by the tree model to obtain feature components of the fused features corresponding to the tree model, the browsing information acquisition model being a composite tree model integrated by multiple tree models; based on feature thresholds of each node in the tree model, determining a leaf node into which the feature components fall in the tree model, and determining a leaf node score corresponding to the leaf node as a contribution degree of the feature components to a browsing situation category; based on multiple contribution degrees output by the multiple tree models, determining a browsing situation category with the highest matching degree to the target account; based on the browsing situation category, determining target browsing information of the target account, the target browsing information being used to represent a browsing situation of the target account for the interaction information of the content item; in a case where the target browsing information meets a target condition, sending a recommendation request for the target account to a recommendation model, the recommendation request being used to request target interaction information matched to the target account from the interaction information of the content item.

2. The method of claim 1, wherein, The method further comprises: in a case where the target browsing information meets the target condition, returning the target interaction information indicated by the recommendation model to the target account; in a case where the target browsing information does not meet the target condition, returning preset interaction information of the content item to the target account.

3. The method according to claim 1 or 2, characterized in that, The account features comprise at least one of attribute features of the target account or behavior features of the target account, the attribute features being used to represent account attributes of the target account, and the behavior features being used to represent features of behaviors of the target account for content items in a preset time period.

4. The method according to claim 1 or 2, characterized in that, The global information features comprise global statistical information of each interaction information of the content item on at least one business indicator.

5. A training method of a browsing information acquisition model, characterized by, The method comprises: performing AB testing on multiple accounts that have initiated historical access requests for any content item, wherein the historical access requests are initiated for interaction information of the content item, in the AB testing, the multiple accounts are divided into an experimental group and a control group, target interaction information matched to the accounts in the experimental group is returned to the accounts in the experimental group, and preset interaction information of the content item is returned to the accounts in the control group; based on historical browsing information of the accounts in the control group, sample accounts that have initiated historical access requests for the interaction information of the content item are screened from the accounts in the experimental group, the sample accounts comprising positive sample accounts and negative sample accounts, a historical browsing time length of the positive sample accounts for the target interaction information being greater than a historical browsing time length of the negative sample accounts for the target interaction information; obtaining account features of the sample account, global information features of the interaction information, trigger scene features of the historical access request, and historical browsing information of the sample account, the trigger scene features being used to represent features of an interface triggering the historical access request; inputting the account features, the global information features, and the trigger scene features into an initial information obtaining model to obtain target browsing information of the sample account; iteratively training the initial information obtaining model based on the historical browsing information and the target browsing information of each sample account of each content item to obtain a browsing information obtaining model, the browsing information obtaining model being used to obtain target browsing information of interaction information of a content item expected by an account.

6. The method of claim 5, wherein, The iteratively training the initial information obtaining model based on the historical browsing information and the target browsing information of each sample account of each content item to obtain a browsing information obtaining model includes: obtaining a loss function value based on the historical browsing information and the target browsing information of each sample account of each content item; iteratively training the initial information obtaining model when the loss function value does not meet a stop iteration condition; stopping training of the initial information obtaining model to obtain the browsing information obtaining model when the loss function value meets the stop iteration condition.

7. The method of claim 5, wherein, The filtering the sample account from the experimental group accounts based on the historical browsing information of the control group accounts includes: determining a plurality of browsing situation categories based on the historical browsing information of the control group accounts, the browsing situation categories being determined based on value intervals of historical browsing time lengths of the control group accounts on the preset interaction information; wherein the browsing situation categories include a first browsing category and a second browsing category, an average browsing time length corresponding to the first browsing category being greater than an average browsing time length corresponding to the second browsing category; determining each experimental group account as a positive sample account when a historical browsing time length of the experimental group account on the target interaction information belongs to the first browsing category; determining the experimental group account as a negative sample account when the historical browsing time length of the experimental group account on the target interaction information belongs to the second browsing category.

8. The method of claim 7, wherein, The first browsing category is located at a first target position in a sequence of the plurality of browsing situation categories in descending order of the average browsing time length, and the second browsing category is located at a second target position in the sequence; For each experimental group account, the method further includes: determining a target browsing category to which a historical browsing time length of the experimental group account on the target interaction information belongs from the plurality of browsing situation categories; determining that the historical browsing time length of the experimental group account on the target interaction information belongs to the first browsing category when the target browsing category is located at the first target position in the sequence; determining that the historical browsing time length of the experimental group account on the target interaction information belongs to the second browsing category when the target browsing category is located at the second target position in the sequence.

9. A request processing apparatus characterized by comprising: includes: The first obtaining unit is configured to obtain account features of the target account, global information features of the interaction information, and trigger scenario features of the access request, in response to an access request of the target account to the interaction information of the content item, the trigger scenario features being used to represent features of an interface triggering the access request; The second obtaining unit is configured to fuse the account features, the global information features, and the trigger scenario features to obtain fused features, input the fused features into each tree model in a browsing information obtaining model, perform feature splitting on the fused features by using the tree model to obtain feature components of the fused features corresponding to the tree model, the browsing information obtaining model being a composite tree model integrated by multiple tree models, determine leaf nodes into which the feature components fall in the tree model based on feature thresholds of each node in the tree model, determine a leaf node score corresponding to the leaf nodes as a contribution degree of the feature components to a browsing situation category, and determine a browsing situation category with a highest matching degree to the target account based on multiple contribution degrees output by the multiple tree models. The target browsing information of the target account is determined based on the browsing situation category, and the target browsing information is used to represent a browsing situation of the target account to the interaction information of the content item. The recommendation unit is configured to send a recommendation request for the target account to a recommendation model in a case where the target browsing information meets a target condition, the recommendation request being used to request target interaction information matched to the target account from the interaction information of the content item.

10. The apparatus of claim 9, wherein, The device further includes: The returning module is configured to return the target interaction information indicated by the recommendation model to the target account in a case where the target browsing information meets the target condition. The returning module is further configured to return preset interaction information of the content item to the target account in a case where the target browsing information does not meet the target condition.

11. The apparatus of claim 9 or 10, wherein, The account features include at least one of attribute features of the target account or behavior features of the target account, the attribute features being used to represent account attributes of the target account, and the behavior features being used to represent features of behaviors of the target account to content items in a preset time period.

12. The apparatus of claim 9 or 10, wherein, The global information features include global statistical information of each interaction information of the content item on at least one business indicator.

13. A training apparatus of a browsing information acquisition model, characterized by comprising: The device further includes: The determining unit is configured to perform AB testing on any content item and a plurality of account numbers that have initiated a historical access request; wherein the historical access request is initiated for interaction information of the content item, in the AB testing, the plurality of account numbers are divided into an experimental group account number and a control group account number, the experimental group account number is returned target interaction information matched with the experimental group account number, and the control group account number is returned preset interaction information of the content item; based on historical browsing information of the control group account number, sample account numbers that have initiated the historical access request for the interaction information of the content item are screened from the experimental group account numbers, the sample account numbers include positive sample account numbers and negative sample account numbers, and a historical browsing time length of the positive sample account numbers for the target interaction information is greater than a historical browsing time length of the negative sample account numbers for the target interaction information; The obtaining unit is configured to perform obtaining account number features of the sample account numbers, global information features of the interaction information, trigger scene features of the historical access request, and historical browsing information of the sample account numbers, and the trigger scene features are used to represent features of an interface triggering the historical access request; The output unit is configured to perform inputting the account number features, the global information features, and the trigger scene features into an initial information obtaining model to obtain target browsing information of the sample account numbers. The training unit is configured to perform iterative training of the initial information obtaining model based on historical browsing information and target browsing information of each sample account number of each content item to obtain a browsing information obtaining model, and the browsing information obtaining model is used to obtain target browsing information of an account number for interaction information of a content item.

14. The apparatus of claim 13, wherein, The training unit is configured to perform: obtaining a loss function value based on historical browsing information and target browsing information of each sample account number of each content item; iterative training of the initial information obtaining model when the loss function value does not meet a stop iteration condition; stopping training of the initial information obtaining model to obtain the browsing information obtaining model when the loss function value meets the stop iteration condition.

15. The apparatus of claim 13, wherein, The determining unit includes: A first determining sub-subunit is configured to perform determining a plurality of browsing situation categories based on historical browsing information of the control group account numbers, the browsing situation categories are determined based on a value interval of a historical browsing time length of the control group account numbers for the preset interaction information; wherein the browsing situation categories include a first browsing category and a second browsing category, and an average browsing time length corresponding to the first browsing category is greater than an average browsing time length corresponding to the second browsing category; A second determining sub-subunit is configured to perform, for each experimental group account number, determining the experimental group account number as a positive sample account number when a historical browsing time length of the experimental group account number for the target interaction information belongs to the first browsing category; The second determining sub-subunit is further configured to perform, for each experimental group account number, determining the experimental group account number as a negative sample account number when a historical browsing time length of the experimental group account number for the target interaction information belongs to the second browsing category.

16. The apparatus of claim 15, wherein, The first browsing category is located in a first target position in a sorting of the plurality of browsing situation categories in descending order of average browsing time length, and the second browsing category is located in a second target position in the sorting; For each experimental group account, the second determining subunit is further configured to perform: From the plurality of browsing situation categories, determine a target browsing category to which a historical browsing time length of the experimental group account for the target interactive information belongs; In a case where the target browsing category is located in the first target position in the sorting, determine that the historical browsing time length of the experimental group account for the target interactive information belongs to the first browsing category; In a case where the target browsing category is located in the second target position in the sorting, determine that the historical browsing time length of the experimental group account for the target interactive information belongs to the second browsing category.

17. A computer device, comprising: Comprise: One or more processors; One or more memories for storing instructions executable by the one or more processors; Wherein the one or more processors are configured to execute the instructions to implement the request processing method of any one of claims 1 to 4; Or, the training method of the browsing information acquisition model of any one of claims 5 to 8.

18. A computer-readable storage medium, characterized in that, When at least one instruction in the computer-readable storage medium is executed by one or more processors of a computer device, the computer device can execute the request processing method of any one of claims 1 to 4; or, the training method of the browsing information acquisition model of any one of claims 5 to 8.

Images