Recommended information delivery method, computer device, and storage medium

Abstract

The application relates to a recommended information delivery method, a computer device and a storage medium, and relates to the technical field of artificial intelligence. The method comprises the following steps: performing delivery testing on a first account set in an account pool to obtain a target account set; obtaining the similarity between the group features of the target account set and the features of each account in a second account set; according to the similarity, screening out to-be-added accounts with similar features from the second account set, adding the to-be-added accounts to the target account set, and obtaining an expanded account set; and in the case that any to-be-recommended account belongs to the expanded account set, delivering recommended information associated with a recommended category to the to-be-recommended account. The method can improve the effectiveness of recommended information delivery.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, computer device, storage medium, and computer program product for delivering recommendation information. Background Technology

[0002] In the field of artificial intelligence, it is necessary to study the impact of interventions on individual user account outcomes. For example, in the process of recommending and delivering marketing information, it is necessary to analyze the incremental impact of issuing coupons and other benefits on users' willingness to use platform services, thereby assisting in decisions on whether to issue such benefits.

[0003] However, current information recommendation and delivery methods require the use of random traffic to verify results, which can damage the effectiveness of online recommendations and result in low effectiveness of information delivery. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for delivering recommendation information that can improve the effectiveness of recommendation information delivery, in order to address the above-mentioned technical problems.

[0005] Firstly, embodiments of this application provide a method for delivering recommendation information. The method includes:

[0006] A campaign test is conducted on the first set of accounts in the account pool to obtain a target set of accounts; the target set of accounts includes accounts in the first set of accounts that have entered interactive operations for at least one recommended category recommended in the campaign test.

[0007] Obtain the similarity between the group characteristics of the target account set and the characteristics of each account in the second account set; the second account set is the account set in the account pool other than the first account set;

[0008] Based on the similarity, accounts with similar characteristics to those in the target account set are selected from the second account set and added to the target account set to obtain the expanded account set.

[0009] For any account to be recommended, if the account belongs to the expanded account set, recommendation information associated with the recommended category is delivered to the account.

[0010] In one embodiment, applied to a live streaming scenario, the method further includes:

[0011] Based on the account information of each account in the target account set, determine the category characteristics corresponding to each of the recommended categories;

[0012] Based on the behavioral information provided by each account in the first account set during the deployment test, behavioral characteristics are determined;

[0013] Based on the content displayed in real time by the terminal in the live streaming scenario during the deployment test, determine the real-time content characteristics;

[0014] The category characteristics, behavioral characteristics, and real-time content characteristics are used as the group characteristics of the target account set.

[0015] In one embodiment, obtaining the similarity between the group characteristics of the target account set and the characteristics of each account in the second account set includes:

[0016] For any account in the second account set, determine the account characteristics of the account based on the account information of the account.

[0017] Calculate sub-similarity between the account characteristics of each account and the category characteristics, the behavioral characteristics, and the real-time content characteristics, respectively;

[0018] The calculated sub-similarity scores are weighted and summed to obtain the total similarity score.

[0019] In one embodiment, determining the category characteristics corresponding to each of the recommended categories based on the account information of each account in the target account set includes:

[0020] Based on the account information of each account in the target account set, determine the account characteristics of each account;

[0021] For each of the recommended product categories, the average value of the account characteristics of each account is taken to obtain the category characteristics corresponding to each of the recommended product categories.

[0022] In one embodiment, the behavioral information includes information on positive behavior and information on negative behavior. The step of determining behavioral characteristics based on the behavioral information fed back by each account in the first account set during the deployment test includes:

[0023] Based on the account information corresponding to the accounts that have positive behavior towards the recommended category in the first account set, positive account characteristics are determined, and the positive account characteristics are averaged to obtain positive behavior characteristics.

[0024] Based on the account information corresponding to the accounts that have negative behavior towards the recommended category in the first account set, negative account characteristics are determined, and the average value of the negative account characteristics is taken to obtain negative behavior characteristics.

[0025] The behavioral characteristics corresponding to the first account set are obtained by subtracting the negative behavioral characteristics from the positive behavioral characteristics.

[0026] In one embodiment, determining the real-time content characteristics based on the content displayed in real-time by the terminal in the live streaming scenario during the deployment test includes:

[0027] Based on the content displayed in real time by the terminal in the live streaming scenario during the deployment test, the characteristics of the multimedia content are determined;

[0028] Obtain account information for each streamer's account in the live streaming scenario;

[0029] Based on the account information of each account in the target account set and the account information of each broadcaster account, the spatial structure of the multimedia content features is fitted to obtain the fitted real-time content features.

[0030] In one embodiment, fitting the spatial structure of the multimedia content features based on the account information of each account in the target account set and the account information of each broadcaster account to obtain the fitted real-time content features includes:

[0031] Obtain the learning objective corresponding to the fitting process; the learning objective includes achieving a preset click-through rate for each account's click-through rate for the recommended information each time.

[0032] Using the learning objective as the fitting direction, the account information of each broadcaster's account and the multimedia content features are input into a multilayer perceptron network for fitting calculation, so that the multimedia content features are fitted to the fitting direction to obtain the fitted real-time content features.

[0033] In one embodiment, delivering recommendation information associated with the recommended category to any of the accounts to be recommended includes:

[0034] If the attention level of any of the accounts to be recommended to the recommended category meets a preset condition, and / or if the cumulative number of times the account to be recommended is delivered to the account to be recommended on a given day for each of the recommended categories meets a preset condition, then the recommended information associated with the recommended category will be delivered to the account to be recommended.

[0035] In one embodiment, the method further includes:

[0036] Obtain a pre-trained threshold testing model; the threshold testing model is a model obtained by training a binary classification task model using a positive sample set and a negative sample set; the positive sample set includes accounts in the first account set where the recommendation information has been displayed and feedback has been received regarding the recommendation information, and accounts in the first account set where the recommendation information has not been displayed; the negative sample set includes accounts in the first account set where the recommendation information has been displayed and feedback has not been received regarding the recommendation information.

[0037] The threshold test model is used to predict the probability value of any account to be recommended, thereby obtaining a probability value for each account to be recommended. The probability value is used to characterize the degree of attention that account to be recommended pays to the recommended category.

[0038] If the probability value is greater than a preset probability threshold, it is determined that the attention level of any account to be recommended to the recommended category meets the preset conditions.

[0039] In one embodiment, applied to a live streaming scenario, delivering recommendation information associated with the recommended category to any of the accounts to be recommended includes:

[0040] Based on the rating ranking of each streamer account under each of the recommended product categories, target streamers under the target product category are selected from each streamer account under each of the recommended product categories; wherein the rating ranking is based on the interaction history data of each streamer account in the live streaming scenario within a preset time period;

[0041] In the live streaming scenario of the target anchor under the target category, recommendation information related to the target category is delivered to any of the accounts to be recommended.

[0042] Secondly, this application also provides a device for distributing recommendation information. The device includes:

[0043] The campaign testing module is used to conduct campaign testing on a first set of accounts in the account pool to obtain a target set of accounts; the target set of accounts includes accounts in the first set of accounts that have entered interactive operations for at least one recommended category recommended in the campaign testing.

[0044] The similarity acquisition module is used to acquire the similarity between the group characteristics of the target account set and the characteristics of each account in the second account set; the second account set is the account set in the account pool other than the first account set.

[0045] An extension module is used to filter out accounts with similar characteristics to those in the target account set from the second account set based on the similarity, and add the accounts to be added to the target account set to obtain an extended account set.

[0046] The delivery module is used to deliver recommendation information associated with the recommended category to any account to be recommended, provided that the account belongs to the expanded account set.

[0047] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0048] A campaign test is conducted on the first set of accounts in the account pool to obtain a target set of accounts; the target set of accounts includes accounts in the first set of accounts that have entered interactive operations for at least one recommended category recommended in the campaign test.

[0049] Obtain the similarity between the group characteristics of the target account set and the characteristics of each account in the second account set; the second account set is the account set in the account pool other than the first account set;

[0050] Based on the similarity, accounts with similar characteristics to those in the target account set are selected from the second account set and added to the target account set to obtain the expanded account set.

[0051] For any account to be recommended, if the account belongs to the expanded account set, recommendation information associated with the recommended category is delivered to the account.

[0052] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0053] A campaign test is conducted on the first set of accounts in the account pool to obtain a target set of accounts; the target set of accounts includes accounts in the first set of accounts that have entered interactive operations for at least one recommended category recommended in the campaign test.

[0054] Obtain the similarity between the group characteristics of the target account set and the characteristics of each account in the second account set; the second account set is the account set in the account pool other than the first account set;

[0055] Based on the similarity, accounts with similar characteristics to those in the target account set are selected from the second account set and added to the target account set to obtain the expanded account set.

[0056] For any account to be recommended, if the account belongs to the expanded account set, recommendation information associated with the recommended category is delivered to the account.

[0057] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps:

[0058] A campaign test is conducted on the first set of accounts in the account pool to obtain a target set of accounts; the target set of accounts includes accounts in the first set of accounts that have entered interactive operations for at least one recommended category recommended in the campaign test.

[0059] Obtain the similarity between the group characteristics of the target account set and the characteristics of each account in the second account set; the second account set is the account set in the account pool other than the first account set;

[0060] Based on the similarity, accounts with similar characteristics to those in the target account set are selected from the second account set and added to the target account set to obtain the expanded account set.

[0061] For any account to be recommended, if the account belongs to the expanded account set, recommendation information associated with the recommended category is delivered to the account.

[0062] The aforementioned method, apparatus, computer equipment, storage medium, and computer program product for delivering recommendation information first conducts a delivery test on a first set of accounts in the account pool to obtain a target set of accounts. Then, it obtains the similarity between the group characteristics of the target set and the characteristics of each account in a second set of accounts. Based on the similarity, it filters out accounts with similar characteristics to those in the target set from the second set of accounts and adds them to the target set, resulting in an expanded set of accounts. Finally, for any account to be recommended, if the account belongs to the expanded set of accounts, it delivers recommendation information related to the recommended category to that account. By eliminating the need for random traffic delivery to online accounts, it avoids losses online. It successfully deploys category delivery functionality based on a gain model without random delivery processes. Furthermore, addressing the category divergence problem inherent in the incremental model, it combines an audience targeting algorithm to expand the target set through similarity, achieving a combination of incremental model and audience targeting algorithms. This results in a large, usable expanded user set, facilitating direct online delivery of recommendation information and improving the effectiveness of recommendation information delivery. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is an application environment diagram of a recommendation information delivery method in one embodiment;

[0065] Figure 2 This is a flowchart illustrating a method for delivering recommendation information in one embodiment;

[0066] Figure 3 This is a flowchart illustrating the method for delivering recommendation information in another embodiment;

[0067] Figure 4 This is an overall flowchart of the method for delivering recommendation information in one embodiment;

[0068] Figure 5 This is a schematic diagram illustrating the fitting principle of a recommendation information delivery method in one embodiment;

[0069] Figure 6 This is a structural block diagram of a device for distributing recommendation information in one embodiment;

[0070] Figure 7 This is an internal structural diagram of a computer device in one embodiment;

[0071] Figure 8 This is a diagram of the internal structure of a computer device in another embodiment. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0073] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data 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.

[0074] The method for delivering recommendation information provided in this application embodiment can be applied to, for example, Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network.

[0075] The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle systems. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0076] In one embodiment, such as Figure 2 As shown, a method for delivering recommendation information is provided, which can be applied to... Figure 1 Taking terminal 102 as an example, the explanation includes the following steps:

[0077] S201, Conduct a deployment test on the first set of accounts in the account pool to obtain the target set of accounts;

[0078] The account pool refers to all accounts associated with the recommended product category that can be used to support data during the campaign. For example, the accounts of all live viewers who have followed live streams in the e-commerce and bullet screen game categories, or all users in the existing database of the live stream platform in the e-commerce and bullet screen game categories.

[0079] The first set of accounts refers to the set of accounts in the account pool that can be used for testing. Generally, the data volume of the first set of accounts accounts for a small proportion of the total data volume of the account pool. In the incremental model algorithm, test deployments can be carried out on users in the first set of accounts, and subsequent work can be carried out based on the test results.

[0080] The target account set includes accounts that have entered interactive actions for at least one recommended category in the first account set for the test campaign. Generally, accounts that have entered interactive actions are identified as accounts that are sensitive to the performance of that category. Therefore, the target account set is a collection of accounts that are sensitive to the performance of the recommended category.

[0081] In addition, recommended categories include, but are not limited to, e-commerce and bullet screen games; interactive operations refer to user interactions with the test input, such as clicking and confirming.

[0082] It should be noted that in the process of acquiring category-sensitive accounts through category-based trial deployments, the user screening criteria vary depending on the category associated with the account data. For example, in the e-commerce category, accounts that click on the recommended information can be identified as sensitive accounts; in the bullet screen game category, accounts that click on the recommended information and are confirmed to participate in the game can be identified as sensitive accounts. This application embodiment considers the different influences of categories on user preferences and applies different screening criteria to sensitive accounts based on the different types of recommended categories in the live streaming scenario, thereby improving the flexibility of sensitive account classification.

[0083] S202, obtain the similarity between the group characteristics of the target account set and the characteristics of each account in the second account set.

[0084] Among them, group characteristics refer to the characteristics possessed by all members of the target account set. Generally, group characteristics are reflected by the average index of each account and used to calculate similarity with other accounts.

[0085] The second account set is the set of accounts in the account pool other than the first account set. Generally, the data volume of the second account set is greater than that of the first account set.

[0086] S203. Based on similarity, select accounts with similar characteristics to those in the target account set from the second account set to be added, and add the accounts to be added to the target account set to obtain the expanded account set.

[0087] The filtering process can be one-by-one selection, that is, determining whether to add each account in the second account set to the target account set based on the similarity of each account. For example, when the similarity reaches 80%, it is taken as an account to be added and added to the target account set. The account set after the addition is the expanded account set.

[0088] For example, in the process of expanding the similarity of sensitive accounts, the lookalike algorithm, which is based on the user similarity of the lookalike algorithm, expands the category-sensitive accounts obtained after the trial investment to obtain a set of similar candidate sensitive accounts based on the user profile similarity, that is, the expanded account set.

[0089] S204, for any account to be recommended, if any account to be recommended belongs to the expanded account set, deliver recommendation information related to the recommended category to any account to be recommended.

[0090] Here, any account to be recommended refers to an account in the online process that is to be recommended information. Before determining whether to deliver the recommended information, it is necessary to determine whether the account belongs to the expanded account set. It can be seen that step S204 belongs to the application process of the online process, while steps S201 to S203 belong to the offline data preparation process.

[0091] For example, in determining whether an online account belongs to a category-sensitive user group, it is necessary to retrieve any account from the online process and determine whether it belongs to the extended sensitive account identified in the offline process. By retrieving data one by one and making real-time judgments, the online and offline processes are connected, making it easier to deploy the extended sensitive accounts obtained from the offline process online and improving the convenience of deployment.

[0092] The above-mentioned method for delivering recommendation information first tests the first set of accounts in the account pool to obtain the target account set. Then, it obtains the similarity between the group characteristics of the target account set and the characteristics of each account in the second account set. Based on the similarity, it filters out accounts with similar characteristics to those in the target account set from the second account set and adds them to the target account set to obtain the expanded account set. Finally, for any account to be recommended, if the account belongs to the expanded account set, it delivers recommendation information related to the recommended category to that account. By eliminating the need for random traffic delivery to online accounts, it avoids losses to the online platform. It successfully deploys the category delivery function based on the gain model without random process delivery. At the same time, to address the category divergence problem caused by the incremental model itself, it combines the audience targeting algorithm to expand the target account set through similarity, realizing the combination of the incremental model algorithm and the audience targeting algorithm. This results in a large number of usable expanded user sets, which facilitates the direct delivery of recommendation information online and improves the effectiveness of recommendation information delivery.

[0093] In one embodiment, applied to a live streaming scenario, the method further includes: determining category characteristics corresponding to each recommended category based on the account information of each account in the target account set; determining behavioral characteristics based on the behavioral information fed back by each account in the first account set during the campaign test; determining real-time content characteristics based on the content displayed in real-time by the terminal in the live streaming scenario during the campaign test; and using the category characteristics, behavioral characteristics, and real-time content characteristics as the group characteristics of the target account set.

[0094] Among them, the live streaming scenario refers to the scenario where recommended information is delivered to accounts watching the live stream under different product categories.

[0095] The target account set's group characteristics include category characteristics, behavioral characteristics, and real-time content characteristics. Specifically, the group characteristics can be represented by embedded features generated through text encoding, i.e., the group characteristics include category embedding, behavioral embedding, and real-time content embedding.

[0096] In this target account set, the account information for each account can be the account ID.

[0097] Among them, the behavioral information fed back during the deployment test refers to the information corresponding to the actions taken by the audience account in response to the recommended information after it is exposed on the live audience's terminal. For example, the recommended information is clicked by the audience account after it is exposed.

[0098] Among them, the terminal is the device on which the viewer's account watches the live broadcast, such as a smartphone; the real-time displayed content refers to the interface of the live broadcast entrance that the viewer's account is shown to the viewer's account on the terminal before entering the live broadcast room. The interface displays live broadcast-related content in real time, which guides the viewer's account to enter the live broadcast room for subsequent interactive operations through clicks and other actions.

[0099] In this embodiment, category features, behavioral features, and real-time content features are aggregated into group features of the target account set by embedding features. Taking into account the characteristics of the anchor's category, the characteristics of the audience's account behavior, and the characteristics of the real-time content displayed in the live broadcast room, a feature representation form that integrates category, anchor, and real-time content is successfully designed. This facilitates the subsequent use of group features for similarity calculation and account set expansion, enriches the data types and evaluation dimensions of group features, improves the accuracy of similarity calculation, and thus ensures the effectiveness of the expanded account set.

[0100] In one embodiment, obtaining the similarity between the group characteristics of the target account set and the characteristics of each account in the second account set includes: for any account in the second account set, determining the account characteristics of any account based on the account information of any account; calculating sub-similarity between the account characteristics of each account and the category characteristics, behavioral characteristics, and real-time content characteristics, respectively; and weighting and summing the calculated sub-similarity to obtain the similarity.

[0101] Here, "any account" refers to any account in the second account set, which is intended to illustrate that the process of determining account features and calculating sub-similarity is performed on an individual account-by-account basis.

[0102] Among them, account information is text information related to the account, and account features are the embedding features obtained based on the text encoding corresponding to the account, i.e., account embedding.

[0103] Sub-similarity refers to the similarity calculated by comparing the account characteristics with the category characteristics, behavioral characteristics, and real-time content characteristics of any account.

[0104] For example, when the sub-similarity between account features and category features is A and the weight coefficient is k, the sub-similarity between account features and behavioral features is B and the weight coefficient is m, and the sub-similarity between account features and real-time content features is C and the weight coefficient is n, the obtained similarity D is k×A + m×B + n×C.

[0105] In this embodiment, the embedding features corresponding to each account in the second account set are determined, and the similarity between the account embedding features and the three different categories of features in the target account set is calculated. Then, the weighted sum is obtained to obtain the final similarity, so as to reflect the similarity between the members of the target account set and the second account set, which facilitates the subsequent expansion of the account set and improves the comprehensiveness of the similarity calculation.

[0106] In one embodiment, determining the category characteristics corresponding to each recommended category based on the account information of each account in the target account set includes: determining the account characteristics of each account based on the account information of each account in the target account set; and taking the average value of the account characteristics of each account for each recommended category to obtain the category characteristics corresponding to each recommended category.

[0107] In this context, the account information of each account in the target account set refers to the text information related to the account, such as the account ID and the category to which the account belongs. The account features are the embedding features obtained based on the text encoding corresponding to the account.

[0108] For example, by parsing the account information of each account, the account ID and category of each account are determined. Then, the account embedding is obtained by encoding the text of the account ID. The category embedding is then averaged according to the category. Specifically, based on the users who clicked on the category, the embedding of all users who clicked on the category is obtained using a global representation, and then the average is taken as the category embedding.

[0109] More specifically, for example, accounts 1 to 10 belong to category A, and accounts 11 to 20 belong to category B. After obtaining the account embeddings for all accounts, it is necessary to average them for category A, taking the average of the account embeddings of accounts 1 to 10 as the category embedding for category A. Similarly, it is necessary to average them for category B, taking the average of the account embeddings of accounts 11 to 20 as the category embedding for category B.

[0110] In this embodiment, by averaging the account characteristics according to categories, we obtain characteristics that can characterize the categories involved in the target account set, thereby improving the richness of the group characteristic assessment.

[0111] In one embodiment, determining behavioral characteristics based on the behavioral information provided by each account in the first account set during the campaign testing includes: determining positive account characteristics based on the account information corresponding to accounts with positive behavior towards the recommended product category in the first account set; averaging the positive account characteristics to obtain positive behavioral characteristics; determining negative account characteristics based on the account information corresponding to accounts with negative behavior towards the recommended product category in the first account set; averaging the negative account characteristics to obtain negative behavioral characteristics; and subtracting the negative behavioral characteristics from the positive behavioral characteristics to obtain the behavioral characteristics corresponding to the first account set.

[0112] The behavioral information includes information on positive behavior and information on negative behavior. Specifically, positive behavior is defined as clicking after exposure, and negative behavior is defined as not clicking after exposure.

[0113] For example, accounts 21 to 30 clicked the recommended link after the exposure, which is considered positive behavior; accounts 31 to 40 did not click the recommended link after the exposure, which is considered negative behavior.

[0114] Specifically, based on the positive and negative behaviors of the broadcasters, the account IDs of the accounts under each behavior are obtained. Then, account embeddings for each behavior are generated based on the account IDs. For accounts with different behaviors, the average value is calculated. That is, the average value of the account embeddings under each behavior is taken to obtain the positive embedding and the negative embedding respectively. Finally, the negative embedding is subtracted from the positive embedding to obtain the behavior embedding, which is the behavior feature corresponding to the first account set.

[0115] In this embodiment, the behavioral features corresponding to the first account set are obtained by averaging and subtracting the features of positive and negative behaviors respectively. This comprehensively considers the impact of positive and negative behaviors on group performance and improves the accuracy of behavioral feature representation.

[0116] In one embodiment, determining real-time content features based on the content displayed in real-time by the terminal in a live-streaming scenario during the deployment test includes: determining multimedia content features based on the content displayed in real-time by the terminal in a live-streaming scenario during the deployment test; obtaining account information of each broadcaster account in the live-streaming scenario; and fitting the spatial structure of the multimedia content features based on the account information of each account in the target account set and the account information of each broadcaster account to obtain the fitted real-time content features.

[0117] In this context, the content displayed in real-time on the terminal refers to the entry interface that guides viewers' accounts to the live stream. This interface displays multimedia content related to the streamer, which can include static images, animated images, and videos. The multimedia content characteristics refer to the embedding obtained after encoding the multimedia content. Correspondingly, the account information for each streamer account refers to the streamer's account identifier, i.e., the streamer ID.

[0118] For example, a general click-through rate task is constructed, which combines user ID, broadcaster ID, and multimedia embedding. The MLP (Multilayer Perceptron) network is used for fitting to obtain the fitted real-time content features, which is the real-time content embedding.

[0119] In this embodiment, multimedia content features are fitted using information from viewer accounts and broadcaster accounts, which unifies the spatial structure of features, facilitates similarity calculation between features, and improves the feasibility of similarity calculation.

[0120] In one embodiment, the spatial structure of multimedia content features is fitted based on the account information of each account in the target account set and the account information of each broadcaster account to obtain the fitted real-time content features. This includes: obtaining the learning objective corresponding to the fitting process; and using the learning objective as the fitting direction, inputting the account information of each broadcaster account and the multimedia content features into a multilayer perceptron network for fitting calculation, so that the multimedia content features are fitted in the fitting direction to obtain the fitted real-time content features.

[0121] The learning objective refers to the learning goal of the transfer learning process, which can be to achieve a preset click-through rate for each account for each recommended information.

[0122] For example, using the click rate of the content related to category information based on the user ID as the learning target, the anchor ID and multimedia embedding are input into an MLP (Multilayer Perceptron) network for calculation, so that the multimedia embedding learns in the direction of the multimedia embedding to obtain the learned real-time content embedding.

[0123] In this embodiment, considering that direct similarity calculation of embedded features from different sources may result in mismatches, a multilayer perceptron network is used for fitting, which facilitates the calculation of similarity.

[0124] In one embodiment, delivering recommendation information associated with a recommended category to any account to be recommended includes: delivering recommendation information associated with a recommended category to any account to be recommended when the level of attention paid to the recommended category by any account to be recommended meets a preset condition, and / or delivering recommendation information associated with a recommended category to any account to be recommended when the cumulative number of deliveries to each recommended category on the same day meets a preset condition.

[0125] The degree of attention any account to be recommended pays to the recommended category can be represented by a probability value. Generally, if the probability value is greater than a preset probability threshold, it can be considered that the degree of attention any account to be recommended pays to the recommended category meets the preset conditions.

[0126] The cumulative number of times an account to be recommended is advertised in each recommended category on a given day refers to the total number of times each account advertises the same category of information online on each category on that day. Specifically, meeting the preset condition can be that the cumulative number of times advertised on a given day is less than a preset frequency threshold. The preset frequency threshold is a threshold set for the number of times ads are advertised.

[0127] For example, during the category frequency control process, the cumulative number of times each category of products is advertised on the same day is obtained. If the cumulative number of times a product in a certain category is advertised on the same day is too high, the advertising of that category of products is suspended.

[0128] In this embodiment, conditions are set in two aspects that can be customized and combined to effectively trigger any account to send recommendation messages associated with the recommended category. These two aspects refer to the level of attention and the cumulative frequency of delivery per day. On the one hand, by setting a threshold to verify the predicted probability value of any online account, sensitive accounts are screened, improving the accuracy of recommendation information delivery. On the other hand, by calculating and judging the cumulative number of deliveries per day, user resentment caused by frequent delivery is avoided, demonstrating the relative real-time nature and effectiveness of the category frequency control process. By controlling the category, the accuracy of online recommendation information delivery is improved.

[0129] In one embodiment, the method further includes: obtaining a pre-trained threshold testing model; predicting any account to be recommended based on the threshold testing model to obtain a probability value corresponding to any account to be recommended; the probability value is used to characterize the degree of attention of any account to be recommended to the recommended category; and determining that the degree of attention of any account to be recommended to the recommended category meets a preset condition if the probability value is greater than a preset probability threshold.

[0130] The threshold test model is a model obtained by training a binary classification task model using positive and negative sample sets.

[0131] The positive sample set includes accounts whose recommended information from the first account set has been displayed and who have received feedback on the recommended information, as well as accounts whose recommended information from the first account set has not been displayed; the negative sample set includes accounts whose recommended information from the first account set has been displayed and who have not received feedback on the recommended information. For example, the positive samples include accounts that have been "exposed and clicked" and accounts that have not been exposed; the negative samples include accounts that have been "exposed but not clicked".

[0132] The probability value is used to characterize the degree of attention that any account to be recommended pays to the recommended category.

[0133] In this embodiment, "unexposed" accounts are added to the positive sample during the threshold test process, preserving their opportunity to be delivered and exposed. By using relatively conservative data sample division rules, the threshold test standard is lowered to avoid losing the opportunity to expose recommended information to accounts due to misjudgment, thereby improving the effectiveness of online recommended information delivery.

[0134] In one embodiment, applied to a live streaming scenario, delivering recommendation information associated with the recommended category to any account to be recommended includes: selecting target streamers for the target category from the streamer accounts under each recommended category based on the rating ranking of each streamer account under each recommended category; wherein the rating ranking is based on the interaction history data of each streamer account in the live streaming scenario within a preset time period; and delivering recommendation information associated with the target category to any account to be recommended in the live streaming scenario of the target streamer under the target category.

[0135] Among them, the target product category is the category finally determined from the recommended product categories, and the target live streamer is the live streamer finally determined from the live streamers in the recommended product categories.

[0136] The preset time period can be within the past month, with the specific launch date as a reference time.

[0137] The interaction history data includes information such as whether the account clicked, viewing time, number of interactions, streamer type, and payment status.

[0138] For example, using a common ranking model, all livestreamers under each category are ranked, their scores are estimated, and then ranked by score. Then, the livestreamer with the highest ranking in a given category is selected.

[0139] In this embodiment, historical data from the past month is used to reflect users' long-term preferences, enabling the rating of multiple livestreamers across various categories. The rating results are then used to filter the target audience and content for delivery, improving the accuracy of the recommended information. Furthermore, considering the limited information display space in a single refresh during livestreaming, the number of selected livestreamers is set to one, enhancing the integration of the ranking and rating process with the livestreaming scenario.

[0140] In another embodiment, such as Figure 3 As shown, a method for delivering recommendation information is provided, including the following steps:

[0141] S301, Determine the account characteristics of each account based on the account information of each account in the target account set.

[0142] S302, for each recommended category, take the average of the account characteristics of each account to obtain the category characteristics corresponding to each recommended category.

[0143] S303, Subtract the negative behavioral features from the positive behavioral features to obtain the behavioral features corresponding to the first account set.

[0144] S304. Based on the account information of each account in the target account set and the account information of each broadcaster account, the spatial structure of the multimedia content features is fitted to obtain the fitted real-time content features.

[0145] S305, for any account in the second account set, determine the account characteristics of any account based on the account information of any account.

[0146] S306 calculates the sub-similarity of each account's account characteristics with category characteristics, behavioral characteristics, and real-time content characteristics.

[0147] S307. The calculated sub-similarity scores are weighted and summed to obtain the similarity score.

[0148] It should be noted that the specific limitations of the above steps can be found in the specific limitations of a method for delivering recommendation information mentioned above, and will not be repeated here.

[0149] In one embodiment, the main method for category targeting or exploration currently utilizes the uplift model. The uplift model works by using a marketing four-quadrant approach to segment audience accounts, primarily into targeting-sensitive, targeting-averse, organically converting, and non-converting audiences. The core focus of targeting is identifying targeting-sensitive and targeting-averse audiences, increasing targeting for targeting-sensitive audiences while avoiding targeting targeting-averse audiences. Uplift is mainly used to predict the incremental impact of interventions on individual behavior, calculating the incremental model score as uplift score = Y(T=1) - Y(T=0). When the uplift score is greater than 0, it indicates a positive effect from the targeting. This embodiment directly uses the uplift prediction method, requiring random traffic to verify the impact of the intervention. It relies on a large number of samples to obtain stable model inferences, which may negatively impact online performance.

[0150] Therefore, in another embodiment of this application, modeling of live-stream sensitive groups is employed, that is, identifying groups that will not convert without advertising and will convert with advertising, to maximize potential users. While this method can identify potential audiences for a target category, its targeting capabilities for specific subcategories are clearly lacking.

[0151] To address the significant lack of targeting capabilities for specific product categories, this application employs a lookalike approach in another embodiment. This approach aims to more accurately identify target users for recommendations at a lower cost, thereby improving the effectiveness of category exploration.

[0152] Based on this, embodiments of this application provide a method for delivering recommendation information. For ease of understanding by those skilled in the art, Figure 4 A flowchart illustrating the overall process of delivering recommendation information is provided. Figure 5 A schematic diagram illustrating the fitting principle of a recommendation information delivery method is provided below; see below for reference. Figure 4 and Figure 5 The following describes in detail a method for delivering recommendation information using a specific embodiment. It is important to understand that the following description is merely illustrative and not intended to limit the scope of the application.

[0153] The recommendation information delivery method provided in this application embodiment, also known as the lookalike-based live streaming category exploration method, enhances the lookalike category exploration effect by combining it with Uplift. This recommendation information delivery method can be applied to software applications such as backing vocals, music, and live streaming. For example, it can be applied to information recommendation for e-commerce live streaming categories and bullet screen games in feed live streaming scenarios, and can also be applied to other games and activities that combine real-time interactive features.

[0154] The overall process of the recommendation information delivery method provided in this application embodiment is as follows: Figure 4 As shown, the process is divided into offline and online processes. The steps of category trial deployment, acquisition of category-sensitive users, and expansion of sensitive users are offline processes, while the other steps are online processes.

[0155] In the offline process, category-sensitive users are identified through small-scale random trials. Then, a unified user representation across all scenarios is used to expand the similarity of these category-sensitive users, resulting in a candidate sensitive user set. This unified user representation across all scenarios is a user profile representation learned from various user clicks, interactions, and other behaviors within the application.

[0156] In the online process, the first step is to determine whether the user is a category-sensitive user. Then, the user's Uplift score is checked for a low threshold. This part is mainly to filter out irrelevant people introduced during the expansion process. Then, category frequency control is implemented to prevent users from feeling annoyed by frequent distribution. Finally, the candidate category items are sorted, and the top-ranked categories are selected for the distribution of recommendation information, such as the top 1 category.

[0157] In the Uplift threshold test, the model is trained using data samples. The trained model is used to predict the probability value of the account’s category preference. A threshold is set for the probability value to classify the accounts. The relationship between the threshold and the probability value is used to filter out accounts with a higher level of category preference.

[0158] Regarding model type, a model capable of performing binary classification tasks is used. As for data samples, the training samples are data from the day before the launch to reflect predictions of user preference scores for that day, thus avoiding bias. The training samples used for the Uplift threshold test include both positive and negative samples. Positive samples include accounts with "exposure and clicks" and "no exposure"; negative samples include accounts with "exposure but no clicks." It is evident that the data samples used differ from the screening criteria used in acquiring category-sensitive accounts, and the account set used is also different.

[0159] For example, a trained binary classification task model outputs a probability value of 0-1, and a fixed parameter set manually is used as a threshold, such as 0.5. When the probability predicted by an account is greater than 0.5, the account is considered to have a strong preference for the category. The account is further identified as a user who is sensitive to the category, and the account is then subjected to a threshold verification process to achieve a second screening of sensitive users through the threshold.

[0160] Compared to the expansion process for category-sensitive accounts, Uplift's threshold screening is based on the model's probability predictions, obtained by training the model using positive and negative samples; while the expansion of sensitive accounts in the offline process is based on the similarity between users. This application combines two evaluation criteria based on user preferences, enabling a process of trial expansion followed by screening, which ensures both the quantity of similar users targeted and the accuracy of targeting.

[0161] During the candidate scoring process, a common ranking model is used to rank all remaining streamers under each category, estimate their scores, and then rank them. Finally, the streamer from the top-ranked category is selected. Categories that can be selected include e-commerce and bullet-screen games, and each category can have at least one streamer.

[0162] The selected model type is any multi-objective model. The model training data consists of interactions corresponding to the account, including whether there are clicks, viewing duration, number of interactions, streamer type, payment status, etc. The sampling period for the training data can be historical data from the past month of online testing to reflect long-term user preferences.

[0163] It should be noted that in a specific live streaming scenario, an account can manually refresh to obtain a streamer's display interface one by one. Each refresh yields a recommended streamer, and if unsatisfied, the account can refresh again. It is understood that in the information recommendation process within a live streaming scenario, the display space for distribution positions in a single account recommendation process is limited. Therefore, this application adopts a candidate scoring and filtering method, selecting only one streamer from a category for recommendation information delivery each time, rather than displaying multiple streamers at once. This achieves a close integration between the recommendation information delivery process and the actual live streaming scenario refresh process, improving the efficiency of computing resource utilization and enhancing the effectiveness of recommendation information delivery in live streaming scenarios.

[0164] The similarity calculation process includes: based on users who clicked on the category, using a global representation, obtaining the embeddings (embedded features) of all users who clicked on the category, and then averaging them as the category embedding. Next, based on the positive and negative behaviors of the broadcaster, obtaining the behavior embeddings: positive behavior is defined as clicking after exposure, and negative behavior is defined as not clicking after exposure. Then, averaging the embeddings of the behavior uid (User ID) yields both the positive and negative embeddings. Finally, the behavior embedding is the positive embedding minus the negative embedding.

[0165] The acquisition of real-time content embedding is based on multimedia embeddings of dynamic images outside the broadcaster's room. These dynamic images refer to the images displayed to users before they enter the broadcaster's live room. However, because these embeddings are not constructed in the same spatial environment as user embeddings, they cannot be used directly. This application constructs a general click-through rate task, combining user ID, broadcaster ID, and multimedia embeddings, and uses an MLP (Multilayer Perceptron) network for fitting. The user embedding and broadcaster embedding inherit initial values ​​from the general embedding, and then learn multimedia information from the network. Finally, the real-time content embedding is obtained based on the broadcaster ID and multimedia embedding.

[0166] It should be noted that, Figure 5 The fitting process can be understood as follows: using the data pairs of user ID (user refers to the person watching the streamer) and streamer ID from a single click event as training data, the model is trained to fit the multimedia embedding, unify the construction space, and facilitate similarity calculation. The embedding layer refers to the embedding layer in the model.

[0167] Finally, the similarity expansion calculates the similarity between the user embedding and the category embedding, behavior embedding, and real-time content embedding, respectively. The three embeddings are then weighted and summed to obtain the final similarity score. A pre-set threshold is used to determine whether the similarity is sufficient.

[0168] The embodiments of this application do not directly target similar extended groups because the global user representation is biased during the expansion process, which may lead to the introduction of some similar users who may not be interested in the category. Uplift, on the other hand, ensures that users are inclined to the category.

[0169] Furthermore, in practical use, it was found that similarity expansion for sensitive users is a difficult point to balance. Therefore, designing a similarity expansion method more suitable for live streaming scenarios has become paramount. This application's embodiment designs a similarity determination and verification method that integrates category, streamer, and real-time content, namely a Lookalike similarity determination and verification method, to calculate the real-time similarity of the streamer's category.

[0170] The recommendation information delivery method provided in this application can deploy a category delivery scheme based on Uplift while avoiding the use of random traffic, which is detrimental to online delivery. At the same time, it addresses the category divergence problem caused by similar Uplift by combining a lookalike audience targeting scheme, and innovatively designs a real-time similarity of anchor categories to adapt to live streaming scenarios, thereby improving the effectiveness of recommendation information delivery.

[0171] The recommendation information delivery method provided in this application can be applied to recommendation feed pages in accompaniment singing scenarios to explore product categories for live streaming of games with bullet comments and e-commerce live streams. After 15 days of category exploration in actual testing, the test results are as follows.

[0172] With ad impressions, overall ad duration, and user retention remaining unchanged, the following results were observed in the live streaming market: View penetration increased by 1.731% (significant), UV penetration increased by 2.394% (significant), and live stream tab click UV penetration increased by 7.536% (significant). View penetration refers to the increase in the proportion of users watching live streams after ad exposure; UV penetration typically refers to the penetration rate of unique visitors, i.e., the change in the number of users participating in a specific activity relative to the total number of users after ad exposure; and live stream tab click UV penetration represents the change in the number of users clicking the live stream tab relative to the total number of users after ad exposure.

[0173] In addition, feed live stream exposure penetration increased by 25.04% (significant), room entry penetration increased by 11.984% (significant), first-time visit to room entry penetration increased by 11.475% (significant), and first-time paid visit penetration increased by 1.627%. Exposure penetration refers to the proportion of users who viewed the content out of those who were exposed to it; room entry penetration rate indicates the proportion of users who entered the live stream in a specific scenario. Specifically, room entry penetration rate indicates how many users actually entered the live stream after seeing your optimized content; first-time visit to room entry penetration rate refers to the proportion of users who entered the live stream through the feed on the first time they retrieved the feed page. This metric measures the attractiveness of live stream content and the rate of rapid conversion.

[0174] The test results show that the experiment under the feed also brought a significant improvement to the live streaming tab scenario. This is because the user activity boosted the atmosphere of the live streaming room and strengthened the initiative of the streamer. The use of the recommendation information delivery method provided in this application embodiment during the test improved the effect of recommendation information delivery.

[0175] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0176] Based on the same inventive concept, this application also provides a device for delivering recommendation information to implement the aforementioned method. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the device for delivering recommendation information provided below can be found in the limitations of the method for delivering recommendation information above, and will not be repeated here.

[0177] In one embodiment, such as Figure 6 As shown, a device for delivering recommendation information is provided, including: a delivery testing module 601, a similarity acquisition module 602, an extension module 603, and a delivery module 604, wherein:

[0178] The campaign testing module 601 is used to conduct campaign testing on the first set of accounts in the account pool to obtain the target set of accounts. The target set of accounts includes accounts in the first set of accounts that have entered interactive operations for at least one recommended category recommended in the campaign testing.

[0179] The similarity acquisition module 602 is used to acquire the similarity between the group characteristics of the target account set and the characteristics of each account in the second account set; the second account set is the account set in the account pool other than the first account set.

[0180] The extension module 603 is used to filter out accounts with similar characteristics to those in the target account set from the second account set based on similarity, and add the accounts to be added to the target account set to obtain the extended account set;

[0181] The delivery module 604 is used to deliver recommendation information related to the recommended category to any account to be recommended, provided that the account belongs to the expanded account set.

[0182] Each module in the aforementioned recommendation information delivery device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0183] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a method for distributing recommendation information.

[0184] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 8As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a method for delivering recommendation information. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0185] Those skilled in the art will understand that Figure 7 and Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0186] In one embodiment, a computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0187] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0188] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0189] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0190] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0191] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for delivering recommendation information, characterized in that, When applied to live streaming scenarios, the method includes: A campaign test is conducted on the first set of accounts in the account pool to obtain a target set of accounts. The target set of accounts includes accounts in the first set of accounts that have entered interactive operations for at least one recommended category recommended in the campaign test. The recommended categories include e-commerce and bullet screen games. The interactive operation corresponding to the e-commerce category includes a click operation. The interactive operation corresponding to the bullet screen game category includes a click operation and a confirmation of participation in the game. Based on the account information of each account in the target account set, determine the category characteristics corresponding to each of the recommended categories; based on the behavioral information fed back by each account in the first account set during the campaign test, determine the behavioral characteristics; based on the content displayed in real time by the terminal in the live streaming scenario during the campaign test, determine the real-time content characteristics. The category characteristics, behavioral characteristics, and real-time content characteristics are used as the group characteristics of the target account set; Obtain the similarity between the group characteristics of the target account set and the characteristics of each account in the second account set; the second account set is the account set in the account pool other than the first account set; Based on the similarity, accounts with similar characteristics to those in the target account set are selected from the second account set and added to the target account set to obtain the expanded account set. For any account to be recommended, if the account to be recommended belongs to the expanded account set and the account to be recommended shows interest in the recommended category according to a preset condition, and / or, if the account to be recommended belongs to the expanded account set and the cumulative number of times the account to be recommended is delivered to each of the recommended categories on the same day is less than a preset frequency threshold, then recommendation information associated with the recommended category will be delivered to the account to be recommended.

2. The method according to claim 1, characterized in that, The step of obtaining the similarity between the group characteristics of the target account set and the characteristics of each account in the second account set includes: For any account in the second account set, determine the account characteristics of the account based on the account information of the account. Calculate sub-similarity between the account characteristics of each account and the category characteristics, the behavioral characteristics, and the real-time content characteristics, respectively; The calculated sub-similarity scores are weighted and summed to obtain the total similarity score.

3. The method according to claim 1, characterized in that, The step of determining the category characteristics corresponding to each of the recommended categories based on the account information of each account in the target account set includes: Based on the account information of each account in the target account set, determine the account characteristics of each account; For each of the recommended product categories, the average value of the account characteristics of each account is taken to obtain the category characteristics corresponding to each of the recommended product categories.

4. The method according to claim 1, characterized in that, The behavioral information includes information on positive behavior and information on negative behavior. The step of determining behavioral characteristics based on the behavioral information reported by each account in the first account set during the deployment test includes: Based on the account information corresponding to the accounts that have positive behavior towards the recommended category in the first account set, positive account characteristics are determined, and the positive account characteristics are averaged to obtain positive behavior characteristics. Based on the account information corresponding to the accounts that have negative behavior towards the recommended category in the first account set, negative account characteristics are determined, and the average value of the negative account characteristics is taken to obtain negative behavior characteristics. The behavioral characteristics corresponding to the first account set are obtained by subtracting the negative behavioral characteristics from the positive behavioral characteristics.

5. The method according to claim 1, characterized in that, The step of determining real-time content characteristics based on the content displayed in real-time by the terminal in the live streaming scenario during the deployment test includes: Based on the content displayed in real time by the terminal in the live streaming scenario during the deployment test, the characteristics of the multimedia content are determined; Obtain account information for each streamer's account in the live streaming scenario; Based on the account information of each account in the target account set and the account information of each broadcaster account, the spatial structure of the multimedia content features is fitted to obtain the fitted real-time content features.

6. The method according to claim 5, characterized in that, The step of fitting the spatial structure of the multimedia content features based on the account information of each account in the target account set and the account information of each broadcaster account to obtain the fitted real-time content features includes: Obtain the learning objective corresponding to the fitting process; the learning objective includes achieving a preset click-through rate for each account's click-through rate for the recommended information each time. Using the learning objective as the fitting direction, the account information of each broadcaster's account and the multimedia content features are input into a multilayer perceptron network for fitting calculation, so that the multimedia content features are fitted to the fitting direction to obtain the fitted real-time content features.

7. The method according to claim 1, characterized in that, The method further includes: Obtain a pre-trained threshold testing model; the threshold testing model is a model obtained by training a binary classification task model using a positive sample set and a negative sample set; the positive sample set includes accounts in the first account set where the recommendation information has been displayed and feedback has been received regarding the recommendation information, and accounts in the first account set where the recommendation information has not been displayed; the negative sample set includes accounts in the first account set where the recommendation information has been displayed and feedback has not been received regarding the recommendation information. The threshold test model is used to predict the probability value of any account to be recommended, thereby obtaining a probability value for each account to be recommended. The probability value is used to characterize the degree of attention that account to be recommended pays to the recommended category. If the probability value is greater than a preset probability threshold, it is determined that the attention level of any account to be recommended to the recommended category meets the preset conditions.

8. The method according to any one of claims 1 to 7, characterized in that, In live streaming scenarios, delivering recommendation information associated with the recommended category to any of the accounts to be recommended includes: Based on the rating ranking of each streamer account under each of the recommended product categories, target streamers under the target product category are selected from each streamer account under each of the recommended product categories; wherein the rating ranking is based on the interaction history data of each streamer account in the live streaming scenario within a preset time period; In the live streaming scenario of the target anchor under the target category, recommendation information related to the target category is delivered to any of the accounts to be recommended.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images