News cold start exposure method and device based on vector mapping

Abstract

The invention discloses a news cold start exposure method and device based on vector mapping. The method comprises the steps that cold start news with the accumulated exposure frequency smaller than a first exposure frequency threshold value is obtained; target hot news similar to the cold start news is determined from the hot news with the accumulated exposure times larger than a second exposure time threshold value; and generating approximate news behavior characteristics of the cold start news according to the news behavior characteristics of the target hot news, and matching with the user according to the approximate news behavior characteristics of the cold start news so as to expose the cold start news to the user. Therefore, according to the scheme, on the premise that matching is not carried out based on the content and the user, the recommendation problem of the cold start news is solved, and the accuracy of exposing the cold start news is improved.

Description

Technical Field

[0001] The present application relates to the field of recommendation technology, and in particular to a news cold start exposure method and device based on vector mapping. Background Art

[0002] Internet content providers can use recommendation systems to determine which news to expose to users, for example, exposing news 1 to user A and news 2 to user B.

[0003] Exposing news to users can be understood as displaying the title and / or summary of the news to be exposed on the terminal used by the user through various forms such as pop-up messages, banner notifications, and drop-down pages. After exposure, the user can choose to click on the displayed title and / or summary to browse the specific content of the news, or may not ignore the displayed title and / or summary and not browse the corresponding content.

[0004] In order to make the exposed news more in line with the user's interest preferences, the recommendation system can determine the behavioral characteristics of the news based on the user feedback after the exposure (that is, whether the user browses the specific content after the exposure), and determine the news exposed to each user based on the matching of the behavioral characteristics and the user.

[0005] However, some recently generated news has a low total exposure time (this type of news can be called cold start news), and the user feedback obtained is also very small, so it is difficult to determine the behavioral characteristics of cold start news based on user feedback. The existing recommendation method for cold start news is to directly match the content of cold start news with users. However, the content of the news is difficult to accurately reflect the user's interest preference for different news. Therefore, this recommendation method for cold start news has low accuracy and is prone to the situation where users are not interested in the exposed news. Summary of the invention

[0006] To this end, this application discloses the following technical solutions:

[0007] The first aspect of the present application provides a news cold start exposure method based on vector mapping, comprising:

[0008] Acquire cold start news, where the cold start news includes news whose cumulative exposure times are less than a first exposure times threshold;

[0009] Determine a plurality of target hot news satisfying a similarity condition with the cold start news from the hot news, wherein the hot news includes news having a cumulative exposure count greater than a second exposure count threshold;

[0010] Generate an approximate news behavior feature of the cold start news according to the news behavior feature of the target hot news, wherein the news behavior feature of the target hot news is obtained by processing the news content feature of the target hot news by a target model, wherein the target model is constructed according to user feedback data of multiple hot news, and the user feedback data of the hot news indicates whether the hot news is browsed by the user after being exposed to the user;

[0011] The cold start news is exposed to users according to the user similarity of the approximate news behavior characteristics of the cold start news, wherein the user similarity represents the similarity between the approximate news behavior characteristics and the user behavior characteristics, and the user behavior characteristics are obtained by processing user information by the target model.

[0012] Optionally, the method for constructing the target model according to user feedback data of a plurality of hot news includes:

[0013] Determine a plurality of samples according to user feedback data of a plurality of hot news, each of the samples comprising user information of a user, news content features of a hot news that has been exposed to the user, and a sample label, the sample label having a first label value or a second label value, the first label value indicating that the user did not browse the hot news after being exposed, and the second label value indicating that the user browsed the hot news after being exposed;

[0014] Processing the user information and news content features of each sample based on the initial model to obtain initial user behavior features corresponding to the user information in the sample and initial news behavior features corresponding to the news content features;

[0015] For each of the samples, a prediction result of the sample is determined according to the initial user behavior characteristics and initial news behavior characteristics corresponding to the sample, wherein the prediction result represents the probability that the user corresponding to the sample browses the hot news corresponding to the sample;

[0016] Obtaining a model loss of the initial model according to a prediction result of the sample and a sample label of the sample;

[0017] The model parameters of the initial model are updated according to the model loss, and the step of processing the user information and news content features of each sample based on the initial model is returned to be executed until the target model whose model loss satisfies the convergence condition is obtained.

[0018] Optionally, obtaining the cold start news includes at least one of the following:

[0019] In response to obtaining user request information, obtaining cold start news;

[0020] In response to satisfying the push condition, cold start news is obtained.

[0021] Optionally, generating the approximate news behavior features of the cold start news according to the news behavior features of the target hot news includes:

[0022] Obtaining news content similarity between the target hot news and the cold start news;

[0023] The approximate news behavior characteristics of the cold start news are determined according to the news content similarities of the plurality of target hot news and the news behavior characteristics of the plurality of target hot news.

[0024] Optionally, exposing the cold start news to the user according to the user similarity of the approximate news behavior characteristics of the cold start news includes:

[0025] When the user similarity between the approximate news behavior feature of the cold start news and the user behavior feature of the first user satisfies a first condition, exposing the cold start news to the first user;

[0026] or,

[0027] A second user whose user similarity satisfies a second condition is determined from among a plurality of candidate users, and the cold start news is exposed to the second user.

[0028] The second aspect of the present application provides a news cold start exposure device based on vector mapping, comprising:

[0029] An obtaining unit, configured to obtain cold start news, wherein the cold start news includes news whose cumulative exposure times are less than a first exposure times threshold;

[0030] a determination unit, configured to determine, from the hot news, a plurality of target hot news satisfying a similarity condition with the cold start news, wherein the hot news includes news having a cumulative exposure count greater than a second exposure count threshold;

[0031] a generating unit, configured to generate an approximate news behavior feature of the cold start news according to the news behavior feature of the target hot news, wherein the news behavior feature of the target hot news is obtained by processing the news content feature of the target hot news by a target model, wherein the target model is constructed according to user feedback data of a plurality of hot news, wherein the user feedback data of the hot news indicates whether the hot news is browsed by the user after being exposed to the user;

[0032] An exposure unit is used to expose the cold start news to users according to the user similarity of the approximate news behavior characteristics of the cold start news, wherein the user similarity represents the similarity between the approximate news behavior characteristics and the user behavior characteristics, and the user behavior characteristics are obtained by processing user information by the target model.

[0033] Optionally, the device further comprises a construction unit, wherein the construction unit is used to:

[0034] Determine a plurality of samples according to user feedback data of a plurality of hot news, each of the samples comprising user information of a user, news content features of a hot news that has been exposed to the user, and a sample label, the sample label having a first label value or a second label value, the first label value indicating that the user did not browse the hot news after being exposed, and the second label value indicating that the user browsed the hot news after being exposed;

[0035] Processing the user information and news content features of each sample based on the initial model to obtain initial user behavior features corresponding to the user information in the sample and initial news behavior features corresponding to the news content features;

[0036] For each of the samples, a prediction result of the sample is determined according to the initial user behavior characteristics and initial news behavior characteristics corresponding to the sample, wherein the prediction result represents the probability that the user corresponding to the sample browses the hot news corresponding to the sample;

[0037] Obtaining a model loss of the initial model according to a prediction result of the sample and a sample label of the sample;

[0038] The model parameters of the initial model are updated according to the model loss, and the step of processing the user information and news content features of each sample based on the initial model is returned to be executed until the target model whose model loss satisfies the convergence condition is obtained.

[0039] Optionally, when the obtaining unit obtains the cold start news, it is configured to perform at least one of the following:

[0040] In response to obtaining user request information, obtaining cold start news;

[0041] In response to satisfying the push condition, cold start news is obtained.

[0042] Optionally, when the generating unit generates the approximate news behavior feature of the cold start news according to the news behavior feature of the target hot news, it is used to:

[0043] Obtaining news content similarity between the target hot news and the cold start news;

[0044] The approximate news behavior characteristics of the cold start news are determined according to the news content similarities of the plurality of target hot news and the news behavior characteristics of the plurality of target hot news.

[0045] Optionally, when the exposure unit exposes the cold start news to the user according to the user similarity of the approximate news behavior feature of the cold start news, it is used to:

[0046] When the user similarity between the approximate news behavior feature of the cold start news and the user behavior feature of the first user satisfies a first condition, exposing the cold start news to the first user;

[0047] or,

[0048] A second user whose user similarity satisfies a second condition is determined from among a plurality of candidate users, and the cold start news is exposed to the second user.

[0049] The method of this embodiment can obtain the news behavior characteristics of hot news based on the user feedback data of hot news, and in view of the fact that there is less user feedback data for cold start news, screen out target hot news similar to the cold start news, use the news behavior characteristics of these target hot news to fit the approximate news behavior characteristics of the cold start news, match the user based on the approximate news behavior characteristics, and thus expose the cold start news to the corresponding user. Therefore, this solution solves the recommendation problem of cold start news without matching based on content and users, and improves the accuracy of exposing cold start news. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0051] Figure 1 It is a flowchart of a news cold start exposure method based on vector mapping provided in an embodiment of the present application;

[0052] Figure 2 It is a flowchart of another news cold start exposure method based on vector mapping provided in an embodiment of the present application;

[0053] Figure 3 is a schematic diagram of the structure of a target model provided in an embodiment of the present application;

[0054] Figure 4 It is a schematic diagram of obtaining approximate news behavior characteristics provided by an embodiment of the present application;

[0055] Figure 5 It is a structural schematic diagram of a news cold start exposure device based on vector mapping provided in an embodiment of the present application. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0057] This application embodiment provides a news cold start exposure method based on vector mapping, see Figure 1 , the method may include the following steps.

[0058] S101, obtaining cold start news, where the cold start news includes news with a cumulative exposure count less than a first exposure count threshold.

[0059] S102, determining a plurality of target hot news satisfying a similarity condition with the cold start news in the hot news, where the hot news includes news having a cumulative exposure count greater than a second exposure count threshold.

[0060] The first exposure number threshold and the second exposure number threshold may be the same or different, and their specific values ​​may be set as needed without limitation.

[0061] S103, generating approximate news behavior features of the cold start news based on the news behavior features of the target hot news, wherein the news behavior features of the target hot news are obtained by processing the news content features of the target hot news by a target model, and the target model is constructed based on user feedback data of multiple hot news, and the user feedback data of the hot news represents whether the hot news is browsed by the user after being exposed to the user.

[0062] S104, based on the user similarity of the approximate news behavior characteristics of the cold start news, the cold start news is exposed to the user, the user similarity represents the similarity between the approximate news behavior characteristics and the user behavior characteristics, and the user behavior characteristics are obtained by processing the user information by the target model.

[0063] Optionally, the user information used for processing in this embodiment can be obtained from a terminal device used by the user on the premise of obtaining the user's authorization in advance.

[0064] The beneficial effects of this embodiment are:

[0065] The problem with directly matching the content of cold start news with users in the related art is that if users have already browsed similar news, they are likely to be uninterested in the subsequent exposure of the same type of news. In contrast, the method of this embodiment can obtain the news behavior characteristics of hot news based on the user feedback data of hot news, and in view of the fact that there is less user feedback data for cold start news, screen out target hot news similar to cold start news, use the news behavior characteristics of these target hot news to fit the approximate news behavior characteristics of cold start news, match users based on the approximate news behavior characteristics, and thus expose the cold start news to the corresponding users. Therefore, this solution solves the recommendation problem of cold start news without matching based on content and users, and improves the accuracy of exposing cold start news.

[0066] The method of this embodiment can be executed by any server for recommending news to users.

[0067] See also Figure 2 , the method of this embodiment can be performed by Figure 2 The user collaborative training module, cold start content mapping module and cold start content exposure module shown are executed, the user collaborative training module is used to build a target model, the cold start content mapping module is used to execute the above steps S102 and S103 to obtain approximate news behavior characteristics, the cold start content exposure module is used to trigger the execution of step S101, and is used to expose cold start news based on user behavior characteristics and approximate news behavior characteristics of cold start news.

[0068] The method provided in this embodiment may be executed when user request information is obtained, or may be executed when a specific push condition is met. Accordingly, step S101 may include at least one of the following:

[0069] In response to obtaining user request information, obtaining cold start news;

[0070] In response to satisfying the push condition, cold start news is obtained.

[0071] For the first case, when logging into a news recommendation system (such as various forums and websites on the Internet), the terminal device can send user request information to the server. The user request information is used to request a number of news for exposure from the server. After receiving the user request information, the server can obtain all or part of the cold start news that has not been exposed to the currently logged-in user in the news database, and execute subsequent steps for each cold start news obtained.

[0072] For the second case, the push condition can be set as needed. For example, it can be set to execute the method of this embodiment once every certain period, or it can be set to execute the method of this embodiment when the total number of cold start news in the news database is greater than a certain threshold. When the push condition is met, the server can obtain all or part of the cold start news in the news database, and execute the subsequent steps for each cold start news obtained.

[0073] For each piece of news, the server sends the title and / or summary of the news to the user terminal and displays it on the user terminal through a pop-up window, message banner, etc., which is called an exposure of the news. Each time a piece of news is exposed, users who see the title and / or summary of the news can choose to browse or not browse the specific content of the news.

[0074] In step S102, the server may obtain multiple hot news from the news database, and then calculate the news content similarity between the cold start news and each hot news according to the content of the cold start news and the news content of the hot news.

[0075] For example, assuming that 100 hot news items are obtained, for the cold start news, the similarities of 100 news items' contents between the cold start news items and the 100 hot news items may be calculated.

[0076] The method for calculating the news content similarity may be to use an existing semantic analysis model to analyze the semantics of the news content, and then calculate the cosine similarity between the semantics of the cold start news and the semantics of the hot news, and use the calculation result as the news content similarity.

[0077] After obtaining the news content similarity, the top N hot news with the highest news content similarity among the multiple hot news can be used as target hot news, or the hot news with news content similarity greater than a preset similarity threshold can be used as the target hot news.

[0078] Continuing with the above example, assuming that N is 20, the 100 hot news items can be sorted from high to low according to the similarity of the corresponding news content, and the top 20 hot news items are selected as the target hot news items corresponding to the cold start news items.

[0079] When there are multiple cold start news, the above steps may be performed on each cold start news to determine the target hot news corresponding to each cold start news.

[0080] In S103, a method of obtaining the approximate news behavior feature may be:

[0081] The news behavior features of all target hot news corresponding to the cold start news are accumulated, and the result is used as the approximate news feature of the cold start news.

[0082] For example, assuming that a cold start news corresponds to n target hot news, and the news behavior features corresponding to the n target hot news are recorded as emb1 to embn in sequence, then the approximate news feature emb of the cold start news can be expressed as: emb=emb1+emb2+…embn.

[0083] In S103, another way to obtain the approximate news behavior feature may be:

[0084] Obtain news content similarity between target hot news and cold start news;

[0085] The approximate news behavior characteristics of the cold start news are determined according to the news content similarity of multiple target hot news and the news behavior characteristics of multiple target hot news.

[0086] In this embodiment, the news content similarity of the target hot news can be used as a weight, and the news behavior features of multiple target hot news can be weighted and summed, and the obtained result can be used as the approximate news behavior feature of the cold start news.

[0087] See also Figure 4 , assuming that the news content similarities of n target hot news are S1 to Sn, then the approximate news behavior feature emb of cold start news can be expressed as: emb=S1*emb1+S2*emb2+…Sn*embn.

[0088] When there are multiple cold start news, the above steps may be performed for each cold start news to obtain the approximate news behavior features corresponding to each cold start news.

[0089] In step S104, the cold start news may be exposed to the user in the following manner 1 or manner 2:

[0090] Method 1: when the user similarity between the approximate news behavior characteristics of the cold start news and the user behavior characteristics of the first user meets the first condition, the cold start news is exposed to the first user;

[0091] Method 2: Determine a second user whose user similarity satisfies the second condition from among multiple candidate users, and expose the cold start news to the second user.

[0092] Method 1 can be applied to the case of exposing cold start news to a specific user. For example, when obtaining user request information, the server can take the user who sent the user request information as the first user, and then expose the cold start news to the first user according to method 1.

[0093] In the first approach, the server may process the user information of the first user using the aforementioned target model to obtain the user behavior characteristics of the first user.

[0094] Then, for each cold start news that has not been exposed to the first user, the cosine similarity between the approximate news behavior feature of the cold start news and the user behavior feature of the first user is calculated, and the calculation result is used as the user similarity of the cold start news.

[0095] Finally, among the plurality of cold start news that have not been exposed to the first user, the cold start news whose user similarity satisfies the first condition is exposed to the first user.

[0096] The first condition may be that the user similarity is greater than a preset threshold, or that the user similarity ranks in the top m among multiple cold start news that have not yet been exposed to the first user, where m may be any preset integer. For example, if the first user sets a maximum of 10 news items to be exposed each time, then m may be set to 10.

[0097] Optionally, when exposing according to method one, for the first user, cold start news whose user similarity with the first user meets the first condition can be retrieved through faiss retrieval, and these cold start news can be added to the cold start news exposure list corresponding to the first user. Then, the cold start news in the list can be exposed to the first user one by one in a certain order.

[0098] The second method can be applied to the case of exposing to multiple users. For example, when the above push conditions are met, the server can determine each user currently logged in to the system as a candidate user.

[0099] After determining multiple candidate users, the server can process the user information of each candidate user using the target model to obtain the user behavior characteristics of each candidate user;

[0100] Then, for each cold start news, the cosine similarity between the approximate news behavior characteristics of the cold start news and the user behavior characteristics of each candidate user can be calculated, and the calculation result can be used as the user similarity between the cold start news and the candidate user. If the user similarity between a cold start news and a candidate user meets the second condition, the server can use the candidate user as the second user, and the cold start news can be exposed to the second user.

[0101] The second condition may be that the user similarity is greater than a preset threshold, or the user similarity between the cold start news and the user is ranked in the top k among multiple candidate users, where k may be any preset integer.

[0102] Optionally, the method of building a target model based on user feedback data of multiple popular news includes:

[0103] A1, determining multiple samples based on user feedback data of multiple hot news, each sample including user information of a user, news content features of a hot news that has been exposed to the user, and a sample label, the sample label having a first label value or a second label value, the first label value indicating that the user did not browse the hot news after being exposed, and the second label value indicating that the user browsed the hot news after being exposed;

[0104] A2, based on the initial model, processes the user information and news content features of each sample to obtain the initial user behavior features corresponding to the user information in the sample and the initial news behavior features corresponding to the news content features;

[0105] A3, for each sample, determine the prediction result of the sample according to the initial user behavior characteristics and initial news behavior characteristics corresponding to the sample, and the prediction result represents the probability that the user corresponding to the sample browses the hot news corresponding to the sample;

[0106] A4, obtain the model loss of the initial model according to the prediction result of the sample and the sample label of the sample;

[0107] A5, updates the model parameters of the initial model according to the model loss, and returns to execute the step of processing the user information and news content features of each sample based on the initial model until a target model whose model loss meets the convergence condition is obtained.

[0108] The initial model used in this embodiment may have the following Figure 3 The model structure shown in Figure 1. The model can be divided into input layer, encoding layer, model layer and output layer. The input layer is used to obtain user information and news content features.

[0109] The encoding layer may include a first embedding layer and a second embedding layer, the model layer may include a first fully connected module (also called a multilayer perceptron, Multilayer Perceptron, MLP) and a second fully connected module, the first embedding layer and the first fully connected module are connected in series to process user information to obtain user behavior features, and the second embedding layer and the second fully connected module are connected in series to process news content features to obtain news behavior features.

[0110] Among them, user information may include information such as the user's gender, age, activity level and long-term interest preferences, and news content features may include news identification (or ID), news category, news topic, and news author.

[0111] The first embedding layer can map each item of user information into a dense low-dimensional vector to obtain multiple low-dimensional vectors with the same dimension corresponding to multiple items of user information, such as mapping gender into a 32-dimensional vector, mapping age into a 32-dimensional vector, mapping activity into a 32-dimensional vector, and so on. Then, multiple low-dimensional vectors can be spliced ​​together end to end to form a spliced ​​vector corresponding to the user information, for example, the four 32-dimensional vectors corresponding to the above four items of user information are spliced ​​into a 4*32-dimensional spliced ​​vector.

[0112] Similarly, the second embedding layer can map each feature of the news content into a dense low-dimensional vector to obtain multiple low-dimensional vectors with the same dimension corresponding to multiple news content features. For example, the news identifier is mapped to a 32-dimensional vector, the news category is mapped to a 32-dimensional vector, the news topic is mapped to a 32-dimensional vector, and so on. The low-dimensional vectors corresponding to the multiple news content features are then concatenated together to form a concatenated vector corresponding to the news content feature.

[0113] The first fully connected module can calculate the concatenated vector corresponding to the user information through multiple layers of full connection to obtain a vector with a specific dimension, which is Figure 3 The second fully connected module can calculate the concatenated vector corresponding to the news content features through multiple layers of full connection to obtain a vector with a specific dimension. This vector is Figure 3 As an example, the dimensions of the user behavior feature and the news behavior feature may be equal, for example, both may be 16, or may be other dimensions, without limitation.

[0114] The output layer can calculate the cosine similarity between the user behavior features and the news behavior features output by the model layer, and use the calculation result as the above prediction result.

[0115] In step A1, the server may obtain multiple hot news from a news database, and for each hot news, determine the users to whom the hot news has been exposed, and then obtain multiple samples based on the hot news, the users to whom the hot news has been exposed, and user feedback data between the hot news and these users.

[0116] Exemplarily, hot news 1 has been exposed to user A and user B. User A browsed the specific content of hot news 1, but user B did not browse it. Thus, sample 1 and sample 2 can be constructed. Sample 1 includes news content features of hot news 1, user A's user information and the second label value. Sample 2 includes news content features of hot news 1, user B's user information and the first label value.

[0117] Optionally, the multiple hot news used to construct the sample in step A1 and the multiple hot news used to filter out the target hot news in step S102 may be the same batch of hot news or two different batches of hot news.

[0118] The initial model used in steps A2 and A3 can be obtained by random initialization or pre-training. Figure 3 Model of the structure shown.

[0119] In steps A2 and A3, for each sample, the initial model can be used to process the user information and news content features contained in the sample in the above-mentioned manner to obtain the initial user behavior features obtained by processing the user information in the sample, and the initial news behavior features corresponding to the news content features in the sample.

[0120] Each time steps A2 and A3 are executed, a batch of samples can be processed to obtain a prediction result for each sample. A batch of samples can include any number of samples, for example, 10, 20 or other numbers of samples.

[0121] In step A4, the first label value of each sample can be set to 0, and the second label value can be set to 1. Then, L1 corresponding to each sample can be calculated according to the following formula (1).

[0122] L1=-[ylogy'+(1-y)log(1-y')], (1).

[0123] Where y represents the label value of the sample. If the sample contains the first label value, y is 0, and if the sample contains the second label value, y is 1. y' represents the prediction result corresponding to the sample, and log represents the logarithm with base 10.

[0124] After calculating the L1 of each sample in a batch in the above manner, all L1s can be accumulated and averaged, and the result obtained is used as the cross entropy loss of the initial model, which is the above model loss.

[0125] In step A5, it can be determined whether the currently obtained model loss meets the convergence condition. If so, the current initial model can be output as the target model;

[0126] If not, the model parameters of the initial model can be updated according to the model loss, and then steps A2 to A5 are executed again based on the updated model. The process is repeated until the current model loss meets the convergence condition when A5 is executed one time.

[0127] The method of updating model parameters according to model loss can be found in the relevant prior art and will not be described in detail.

[0128] The convergence condition can be defined as the model loss being less than or equal to a preset loss threshold.

[0129] In some optional embodiments, in order to enable the above-mentioned target model to obtain more accurate news behavior characteristics, the server can periodically update the model parameters of the target model using the latest hot news according to the above-mentioned method of constructing the target model. When updating, the target model before the update can replace the initial model in the above-mentioned construction method. The other steps are consistent with the above-mentioned construction method and will not be repeated here.

[0130] based on Figure 3 The model structure shown, when using the target model to obtain news behavior features, the second embedding layer and the second fully connected module therein can be used to process news content features to obtain corresponding news behavior features. When using the target model to obtain user behavior features, the first embedding layer and the first fully connected module therein can be used to process user information to obtain corresponding user behavior features.

[0131] Compared with some existing technologies, the method of this embodiment has a significant improvement in the accuracy of exposure. That is to say, the cold start news exposed to users according to the method of this embodiment is more likely to be browsed by users for specific content. In other words, the cold start news exposed to users according to the method of this embodiment is more likely to meet the user's interest preferences.

[0132] Specifically, when evaluated by the offline recall rate indicator, the offline recall rate of the method of this embodiment can reach 0.22 when exposing cold start news. Compared with the offline recall rate of 0.20 of the generalized attribute recommendation method, the offline recall rate of the method of this embodiment can be improved by 10%. Compared with the offline recall rate of 0.19 of the content vector recommendation method, the offline recall rate of the method of this embodiment can be improved by 15.8%.

[0133] After online experiments, the application of the method of this embodiment can increase the overall click-through rate of cold start news after exposure by 2%, and at the same time increase the exposure ratio of cold start content by 5%.

[0134] Among them, the generalized attribute recommendation method refers to generalizing the cold start news to broader attributes such as categories, tags, and topics, and making recommendations based on broader attributes. For example, newly released military information can be exposed to people who like military categories, people who follow the author of the release, and people who like this topic.

[0135] The content vector recommendation method refers to using multimodal information such as pictures, texts, and videos to train a model to obtain the content vector of cold start news. Based on the similarity calculation of the vectors, the cold start news whose content vectors are similar to the user's historical behavior is exposed to the user.

[0136] This application embodiment provides a news cold start exposure device based on vector mapping, see Figure 5 , the device may include the following units.

[0137] An obtaining unit 501 is used to obtain cold start news, where the cold start news includes news whose cumulative exposure times are less than a first exposure times threshold;

[0138] A determination unit 502 is used to determine multiple target hot news that meet a similarity condition with the cold start news in the hot news, and the hot news includes news with a cumulative exposure number greater than a second exposure number threshold;

[0139] A generating unit 503 is used to generate an approximate news behavior feature of the cold start news according to the news behavior feature of the target hot news, wherein the news behavior feature of the target hot news is obtained by processing the news content feature of the target hot news by a target model, and the target model is constructed according to user feedback data of multiple hot news, and the user feedback data of the hot news indicates whether the hot news is browsed by the user after being exposed to the user;

[0140] The exposure unit 504 is used to expose the cold start news to the user according to the user similarity of the approximate news behavior characteristics of the cold start news. The user similarity represents the similarity between the approximate news behavior characteristics and the user behavior characteristics. The user behavior characteristics are obtained by processing the user information by the target model.

[0141] Optionally, the device further includes a construction unit 505, the construction unit being configured to:

[0142] Determine multiple samples based on user feedback data of multiple hot news, each sample includes user information of a user, news content features of a hot news that has been exposed to the user, and a sample label, the sample label has a first label value or a second label value, the first label value indicates that the user did not browse the hot news after exposure, and the second label value indicates that the user browsed the hot news after exposure;

[0143] Based on the initial model, the user information and news content features of each sample are processed to obtain the initial user behavior features corresponding to the user information in the sample and the initial news behavior features corresponding to the news content features;

[0144] For each sample, the prediction result of the sample is determined according to the initial user behavior characteristics and initial news behavior characteristics corresponding to the sample. The prediction result represents the probability that the user corresponding to the sample browses the hot news corresponding to the sample.

[0145] Obtain the model loss of the initial model based on the prediction results of the samples and the sample labels of the samples;

[0146] The model parameters of the initial model are updated according to the model loss, and the step of processing the user information and news content features of each sample based on the initial model is returned to execute until a target model whose model loss meets the convergence condition is obtained.

[0147] Optionally, when the obtaining unit 501 obtains the cold start news, it is configured to perform at least one of the following:

[0148] In response to obtaining user request information, obtaining cold start news;

[0149] In response to satisfying the push condition, cold start news is obtained.

[0150] Optionally, when the generating unit 503 generates the approximate news behavior features of the cold start news according to the news behavior features of the target hot news, it is used to:

[0151] Obtain news content similarity between target hot news and cold start news;

[0152] The approximate news behavior characteristics of the cold start news are determined according to the news content similarity of multiple target hot news and the news behavior characteristics of multiple target hot news.

[0153] Optionally, when exposing the cold start news to the user according to the user similarity of the approximate news behavior characteristics of the cold start news, the exposure unit 504 is used to:

[0154] When the user similarity between the approximate news behavior feature of the cold start news and the user behavior feature of the first user satisfies a first condition, exposing the cold start news to the first user;

[0155] or,

[0156] A second user whose user similarity satisfies a second condition is determined from among the multiple candidate users, and the cold start news is exposed to the second user.

[0157] The working principle of the news cold-start exposure device based on vector mapping in this embodiment can refer to the relevant steps of the news cold-start exposure method based on vector mapping provided in any embodiment of the present application, and will not be repeated here.

[0158] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0159] For the convenience of description, the above system or device is described as being divided into various modules or units according to their functions. Of course, when implementing the present application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0160] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0161] Finally, it should be noted that, in this article, relational terms such as first, second, third and fourth are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0162] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A news cold start exposure method based on vector mapping, characterized in that: include: Acquire cold start news, where the cold start news includes news whose cumulative exposure times are less than a first exposure times threshold; Determine a plurality of target hot news satisfying a similarity condition with the cold start news from the hot news, wherein the hot news includes news having a cumulative exposure count greater than a second exposure count threshold; Generate an approximate news behavior feature of the cold start news according to the news behavior feature of the target hot news, wherein the news behavior feature of the target hot news is obtained by processing the news content feature of the target hot news by a target model, wherein the target model is constructed according to user feedback data of multiple hot news, and the user feedback data of the hot news indicates whether the hot news is browsed by the user after being exposed to the user; The cold start news is exposed to users according to the user similarity of the approximate news behavior characteristics of the cold start news, wherein the user similarity represents the similarity between the approximate news behavior characteristics and the user behavior characteristics, and the user behavior characteristics are obtained by processing user information by the target model.

2. The method according to claim 1, characterized in that The method for constructing the target model according to user feedback data of multiple popular news includes: Determine a plurality of samples according to user feedback data of a plurality of hot news, each of the samples comprising user information of a user, news content features of a hot news that has been exposed to the user, and a sample label, the sample label having a first label value or a second label value, the first label value indicating that the user did not browse the hot news after being exposed, and the second label value indicating that the user browsed the hot news after being exposed; Processing the user information and news content features of each sample based on the initial model to obtain initial user behavior features corresponding to the user information in the sample and initial news behavior features corresponding to the news content features; For each of the samples, a prediction result of the sample is determined according to the initial user behavior characteristics and initial news behavior characteristics corresponding to the sample, wherein the prediction result represents the probability that the user corresponding to the sample browses the hot news corresponding to the sample; Obtaining a model loss of the initial model according to a prediction result of the sample and a sample label of the sample; The model parameters of the initial model are updated according to the model loss, and the step of processing the user information and news content features of each sample based on the initial model is returned to be executed until the target model whose model loss satisfies the convergence condition is obtained.

3. The method according to claim 1, characterized in that The obtaining of cold start news includes at least one of the following: In response to obtaining user request information, obtaining cold start news; In response to satisfying the push condition, cold start news is obtained.

4. The method according to claim 1, characterized in that: The step of generating the approximate news behavior features of the cold start news according to the news behavior features of the target hot news includes: Obtaining news content similarity between the target hot news and the cold start news; The approximate news behavior characteristics of the cold start news are determined according to the news content similarities of the plurality of target hot news and the news behavior characteristics of the plurality of target hot news.

5. The method according to claim 1, characterized in that The exposing the cold start news to the user according to the user similarity of the approximate news behavior characteristics of the cold start news includes: When the user similarity between the approximate news behavior feature of the cold start news and the user behavior feature of the first user satisfies a first condition, exposing the cold start news to the first user; or, A second user whose user similarity satisfies a second condition is determined from among a plurality of candidate users, and the cold start news is exposed to the second user.

6. A news cold start exposure device based on vector mapping, characterized in that: include: An obtaining unit, configured to obtain cold start news, wherein the cold start news includes news whose cumulative exposure times are less than a first exposure times threshold; a determination unit, configured to determine, from the hot news, a plurality of target hot news satisfying a similarity condition with the cold start news, wherein the hot news includes news having a cumulative exposure count greater than a second exposure count threshold; a generating unit, configured to generate an approximate news behavior feature of the cold start news according to the news behavior feature of the target hot news, wherein the news behavior feature of the target hot news is obtained by processing the news content feature of the target hot news by a target model, wherein the target model is constructed according to user feedback data of a plurality of hot news, wherein the user feedback data of the hot news indicates whether the hot news is browsed by the user after being exposed to the user; An exposure unit is used to expose the cold start news to users according to the user similarity of the approximate news behavior characteristics of the cold start news, wherein the user similarity represents the similarity between the approximate news behavior characteristics and the user behavior characteristics, and the user behavior characteristics are obtained by processing user information by the target model.

7. The device according to claim 6, characterized in that The device further comprises a construction unit, wherein the construction unit is used to: Determine a plurality of samples according to user feedback data of a plurality of hot news, each of the samples comprising user information of a user, news content features of a hot news that has been exposed to the user, and a sample label, the sample label having a first label value or a second label value, the first label value indicating that the user did not browse the hot news after being exposed, and the second label value indicating that the user browsed the hot news after being exposed; Processing the user information and news content features of each sample based on the initial model to obtain initial user behavior features corresponding to the user information in the sample and initial news behavior features corresponding to the news content features; For each of the samples, a prediction result of the sample is determined according to the initial user behavior characteristics and initial news behavior characteristics corresponding to the sample, wherein the prediction result represents the probability that the user corresponding to the sample browses the hot news corresponding to the sample; Obtaining a model loss of the initial model according to a prediction result of the sample and a sample label of the sample; The model parameters of the initial model are updated according to the model loss, and the step of processing the user information and news content features of each sample based on the initial model is returned to be executed until the target model whose model loss satisfies the convergence condition is obtained.

8. The device according to claim 6, characterized in that When the obtaining unit obtains the cold start news, it is used to perform at least one of the following: In response to obtaining user request information, obtaining cold start news; In response to satisfying the push condition, cold start news is obtained.

9. The device according to claim 6, characterized in that When the generating unit generates the approximate news behavior feature of the cold start news according to the news behavior feature of the target hot news, it is used to: Obtaining news content similarity between the target hot news and the cold start news; The approximate news behavior characteristics of the cold start news are determined according to the news content similarities of the plurality of target hot news and the news behavior characteristics of the plurality of target hot news.

10. The device according to claim 6, characterized in that When the exposure unit exposes the cold start news to the user according to the user similarity of the approximate news behavior feature of the cold start news, it is used to: When the user similarity between the approximate news behavior feature of the cold start news and the user behavior feature of the first user satisfies a first condition, exposing the cold start news to the first user; or, A second user whose user similarity satisfies a second condition is determined from among a plurality of candidate users, and the cold start news is exposed to the second user.

Images