Content item recommendation method, apparatus, server, and storage medium

Abstract

The present disclosure provides a content item recommendation method and device, a server and a storage medium, and belongs to the technical field of Internet. The method comprises the following steps: obtaining a popularity preference score of a target user and a popularity score of each candidate content item; determining a popularity preference matching score of each candidate content item according to the popularity preference score and the popularity score of each candidate content item; determining a final ranking score of each candidate content item according to an initial ranking score of each candidate content item and the popularity preference matching score; and recommending a plurality of candidate content items to the target user based on a recommendation order determined by the final ranking scores of the plurality of candidate content items. The present disclosure combines the preference of the user for popular content items and the popularity of each content item, fully considers the influence of popularity on the recommendation result in different aspects, and the recommendation result is better than simply relying on popularity or user preference.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of Internet, and particularly relates to a content item recommendation method and device, a server and a storage medium. BACKGROUND

[0002] In the technical field of Internet, popularity bias refers to a phenomenon that popular content items can obtain more recommendation opportunities and become more and more popular, and long-tail content items can only obtain fewer recommendation opportunities and become more and more unpopular. As a common problem of recommendation systems, popularity bias not only damages the personalized experience of users, but also is not conducive to the healthy development of a recommendation platform.

[0003] In order to eliminate the influence of popularity bias on recommendation results, related technologies train a content item recommendation model based on a popularity implicit representation vector and an interest implicit representation vector in a model training stage, and only perform recommendation of candidate content items based on the interest implicit representation vector of a user in a content item recommendation stage.

[0004] However, decoupling the recommendation process from popularity, although achieving the purpose of eliminating popularity bias, completely ignores the key role of popularity in user selection. In particular, in the cold start stage, popular content items have a good guiding effect on many new users, and related technologies cannot obtain better recommendation results by completely eliminating popularity bias. Therefore, in order to improve the recommendation effect of content items, it is urgent to provide a new content item recommendation method. SUMMARY

[0005] Embodiments of the present disclosure provide a content item recommendation method, device, server and storage medium, which can improve the recommendation effect of content items. The technical solution is as follows:

[0006] In a first aspect, a content item recommendation method is provided, and the method comprises:

[0007] In response to a content item acquisition request of a target user, a plurality of candidate content items are obtained, each candidate content item corresponding to an initial ranking score;

[0008] A popularity preference score of the target user and a popularity score of each candidate content item are obtained;

[0009] A popularity preference matching score of each candidate content item is determined according to the popularity preference score and the popularity score of each candidate content item, and the popularity preference matching score is used to represent the matching degree between the candidate content item and the popular content item preferred by the target user;

[0010] A final ranking score of each candidate content item is determined according to the initial ranking score and the popularity preference matching score of each candidate content item;

[0011] recommend the plurality of candidate content items to the target user in a recommended order determined based on final ranking scores of the plurality of candidate content items.

[0012] In another embodiment of the present disclosure, the obtaining the popularity preference score of the target user comprises:

[0013] When the target user is a cold-start user, a preset popularity preference score is determined as the popularity preference score.

[0014] In another embodiment of the present disclosure, the obtaining the popularity preference score of the target user comprises:

[0015] When the target user is a non-cold-start user, a positive feedback popular content item sequence, a negative feedback popular content item sequence and a user portrait of the target user are obtained, the positive feedback popular content item sequence comprising a plurality of popular content items on which the target user has performed positive operations, and the negative feedback content item sequence comprising a plurality of popular content items on which the target user has performed negative operations.

[0016] positive feedback popular content item features, negative feedback popular content item features and user features are extracted from the positive feedback popular content item sequence, the negative feedback popular content item sequence and the user portrait respectively;

[0017] Based on the positive feedback popular content item features, the negative feedback popular content item features and the user features, a popularity preference feature vector of the target user is generated.

[0018] A user popularity preference model is invoked to process the popularity preference feature vector to obtain the popularity preference score of the target user, the user popularity preference model being configured to determine a popularity preference score of any user.

[0019] In another embodiment of the present disclosure, a training process of the user popularity preference model is as follows:

[0020] A sample positive feedback popular content item sequence, a sample negative feedback popular content item sequence and a sample user portrait of a plurality of sample users are obtained, the sample positive feedback popular content item sequence comprising a plurality of popular content items on which a sample user has performed positive operations, and the sample negative feedback content item sequence comprising a plurality of popular content items on which the sample user has performed negative operations.

[0021] Sample positive feedback popular content item features, sample negative feedback popular content item features and sample user features are extracted from the sample positive feedback popular content item sequence, the sample negative feedback popular content item sequence and the sample user portrait of each sample user respectively.

[0022] generate a popularity preference feature vector of each sample user based on the sample positive feedback popular content item feature, the sample negative feedback popular content item feature and the sample user feature of each sample user;

[0023] train the initial user popularity preference model based on the sample popularity preference feature vectors of the plurality of sample users to obtain the user popularity preference model.

[0024] In another embodiment of the present disclosure, the obtaining of the popularity score of each candidate content item comprises:

[0025] obtaining content information and display information of each candidate content item;

[0026] extracting content features and display features from the content information and the display information of each candidate content item, respectively;

[0027] generating a popularity feature vector of each candidate content item based on the content features and the display features of each candidate content item;

[0028] calling the popularity estimation model to process the popularity feature vector of each candidate content item to obtain the popularity score of each candidate content item, wherein the popularity estimation model is used to determine the popularity score of any content item.

[0029] In another embodiment of the present disclosure, the training process of the popularity estimation model comprises:

[0030] obtaining a plurality of sample popular content items, each sample popular content item comprising sample content information and sample display information;

[0031] extracting sample content features and sample display features from the sample content information and the sample display information of each sample popular content item, respectively;

[0032] generating a sample popularity feature vector of each sample popular content item based on the sample content features and the sample display features of each sample popular content item;

[0033] training the initial popularity estimation model based on the sample popularity feature vectors of the plurality of sample popular content items to obtain the popularity estimation model.

[0034] In another embodiment of the present disclosure, the determining of the popularity preference matching score of each candidate content item based on the popularity preference score and the popularity score of each candidate content item comprises:

[0035] Based on the popularity preference score and the popularity score of each candidate content item, a function value of a popularity preference matching function is calculated as a popularity preference matching score of each candidate content item, the popularity preference matching function being used to represent a corresponding relationship between the popularity preference score of the user, the popularity score of the content item and the popularity preference matching score of the content item.

[0036] In another embodiment of the present disclosure, the popularity preference score includes popularity preference scores of the target user for different types of popular content items, and the determining of the popularity preference matching score of each candidate content item according to the popularity preference score and the popularity score of each candidate content item includes:

[0037] According to the type to which each candidate content item belongs, a same-type popularity preference score corresponding to each candidate content item is obtained from the popularity preference score;

[0038] Based on the popularity score of each candidate content item and the same-type popularity preference score, a function value of a popularity preference matching function is calculated as a popularity preference matching score of each candidate content item, the popularity preference matching function being used to represent a corresponding relationship between the popularity preference score of the user, the popularity score of the content item and the popularity preference matching score of the content item.

[0039] In another embodiment of the present disclosure, the determining of the final ranking score of each candidate content item according to the initial ranking score and the popularity preference matching score of each candidate content item includes:

[0040] The initial ranking score and the popularity preference matching score corresponding to each candidate content item are added as the final ranking score of each candidate content item; or,

[0041] The initial ranking score and the popularity preference matching score corresponding to each candidate content item are multiplied as the final ranking score of each candidate content item.

[0042] In a second aspect, a content item recommendation device is provided, the device comprising:

[0043] A first obtaining module is configured to obtain a plurality of candidate content items in response to a content item obtaining request of a target user, each candidate content item corresponding to an initial ranking score;

[0044] The first obtaining module is further configured to obtain a popularity preference score of the target user and a popularity score of each candidate content item;

[0045] determine a popularity preference matching score of each candidate content item according to the popularity preference score and the popularity score of each candidate content item, the popularity preference matching score being used to represent a matching degree between the candidate content item and the popular content item preferred by the target user;

[0046] The determining module is further configured to determine a final ranking score of each candidate content item according to the initial ranking score and the popularity preference matching score of each candidate content item.

[0047] The recommending module is configured to recommend the plurality of candidate content items to the target user based on a recommendation order determined according to the final ranking scores of the plurality of candidate content items.

[0048] In another embodiment of the present disclosure, the first obtaining module is configured to determine a preset popularity preference score as the popularity preference score when the target user is a cold-start user.

[0049] In another embodiment of the present disclosure, the first obtaining module is configured to, when the target user is a non-cold-start user, obtain a positive feedback popular content item sequence, a negative feedback popular content item sequence and a user portrait of the target user, the positive feedback popular content item sequence including a plurality of popular content items on which the target user has performed positive operations, the negative feedback content item sequence including a plurality of popular content items on which the target user has performed negative operations; extract positive feedback popular content item features, negative feedback popular content item features and user features from the positive feedback popular content item sequence, the negative feedback popular content item sequence and the user portrait respectively; generate a popularity preference feature vector of the target user based on the positive feedback popular content item features, the negative feedback popular content item features and the user features; and invoke a user popularity preference model to process the popularity preference feature vector to obtain the popularity preference score of the target user, the user popularity preference model being used to determine the popularity preference score of any user.

[0050] In another embodiment of the present disclosure, an apparatus for training the user popularity preference model comprises:

[0051] The second obtaining module is configured to obtain a sample positive feedback popular content item sequence, a sample negative feedback popular content item sequence and a sample user portrait of a plurality of sample users, the sample positive feedback popular content item sequence including a plurality of popular content items on which a sample user has performed positive operations, the sample negative feedback content item sequence including a plurality of popular content items on which a sample user has performed negative operations.

[0052] The first extraction module is configured to extract sample positive feedback popular content item features, sample negative feedback popular content item features and sample user features from the sample positive feedback popular content item sequence, the sample negative feedback popular content item sequence and the sample user portrait of each sample user respectively.

[0053] The first generation module is configured to generate a popularity preference feature vector of each sample user based on the sample positive feedback popular content item features, the sample negative feedback popular content item features and the sample user features of the sample user.

[0054] The first training module is configured to train an initial user popularity preference model based on the sample popularity preference feature vectors of the plurality of sample users to obtain the user popularity preference model.

[0055] In another embodiment of the present disclosure, the first acquisition module is configured to acquire content information and display information of each candidate content item, extract content features and display features from the content information and the display information of each candidate content item respectively, and generate a popularity feature vector of each candidate content item based on the content features and the display features of each candidate content item. The popularity estimation model is called to process the popularity feature vector of each candidate content item to obtain a popularity score of each candidate content item, and the popularity estimation model is configured to determine the popularity score of any content item.

[0056] In another embodiment of the present disclosure, the device for training the popularity estimation model comprises:

[0057] The third acquisition module is configured to acquire a plurality of sample popular content items, each sample popular content item comprising sample content information and sample display information.

[0058] The second extraction module is configured to extract sample content features and sample display features from the sample content information and the sample display information of each sample popular content item respectively.

[0059] The second generation module is configured to generate a sample popularity feature vector of each sample popular content item based on the sample content features and the sample display features of the sample popular content item.

[0060] The second training module is configured to train an initial popularity estimation model based on the sample popularity feature vectors of the plurality of sample popular content items to obtain the popularity estimation model.

[0061] In another embodiment of the present disclosure, the determining module is configured to calculate a function value of a popularity preference matching function as a popularity preference matching score of each candidate content item, based on the popularity preference score and the popularity score of each candidate content item, the popularity preference matching function being configured to represent a corresponding relationship between the popularity preference score of the user, the popularity score of the content item, and the popularity preference matching score of the content item.

[0062] In another embodiment of the present disclosure, the popularity preference score comprises popularity preference scores of the target user for different types of popular content items, and the determining module is configured to obtain, from the popularity preference score, a same-type popularity preference score corresponding to each candidate content item according to a type to which each candidate content item belongs, and calculate a function value of a popularity preference matching function as a popularity preference matching score of each candidate content item, based on the popularity score of each candidate content item and the same-type popularity preference score, the popularity preference matching function being configured to represent a corresponding relationship between the popularity preference score of the user, the popularity score of the content item, and the popularity preference matching score of the content item.

[0063] In a third aspect, a server is provided, the server comprising a processor and a memory, the memory storing at least one program code, the at least one program code being loaded and executed by the processor to implement the content item recommendation method according to the first aspect.

[0064] In a fourth aspect, a computer-readable storage medium is provided, the storage medium storing at least one program code, the at least one program code being loaded and executed by a processor to implement the content item recommendation method according to the first aspect.

[0065] In a fifth aspect, a computer program product is provided, the computer program product comprising computer program code stored in a computer-readable storage medium, a processor of a server reading the computer program code from the computer-readable storage medium, and the processor executing the computer program code to cause the server to perform the content item recommendation method according to the first aspect.

[0066] The technical scheme provided by the embodiments of the present disclosure has the following beneficial effects:

[0067] After obtaining the plurality of candidate content items to be recommended to the target user, a popularity preference score for representing the target user's preference degree for popular content items is obtained, and a popularity score for representing the popularity of each candidate content item is obtained, and then a popularity preference matching score for representing the matching degree of the candidate content item with the popular content item preferred by the target user is calculated based on the popularity preference score of the target user and the popularity score of each candidate content item, and then the initial ranking score determined by the recommendation system for each candidate content item is weighted and calculated by taking the popularity preference matching score as a weight value to obtain the final ranking score of each candidate content item, and then the plurality of candidate content items are recommended according to the recommendation order determined by the final ranking score of each candidate content item. This method combines the user's preference for popular content items and the popularity of each content item, fully considers the influence of popularity on the recommendation result in different aspects, and the recommendation result is better than simply relying on popularity or user preference. BRIEF DESCRIPTION OF DRAWINGS

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present disclosure, and other drawings can be obtained by those skilled in the art without creative labor.

[0069] Figure 1 is a schematic diagram of an implementation environment involved in a content item recommendation method provided by an embodiment of the present disclosure;

[0070] Figure 2 is a flowchart of a content item recommendation method provided by an embodiment of the present disclosure;

[0071] Figure 3 is a flowchart of a user popularity preference model training method provided by an embodiment of the present disclosure;

[0072] Figure 4 is a flowchart of a popularity estimation model training method provided by an embodiment of the present disclosure;

[0073] Figure 5 is a flowchart of another content item recommendation method provided by an embodiment of the present disclosure;

[0074] Figure 6 is a schematic diagram of a content item recommendation process provided by an embodiment of the present disclosure;

[0075] Figure 7 is a structural schematic diagram of a content item recommendation device provided by an embodiment of the present disclosure;

[0076] Figure 8A server for content item recommendation according to an example embodiment. DETAILED DESCRIPTION

[0077] For the purpose of the present disclosure, technical solutions and advantages, the following will be further described in detail with reference to the drawings.

[0078] It can be understood that the terms "each", "multiple" and "any" used in the embodiments of the present disclosure include two or more, each refers to each of the corresponding multiple, and any refers to any one of the corresponding multiple. For example, multiple words include 10 words, and each word refers to each of the 10 words, and any word refers to any one of the 10 words.

[0079] In the specific embodiments of the present disclosure, the related data of the user is involved, and when the above embodiments of the present disclosure are applied to specific products or technologies, the user permission or consent needs to be obtained, and the collection, use and processing of the related data need to comply with the relevant laws, regulations and standards of the country and region.

[0080] Before the embodiments of the present disclosure are executed, first, the terms involved in the embodiments of the present disclosure are explained.

[0081] Cold start refers to that for a new registered user, since the operation behavior is less or none, the original recommendation method based on user operation behavior is not applicable to the new registered user, at this time, how to recommend content items to the new registered user so that the new user can be satisfied, that is, the cold start problem.

[0082] The content item is various information browsed by the user, which refers to content that can be electrified (or digitized) by any electronic processing mode such as scanning, and content that has been electrified. The content item includes documents, images, audio and video of moving images formed by characters and (still) images. The content of the content item is very rich, which can be music, information, goods, merchants and the like.

[0083] The user portrait, also known as the crowd portrait, refers to a labeled portrait abstracted according to user demographic information, social relationship, preference habit, consumption behavior and the like. The core work of constructing the user portrait is to "label" the user, and part of the label is directly obtained according to the behavior data of the user, and part is obtained by a series of algorithms or rules.

[0084] Popularity bias is usually a fairness problem for content item parties, which is embodied in that unpopular content items have less chance to be recommended (exhibited), and accordingly, the recommendation system prefers to recommend popular content items. In the long run, popular content items become more and more popular, and unpopular content items become more and more unpopular, which is the Matthew effect in the recommendation system. Popularity bias is unfair to unpopular content items first, which can be compared to a newly opened store. The store may have good quality goods, but it is difficult to get the recommendation of the recommendation system, so the store owner has to choose another platform. Secondly, it is also not conducive to the recommendation platform. Increasing the exposure of unpopular content items can provide users with more choices, improve the diversity of recommended content items, and make the recommendation system better achieve personalization. On the other hand, it can attract more new merchants to expand the platform scale. Conversely, if the recommendation result contains a large number of popular content items, the recommendation result tends to be the same, and users cannot get a good personalized experience. Moreover, these popular content items can be discovered by users without the recommendation system, and this kind of recommendation method may also be a waste of resources.

[0085] In order to reduce the influence of popularity bias on the recommendation result, and at the same time take into account the effect of popularity on different users, the embodiment of the present disclosure provides a content item recommendation method. The method optimizes the popularity to each user according to the commonness and individuality of each user and the focus of each person on commonness and individuality. When recommending content items to each user, not only the popularity of each candidate content item is considered, but also the individualized preference of each user is considered, so that the recommended content items can better meet the needs of users. For cold start users, the method retains the positive effect of popularity on user behavior, guides users to perform more behaviors, gradually explores the interests of users, prolongs the use time of users, and improves the satisfaction of users to the recommendation result. For users who pay more attention to individualized interests, a larger proportion of similar items based on user positive feedback are exposed, reducing the damage of high popularity to user experience, and improving the use time of users.

[0086] Please refer to Figure 1 which shows an implementation environment related to the content item recommendation method provided by the embodiment of the present disclosure. The implementation environment includes a terminal 101 and a server 102.

[0087] The terminal 101 is installed with an application having a content item recommendation function, which can be a browser application, a news application, a shopping application, a social application, an audio playing application, a video playing application, etc. The terminal 101 has a display screen, can send a content item acquisition request to the server, and display a plurality of content items recommended by the server 102. The terminal 101 can be a smart phone, a tablet computer, a notebook computer, etc., and the product type of the terminal 101 is not specifically limited in the embodiments of the present disclosure.

[0088] The server 102 is a background server of the application having the content item recommendation function. The server 102 can be a stand-alone physical server, or a server cluster or a distributed system composed of a plurality of physical servers, etc. The server 102 is installed with a content item recommendation model, which includes a rough ranking model and a fine ranking model. The rough ranking model is used to sort a large number of content items recalled from the Internet, so as to reduce the large number of content items to a reasonable number. The fine ranking model is used to sort the content items after rough ranking, so as to further reduce the content items after rough ranking, and obtain a plurality of candidate content items. The server 102 is also installed with a user popularity preference model and a popularity estimation model. The user popularity preference model is used to determine a user popularity preference score. The popularity estimation model is used to determine a popularity score of each candidate content item. Based on the user popularity preference score and the popularity score of each candidate content item, the server determines a popularity preference matching score of each candidate content item, and then performs a weighted calculation on an initial ranking score output by the fine ranking model based on the popularity preference matching score of each candidate content item, to obtain a final ranking score of each candidate content item. Thus, the recommendation of the candidate content items is performed based on the final ranking scores of the plurality of candidate content items. The above content item recommendation model, user popularity preference model and popularity estimation model can be trained by the server 102, or trained by other servers, and the embodiments of the present disclosure do not specifically limit this. When the above content item recommendation model, user popularity preference model and popularity estimation model are trained by the server 102, the server 102 can load the application in the online stage after the training in the offline stage is completed. When the above content item recommendation model, user popularity preference model and popularity estimation model are trained by other servers, the server 102 can load and use after the training of the other servers is completed.

[0089] The above terminal 101 and server 102 can be directly or indirectly connected through wired or wireless communication, which is not limited in the embodiments of the present disclosure.

[0090] The embodiments of the present disclosure provide a content item recommendation method, which comprises the following steps. Figure 1 The server in the above embodiment of the present disclosure is taken as an example, and the following is described with reference toFigure 2 The method provided by the embodiments of the present disclosure comprises the following steps:

[0091] 201. In response to a content item acquisition request of a target user, a plurality of candidate content items are acquired.

[0092] The target user is a user who sends a content item acquisition request to the server by using a terminal at present. When the content item acquisition request of the target user is received, the server acquires a plurality of candidate content items, and each candidate content item in the plurality of candidate content items corresponds to an initial ranking score.

[0093] 202. A popularity preference score of the target user and a popularity score of each candidate content item are acquired.

[0094] The popularity preference score of the target user is used to represent the preference degree of the target user to popular content items. The higher the popularity preference score is, the greater the preference degree of the target user to popular content items is; on the contrary, the lower the popularity preference score is, the smaller the preference degree of the target user to popular content items is.

[0095] The popularity score of the candidate content item is used to represent the popularity of the candidate content item. The higher the popularity score is, the more popular the candidate content item is; on the contrary, the lower the popularity score is, the less popular the candidate content item is.

[0096] 203. According to the popularity preference score and the popularity score of each candidate content item, a popularity preference matching score of each candidate content item is determined.

[0097] The popularity preference matching score is used to represent the matching degree between the candidate content item and the popular content item preferred by the target user. The higher the popularity preference matching score is, the higher the degree that the candidate content item meets the preference of the target user to popular content items is; the lower the popularity preference matching score is, the lower the degree that the candidate content item meets the preference of the target user to popular content items is.

[0098] 204. According to the initial ranking score and the popularity preference matching score of each candidate content item, a final ranking score of each candidate content item is determined.

[0099] Based on the determined popularity preference matching score of each candidate content item, the server performs weighted calculation on the initial ranking score of each candidate content item by taking the popularity preference matching score of each candidate content item as a weight value, to obtain the final ranking score of each candidate content item.

[0100] 205. According to a recommended order determined based on the final ranking scores of the plurality of candidate content items, the plurality of candidate content items are recommended to the target user.

[0101] The server ranks the plurality of candidate content items in descending order of final ranking scores to obtain a recommendation order, and further recommends the candidate content items to the target user according to the recommendation order.

[0102] The method provided in the embodiments of the present disclosure obtains the popularity preference score for representing the preference degree of the target user for popular content items and the popularity score for representing the popularity of each candidate content item after obtaining the plurality of candidate content items to be recommended to the target user, and further calculates the popularity preference matching score for representing the matching degree of the candidate content item with the popular content item preferred by the target user based on the popularity preference score of the target user and the popularity score of each candidate content item, and further performs weighted calculation on the initial ranking score determined by the recommendation system for each candidate content item by taking the popularity preference matching score as a weight value to obtain the final ranking score of each candidate content item, and further recommends the plurality of candidate content items according to the recommendation order determined by the final ranking score of each candidate content item. The method combines the preference of the user for popular content items and the popularity of each content item, fully considers the influence of popularity on the recommendation result in different aspects, and the recommendation result is better than simply relying on popularity or user preference.

[0103] The embodiments of the present disclosure provide a training method of a user popularity preference model, which is executed by a server. Figure 1 The server shown in FIG. 1, or other servers, see Figure 3 The method provided in the embodiments of the present disclosure includes the following steps.

[0104] 301. The server obtains a sample positive feedback popular content item sequence, a sample negative feedback popular content item sequence of a plurality of sample users, and a sample user portrait of the plurality of sample users.

[0105] In the embodiments of the present disclosure, the operation behavior logs of a plurality of users are stored on the server, and the operation behavior logs include various behaviors of the users such as clicking, liking, collecting, deleting, and complaining, and the content items corresponding to each behavior, etc. Based on the stored operation behavior logs of the plurality of users, the server selects a plurality of sample users from the plurality of users, and obtains the operation behavior logs of the plurality of sample users, and further screens the popular content items on which each sample user has performed operations from the operation behavior logs of each sample user, and further obtains the sample positive feedback popular content item sequence and the sample negative feedback content item sequence of each sample user according to the operation behaviors performed by each sample user.

[0106] The sample positive feedback popular content item sequence includes a plurality of popular content items on which the sample user has performed positive operations. The positive operation refers to an operation that positively reflects the preference of the sample user, including clicking, liking, collecting, and the like. For example, if the sample user performs a clicking operation on a popular content item, it can be determined that the sample user is interested in this type of popular content item.

[0107] The sample negative feedback popular content item sequence includes a plurality of popular content items on which the sample user has performed negative operations. The negative operation refers to an operation that negatively reflects the preference of the sample user, including deleting, complaining, and the like. For example, if the sample user performs a deleting operation on a popular content item, it can be determined that the sample user is not interested in this type of popular content item.

[0108] 302. The server extracts sample positive feedback popular content item features, sample negative feedback popular content item features, and sample user features from the sample positive feedback popular content item sequence, the sample negative feedback popular content item sequence, and the sample user portrait of each sample user, respectively.

[0109] For any sample user in the plurality of sample users, the server obtains content features of each popular content item from each popular content item included in the sample positive feedback popular content item sequence of the sample user, and processes the content features of each popular content item, such as splicing, combining, merging, and the like, to obtain the sample positive feedback popular content item features of the sample user. The server obtains content features of each popular content item from each popular content item included in the sample negative feedback popular content item sequence of the sample user, and processes the content features of each popular content item, such as splicing, combining, merging, and the like, to obtain the sample negative feedback popular content item features of the sample user. The server extracts the sample user features from the sample user portrait of the sample user, which include age, gender, place of origin, education level, and the like.

[0110] It should be noted that the above describes taking any sample user in the plurality of sample users as an example. For other sample users in the plurality of sample users, the server processes the other sample users according to the processing manner of the one sample user described above, and finally obtains the sample positive feedback popular content item features, the sample negative feedback popular content item features, and the sample user features of each sample user.

[0111] 303. Based on the sample positive feedback popular content item features, the sample negative feedback popular content item features, and the sample user features of each sample user, the server generates a popularity preference feature vector of each sample user.

[0112] The server can splice the sample positive feedback popular content item features, the sample negative feedback popular content item features, and the sample user features of each sample user to obtain a popularity preference feature vector of each sample user. Before splicing the sample positive feedback popular content item features, the sample negative feedback popular content item features, and the sample user features of each sample user, the server can preprocess the features to be spliced to remove useless features and reduce the dimension of the generated popularity preference feature vector, thereby reducing the subsequent calculation amount.

[0113] 304. Based on the sample popularity preference feature vectors of the plurality of sample users, the server trains the initial user popularity preference model to obtain a user popularity preference model.

[0114] The server sets initial model parameters for the initial user popularity preference model, and inputs the popularity preference feature vectors of the plurality of sample users into the initial user popularity preference model to output the recognition results of the popularity preference scores of the plurality of sample users. Then, the server inputs the labeled results and the recognition results of the plurality of sample users into a first target loss function constructed in advance to obtain a function value of the first target loss function. When the function value of the first target loss function does not satisfy a first threshold condition, the server continuously adjusts the model parameters of the initial user popularity preference model until the function value of the first target loss function satisfies the first threshold condition. The server obtains the model parameters when the first threshold condition is satisfied, and then takes the initial user popularity preference model corresponding to the model parameters as the trained initial user popularity preference model. The first threshold condition can be set according to the processing accuracy of the server.

[0115] The embodiments of the present disclosure provide a training method of a popularity estimation model, which is executed by a server. Figure 1 The server shown in FIG. 1 can also be a third-party server, as shown in FIG. 2. Figure 4 The method provided in the embodiments of the present disclosure includes the following steps:

[0116] 401. The server obtains a plurality of sample popular content items.

[0117] The server can obtain popular content items in a preset time period from the network as a plurality of sample popular content items according to the recommendation function to be implemented. For example, the recommendation function to be implemented is an audio file recommendation function, and the obtained sample popular content items are sample popular audio files; for another example, the recommendation function to be implemented is an item recommendation function, and the obtained sample popular content items are sample item information, and the like. The preset time period can be 10 days, 20 days, 30 days, and the like, and the preset time period is not limited specifically in the embodiments of the present disclosure. Each sample popular content item includes sample content information and sample display information, and the like. The sample content information includes the type of the sample popular content item, the category of the sample popular content item, and the like. The sample display information includes pictures, texts, and the like displayed in the sample popular content item.

[0118] 402. The server extracts sample content features and sample display features from the sample content information and the sample display information of each sample popular content item, respectively.

[0119] The server extracts sample content features from the sample content information of each sample popular content item, and the sample content features are used to reflect the content attributes of the sample popular content item. The server extracts sample display features from the sample display information of each sample popular content item, and the sample display features are used to reflect the display form of the sample popular content item.

[0120] 403. The server generates a sample popularity feature vector of each sample popular content item based on the sample content features and the sample display features of the sample popular content item.

[0121] The server obtains the sample popularity feature vector of each sample popular content item by splicing the sample content features and the sample display features of the sample popular content item based on the sample content features and the sample display features of each sample popular content item. Before splicing the sample content features and the sample display features of each sample popular content item, the server can pre-process the sample content features and the sample display features of each sample popular content item to remove useless features, reduce the dimension of the sample popularity feature vector, and reduce the subsequent calculation amount.

[0122] 404. The server trains the initial popularity estimation model based on the sample popularity feature vectors of the plurality of sample popular content items to obtain a popularity estimation model.

[0123] The server sets initial model parameters for the initial popularity estimation model, and inputs sample popularity feature vectors of a plurality of sample popular content items into the initial popularity estimation model, and outputs identification results of popularity scores of the plurality of sample popular content items. Then the server inputs the labeled results and the identification results of the plurality of sample popular content items into the second target loss function constructed in advance, to obtain a function value of the second target loss function. When the function value of the second target loss function does not satisfy a second threshold condition, the server continuously adjusts the model parameters of the initial popularity estimation model until the function value of the second target loss function satisfies the second threshold condition. The server obtains the model parameters when the second threshold condition is satisfied, and then takes the initial popularity estimation model corresponding to the model parameters as the trained popularity estimation model. The second threshold condition can be set according to the processing accuracy of the server.

[0124] The embodiment of the present disclosure provides a content item recommendation method, which comprises the following steps: Figure 1 The server in the embodiment of the present disclosure is taken as an example, and the method provided by the embodiment of the present disclosure comprises the following steps: Figure 5 The method provided by the embodiment of the present disclosure comprises the following steps:

[0125] 501. In response to a content item acquisition request of a target user, the server acquires a plurality of candidate content items.

[0126] In response to a content item acquisition request of a target user, the server calls a coarse ranking model to score a plurality of content items acquired from the Internet, and then filters a plurality of content items from the recalled plurality of content items based on the scoring results, and then calls a fine ranking model to score the plurality of content items filtered by the coarse ranking model, and then filters a plurality of candidate content items from the plurality of content items filtered by the coarse ranking model based on the scoring results, each candidate content item corresponding to an initial ranking score, which is the scoring result of the fine ranking model.

[0127] It should be noted that the fine ranking model used in the embodiment of the present disclosure can be any fine ranking model currently used in the market, that is, the fine ranking model can be a model trained based on a user's interest implicit representation vector and a popularity implicit representation vector, or a model trained based on a user's interest implicit representation vector. In the recommendation stage, the fine ranking model can score based on the interest implicit representation vector of the target user, or score based on the interest implicit representation vector and the popularity implicit representation vector of the target user, and the embodiment of the present disclosure does not make specific limitation.

[0128] 502. The server obtains a popularity preference score of the target user and a popularity score of each candidate content item.

[0129] In the embodiments of the present disclosure, the target user can be a cold start user or a non-cold start user. The server adopts different methods to obtain the popularity preference score of the target user according to the different target users.

[0130] In a possible implementation, when the target user is a cold start user, the server does not store any information of the target user. In this case, the server can determine a preset popularity preference score as the popularity preference score of the target user. The preset popularity preference score is a score that can regard a popular content item as a preferred content item of the target user.

[0131] In another possible implementation, when the target user is a non-cold start user, the server can obtain the popularity preference score of the target user based on the user popularity preference model trained in the above-mentioned embodiments.

[0132] The specific steps are as follows:

[0133] In the first step, the server obtains the positive feedback popular content item sequence, the negative feedback popular content item sequence and the user portrait of the target user.

[0134] The server obtains the operation behavior log of the target user, and then obtains the popular content items from the content items on which the target user has performed operation behaviors, and then obtains the positive feedback popular content item sequence and the negative feedback popular content item sequence of the target user according to the operation behaviors performed by the target user. The positive feedback popular content item sequence includes a plurality of popular content items on which the target user has performed positive operations, and the negative feedback content item sequence includes a plurality of popular content items on which the target user has performed negative operations.

[0135] In the second step, the server extracts the positive feedback popular content item feature, the negative feedback popular content item feature and the user feature from the positive feedback popular content item sequence, the negative feedback popular content item sequence and the user portrait respectively.

[0136] The server extracts the positive feedback popular content item feature from the positive feedback popular content item sequence, extracts the negative feedback popular content item feature from the negative feedback popular content item sequence, and extracts the user feature from the user portrait.

[0137] In the third step, the server generates the popularity preference feature vector of the target user based on the positive feedback popular content item feature, the negative feedback popular content item feature and the user feature.

[0138] The server splices the positive feedback popular content item feature, the negative feedback popular content item feature, and the user feature to obtain a popularity preference feature vector of the target user. Before splicing the positive feedback popular content item feature, the negative feedback popular content item feature, and the user feature of the target user, the server can pre-process the features to be spliced to remove useless features and reduce the dimension of the generated popularity preference feature vector.

[0139] In the fourth step, the server calls a user popularity preference model to process the popularity preference feature vector to obtain a popularity preference score of the target user.

[0140] When the server obtains the popularity score of each candidate content item, the following method can be used:

[0141] In the first step, the server obtains content information and display information of each candidate content item.

[0142] In the second step, the server extracts content features and display features from the content information and the display information of each candidate content item.

[0143] In the third step, the server generates a popularity feature vector of each candidate content item based on the content features and the display features of each candidate content item.

[0144] The server splices the content features and the display features of each candidate content item to obtain a popularity feature vector of each candidate content item. Before splicing, the server can pre-process the features to be spliced to remove useless features, reduce the dimension of the popularity feature vector, and reduce the subsequent calculation amount.

[0145] In the fourth step, the server calls a popularity estimation model to process the popularity feature vector of each candidate content item to obtain a popularity score of each candidate content item.

[0146] 503. The server determines a popularity preference matching score of each candidate content item according to the popularity preference score and the popularity score of each candidate content item.

[0147] In the embodiments of the present disclosure, the server can construct a popularity preference matching function according to specific business needs, where the popularity preference matching function is used to represent the corresponding relationship between the popularity preference score of the user, the popularity score of the content item, and the popularity preference matching score of the content item. Based on the constructed popularity preference matching function, the server determines the popularity preference matching score of each candidate content item, including but not limited to the following two cases:

[0148] In the first case, the popularity preference score is used to measure the preference degree of the target user for popular content items as a whole.

[0149] In this case, the server calculates, as a popularity preference matching score of each candidate content item, a function value of a popularity preference matching function by inputting the popularity preference score and the popularity score of each candidate content item into the popularity preference matching function pre-constructed.

[0150] The second case is that the popularity preference score is used to measure the preference degree of the target user for different types of popular content items respectively, and the popularity preference score includes popularity preference scores of the target user for different types of popular content items.

[0151] In this case, the server can obtain, from the popularity preference score, a same-type popularity preference score corresponding to each candidate content item according to a type to which each candidate content item belongs, and then calculate, as a popularity preference matching score of each candidate content item, a function value of a popularity preference matching function based on the popularity score of each candidate content item and the same-type popularity preference score.

[0152] 504. The server determines a final ranking score of each candidate content item according to the initial ranking score and the popularity preference matching score of each candidate content item.

[0153] In a possible implementation, based on the initial ranking score and the popularity preference matching score of each candidate content item, the server can add the initial ranking score and the popularity preference matching score corresponding to each candidate content item to obtain a final ranking score of each candidate content item.

[0154] In another possible implementation, based on the initial ranking score and the popularity preference matching score of each candidate content item, the server can multiply the initial ranking score and the popularity preference matching score corresponding to each candidate content item to obtain a final ranking score of each candidate content item.

[0155] 505. The server determines a recommendation order based on the final ranking scores of the plurality of candidate content items.

[0156] Based on the final ranking scores of the plurality of candidate content items, the server ranks the candidate content items in a descending order of the final ranking scores to obtain a recommendation order of the plurality of candidate content items.

[0157] 506. Based on the determined recommendation order, the server recommends the plurality of candidate content items to the target user.

[0158] Based on the determined recommendation order, the server selects candidate content items ranked in the top N positions from the candidate content items, and then recommends the selected candidate content items to the target user. N can be 20, 50, or the like.

[0159] Figure 6 The whole recommendation process of the content item recommendation method provided by the embodiments of the present disclosure is shown, see Figure 6 When receiving the content item acquisition request sent by the user, the server acquires a plurality of candidate content items, and based on the stored operation behavior log of the user, acquires the positive feedback popular content item sequence, the negative feedback popular content item sequence and the user portrait of the user, then extracts the positive feedback popular content item feature, the negative feedback popular content item feature and the user feature from the positive feedback popular content item sequence, the negative feedback popular content item sequence and the user portrait respectively, generates a popularity preference feature vector based on the extracted positive feedback popular content item feature, negative feedback popular content item feature and user feature, then calls the user popularity preference model to process the popularity preference feature vector to obtain the popularity preference score of the user. After acquiring the plurality of candidate content items, the server further extracts the content feature and the display feature from the content information and the display information of each candidate content item respectively, and generates the popularity feature vector of each candidate content item based on the content feature and the display feature of each candidate content item, then calls the popularity estimation model to process the popularity feature vector of each candidate content item to obtain the popularity score of each candidate content item. Then, the server inputs the popularity preference score of the user and the popularity score of each candidate content item into the popularity preference matching function to obtain the popularity preference matching score of each candidate content item, and further determines the final ranking score of each candidate content item according to the initial ranking score and the popularity preference matching score of each candidate content item, and finally recommends the plurality of candidate content items to the target user based on the recommendation order determined by the final ranking scores of the plurality of candidate content items.

[0160] The method provided by the embodiments of the present disclosure, after acquiring the plurality of candidate content items to be recommended to the target user, acquires the popularity preference score for representing the preference degree of the target user to popular content items, and the popularity score for representing the popularity of each candidate content item, and further calculates the popularity preference matching score for representing the matching degree of the candidate content item to the preferred popular content item of the target user based on the popularity preference score of the target user and the popularity score of each candidate content item, and further performs weighted calculation on the initial ranking score of each candidate content item determined by the recommendation system by taking the popularity preference matching score as a weight value to obtain the final ranking score of each candidate content item, and further recommends the plurality of candidate content items according to the recommendation order determined by the final ranking score of each candidate content item. The method combines the preference of the user to popular content items and the popularity of each content item, fully considers the influence of popularity on the recommendation result in different aspects, and the recommendation result is better than simply relying on popularity or user preference.

[0161] See Figure 7The embodiment of the present disclosure provides a content item recommendation device, which comprises:

[0162] The first acquisition module 701 is configured to acquire a plurality of candidate content items in response to a content item acquisition request of a target user, each candidate content item corresponding to an initial ranking score;

[0163] The first acquisition module 701 is further configured to acquire a popularity preference score of the target user and a popularity score of each candidate content item;

[0164] The determination module 702 is configured to determine a popularity preference matching score of each candidate content item according to the popularity preference score and the popularity score of each candidate content item, the popularity preference matching score being used to represent a matching degree between the candidate content item and a popular content item preferred by the target user;

[0165] The determination module 702 is further configured to determine a final ranking score of each candidate content item according to the initial ranking score and the popularity preference matching score of each candidate content item;

[0166] The recommendation module 703 is configured to recommend the plurality of candidate content items to the target user based on a recommendation order determined by the final ranking scores of the plurality of candidate content items.

[0167] In another embodiment of the present disclosure, the first acquisition module is configured to determine a preset popularity preference score as the popularity preference score when the target user is a cold start user.

[0168] In another embodiment of the present disclosure, the first acquisition module is configured to acquire a positive feedback popular content item sequence, a negative feedback popular content item sequence and a user portrait of the target user when the target user is a non-cold start user, the positive feedback popular content item sequence comprising a plurality of popular content items on which the target user has implemented positive operations, the negative feedback content item sequence comprising a plurality of popular content items on which the target user has implemented negative operations; extract positive feedback popular content item features, negative feedback popular content item features and user features from the positive feedback popular content item sequence, the negative feedback popular content item sequence and the user portrait respectively; generate a popularity preference feature vector of the target user based on the positive feedback popular content item features, the negative feedback popular content item features and the user features; call a user popularity preference model to process the popularity preference feature vector to obtain the popularity preference score of the target user, the user popularity preference model being used to determine the popularity preference score of any user.

[0169] In another embodiment of the present disclosure, a device for training a user popularity preference model comprises:

[0170] The second obtaining module is configured to obtain a sample positive feedback popular content item sequence, a sample negative feedback popular content item sequence and a sample user portrait of a plurality of sample users, the sample positive feedback popular content item sequence comprising a plurality of popular content items on which a sample user has performed a positive operation, and the sample negative feedback popular content item sequence comprising a plurality of popular content items on which the sample user has performed a negative operation.

[0171] The first extraction module is configured to extract a sample positive feedback popular content item feature, a sample negative feedback popular content item feature and a sample user feature from the sample positive feedback popular content item sequence, the sample negative feedback popular content item sequence and the sample user portrait of each sample user, respectively.

[0172] The first generation module is configured to generate a popularity preference feature vector of each sample user based on the sample positive feedback popular content item feature, the sample negative feedback popular content item feature and the sample user feature of the sample user.

[0173] The first training module is configured to train an initial user popularity preference model based on the sample popularity preference feature vectors of the plurality of sample users to obtain the user popularity preference model.

[0174] In another embodiment of the present disclosure, the first obtaining module is configured to obtain content information and display information of each candidate content item, extract a content feature and a display feature from the content information and the display information of each candidate content item, and generate a popularity feature vector of each candidate content item based on the content feature and the display feature of each candidate content item. The popularity estimation model is called to process the popularity feature vector of each candidate content item to obtain a popularity score of each candidate content item, and the popularity estimation model is configured to determine the popularity score of any content item.

[0175] In another embodiment of the present disclosure, the device for training the popularity estimation model comprises:

[0176] The third obtaining module is configured to obtain a plurality of sample popular content items, each sample popular content item comprising sample content information and sample display information.

[0177] The second extraction module is configured to extract a sample content feature and a sample display feature from the sample content information and the sample display information of each sample popular content item.

[0178] The second generation module is configured to generate a sample popularity feature vector of each sample popular content item based on the sample content feature and the sample display feature of each sample popular content item.

[0179] The second training module is configured to train an initial popularity estimation model based on the sample popularity feature vectors of the plurality of sample popular content items to obtain the popularity estimation model.

[0180] In another embodiment of the present disclosure, the determining module is configured to calculate, as the popularity preference matching score of each candidate content item, a function value of a popularity preference matching function based on the popularity preference score and the popularity score of each candidate content item, the popularity preference matching function being configured to represent a corresponding relationship between the popularity preference score of the user, the popularity score of the content item and the popularity preference matching score of the content item.

[0181] In another embodiment of the present disclosure, the popularity preference score includes popularity preference scores of different types of popular content items by the target user, and the determining module is configured to obtain, from the popularity preference score, a same-type popularity preference score corresponding to each candidate content item according to a type to which each candidate content item belongs, and calculate, as the popularity preference matching score of each candidate content item, a function value of a popularity preference matching function based on the popularity score of each candidate content item and the same-type popularity preference score, the popularity preference matching function being configured to represent a corresponding relationship between the popularity preference score of the user, the popularity score of the content item and the popularity preference matching score of the content item.

[0182] In summary, the apparatus provided by the embodiments of the present disclosure, after obtaining a plurality of candidate content items to be recommended to a target user, obtains a popularity preference score representing a preference degree of the target user for popular content items and a popularity score representing a popularity of each candidate content item, and then calculates a popularity preference matching score representing a matching degree of the candidate content item with the popular content item preferred by the target user based on the popularity preference score of the target user and the popularity score of each candidate content item, and then performs weighted calculation on an initial ranking score determined by the recommendation system for each candidate content item by taking the popularity preference matching score as a weight value, to obtain a final ranking score of each candidate content item, and then recommends the plurality of candidate content items according to a recommendation order determined by the final ranking score of each candidate content item. This method combines the preference of the user for popular content items and the popularity of each content item, fully considers the influence of popularity on the recommendation result in different aspects, and provides a better recommendation result compared with simply relying on popularity or user preference.

[0183] Figure 8 is a server for content item recommendation according to an exemplary embodiment. Referring to Figure 8The server 800 includes a processing component 822, further including one or more processors, and a memory resource represented by the memory 832, for storing instructions, such as an application program, executable by the processing component 822. The application program stored in the memory 832 can include one or more than one module each corresponding to a set of instructions. In addition, the processing component 822 is configured to execute the instructions to perform the functions of the server in the content item recommendation method described above.

[0184] The server 800 can further include a power supply component 826 configured to perform power management of the server 800, a wired or wireless network interface 850 configured to connect the server 800 to a network, and an input output (I / O) interface 858. The server 800 can operate based on an operating system stored in the memory 832, such as Windows Server TM , Mac OSX TM , Unix TM , Linux TM , FreeBSD TM or the like.

[0185] The embodiments of the present disclosure provide a computer readable storage medium, the storage medium stores at least one program code, the at least one program code is loaded and executed by a processor to implement a content item recommendation method. The computer readable storage medium can be non-transitory. For example, the computer readable storage medium can be a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a CD-ROM, a magnetic tape, a floppy disk and an optical data storage device, etc.

[0186] The embodiments of the present disclosure provide a computer program product, the computer program product includes computer program code stored in a computer readable storage medium, a processor of a server reads the computer program code from the computer readable storage medium, and the processor executes the computer program code to make the server perform a content item recommendation method.

[0187] Those of ordinary skill in the art can understand that all or part of the steps of the above-mentioned embodiments can be completed by hardware, or by programs instructing relevant hardware to complete, and the programs can be stored in a computer readable storage medium, and the above-mentioned storage medium can be a read-only memory, a magnetic disk or an optical disk, etc.

[0188] The above only describes optional embodiments of the present disclosure, and does not limit the present disclosure, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A content item recommendation method characterized by, The method comprises: in response to a content item acquisition request of a target user, calling a coarse ranking model to acquire a plurality of content items for scoring, and screening a plurality of content items; calling a fine ranking model to score the plurality of content items, and screening a plurality of candidate content items from the plurality of content items based on the scoring results, each candidate content item corresponding to an initial ranking score; acquiring a popularity preference score of the target user and a popularity score of each candidate content item, the popularity preference score being used to represent the preference degree of the target user to popular content items, the higher the popularity preference score, the greater the corresponding preference degree, the popularity preference score of a cold start user being a score capable of taking popular content items as the preferred content items of the cold start user, and the popularity preference score of a non-cold start user being determined based on the operation behavior of the non-cold start user to popular content items and the user characteristics of the non-cold start user; determining a popularity preference matching score of each candidate content item according to the popularity preference score and the popularity score of each candidate content item, the popularity preference matching score being used to represent the matching degree between the candidate content item and the popular content item preferred by the target user, the higher the popularity preference matching score, the higher the degree of the corresponding candidate content item meeting the preference of the target user to popular content items; determining a final ranking score of each candidate content item according to the initial ranking score and the popularity preference matching score of each candidate content item; recommending the plurality of candidate content items to the target user based on the recommendation order determined based on the final ranking scores of the plurality of candidate content items.

2. The method of claim 1, wherein, The method comprises: when the target user is a non-cold start user, acquiring a positive feedback popular content item sequence, a negative feedback popular content item sequence and a user portrait of the target user, the positive feedback popular content item sequence comprising a plurality of popular content items on which the target user has implemented positive operations, and the negative feedback popular content item sequence comprising a plurality of popular content items on which the target user has implemented negative operations; extracting positive feedback popular content item features, negative feedback popular content item features and user features from the positive feedback popular content item sequence, the negative feedback popular content item sequence and the user portrait respectively; generating a popularity preference feature vector of the target user based on the positive feedback popular content item features, the negative feedback popular content item features and the user features; calling a user popularity preference model to process the popularity preference feature vector to obtain the popularity preference score of the target user, the user popularity preference model being used to determine the popularity preference score of any user.

3. The method of claim 2, wherein, The training process of the user popularity preference model comprises: acquiring sample positive feedback popular content item sequences, sample negative feedback popular content item sequences and sample user portraits of a plurality of sample users, the sample positive feedback popular content item sequences comprising a plurality of popular content items on which the sample users have implemented positive operations, and the sample negative feedback popular content item sequences comprising a plurality of popular content items on which the sample users have implemented negative operations; extracting sample positive feedback popular content item features, sample negative feedback popular content item features and sample user features from the sample positive feedback popular content item sequence, the sample negative feedback popular content item sequence and the sample user portrait of each sample user respectively; generating a popularity preference feature vector of each sample user based on the sample positive feedback popular content item features, the sample negative feedback popular content item features and the sample user features of each sample user; training an initial user popularity preference model based on the sample popularity preference feature vectors of the plurality of sample users to obtain the user popularity preference model.

4. The method of claim 1, wherein, obtaining a popularity score of each candidate content item, including: obtaining content information and display information of each candidate content item; extracting content features and display features from the content information and the display information of each candidate content item respectively; generating a popularity feature vector of each candidate content item based on the content features and the display features of each candidate content item; calling a popularity estimation model to process the popularity feature vector of each candidate content item to obtain the popularity score of each candidate content item, the popularity estimation model being configured to determine the popularity score of any content item.

5. The method of claim 4, wherein, The training process of the popularity estimation model includes: obtaining a plurality of sample popular content items, each sample popular content item including sample content information and sample display information; extracting sample content features and sample display features from the sample content information and the sample display information of each sample popular content item respectively; generating a sample popularity feature vector of each sample popular content item based on the sample content features and the sample display features of each sample popular content item; training an initial popularity estimation model based on the sample popularity feature vectors of the plurality of sample popular content items to obtain the popularity estimation model.

6. The method of claim 1, wherein, The determination of the popularity preference matching score of each candidate content item based on the popularity preference score and the popularity score of each candidate content item includes: calculating a function value of a popularity preference matching function as the popularity preference matching score of each candidate content item based on the popularity preference score and the popularity score of each candidate content item, the popularity preference matching function being configured to represent the corresponding relationship between the popularity preference score of the user, the popularity score of the content item and the popularity preference matching score of the content item.

7. The method of claim 1, wherein, The popularity preference score includes popularity preference scores of the target user for different types of popular content items, and the determination of the popularity preference matching score of each candidate content item based on the popularity preference score and the popularity score of each candidate content item includes: obtaining the same type of popularity preference score corresponding to each candidate content item from the popularity preference score according to the type to which each candidate content item belongs. Based on the popularity score of each candidate content item and the popularity preference score of the same type, a function value of a popularity preference matching function is calculated as a popularity preference matching score of each candidate content item, the popularity preference matching function being used to represent a corresponding relationship between the popularity preference score of the user, the popularity score of the content item and the popularity preference matching score of the content item.

8. The method of claim 1, wherein, The determining of the final ranking score of each candidate content item according to the initial ranking score and the popularity preference matching score of each candidate content item comprises: adding the initial ranking score and the popularity preference matching score of each candidate content item as the final ranking score of each candidate content item; or multiplying the initial ranking score and the popularity preference matching score of each candidate content item as the final ranking score of each candidate content item.

9. A content item recommendation apparatus characterized by comprising: The apparatus comprises: The first obtaining module is configured to, in response to a content item obtaining request of a target user, call a coarse ranking model to obtain a plurality of content items for scoring, and filter a plurality of content items; call a fine ranking model to score the plurality of content items, and filter a plurality of candidate content items from the plurality of content items based on the scoring result, each candidate content item corresponding to an initial ranking score; The first obtaining module is further configured to obtain a popularity preference score of the target user and a popularity score of each candidate content item, the popularity preference score being used to represent a preference degree of the target user to popular content items, the higher the popularity preference score, the greater the corresponding preference degree, the popularity preference score of a cold start user being a score of popular content items that can be preferred by the cold start user, and the popularity preference score of a non-cold start user being determined based on operation behaviors of the non-cold start user to popular content items and user features of the non-cold start user; The determining module is configured to determine a popularity preference matching score of each candidate content item according to the popularity preference score and the popularity score of each candidate content item, the popularity preference matching score being used to represent a matching degree between the candidate content item and popular content items preferred by the target user, the higher the popularity preference matching score, the higher the degree to which the corresponding candidate content item meets the preference of the target user to popular content items. The determining module is further configured to determine a final ranking score of each candidate content item according to the initial ranking score and the popularity preference matching score of each candidate content item. The recommendation module is configured to recommend the plurality of candidate content items to the target user based on a recommendation order determined based on the final ranking scores of the plurality of candidate content items.

10. A server, characterized by The server comprises a processor and a memory, the memory storing at least one program code, the at least one program code being loaded and executed by the processor to implement the content item recommendation method according to any one of claims 1 to 8.

11. A computer readable storage medium, characterized in that, The storage medium stores at least one program code, the at least one program code being loaded and executed by the processor to implement the content item recommendation method according to any one of claims 1 to 8.

12. A computer program product, characterised in that, The computer program product comprises computer program code stored in a computer readable storage medium, the computer program code being read by a processor of a server from the computer readable storage medium, the processor executing the computer program code causing the server to perform the content item recommendation method according to any one of claims 1 to 8.

Images