Recommended method, device, storage medium and electronic equipment

Abstract

Embodiments of the present disclosure provide a recommendation method, a recommendation device, a medium and an equipment; the method comprises: establishing a user preference network based on sample user features and sample item features of items associated with the sample user; establishing a user bias network based on sample user bias features; establishing an item bias network based on sample item features; obtaining a recommendation system according to the user preference network, the user bias network and the item bias network; determining a target item to be recommended through the recommendation system, and recommending and displaying the target item in a graphical user interface. It can be seen that the technical solution of the embodiments of the present disclosure can improve the accuracy of recommending information to users.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of information processing, and in particular, to a recommendation method, a recommendation device, a computer readable storage medium and an electronic device. BACKGROUND

[0002] With the development of computer technology, network is also indispensable in life, learning and work. In these processes, users need to receive a large amount of information and browse a large amount of items on the network.

[0003] The current scheme analyzes the behavior of the user and the characteristics of the item to capture the preference relationship between the user and the item, so as to recommend different items to different users.

[0004] The above scheme only considers the user's own behavior and the characteristics of the item to analyze the preference relationship, which cannot comprehensively analyze the influencing factors and cannot achieve higher accuracy when recommending items.

[0005] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] The purpose of the embodiments of the present disclosure is to provide a recommendation method, a recommendation device, a computer readable storage medium and an electronic device, thereby at least partially solving the problem of inaccurate information pushing and improving the accuracy of information or item pushing.

[0007] A first aspect of the embodiments of the present disclosure provides a recommendation method, the method comprising:

[0008] establishing a user preference network based on sample user features and sample item features of items associated with the sample user;

[0009] establishing a user bias network based on sample user bias features;

[0010] establishing an item bias network based on sample item features;

[0011] obtaining a recommendation system according to the user preference network, the user bias network and the item bias network;

[0012] determining a target item to be recommended by the recommendation system, and recommending and displaying the target item in a graphical user interface.

[0013] In an exemplary embodiment of the present disclosure, the step of determining a target item by a recommendation system comprises:

[0014] The comprehensive probability is obtained by inputting the probability output by each network into the following formula, and the target item is determined according to the comprehensive probability of the item;

[0015] y r =y prefer +αy user ·y item

[0016] Wherein, y r is the comprehensive probability of the recommendation system, y prefer is the probability output by the user preference network, y user is the probability output by the user bias network, y item is the probability output by the item bias network, and alpha is a learning parameter.

[0017] In an exemplary embodiment of the present disclosure, the step of establishing a user preference network based on sample user features and sample item features associated with the user comprises:

[0018] Obtaining user features and sample item features, and inputting the user features and the sample item features into a user preference model;

[0019] Training to obtain a user preference network, wherein the user features include at least one of the following: valid click sequence, valid play sequence, like, comment; the sample item features include at least one of the following: number of times the item is clicked, click rate of the item.

[0020] In an exemplary embodiment of the present disclosure, the step of establishing a user bias network based on sample user bias features comprises:

[0021] Obtaining sample user bias features, and inputting the sample user bias features into a user bias model;

[0022] Training to obtain a user bias network, wherein the sample user bias features include at least one of the following: age, geographic location, accidental touch behavior, induced behavior.

[0023] In an exemplary embodiment of the present disclosure, the step of establishing an item bias network based on sample item bias features comprises:

[0024] Obtaining sample item bias features, and inputting the item bias features into an item bias model;

[0025] Training to obtain an item bias network, wherein the sample item bias features include at least one of the following: item popularity, item display position.

[0026] In an exemplary embodiment of the present disclosure, the method further comprises:

[0027] In response to the switching instruction of the user recommendation scene, a target user preference model, a target user bias model, and a target item bias model corresponding to the switched user recommendation scene are determined;

[0028] The sample user features and sample item features are input into the target user preference model, and a target user preference network is trained;

[0029] The sample user bias features are input into the target user bias model, and a target user bias network is trained;

[0030] The sample item features are input into the target item bias model, and a target item bias network is trained.

[0031] In an example embodiment of the present disclosure, the step of training the user bias network comprises:

[0032] According to the user preference network, the user bias network, and the item bias network, a comprehensive probability is obtained, and mean square error loss calculation is performed on the comprehensive probability;

[0033] The mean square error is minimized through back propagation to optimize the parameters of the network;

[0034] When the change of the loss value or the area under the receiver operating characteristic curve meets the preset convergence condition, the network parameters of the user preference network are determined.

[0035] In an example embodiment of the present disclosure, the step of training the item bias network comprises:

[0036] According to the user preference network, the user bias network, and the item bias network, a comprehensive probability is obtained, and mean square error loss calculation is performed on the comprehensive probability;

[0037] The mean square error is minimized through back propagation to optimize the parameters of the network;

[0038] When the change of the loss value or the area under the receiver operating characteristic curve meets the preset convergence condition, the network parameters corresponding to the item bias network are determined.

[0039] In an example embodiment of the present disclosure, the method further comprises:

[0040] An output vector of the user bias network and the item bias network is obtained;

[0041] The output vector is taken as a bias feature, and user classification and / or item clustering are performed based on the bias feature.

[0042] According to a second aspect of the embodiments of the present disclosure, a recommendation device is provided, and the device comprises:

[0043] a user preference module configured to establish a user preference network based on sample user features and sample item features of items associated with the sample user;

[0044] a user bias module configured to establish a user bias network based on sample user bias features;

[0045] an item bias module configured to establish an item bias network based on item bias features;

[0046] a recommendation system module configured to obtain a recommendation system according to the user preference network, the user bias network, and the item bias network;

[0047] an item recommendation module configured to determine a target item to be recommended by the recommendation system, and display the target item to a user in a graphical user interface.

[0048] In an example embodiment of the present disclosure, the item recommendation module is configured to obtain a comprehensive probability by inputting the probabilities output by the networks into the following formula, and determine the target item according to the comprehensive probability of the item:

[0049] y r = y prefer + a y user · y item

[0050] wherein y r is the comprehensive probability of the recommendation system, y prefer is the probability output by the user preference network, y user is the probability output by the user bias network, y item is the probability output by the item bias network, and a is a learning parameter.

[0051] In an example embodiment of the present disclosure, the user preference module is configured to obtain user features and sample item features, and input the user features and the sample item features into a user preference model.

[0052] and train the user preference network, wherein the user features include at least one of the following: an effective click sequence, an effective play sequence, a like, and a comment; and the sample item features include at least one of the following: a number of clicks on the item, and a click rate of the item.

[0053] In an example embodiment of the present disclosure, the user bias module is configured to

[0054] obtain sample user bias features, and input the sample user bias features into a user bias model.

[0055] The user bias network is trained, and the sample user bias features include at least one of the following: age, geographical location, accidental touch behavior, and induced behavior.

[0056] In an example embodiment of the present disclosure, the item bias module is configured to

[0057] The sample item bias features are obtained, and the item bias features are input into the item bias model.

[0058] The item bias network is trained, and the sample item bias features include at least one of the following: item popularity and item display position.

[0059] In an example embodiment of the present disclosure, the device further includes:

[0060] The scene switching module is configured to, in response to a switching instruction for a user recommendation scene, determine a target user preference model, a target user bias model, and a target item bias model corresponding to the switched user recommendation scene.

[0061] The sample user features and sample item features are input into the target user preference model, and a target user preference network is trained.

[0062] The sample user bias features are input into the target user bias model, and a target user bias network is trained.

[0063] The sample item features are input into the target item bias model, and a target item bias network is trained.

[0064] In an example embodiment of the present disclosure, the user bias module is configured to obtain a comprehensive probability according to the user preference network, the user bias network, and the item bias network, and perform mean square error loss calculation on the comprehensive probability.

[0065] The mean square error is minimized through back propagation to optimize the parameters of the network.

[0066] When the change in the loss value or the area under the receiver operating characteristic curve satisfies a preset convergence condition, the network parameters of the user preference network are determined.

[0067] In an example embodiment of the present disclosure, the item bias module is configured to obtain a comprehensive probability according to the user preference network, the user bias network, and the item bias network, and perform mean square error loss calculation on the comprehensive probability.

[0068] The mean square error is minimized through back propagation to optimize the parameters of the network.

[0069] When the change of the loss value or the area under the receiver operating characteristic curve meets a preset convergence condition, the network parameter corresponding to the item bias network is determined.

[0070] In an example embodiment of the present disclosure, the apparatus further includes:

[0071] The classification clustering module is configured to obtain output vectors of the user bias network and the item bias network.

[0072] The output vectors are taken as bias features, and user classification and / or item clustering are performed based on the bias features.

[0073] According to a third aspect of the embodiments of the present disclosure, a computer readable storage medium is provided, which stores a computer program. The program is executed by a processor to implement the recommendation method according to the first aspect of the above-described embodiments.

[0074] According to a fourth aspect of the embodiments of the present disclosure, an electronic device is provided, which includes one or more processors, and a storage device configured to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the recommendation method according to the first aspect of the above-described embodiments.

[0075] According to a sixth aspect of the present disclosure, a computer program product or computer program is provided, which includes computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to cause the computer device to perform the method provided in the various optional implementation manners described above.

[0076] The technical solutions provided by the embodiments of the present disclosure can include the following beneficial effects:

[0077] In the technical solutions provided by some embodiments of the present disclosure, the user preference network can be established based on the sample user features and the sample item features; the user bias network and the item bias network can be respectively established based on the sample user bias features and the sample item features; the recommendation system can be obtained according to the user preference network, the user bias network and the item bias network; the probability of the item can be determined by the recommendation system, and the target item can be recommended and displayed to the user. By implementing the embodiments of the present disclosure, on the one hand, the comprehensiveness of the item recommendation analysis is improved by considering the user features, the sample item features, the user bias features, and establishing the corresponding models; on the other hand, by separately setting the user preference network, the user bias network and the item bias network, the direct relationship between the bias term and the prediction probability is avoided to be ignored, the influence of the user features and the item features on the bias term is fully considered, and the accuracy of recommending the item to the user is improved.

[0078] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0079] The accompanying drawings, which are incorporated in and form a part of the specification, illustrate one embodiment of the present disclosure and, together with the description, serve to explain the principles of the present disclosure. It is appreciated that the drawings described below are only some embodiments of the present disclosure, and other drawings can be obtained according to these drawings without creative labor for those skilled in the art.

[0080] Figure 1 A schematic diagram of an exemplary terminal device is shown, which can apply the recommendation method and the recommendation device according to the embodiments of the present disclosure;

[0081] Figure 2 A flowchart of a recommendation method according to an embodiment of the present disclosure is shown schematically;

[0082] Figure 3 A structural diagram of a recommendation system according to an embodiment of the present disclosure is shown schematically;

[0083] Figure 4 A flowchart of a method after switching scenarios according to an embodiment of the present disclosure is shown schematically;

[0084] Figure 5 An effect diagram of a comparison between a curve of eliminating bias and a curve of not eliminating bias according to an embodiment of the present disclosure is shown schematically;

[0085] Figure 6 A structural block diagram of a recommendation device according to an embodiment of the present disclosure is shown schematically;

[0086] Figure 7 A structural diagram of a computer system of a terminal device suitable for implementing the embodiments of the present disclosure is shown schematically. DETAILED DESCRIPTION

[0087] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations can be implemented in any of various forms, and are not necessarily limited to the specific implementations set forth herein; rather, the specific implementations set forth herein are provided so as to enable thorough and complete disclosure of example implementations, and to fully convey the scope thereof to those skilled in the art. Many of the described features, structures, or characteristics can be combined in any suitable manner in one or more implementations. In the following description, numerous specific details are provided to enable sufficient understanding of example implementations. However, example implementations can be practiced without one or more of these specific details. In other instances, well-known structures have not been described in detail so as not to unnecessarily obscure aspects of the

[0088] In addition, the drawings are to be regarded as being schematic only and therefore are not intended to portray the relative actual size or proportions of the various components depicted therein. Like reference numerals are intended to represent like or similar components throughout the specification. Some of the blocks depicted in the drawings can be functional blocks that do not necessarily have a corresponding physical or logical structure in an implementation, and thus the various components described herein can be implemented in a variety of forms, and should not be construed as necessarily requiring the particular forms set forth herein. The functional blocks can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0089] In scenarios involving information flow pushing, the Matthew Effect is likely to have a significant impact on the recommendation results. In pushing, the causes of the Matthew Effect include the high popularity of the objects to be recommended, which have high click rates and high browsing probabilities; and the user's own click behavior, including curious clicks, accidental clicks, and the like.

[0090] Currently, to eliminate the Matthew Effect, many models use a bias term added directly to the Logit layer of the model output, so that the model learns the value of the bias term according to the data distribution and applies it to the score prediction layer. The specific formula is as follows:

[0091] y = σ (Logit + b user +b item +b)

[0092] where y is the final output prediction probability, σ (…) is a sigmoid nonlinear function, Logit is the unactivated value after forward propagation of the model, b user is the user bias term, b item is the item bias term, and b is the global bias term.

[0093] The current method, although to some extent can be according to the actual preference of the user modeling, so that the recommended results more in line with the user's wishes than other models. But, ignore the bias and the direct relationship between the predicted results, and the influence of user characteristics and item characteristics on the bias, so the accuracy is still not very high.

[0094] Figure 1 A schematic diagram of an exemplary terminal device is shown, which can apply the recommendation method and the recommendation device according to the embodiments of the present disclosure.

[0095] As Figure 1 shown, the terminal device can include one or more of the terminal devices 101, 102, 103. The terminal devices 101, 102, 103 can be various electronic devices with a display screen, including but not limited to game consoles, desktop computers, portable computers, smartphones, and tablet computers, and the like.

[0096] It should be noted that the computer readable storage medium shown in the present disclosure can be a computer readable signal medium or a computer readable storage medium or any combination of the above two. The computer readable storage medium may, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination of the above. More specific examples of computer readable storage medium can include, but are not limited to, electrical connections with one or more conductive wires, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above. In the present disclosure, the computer readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or component. In the present disclosure, the computer readable signal medium can include a data signal carried in a baseband or as part of a carrier wave, which carries computer readable program code. Such a propagated data signal can take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer readable signal medium can also be any computer readable storage medium other than the computer readable storage medium, which can send, propagate or transmit programs for use by or in conjunction with an instruction execution system, device or component. The program code contained on the computer readable storage medium can be transmitted by any suitable medium, including but not limited to wireless, wire, optical cable, etc., or any suitable combination thereof.

[0097] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings.

[0098] This example implementation provides a recommended method, see reference. Figure 2 As shown, a graphical user interface (GUI) is provided through a terminal device. The GUI includes a virtual scene containing at least two virtual characters. The user controls these virtual characters to use virtual items via the terminal device. The method includes:

[0099] Step S210: Establish a user preference network based on the characteristics of the sample users and the characteristics of the sample items associated with the sample users;

[0100] Step S220: Based on the sample user bias features, establish a user bias network;

[0101] Step S230: Based on the features of the sample items, establish an item bias network;

[0102] Step S240: Obtain a recommendation system based on the user preference network, the user bias network, and the item bias network;

[0103] Step S250: Determine the target item to be recommended through the recommendation system, and recommend and display the target item to the user in the graphical user interface.

[0104] Implementation Figure 2 The recommendation method described herein can be achieved by establishing a user preference network based on user characteristics and sample item characteristics; establishing a user bias network and an item bias network based on sample user bias characteristics and item characteristics, respectively; obtaining a recommendation system based on the user preference network, user bias network, and item bias network; and determining target items through the recommendation system and recommending the target items to users. Implementing the embodiments of this disclosure improves the comprehensiveness of item recommendation analysis by considering user characteristics, sample item characteristics, user bias characteristics, and item bias characteristics and establishing corresponding models. Furthermore, compared to current bias term strategies that eliminate the Matthew effect, by separately setting up user preference networks, user bias networks, and item bias networks, the direct relationship between bias terms and prediction probabilities is avoided, and the influence of user characteristics and item characteristics on bias terms is fully considered, thus improving the accuracy of recommending items to users.

[0105] Next, the above steps of the present example embodiment are described in more detail.

[0106] In step S210, a user preference network is established based on sample user features and sample item features of items associated with the sample user.

[0107] In the embodiments of the present disclosure, the user preference network can represent the real preference degree of the user for the item, and the network input layer of the user preference network contains both user features and sample item features. The sample item features of items associated with the sample user refer to sample item features corresponding to the sample user. For example, user A selects to watch sports video A, the duration of video A is 5 minutes, and the sports content of video A is basketball, so the sample item features associated with user A can be sports, basketball, and duration of 5 minutes.

[0108] It can be understood that the content of the sample user features and the content of the sample item features are not limited, and can be designed according to different needs of the scene of recommending items to users. When designing user behavior features, it needs to be noted that the user preference network needs to learn the real preference of the user for the item, so the input layer of the user preference network should contain user features that can reflect the real preference degree of the user for the item.

[0109] For example, in the field of short videos, when recommending short videos as items to users. The selected features such as the effective click sequence, the effective play sequence, the like, the comment, the attention sequence, the number of collections, the field to which the video belongs, and the like can express the real preference of the user for the video, and using these features as the input of the user preference network can model the real preference of the user for the item. Among them, the effective click sequence and the effective play sequence can be judged as effective or not based on a preset rule, for example, the effective click sequence is judged as whether it is watched for more than 10 seconds after clicking, if the condition is met, it is determined that this click is an effective click, otherwise it is an invalid click; the effective play sequence is determined by selecting the video sequence watched by the user for more than 10 seconds as the effective play sequence of the user, otherwise it is an invalid play sequence. As for the like and the comment as feature input, the input feature vector can include (has liked, comment word number 5), (has liked, comment word number 8), (has not liked, comment word number 2), …, and other sample item features including the number of clicks of the item, the click rate of the item, and the like, which are not limited in the embodiments of the present disclosure.

[0110] In step S220, a user bias network is established based on sample user bias features.

[0111] In the embodiments of the present disclosure, the user bias network can represent the behavior of the user due to the external environmental factors due to the user's own appearance, which is irrelevant to the user's real preference. Therefore, the network input layer of the user bias network can only contain some bias features that can reflect the real preference of the user. For example, the mis-touch behavior and the induced behavior, including the user's click or browse behavior induced by the special content of the cover picture (cover party) and the user's click or browse behavior induced by the serious inconsistency between the title and the actual content (title party).

[0112] In judging these behaviors irrelevant to the user's real preference, the actual viewing time after the user clicks and browses can be judged. If the browsing time is shorter than the preset time length, it is considered that the behavior irrelevant to the user's real preference has occurred. Or other operations, for example, after entering the browse, the user can mark the related content as not liking, cover party and title party, etc., and it is considered that the click and browse behavior is actually irrelevant to the user's real preference.

[0113] Correspondingly, the sample user bias feature can be the video file sequence clicked by the user and watched for less than a preset time length, and can also include the user's own information such as age and geographical location, etc. The embodiments of the present disclosure do not limit here.

[0114] For example, in the field of short videos, the user bias feature can include the following features: user's own information such as region, age, etc., user behavior features such as recent invalid behavior sequence features (recent invalid click sequence, recent invalid play sequence, etc.), and some features for representing the user's own stickiness to the platform, such as activity level, etc. These features can express the preference degree of the user for short videos due to the "self bias" of the user. Using these user bias features to establish a user bias network for the user to reflect the influence of the user's own bias information (such as mis-touch, curiosity, etc.) on the click rate, i.e. whether the user will or the recommendation will be successful. Taking region, age and activity level as feature inputs, the input feature vector can include (A place, 15, activity level 85), (A place, 30, activity level 70), (B place, 50, activity level 55)…….

[0115] Recent invalid click sequence, for example, by collecting all videos clicked by the user in the past 1 week, filtering out videos with <10s viewing time after clicking, and connecting these videos in the order of clicking time to get the recent invalid click sequence of the user.

[0116] Recent invalid play sequence, for example, by collecting all videos watched by the user in the past 1 week, filtering out videos with <30s viewing time, and connecting these videos in the order of starting viewing time to form the recent invalid play sequence of the user.

[0117] User stickiness features can collect the user's active days in the past month. Active is defined as the user logging into the platform on the same day, and inactive is defined as not logging in. The active days in the past month / 30 days in a month are used as the user's activity in the past month. It can be understood that different judgment criteria can be used to determine stickiness features, and the embodiments of the present disclosure are not limited herein.

[0118] It can be understood that the application scenarios with information flow pushing are not limited to live broadcast or short video fields and can be correspondingly applied. For example, it can also be an online shopping scenario. User bias features can include user invalid click sequences, which can be sequences of goods that the user browses for less than 5s after clicking the goods, user's own regional age information, etc. The goods bias features can include the current popularity information of the goods, the exposure position characteristics of the goods, etc. The embodiments of the present disclosure are not limited to the scenarios that can be applied herein.

[0119] In step S230, an item bias network is established based on the sample item features.

[0120] In the embodiments of the present disclosure, the item bias network can represent the influence on the click rate due to the item's own heat deviation or item display position deviation. In actual use scenarios, items displayed on the network platform are usually items with high heat or high position, which are more likely to be clicked by users, but these items may not be what the user really wants and do not have the user's actual preferences. In the item bias network, the sample item features are input alone, and the network input layer only includes the sample item features. The sample item features can be, for example, basic attribute feature information such as current heat, display position, display time, etc. Taking display position and heat as feature input, the input feature vector can include (interface middle, heat 10), (interface top, heat 3), (interface bottom, heat 6) ……, etc. The interface can be divided and numbered for image user interface to represent different positions of the interface. The heat feature can be measured based on various evaluation methods. For example, it can include the number of clicks, the number of shares, the number of repeated plays, etc.

[0121] In step S240, a recommendation system is obtained according to the user preference network, the user bias network, and the item bias network.

[0122] In this embodiment, the user preference network, user bias network, and item bias network are three independently constructed networks. The user preference network directly predicts the user's click-through rate, i.e., the likelihood of clicking on recommended items. The user bias network and item bias network reduce the influence of interfering factors by learning the user's own bias characteristics and the item's bias characteristics. By treating the user preference network, user bias network, and item bias network as a whole, a recommendation system that comprehensively considers the influencing factors of item recommendations can be obtained.

[0123] For example, such as Figure 3 As shown, modeling y based on an existing Click-Through-Rate (CTR) model. pred Based on this, the item bias network structure and user bias network structure are extended for modeling y. user and y item This includes a user preference network 301, a user bias network 302, and an item bias network 303. User preference features and sample item features are input into the user preference network 301; user bias features are input into the user bias network 302; and sample item features are input into the item bias network 303.

[0124] In step S250, the recommendation system determines the target item to be recommended, and the target item is recommended and displayed to the user in the graphical user interface.

[0125] In this embodiment, a recommendation system predicts the click probability of multiple items for each user. A higher click probability for an item indicates a greater likelihood that the user will click on it. The features of different items are input into the recommendation system to obtain recommendation results. Items can be displayed on the user's graphical user interface (GUI) based on their click rates. For example, items can be arranged from top to bottom according to their click rates; alternatively, different display sizes can be set for different items on the GUI based on their click rates. This embodiment does not limit how recommendations are made to the user.

[0126] This disclosure also proposes an implementation of a recommendation method, which involves determining the target item through a recommendation system, including:

[0127] The overall probability is obtained by inputting the probabilities of each network output into the formula below. The target item is determined based on the overall probability of the item.

[0128] y r =y prefer +αy user ·y item

[0129] Among them, y r Let y be the overall probability of the recommender system.prefer y is the probability of the user preference network output user y is the probability of the user bias network output item a is the probability of the item bias network output, and a is a learning parameter.

[0130] In the embodiments of the present disclosure, y r y is the comprehensive probability, that is, the output of the overall model, which is obtained by adding the user preference item and the related bias item: y prefer y is the user preference probability, that is, the user preference item, which is used to measure the actual preference degree of the user to the item; y user y is the user bias probability, y item a is the item bias probability, both of which are related bias items. Since the biases are independent, there is a multiplication relationship between the user bias and the item bias.

[0131] a is a learning parameter, representing the influence degree of the bias on the comprehensive score. The value range of a is [0, 1], and the specific value can be obtained by grid search between 0 and 1. By comparing the AUC (Area Under Curve, the area surrounded by the coordinate axis under the receiver operating characteristic curve) values of different a values on the validation set, the a that makes the AUC of the validation set highest is taken as the final learning parameter. a can also be taken as a parameter for network self-adaptive learning, and its value is dynamically adjusted during network training, and finally self-adaptively converges to the optimal value.

[0132] On different platforms, the items recommended to the user can be completely different. For example, on a literature platform, the user can be recommended to the user a classic work or a science fiction novel that the user can be interested in; on a music platform, the user can be recommended to the user pop music or recitation that the user can be interested in; and on a video platform, the user can be pushed to a TV series or a suspense short film. Different platforms correspond to different scenarios, so the applicable model is correspondingly different.

[0133] Based on this, an implementation manner of a recommendation method is also proposed in the present disclosure, as shown in Figure 4 The implementation manner includes the following steps:

[0134] S410, in response to a switching instruction of a user recommendation scenario, determining a target user preference model, a target user bias model, and a target item bias model corresponding to the switched user recommendation scenario;

[0135] S420, inputting the sample user features and sample item features to the target user preference model to train a target user preference network;

[0136] S430, inputting the sample user bias features to the target user bias model to train a target user bias network;

[0137] S440, input the sample item feature into the target item bias model, and train to obtain a target item bias network.

[0138] In the embodiments of the present disclosure, the user preference network can use a Click Through Rate (CTR) model, or other model structures. The present disclosure does not limit the replaceable model structures herein.

[0139] The user bias model and the item bias model can also be similar to the user preference model, and different model structures can be replaced according to different scene requirements to achieve more accurate prediction and recommend items to users.

[0140] Implementing the embodiments of the present disclosure can construct corresponding network structures and model structures by migrating different model structures in different scenes for the user bias model, the item bias model and the user preference model according to different recommendation scenes. This can improve the accuracy of predicted click rates in different item recommendation scenes, and further improve the accuracy of recommendations.

[0141] In the present disclosure, an implementation of a recommendation method is also proposed. The step of training the user bias network includes:

[0142] According to the user preference network, the user bias network and the item bias network, a comprehensive probability is obtained, and mean square error loss calculation is performed on the comprehensive probability.

[0143] The mean square error is minimized by back propagation to optimize the parameters of the network.

[0144] When the change of the loss value or the area under the receiver operating characteristic curve meets the preset convergence condition, the network parameters of the user preference network are determined.

[0145] In the embodiments of the present disclosure, the training process of the user bias network, the user preference network and the item bias network can be joint training, sequential training and alternating training.

[0146] Sequential training is that the knowledge graph first obtains entity and relationship vectors through embedding, and then introduces a recommendation system to learn user vectors and item vectors for model training. For example, a Deep Knowledge-aware Network (DKN) can be used. Alternating training can also be used, which alternately trains the feature learning task of the knowledge graph and the click rate prediction task of the model, and can use a Multi-task Learning for KG enhanced Recommendation (MKR) method.

[0147] For example, in the joint training manner, the training processes of the user bias network, the user preference network, and the item bias network are similar. The output scores of the user bias network, the output scores of the item bias network, and the output scores of the user preference network are added to the network outputs of the above-mentioned probability model to obtain a comprehensive probability of a final network output. The comprehensive probability is subjected to mean square error loss calculation with a real label value of an item to obtain a final sample loss value. The error loss is minimized through back propagation, so as to continuously optimize network parameters. When the AUC value on a verification set or the loss value of the network has converged and is not changed much, it is considered that a preset convergence condition is met, and the network has converged. At this time, the respective network parameters are the parameters trained when the network converges.

[0148] By training the user bias network in a joint manner, the embodiment of the present disclosure can reflect the influence of user bias reasons, such as mis-touch and being induced, on the click probability in the information or item recommendation process. The probability influencing factors can be more comprehensively analyzed, and the accuracy of recommended information or items can be improved.

[0149] In the present disclosure, an implementation manner of a recommendation method is also provided. The step of training the item bias network comprises:

[0150] The comprehensive probability is obtained according to the user preference network, the user bias network, and the item bias network, and the comprehensive probability is subjected to mean square error loss calculation.

[0151] The parameters of the network are optimized by minimizing the mean square error through back propagation.

[0152] When the loss value or the change of the area under the receiver operating characteristic curve meets the preset convergence condition, the network parameters corresponding to the item bias network are determined.

[0153] In the embodiment of the present disclosure, the item bias network is trained in a joint manner based on a similar manner of training the user bias network. The embodiment of the present disclosure does not repeat the description here.

[0154] In the present disclosure, an implementation manner of a recommendation method is also provided. The method further comprises:

[0155] The output vectors of the user bias network and the item bias network are obtained.

[0156] The output vectors are taken as bias features, and user classification and / or item clustering are performed based on the bias features.

[0157] In the embodiments of the present disclosure, the output vectors of the last layers of the user bias network and the item bias network can be used for clustering, and the output vectors are taken as the final bias features of the users or the items, and clustering is performed through the features. For example, clustering can be performed by calculating the Euclidean distance between the features, and then using the k-means (a clustering algorithm based on the Euclidean distance) to obtain different user clusters and item clusters, and data analysis is performed based on the user clusters and the item clusters.

[0158] The embodiments of the present disclosure are implemented by classifying users and / or clustering items according to the output vectors of the user bias network and the item bias network. The characteristics of the users and the items can be macroscopically evaluated, data accumulation is provided for analysis of platform operation and recommendation upgrade, and the accuracy of item recommendation is further improved.

[0159] In the actual live home page pushing scene, the model of the user bias network structure and the item bias network structure is expanded on the control model as the experiment model, and the following offline validation set AUC can be obtained, and the pushing accuracy is improved by about 0.5%. As shown in Figure 5 , the upper light curve without bias elimination and the lower dark curve with bias elimination are included.

[0160] Further, in the example embodiment, a recommendation device is also provided. As shown in Figure 6 , the recommendation device 600 can include:

[0161] The user preference module 601 is configured to establish a user preference network based on a sample user feature and a sample item feature of an item associated with the sample user.

[0162] The user bias module 602 is configured to establish a user bias network based on a sample user bias feature.

[0163] The item bias module 603 is configured to establish an item bias network based on an item bias feature.

[0164] The recommendation system module 604 is configured to obtain a recommendation system according to the user preference network, the user bias network, and the item bias network.

[0165] The item recommendation module 605 is configured to determine a target item to be recommended through the recommendation system, and display the target item to a user in a graphical user interface.

[0166] In an example embodiment of the present disclosure, the item recommendation module is configured to obtain a comprehensive probability by inputting the probabilities output by the networks into the following formula, and determine the target item according to the comprehensive probability of the item.

[0167] y r = y prefer + a y user · y item

[0168] wherein, y r is a comprehensive probability of the recommendation system, y prefer is a probability of the user preference network output, y user is a probability of the user bias network output, y item is a probability of the item bias network output, and a is a learning parameter.

[0169] In an example embodiment of the present disclosure, the user preference module is configured to obtain a user feature and a sample item feature, and input the user feature and the sample item feature into a user preference model.

[0170] The user preference network is trained, wherein the user feature comprises at least one of the following: an effective click sequence, an effective play sequence, a like, and a comment; and the sample item feature comprises at least one of the following: a number of clicks on the item and a click rate of the item.

[0171] In an example embodiment of the present disclosure, the user bias module is configured to

[0172] obtain a sample user bias feature, and input the sample user bias feature into a user bias model.

[0173] The user bias network is trained, wherein the sample user bias feature comprises at least one of the following: an age, a geographic location, a mis-touch behavior, and an induced behavior.

[0174] In an example embodiment of the present disclosure, the item bias module is configured to

[0175] obtain a sample item bias feature, and input the item bias feature into an item bias model.

[0176] The item bias network is trained, wherein the sample item bias feature comprises at least one of the following: an item popularity and an item display position.

[0177] In an example embodiment of the present disclosure, the apparatus further comprises:

[0178] a scene switching module configured to, in response to a switching instruction for a user recommendation scene, determine a target user preference model, a target user bias model, and a target item bias model corresponding to the switched user recommendation scene;

[0179] input the sample user feature and the sample item feature into the target user preference model, and train a target user preference network.

[0180] inputting the sample user bias feature into the target user bias model to obtain a target user bias network through training;

[0181] inputting the sample item feature into the target item bias model to obtain a target item bias network through training.

[0182] In an example embodiment of the present disclosure, a user bias module is configured to obtain a comprehensive probability according to the user preference network, the user bias network and the item bias network, and perform mean square error loss calculation on the comprehensive probability.

[0183] The mean square error is minimized through back propagation to optimize the parameters of the network.

[0184] When the change of the loss value or the area under the receiver operating characteristic curve meets a preset convergence condition, the network parameters of the user preference network are determined.

[0185] In an example embodiment of the present disclosure, an item bias module is configured to obtain a comprehensive probability according to the user preference network, the user bias network and the item bias network, and perform mean square error loss calculation on the comprehensive probability.

[0186] The mean square error is minimized through back propagation to optimize the parameters of the network.

[0187] When the change of the loss value or the area under the receiver operating characteristic curve meets a preset convergence condition, the network parameters corresponding to the item bias network are determined.

[0188] In an example embodiment of the present disclosure, the apparatus further comprises:

[0189] A classification and clustering module is configured to obtain output vectors of the user bias network and the item bias network.

[0190] The output vectors are taken as bias features, and user classification and / or item clustering are performed based on the bias features.

[0191] Figure 7 A structural schematic diagram of a computer system of a terminal device suitable for implementing the embodiments of the present disclosure is shown.

[0192] It should be noted that, Figure 7 The computer system of the terminal device shown is only an example, and should not impose any limitation on the functions and use range of the embodiments of the present disclosure.

[0193] As Figure 7As shown, the computer system includes a central processing unit CPU which can perform various appropriate actions and processes in accordance with programs stored in a read only memory (ROM) or programs loaded from a storage section into a random access memory RAM. In the RAM, various programs and data required for the operation of the system are also stored. The CPU, the ROM, and the RAM are connected to each other through a bus. An input / output (I / O) interface is also connected to the bus.

[0194] Connected to the (I / O) interface are an input section including a keyboard, a mouse, etc.; an output section including a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section including a hard disk, etc.; and a communication section including a network interface card such as a LAN card, a modem, etc. The communication section performs communication processing via a network such as the Internet. A drive is also connected to the (I / O) interface as necessary. A removable medium such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is attached to the drive as necessary, so that a computer program read out therefrom is installed into the storage section as necessary.

[0195] In particular, the processes described above with reference to the flowcharts can be implemented as a computer software program in accordance with embodiments of the present disclosure. For example, embodiments of the present disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for executing the methods illustrated by the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via the communication section, and / or installed from a removable medium. When the computer program is executed by the central processing unit, various functions defined in the methods and apparatuses of the present disclosure are executed. It should be noted that although several modules or units of the apparatus for action execution are mentioned in the above detailed description, such division is not mandatory. In fact, features and functions of two or more modules or units described above can be embodied in one module or unit in accordance with embodiments of the present disclosure. Conversely, features and functions of one module or unit described above can be further divided into embodied by multiple modules or units.

[0196] Since the various functional modules of the recommendation apparatus of the example embodiments of the present disclosure correspond to the steps of the example embodiments of the recommendation method described above, for details and effects not disclosed in the apparatus embodiments of the present disclosure, please refer to the above-described embodiments of the recommendation method of the present disclosure.

[0197] The modules above can be one or more integrated circuits configured to implement the methods above, for example, one or more Application Specific Integrated Circuits (ASICs), or one or more Digital Signal Processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs), etc. For another example, when a certain module above is implemented in the form of a processing element scheduling code, the processing element can be a general purpose processor, such as a Central Processing Unit (CPU) or other processor capable of invoking code. For another example, the modules can be integrated together to be implemented in the form of a System-on-a-chip (SOC).

[0198] The integrated units implemented in the form of software functional units above can be stored in a computer readable storage medium. The software functional units are stored in a storage medium, including a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor (English: processor) execute part of the steps of the methods described in various embodiments of the present application.

[0199] The above merely describes specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A recommendation method characterized by comprising: The method comprises: establishing a user preference network based on sample user features and sample item features associated with the sample user; establishing a user bias network based on sample user bias features; establishing an item bias network based on sample item bias features; obtaining a recommendation system according to the user preference network, the user bias network and the item bias network; determining a target item according to the comprehensive probability of the item, and recommending and displaying the target item in a graphical user interface; wherein, is the overall probability for the recommendation system, is the probability of the user preference network output, is the probability of the user bias network output, is the probability of the item bias network output, is the learning parameter.

2. The method of claim 1, wherein, The step of establishing a user preference network based on sample user features and sample item features associated with the sample user comprises: obtaining sample user features and sample item features, and inputting the sample user features and the sample item features into a user preference model; training to obtain a user preference network, wherein the sample user features include at least one of the following: effective click sequence, effective play sequence, like, comment; and the sample item features include at least one of the following: number of times of clicking an item, click rate of the item.

3. The method of claim 1, wherein, The step of establishing a user bias network based on sample user bias features comprises: obtaining sample user bias features, and inputting the sample user bias features into a user bias model; training to obtain a user bias network, wherein the sample user bias features include at least one of the following: age, geographic location, accidental touch behavior, induced behavior.

4. The method of claim 1, wherein, The step of establishing an item bias network based on sample item bias features comprises: obtaining sample item bias features, and inputting the sample item bias features into an item bias model; training to obtain an item bias network, wherein the sample item bias features include at least one of the following: item popularity, item display position.

5. The method of claim 2, wherein, The method further comprises: in response to a switching instruction of a user recommendation scenario, determining a target user preference model, a target user bias model and a target item bias model corresponding to the switched user recommendation scenario; inputting the sample user features and the sample item features into the target user preference model, and training to obtain a target user preference network; inputting the sample user bias features into the target user bias model, and training to obtain a target user bias network; inputting the sample item bias features into the target item bias model, and training to obtain a target item bias network.

6. The method of claim 3, wherein, The step of training to obtain a user bias network comprises: obtaining a comprehensive probability according to the user preference network, the user bias network and the item bias network, and performing mean square error loss calculation on the comprehensive probability; minimizing the mean square error through back propagation to optimize the parameters of the network; when the change of the loss value or the area under the receiver operating characteristic curve meets a preset convergence condition, determining the network parameters of the user preference network.

7. The method of claim 4, wherein, The step of training to obtain an item bias network comprises: obtaining a comprehensive probability according to the user preference network, the user bias network and the item bias network, and performing mean square error loss calculation on the comprehensive probability; minimizing the mean square error through back propagation to optimize the parameters of the network; When the change of the loss value or the area under the receiver operating characteristic curve meets a preset convergence condition, the network parameters corresponding to the item bias network are determined.

8. The method of claim 1, wherein, The method further includes: obtaining output vectors of the user bias network and the item bias network; using the output vectors as bias features, performing user classification and / or item clustering based on the bias features.

9. A recommendation apparatus characterized by comprising: The device includes: a user preference module configured to establish a user preference network based on sample user features and sample item features of items associated with the sample user; a user bias module configured to establish a user bias network based on sample user bias features; an item bias module configured to establish an item bias network based on sample item bias features; a recommendation system module configured to obtain a recommendation system according to the user preference network, the user bias network and the item bias network; an item recommendation module configured to obtain a comprehensive probability by inputting probabilities output by the networks into a formula, determine a target item according to the comprehensive probability of the item, and display the target item to the user in a graphical user interface. wherein, is the probability of the recommendation system, is the probability of the user preference network output, is the probability of the user bias network output, is the probability of the item bias network output, is the learning parameter.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by a processor, implements the recommendation method of any one of claims 1-8.

11. An electronic device, comprising: includes: one or more processors; a storage device configured to store one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement the recommendation method of any one of claims 1-8.

Images