Media recommendation model training and recommendation method and device, and storage medium

Abstract

The application discloses a media recommendation model training and recommendation method and device and a storage medium. The training method comprises the following steps: training a media recommendation model based on a sample data set to obtain a trained media recommendation model; wherein each sample data in the sample data set comprises user feature representation information, trigger media representation information, candidate media representation information, behavior sequence representation information and trigger context representation information of a user, and the candidate media representation information has a label value; the media recommendation model comprises a confused user identification model for confused user identification, a real-time interest perception model for identifying a real-time interest representation of a current user, and a prediction model for predicting a click probability value of a candidate media based on a first identification result of the confused user identification model and a second identification result of the real-time interest perception model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, and particularly relates to a training and recommendation method and device of a media asset recommendation model and a storage medium. BACKGROUND

[0002] With the development of the Internet era, the information around people grows exponentially. In order to solve the problem of information overload, researchers have proposed a personalized recommendation system to solve the problem of information overload. With the development of the industry, the personalized recommendation system increases user activity and improves product market competitiveness, which brings huge commercial value to companies and individuals.

[0003] In the related art, a user can play media assets that the user wants to watch on a smart television at any time. However, with the increasing number of media assets, how to help the user quickly find the media assets of interest becomes a technical problem to be solved. SUMMARY

[0004] Therefore, the embodiments of the present application provide a training and recommendation method, device and storage medium of a media asset recommendation model, aiming to improve the accuracy of media asset recommendation.

[0005] The technical solutions of the embodiments of the present application are as follows:

[0006] In a first aspect, the embodiments of the present application provide a training method of a media asset recommendation model, comprising:

[0007] training the media asset recommendation model based on a sample data set to obtain a trained media asset recommendation model;

[0008] The sample data in the sample data set includes user feature representation information of a user, trigger media asset representation information, candidate media asset representation information, behavior sequence representation information and trigger context representation information, and the candidate media asset representation information has a label value. The media asset recommendation model includes a confused user identification model for confused user identification, a real-time interest perception model for identifying a real-time interest representation of a current user, and a prediction model for predicting a click probability value of a candidate media asset based on a first identification result of the confused user identification model and a second identification result of the real-time interest perception model.

[0009] In the above solution, the training of the media asset recommendation model based on the sample data set to obtain the trained media asset recommendation model comprises:

[0010] training the media asset recommendation model based on the sample data in the sample data set to obtain a click probability value of a candidate media asset prediction;

[0011] calculating a loss value based on a loss function for the click probability value and the label value;

[0012] determining that the loss value converges or the number of iterations of training reaches a set number of times, to obtain the trained media recommendation model.

[0013] In the foregoing solution, the training of the media recommendation model based on the sample data in the sample data set and the obtaining of the predicted click probability value of the candidate media include:

[0014] obtaining the first identification result based on the user feature representation information, the trigger media representation information, the trigger context representation information and the behavior sequence representation information of the sample data and the confused user identification model;

[0015] obtaining the second identification result based on the behavior sequence representation information of the sample data and the real-time interest perception model;

[0016] obtaining the predicted click probability value of the candidate media based on the first identification result, the second identification result and the prediction model.

[0017] In the foregoing solution, the obtaining of the first identification result based on the user feature representation information, the trigger media representation information, the trigger context representation information and the behavior sequence representation information of the sample data and the confused user identification model includes:

[0018] generating trigger environment representation information based on the trigger media representation information and the trigger context representation information of the sample data;

[0019] processing the trigger environment representation information and the behavior sequence representation information based on an attention mechanism to obtain guided user feature representation information;

[0020] performing confused user identification on the user feature representation information and the guided user feature representation information based on a feedforward neural network to obtain the first identification result.

[0021] In the foregoing solution, the obtaining of the second identification result based on the behavior sequence representation information of the sample data and the real-time interest perception model includes:

[0022] processing the behavior sequence representation information based on a multi-head self-attention mechanism to obtain multi-head representation information;

[0023] performing entanglement disentangling on the multi-head representation information and the trigger environment representation information to obtain disentangled interest head representation information;

[0024] performing global weight calculation on the interest head representation information to obtain the second identification result.

[0025] In the solution, the click probability value of the candidate media asset is obtained based on the first recognition result, the second recognition result and the prediction model, and the solution comprises the following steps:

[0026] The first recognition result, the second recognition result, the user feature representation information, the trigger media asset representation information, the trigger context representation information, the behavior sequence representation information and the candidate media asset representation information are spliced to obtain a spliced result.

[0027] The spliced result is input into a multi-layer perception machine to obtain the click probability value of the candidate media asset prediction.

[0028] In a second aspect, an embodiment of the present application provides a media asset recommendation method, which comprises the following steps:

[0029] User feature representation information, trigger media asset representation information, behavior sequence representation information, trigger context representation information and candidate media asset representation information of a user are obtained.

[0030] The user feature representation information, the trigger media asset representation information, the behavior sequence representation information, the trigger context representation information and the candidate media asset representation information are input into a media asset recommendation model trained by the method of the first aspect of the present application to obtain a click probability value of each candidate media asset prediction.

[0031] In the solution, the method further comprises the following steps:

[0032] The click probability values are sorted.

[0033] The first N candidate media assets in a descending order of the click probability values are selected, and the first N candidate media assets are determined as the current recommended candidate media assets.

[0034] Wherein, N is a natural number greater than 1.

[0035] In a third aspect, an embodiment of the present application provides a training device of a media asset recommendation model, which comprises the following steps:

[0036] A training module is configured to train a media asset recommendation model based on a sample data set to obtain a trained media asset recommendation model.

[0037] The sample data in the sample data set includes user feature representation information of a user, trigger media asset representation information, candidate media asset representation information, behavior sequence representation information, and trigger context representation information, and the candidate media asset representation information has a label value.

[0038] In a fourth aspect, an embodiment of the present application provides a media asset recommendation device, including:

[0039] The acquisition module is configured to acquire user feature representation information of a user, trigger media asset representation information, behavior sequence representation information, trigger context representation information, and candidate media asset representation information.

[0040] The prediction module is configured to input the user feature representation information, the trigger media asset representation information, the behavior sequence representation information, the trigger context representation information, and the candidate media asset representation information into the media asset recommendation model trained by the training device in the third aspect of the present application, to obtain a predicted click probability value of each candidate media asset.

[0041] In a fifth aspect, an embodiment of the present application provides an electronic device, including a processor and a memory for storing a computer program capable of running on the processor, wherein the processor is configured to run the computer program to perform the steps of the method in the first aspect or the second aspect of the present application.

[0042] In a sixth aspect, an embodiment of the present application provides a computer storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the method in the first aspect or the second aspect of the present application.

[0043] The sample data in the sample data set includes user feature representation information of a user, trigger media asset representation information, candidate media asset representation information, behavior sequence representation information, and trigger context representation information, and the candidate media asset representation information has a label value. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 a flowchart of a training method of a media asset recommendation model in an embodiment of the present application;

[0045] Figure 2 a flowchart of a media asset recommendation method in an embodiment of the present application;

[0046] Figure 3 a schematic diagram of a principle of a media asset recommendation model in an application example of the present application;

[0047] Figure 4 a structural schematic diagram of a training device of a media asset recommendation model in an embodiment of the present application;

[0048] Figure 5 a structural schematic diagram of a media asset recommendation device in an embodiment of the present application;

[0049] Figure 6 a structural schematic diagram of an electronic device in an embodiment of the present application. DETAILED DESCRIPTION

[0050] The present application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0051] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application.

[0052] In the related art, in a large-screen media asset detail page recommendation scenario, multiple family users often share a family large screen, and a key problem is the confusion of user identification. Since the smart television is deployed in a family terminal, the interaction behavior of the television media asset is often generated by multiple family members, and the recommendation system cannot accurately identify which family member is currently using the smart television. Another key problem is how to accurately perceive the current real-time interest of the user, and then personalized recommend media assets that the user may be interested in.

[0053] In the related art, user identification is a great challenge. First, if the user enters the large screen detail page, the current media trigger play behavior is most likely to be generated by the same user, because it is common for the same user to operate in the same period. For example: a TV series media is currently playing, and the user who triggers the entry of the detail page recommendation scene and the user who triggers the click to play the media are most likely the same person. Secondly, due to the existence of shared accounts, a simple and mixed representation of user features is not appropriate. Especially in the large screen media detail page scene, the user has a strong personal preference at this moment, and only appropriate user representation can mine the user's current personal interest. For example: a family includes an old man and a child, and when the child triggers the click of the media, the subsequent detail page recommendation should focus on the interest mining of the child. Therefore, user identification is a particularly important thing in large screen media detail page recommendation. Finally, the current user identification technology ignores the trigger environment information, but the trigger environment information is important in user identification. Trigger environment information can help the model more accurately identify which type of user is currently operating the large screen TV, and then guide the model to recommend the interests of which user in the family.

[0054] In addition, in the related art, the user behavior sequence is often expressed in sequence to obtain the current real-time interest of the user, but the negative influence of the negative interest in the user behavior sequence opposite to the current real-time interest on the recommendation performance is ignored. Generally, a user can show multiple interests from historical behaviors, especially in the TV large screen under the family market, multiple members often share an account, and their historical behaviors are mixed together. For example: animation, variety, movie, animation, etc. In the unrelated scene, the sequence modeling method is often used to capture the dynamic interest of the user from the historical behavior of the user. For example: Long short-term memory (LSTM) model, Deep Interest Network (DIN). But in the detail page recommendation scene, the above recommendation method is not so effective. For example, the category of the trigger item clicked is variety, which means that the user is interested in the variety media at that moment. Therefore, when predicting the future real-time interest of the user using the user historical behavior sequence, not only the positive interest related to the current real-time interest in the historical behavior needs to be considered, but also how to weaken the negative interest unrelated to the current real-time interest of the user needs to be considered.

[0055] Therefore, in various embodiments of the present application, a media recommendation model training and recommendation method based on user identification and real-time interest perception is provided, which effectively improves the accuracy of media recommendation in large screen media recommendation and other media recommendation scenes.

[0056] The embodiment of the present application provides a training method of a media asset recommendation model, as shown in the following formula (1), the training method comprises the following steps: Figure 1

[0057] Step 101, training the media asset recommendation model based on a sample data set to obtain a trained media asset recommendation model; wherein each sample data in the sample data set comprises user feature representation information of a user, trigger media asset representation information, candidate media asset representation information, behavior sequence representation information and trigger context representation information, the candidate media asset representation information has a label value; the media asset recommendation model comprises a confused user identification model used for confused user identification, a real-time interest perception model used for identifying a real-time interest representation of a current user and a prediction model based on a first identification result of the confused user identification model and a second identification result of the real-time interest perception model, for predicting a click probability value of a candidate media asset.

[0058] It can be understood that in the embodiment of the present application, each sample data in the constructed sample data set comprises user feature representation information of a user, trigger media asset representation information, candidate media asset representation information, behavior sequence representation information and trigger context representation information, which can comprehensively reflect the feature information related to the user, and the media asset recommendation model comprises a confused user identification model used for confused user identification, a real-time interest perception model used for identifying a real-time interest representation of a current user and a prediction model based on a first identification result of the confused user identification model and a second identification result of the real-time interest perception model, for predicting a click probability value of a candidate media asset, so that the accuracy of media asset recommendation in a media asset recommendation scene such as a large-screen media asset can be effectively improved based on confused user identification and real-time interest perception.

[0059] Here, the user feature representation information can be understood as information constructed based on a plurality of attribute features (also referred to as user information) of a user, for example, a plurality of user information is subjected to feature representation by using a feedforward neural network based on MLP (Multi-Layer Perceptron, multi-layer perceptron) to obtain the user feature representation information.

[0060] In an application example, it is assumed that the user information X u includes user total viewing time, user frequency viewing time, user channel preference and the like, the user information is spliced together and sent into the feedforward neural network based on MLP to output the user feature representation information E u , and the specific formula is as shown in the following formula (2):

[0061] X u = [x u,1 , x u,2 ,..., x u,n ] (1) ​

[0062] E u =f u (W u X u +b u ) (2)

[0063] where n denotes the number of user information, W u and b u denote trainable parameters of the MLP-based feedforward neural network, f u denotes the activation function relu, E u is the user feature representation information.

[0064] Here, the trigger media representation information is used for feature representation of the trigger media, which can be understood as the currently clicked media. The candidate media representation information is used for feature representation of the candidate media, i.e., the media to be recommended.

[0065] Exemplarily, both the trigger media representation information and the candidate media representation information can utilize the media representation network based on embedding representation to perform feature representation on media features.

[0066] In an application example, it is assumed that the attribute information of the media includes: media id, media duration, media language, media region, media recent 1 day / 7 day / 15 day / 30 day click times, media recent 1 day / 7 day / 15 day / 30 day exposure times, and the like. First, the media id is mapped, and the complex media id is mapped into a digital index id, and then the index id is sent into the embedding representation network to convert the index id into an embedding vector. Finally, the other attribute information is sent into the feedforward neural network, and the obtained result is spliced with the embedding vector to obtain the media representation information, and the specific formula is as follows:

[0067] index i =look_up(id i ) (3)

[0068] emd i =embedding(index i ) (4)

[0069] X i =[x i,1 ,x i,2 ,...,x i,m ] (5)

[0070] E i =concat[emd i ,f i (W i X i +bi )] (6)

[0071] where id i represents the media id, look_up() represents a mapping function, index i represents an index id after mapping by the mapping function, embedding() represents an embedding function, emd i represents an embedding vector after the embedding function, X i represents all attribute information of the media id, m represents the number of attribute information of the media, W i and b i represent trainable parameters of the feedforward neural network, f i represents an activation function relu, concat[] represents a concatenation function, E i is the media representation information.

[0072] It can be understood that the trigger media and the candidate media can both be converted and processed by the above process to obtain corresponding media representation information, and the difference lies in that the attribute information of the trigger media and the candidate media is different.

[0073] Here, the behavior sequence representation information can use the embedding representation-based media representation network to perform feature representation on the media features in the order of the user's behavior sequence, and then perform concatenation to form the user's behavior sequence representation.

[0074] In an application example, all the click-to-play media of the user in the last 30 days are concatenated in time order to form the user's click-to-play behavior sequence. For each media, the attribute information of the media selected by the foregoing media representation information is exactly the same, that is, it includes: media id, media duration, media language, media region, media click count in the last 1 day / 7 days / 15 days / 30 days, media exposure count in the last 1 day / 7 days / 15 days / 30 days, and the like.

[0075] X b = [item b,1 ; item b,2 ;... ; item b,t ] (7)

[0076] where t represents the length of the user behavior, and exemplarily, the maximum length is 50, and if the maximum length is exceeded, truncation processing is performed. item b,t represents the tthmedia in the order of the user's click-to-play media behavior, X b is the user's click-to-play behavior sequence. For each media in the sequence, it is expressed according to the foregoing formulas (3), (4), (5) and (6), and finally concatenated together, and the specific formula is as follows:

[0077] E b =concat[E b,1 ;E b,2 ;...;E b,t ] (8)

[0078] wherein E b is the behavior sequence representation information, concat[] represents the concatenation function, and E b,t is the tth media representation information sorted according to the user's click-to-play media behavior.

[0079] Here, the trigger context representation information can utilize an MLP-based feedforward neural network to perform feature representation on the trigger context information.

[0080] In an application example, the trigger context information includes the time, province, played duration of the played media that triggers the media, and the time since the last viewing of the media of the same channel, and the like. These information are concatenated together and sent to the MLP-based feedforward neural network to output the trigger context representation information E c , and the specific formula is as follows:

[0081] X c =[x c,1 ,x c,2 ,...,x c,k ] (9)

[0082] E c =f c (W c X c +b c ) (10)

[0083] wherein k represents the number of trigger context information, W c and b c represent the trainable parameters of the MLP-based feedforward neural network, f c represents the activation function relu, and E c is the trigger context representation information.

[0084] Exemplarily, the training of the media recommendation model based on the sample data set to obtain the trained media recommendation model includes:

[0085] training the media recommendation model based on the sample data in the sample data set and obtaining the predicted click probability value of the candidate media;

[0086] calculating the loss value based on the loss function and the label value;

[0087] determining that the loss value converges or the number of iterations of training reaches a set number of times, to obtain the trained media recommendation model.

[0088] Exemplarily, the loss function can adopt a binary cross-entropy loss function, the media recommendation model is trained by using sample data with label values, the iterative optimization of model parameters is realized, the loss value is calculated based on the loss function between the predicted click probability value and the label value, the iteration termination condition is determined, and the trained media recommendation model is obtained.

[0089] Exemplarily, the training of the media recommendation model based on the sample data in the sample data set and the obtaining of the predicted click probability value of the candidate media include:

[0090] obtaining the first identification result based on the user feature representation information, the trigger media representation information, the trigger context representation information and the behavior sequence representation information of the sample data and the confusion user identification model;

[0091] obtaining the second identification result based on the behavior sequence representation information of the sample data and the real-time interest perception model;

[0092] obtaining the predicted click probability value of the candidate media based on the first identification result, the second identification result and the prediction model.

[0093] Here, the confusion user identification model can be understood as a model that detects the user currently operating based on trigger environment information by using an attention mechanism and performs feature representation.

[0094] Here, the real-time interest perception model can be understood as a model that perceives the real-time interest of the current user based on disentanglement.

[0095] Here, the prediction model can be understood as a model that predicts the click probability value of the candidate media by the user based on the first identification result of the confusion user identification model and the second identification result of the real-time interest perception model.

[0096] It can be understood that the media recommendation model of the embodiments of the present application can realize confusion user identification and real-time interest perception by integrating the confusion user identification model, the real-time interest perception model and the prediction model, and combining the user feature representation information, the trigger media representation information, the candidate media representation information, the behavior sequence representation information and the trigger context representation information of the user constructed by the embodiments of the present application, and can effectively improve the accuracy of media recommendation in the media recommendation scene such as large-screen media recommendation.

[0097] Exemplarily, the first recognition result is obtained based on the user feature representation information, the trigger media representation information, the trigger context representation information and the behavior sequence representation information of the sample data, and the confusion user identification model, and the obtaining comprises:

[0098] trigger environment representation information is generated based on the trigger media representation information and the trigger context representation information of the sample data;

[0099] attention mechanism is used to process the trigger environment representation information and the behavior sequence representation information to obtain guided user feature representation information;

[0100] a feedforward neural network is used to identify the user feature representation information and the guided user feature representation information to obtain the first recognition result.

[0101] It can be understood that the confusion user identification model of the embodiments of the present application can realize the functions of generating trigger environment representation information, obtaining guided user feature representation information based on attention mechanism, and identifying confusion user based on feedforward neural network. Since the trigger environment representation information is constructed, the accuracy of confusion user identification can be effectively improved based on the trigger environment representation information, and the feature representation of the current user can be accurately determined.

[0102] Exemplarily, the second recognition result is obtained based on the behavior sequence representation information of the sample data and the real-time interest perception model, and the obtaining comprises:

[0103] multi-head self-attention mechanism is used to process the behavior sequence representation information to obtain multi-head representation information;

[0104] the multi-head representation information and the trigger environment representation information are disentangled to obtain disentangled interest head representation information;

[0105] global weight calculation is performed on the interest head representation information to obtain the second recognition result.

[0106] It can be understood that the real-time interest perception model can realize the functions of processing the behavior sequence representation information based on the multi-head self-attention mechanism, obtaining multi-head representation information, disentangling the multi-head representation information and the trigger environment representation information to obtain disentangled interest head representation information, and performing global weight calculation on the interest head representation information. The real-time interest perception model obtains multi-head representation information based on the multi-head self-attention mechanism, and disentangles the multi-head representation information and the trigger environment representation information, so as to suppress negative interest and strengthen positive interest in the behavior sequence of the user, and further accurately realize real-time interest perception of the user.

[0107] Exemplarily, the obtaining of the candidate media prediction click probability value based on the first identification result, the second identification result and the prediction model comprises:

[0108] The first identification result, the second identification result, the user feature representation information, the trigger media representation information, the trigger context representation information, the behavior sequence representation information and the candidate media representation information are spliced to obtain a splicing result.

[0109] The splicing result is input into a multi-layer perception machine to obtain a candidate media prediction click probability value.

[0110] It can be understood that the prediction model has the splicing function and the prediction function of predicting the probability based on the splicing result. In this way, the prediction of the candidate media click probability value of the user can be realized, and based on the predicted click probability value, the recommendation accuracy of the candidate media can be effectively improved.

[0111] Exemplarily in the embodiment, the embodiment of the application further provides a media recommendation method, as shown in Figure 2 The method comprises:

[0112] In step 201, user feature representation information, trigger media representation information, behavior sequence representation information, trigger context representation information and candidate media representation information of a user are obtained.

[0113] In step 202, the user feature representation information, the trigger media representation information, the behavior sequence representation information, the trigger context representation information and the candidate media representation information are input into a media recommendation model trained by the training method of the media recommendation model to obtain a candidate media prediction click probability value.

[0114] Exemplarily, the user feature representation information, the trigger media representation information, the behavior sequence representation information, the trigger context representation information and the candidate media representation information can be constructed by referring to the foregoing method, and the embodiment of the application will not be described here.

[0115] It can be understood that the media asset recommendation method of the embodiments of the present application can realize confusion user identification and real-time interest perception based on the trained media asset recommendation model and the constructed user feature representation information, trigger media representation information, behavior sequence representation information, trigger context representation information and candidate media representation information, thereby effectively improving the accuracy of media asset recommendation in media asset recommendation scenarios such as large-screen media asset recommendation.

[0116] Exemplarily, the method further comprises:

[0117] sorting the plurality of click probability values;

[0118] selecting the first N candidate media assets in the descending order of the click probability values, and determining the first N candidate media assets as the currently recommended candidate media assets;

[0119] wherein N is a natural number greater than 1.

[0120] The present application will be further described in detail below in combination with application examples.

[0121] The present application example is applied in a large-screen media asset detail page recommendation scenario. It can be understood that the user feature representation information E u , trigger media representation information, candidate media representation information, behavior sequence representation information E b , and trigger context representation information E c of the user can be constructed by using the foregoing method, and details are not repeated here.

[0122] In the present application example, as shown in Figure 3 , the constructed media asset recommendation model comprises a confusion user identification model, a real-time interest perception model and a prediction model. The training process of the media asset recommendation model of the present application example will be exemplarily described in combination with the above-mentioned models.

[0123] In the large-screen media asset recommendation detail page recommendation scenario, the user can explicitly express which user in the family by triggering a click on the media asset. In order to accurately identify which user is operating the large screen, a confusion user identification model based on trigger environment representation information is used in the present application example to identify the confusion user. The confusion user identification model needs to identify the confusion user based on the user feature representation information, the behavior sequence representation information, the trigger media representation information and the trigger context representation information, and obtain a first identification result.

[0124] Exemplarily, the trigger environment representation information is generated based on the trigger media representation information and the trigger context representation information, and the specific formula is as follows:

[0125] E e = f1(W1E c +W2Ei,seed +b1) (11)

[0126] wherein E e is the trigger environment representation information, W1, W2 and b1 represent trainable parameters of the feedforward neural network, f1 represents an activation function relu, E i,seed is the trigger media representation information, E c is the trigger context representation information.

[0127] In the traditional recommendation algorithm, the user behavior used is implicitly assumed to come from the same user, however this is not the case in the large screen recommendation scenario, because all people in the family are sharing the same account. In the application example, the trigger environment representation information and the behavior sequence representation information can be processed based on the attention mechanism to obtain the guided user feature representation information, and the specific formula is as follows:

[0128]

[0129] wherein V u represents the user feature representation information guided by the trigger environment representation information (i.e. the guided user feature representation information), att() represents an attention mechanism function, and a() is a feedforward network whose output is the activation weight w i .

[0130] Then, the user feature representation information and the guided user feature representation information can be processed based on the feedforward neural network to obtain the first identification result, and the specific formula is as follows:

[0131] E h = f h (W h,1 V u +W h,2 E u +b h ) (13)

[0132] wherein E h represents the confused user identification representation information (i.e. the first identification result), W h,1 , W h,2 and b h represent trainable parameters of the feedforward neural network, and f h represents an activation function relu.

[0133] The real-time interest perception model can use multi-head self-attention (MHSA) to capture the dependency relationship between media representation pairs and establish the sequential relationship between behavior sequences. The specific formula is as follows:

[0134] MHSA(F l )=[head1,head2,...,head h ]W o (14)

[0135] Q=F l W i Q (15)

[0136] K=F l W i K (16)

[0137] V=F l W i V (17)

[0138]

[0139] where F l represents the input of the l-th layer, W o represents the learnable parameters of each head, represents a normalized scale factor, and represents a mapping matrix.

[0140] Exemplarily, the behavior sequence representation information E b of the user can be input into the above formulas (14), (15), (16) and (17), and the multi-head attention mechanism is used to capture different interests of the user behavior and represent them in different heads to obtain HE b,i (i.e., multi-head representation information), where i represents the i-th head.

[0141] Exemplarily, in order to accurately perceive the real-time interest of the current user, the multi-head representation information and the trigger environment representation information are disentangled, the negative interest in the behavior sequence is suppressed, and the positive interest in the behavior sequence is strengthened, and finally the user real-time interest representation is obtained. The specific formula is as follows:

[0142]

[0143] where V b,i represents the i-th interest head after disentanglement, att() represents an attention mechanism function, a() is a feedforward network, and the output is an activation weight hw i,j .

[0144] Exemplarily, the interest head representation information V bThe global weight calculation is performed to obtain the second recognition result. For example, the disentangled interest head representation information V b is input into a global weight calculation unit to obtain the real-time interest representation E rt of the current user, and the specific formula is as follows:

[0145]

[0146] where W h and b h represent trainable parameters, H represents the number of interest heads of the multi-head attention mechanism, and E rt is the real-time interest representation information of the current user.

[0147] By way of example, the processing of the prediction model includes the following processes:

[0148] The first recognition result, the second recognition result, the user feature representation information, the trigger media representation information, the trigger context representation information, the behavior sequence representation information, and the candidate media representation information are input into a concatenate layer for splicing to obtain a splicing result.

[0149] The splicing result is input into a multi-layer perceptron, and the multi-layer perceptron is used to map multi-dimensional data into a single output. The output layer of the multi-layer perceptron has only one neuron, which represents the click probability value of the user on the candidate media, that is, the higher the result obtained, the more interested the user is in the candidate media, and the model should recommend the candidate media to the user more. The activation function of the output layer is a sigmoid function.

[0150] During the training process of the model, the output result can be compared with the real label in the data set, and the binary_cross_entropy is calculated as the loss function of the training process, and the model is trained with the goal of minimizing the loss function. The binary_cross_entropy loss function is represented as follows:

[0151]

[0152] wherein, is the real result of the user of the i-th data in the data set; y (i) is the click probability value predicted for the i-th data; log() represents a logarithmic function with base 2; N represents the data amount of the training data, and the value range of N is 500000-1000000.

[0153] To implement the method of the embodiments of the present application, the embodiments of the present application further provide a media asset recommendation model training device, which corresponds to the media asset recommendation model training method described above, and each step in the embodiments of the media asset recommendation model training method is fully applicable to the embodiments of the media asset recommendation model training device.

[0154] As shown in Figure 4 the media asset recommendation model training device includes a training module 401 configured to train a media asset recommendation model based on a sample data set to obtain a trained media asset recommendation model; wherein each sample data in the sample data set includes user feature representation information of a user, trigger media asset representation information, candidate media asset representation information, behavior sequence representation information, and trigger context representation information, the candidate media asset representation information has a label value; the media asset recommendation model includes a confused user identification model for confused user identification, a real-time interest perception model for identifying a real-time interest representation of a current user, and a prediction model for predicting a click probability value of a candidate media asset based on a first identification result of the confused user identification model and a second identification result of the real-time interest perception model.

[0155] For example, the training module 401 is specifically configured to:

[0156] train the media asset recommendation model based on sample data in the sample data set and obtain a click probability value of a candidate media asset prediction;

[0157] calculate a loss value based on the loss function and the click probability value and the label value;

[0158] determine that the loss value converges or the number of iterations of training reaches a set number of times to obtain the trained media asset recommendation model.

[0159] For example, the training module 401 trains the media asset recommendation model based on sample data in the sample data set and obtains a click probability value of a candidate media asset prediction, including:

[0160] obtain the first identification result based on the user feature representation information, the trigger media asset representation information, the trigger context representation information, the behavior sequence representation information of the sample data, and the confused user identification model;

[0161] obtain the second identification result based on the behavior sequence representation information of the sample data and the real-time interest perception model;

[0162] obtain a click probability value of a candidate media asset prediction based on the first identification result, the second identification result, and the prediction model.

[0163] Illustratively, the first identification result is obtained based on the user feature representation information, the trigger media representation information, the trigger context representation information and the behavior sequence representation information of the sample data and the confusion user identification model, and the method comprises the following steps:

[0164] Trigger environment representation information is generated based on the trigger media representation information and the trigger context representation information of the sample data;

[0165] The trigger environment representation information and the behavior sequence representation information are processed based on an attention mechanism to obtain guided user feature representation information;

[0166] The user feature representation information and the guided user feature representation information are subjected to confusion user identification based on a feedforward neural network to obtain the first identification result.

[0167] Illustratively, the second identification result is obtained based on the behavior sequence representation information of the sample data and the real-time interest perception model, and the method comprises the following steps:

[0168] The behavior sequence representation information is processed based on a multi-head self-attention mechanism to obtain multi-head representation information;

[0169] The multi-head representation information and the trigger environment representation information are disentangled to obtain disentangled interest head representation information;

[0170] Global weight calculation is performed on the interest head representation information to obtain the second identification result.

[0171] Illustratively, the click probability value of the candidate media prediction is obtained based on the first identification result, the second identification result and the prediction model, and the method comprises the following steps:

[0172] The first identification result, the second identification result, the user feature representation information, the trigger media representation information, the trigger context representation information, the behavior sequence representation information and the candidate media representation information are spliced to obtain a splicing result;

[0173] The splicing result is input into a multi-layer perceptron to obtain the click probability value of the candidate media prediction.

[0174] In actual application, the training module 401 can be realized by a processor in a media recommendation model training device. Of course, the processor needs to run a computer program in the memory to realize its functions.

[0175] To implement the method of the embodiments of the present application, the embodiments of the present application further provide a media asset recommendation device corresponding to the media asset recommendation method, and each step in the media asset recommendation method embodiment is fully applicable to the media asset recommendation device embodiment.

[0176] As shown in Figure 5 , the media asset recommendation device comprises an acquisition module 501 and a prediction module 502. The acquisition module 501 is configured to acquire user feature representation information of a user, trigger media asset representation information, behavior sequence representation information, trigger context representation information and candidate media asset representation information. The prediction module 502 is configured to input the user feature representation information, the trigger media asset representation information, the behavior sequence representation information, the trigger context representation information and the candidate media asset representation information into the media asset recommendation model trained by the training device in the third aspect of the embodiments of the present application, to obtain a click probability value predicted for each candidate media asset.

[0177] Exemplarily, the media asset recommendation device further comprises a determination module 503 configured to sort a plurality of the click probability values, and select the first N candidate media assets in the descending order of the click probability values as the candidate media assets to be currently recommended, wherein N is a natural number greater than 1.

[0178] In actual application, the acquisition module 501, the prediction module 502 and the determination module 503 can be implemented by a processor in the media asset recommendation device. Of course, the processor needs to run a computer program in the memory to realize its functions.

[0179] Based on the hardware implementation of the above program modules, and to implement the method of the embodiments of the present application, the embodiments of the present application further provide an electronic device. Figure 6 Only exemplary structures of the electronic device are shown, not all structures, and part or all of the structures can be implemented according to needs. Figure 6

[0180] As shown in Figure 6 , the electronic device 600 provided by the embodiments of the present application comprises at least one processor 601, a memory 602, a user interface 603 and at least one network interface 604. Each component in the electronic device 600 is coupled together through a bus system 605. It can be understood that the bus system 605 is used to realize the connection and communication between the components. The bus system 605 includes not only a data bus, but also a power bus, a control bus and a status signal bus. However, for the purpose of clear illustration, all kinds of buses are marked as the bus system 605 in the Figure 6 .

[0181] ​The user interface 603 can include a display, a keyboard, a mouse, a trackball, a click wheel, a keypad, a button, a touchpad, or a touchscreen, etc.

[0182] The memory 602 in the embodiments of the present application is used to store various types of data to support the operation of the electronic device. Examples of the data include any computer programs used to operate on the electronic device.

[0183] The method disclosed in the embodiments of the present application can be applied in the processor 601 or implemented by the processor 601. The processor 601 can be an integrated circuit chip having a signal processing capability. In the implementation process, each step of the method disclosed in the embodiments of the present application can be completed by the integrated logic circuit or the instruction in the software form of the hardware in the processor 601. The processor 601 described above can be a general-purpose processor, a digital signal processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc. The processor 601 can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor can be a microprocessor or any conventional processor, etc. In combination with the steps of the method disclosed in the embodiments of the present application, the hardware decoding processor can be directly embodied to execute the steps, or the hardware and software modules in the decoding processor can be combined to execute the steps. The software module can be located in the storage medium, which is located in the memory 602. The processor 601 reads the information in the memory 602 and combines the hardware to complete the steps of the method provided in the embodiments of the present application.

[0184] In the exemplary embodiments, the electronic device can be implemented by one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field programmable gate arrays (FPGAs), general-purpose processors, controllers, micro controllers (MCUs), microprocessors (Microprocessors), or other electronic elements, for executing the foregoing method.

[0185] It can be appreciated that the memory 602 can be a volatile memory or a nonvolatile memory, and can also include both volatile and nonvolatile memory. Among them, the nonvolatile memory can be a Read Only Memory (ROM), a Programmable Read-Only Memory (PROM), an Erasable Programmable Read-Only Memory (EPROM), an Electrically Erasable Programmable Read-Only Memory (EEPROM), a ferromagnetic random access memory (FRAM), a Flash Memory, a magnetic surface memory, an optical disc, or a Compact Disc Read-Only Memory (CD-ROM); the magnetic surface memory can be a disk memory or a tape memory. The volatile memory can be a Random Access Memory (RAM) used as an external cache. By way of example but not limitation, many forms of RAM can be used, such as Static Random Access Memory (SRAM), Synchronous Static Random Access Memory (SSRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDR SDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), Sync Link Dynamic Random Access Memory (SLDRAM), and Direct Rambus Random Access Memory (DRRAM).The memory described in the embodiments of the present application is intended to include, but not limited to, these and any other suitable types of memory.

[0186] In the example embodiments, the embodiments of the present application also provide a computer storage medium, which can be specifically a computer readable storage medium, for example, a memory 602 storing a computer program, which can be executed by the processor 601 of the electronic device 600 to complete the steps described in the method of the embodiments of the present application. The computer readable storage medium can be a ROM, a PROM, an EPROM, an EEPROM, a Flash Memory, a magnetic surface memory, an optical disc, or a CD-ROM memory, etc.

[0187] It should be noted that "first", "second", etc. are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0188] In addition, the technical solutions described in the embodiments of the present application can be combined arbitrarily without conflict.

[0189] The above is only a specific implementation 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 range disclosed in the present application, which should be covered within 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 training method for a media asset recommendation model, characterized in that, include: Based on the sample dataset, a media asset recommendation model is trained to obtain a well-trained media asset recommendation model. The sample dataset includes user feature representation information, trigger media asset representation information, candidate media asset representation information, behavior sequence representation information, and trigger context representation information, wherein the candidate media asset representation information has a label value; the media asset recommendation model includes: a confused user identification model for confusing user identification, a real-time interest perception model for identifying the real-time interest representation of the current user, and a prediction model for predicting the click probability value of candidate media assets based on the first identification result of the confused user identification model and the second identification result of the real-time interest perception model; The step of training a media asset recommendation model based on a sample dataset to obtain a trained media asset recommendation model includes: Based on the triggering media representation information and the triggering context representation information of the sample data, triggering environment representation information is generated; The triggering environment representation information and the behavior sequence representation information are processed based on the attention mechanism to obtain the user feature representation information after guidance; Based on the feedforward neural network, the user feature representation information and the guided user feature label representation information are used to perform confused user identification to obtain the first identification result; Based on the behavioral sequence representation information of the sample data and the real-time interest perception model, the second recognition result is obtained; Based on the first identification result, the second identification result, and the prediction model, the click probability value of the candidate media asset is obtained; The loss value is calculated based on the click probability value and the tag value using a loss function. Once the loss value converges or the number of training iterations reaches a set number, the trained media asset recommendation model is obtained.

2. The method according to claim 1, characterized in that, Based on the behavioral sequence representation information of the sample data and the real-time interest perception model, the second recognition result is obtained, including: The behavior sequence representation information is processed based on a multi-head self-attention mechanism to obtain multi-head representation information; The multi-head representation information and the triggering environment representation information are de-entangled to obtain the de-entangled head of interest representation information; The second recognition result is obtained by performing global weight calculation on the interest header representation information.

3. The method according to claim 1, characterized in that, The step of obtaining the click probability value of the candidate media asset prediction based on the first identification result, the second identification result, and the prediction model includes: The first identification result, the second identification result, the user feature representation information, the trigger media asset representation information, the trigger context representation information, the behavior sequence representation information, and the candidate media asset representation information are concatenated to obtain the concatenation result; The splicing result is input into a multilayer perceptron to obtain the click probability value of the candidate media asset prediction.

4. A media asset recommendation method, characterized in that, include: Obtain user feature representation information, trigger media asset representation information, behavior sequence representation information, trigger context representation information, and candidate media asset representation information; The user feature representation information, the trigger media asset representation information, the behavior sequence representation information, the trigger context representation information, and the candidate media asset representation information are input into the media asset recommendation model trained by the method described in any one of claims 1 to 3 to obtain the click probability value predicted for each candidate media asset.

5. The method according to claim 4, characterized in that, The method further includes: Sort the multiple click probability values; Select the top N candidate media assets with the click probability values ​​sorted from largest to smallest, and determine the top N candidate media assets as the currently recommended candidate media assets; Where N is a natural number greater than 1.

6. A training device for a media asset recommendation model, characterized in that, include: The training module is used to train the media asset recommendation model based on the sample dataset, and obtain the trained media asset recommendation model. The sample dataset includes user feature representation information, trigger media asset representation information, candidate media asset representation information, behavior sequence representation information, and trigger context representation information, wherein the candidate media asset representation information has a label value; the media asset recommendation model includes: a confused user identification model for confusing user identification, a real-time interest perception model for identifying the real-time interest representation of the current user, and a prediction model for predicting the click probability value of candidate media assets based on the first identification result of the confused user identification model and the second identification result of the real-time interest perception model; The training module is specifically used for: generating trigger environment representation information based on the trigger media asset representation information and the trigger context representation information of the sample data; processing the trigger environment representation information and the behavior sequence representation information based on an attention mechanism to obtain guided user feature representation information; performing confused user identification based on the user feature representation information and the guided user feature label representation information using a feedforward neural network to obtain the first identification result; obtaining the second identification result based on the behavior sequence representation information of the sample data and the real-time interest perception model; obtaining the click probability value of the candidate media asset prediction based on the first identification result, the second identification result, and the prediction model; calculating the loss value based on the click probability value and the label value using a loss function; determining that the loss value converges or the number of training iterations reaches a set number to obtain the trained media asset recommendation model.

7. A media asset recommendation device, characterized in that, include: The acquisition module is used to acquire user feature representation information, trigger media asset representation information, behavior sequence representation information, trigger context representation information, and candidate media asset representation information; The prediction module is used to input the user feature representation information, the trigger media asset representation information, the behavior sequence representation information, the trigger context representation information, and the candidate media asset representation information into the media asset recommendation model trained by the training device as described in claim 6, and obtain the predicted click probability value of each candidate media asset.

8. An electronic device, characterized in that, include: The processor and memory for storing computer programs that can run on the processor, wherein, The processor, when running a computer program, performs the steps of the method according to any one of claims 1 to 3 or claims 4 to 5.

9. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images