Data processing method and device, equipment and medium

Abstract

The embodiment of the invention provides a data processing method and device, equipment and a medium. The method comprises the steps of obtaining a sample negative feedback feature of sample negative feedback data in multimedia sample data, and obtaining a sample trigger feature of sample trigger data in the multimedia sample data; performing feature transformation on the negative feedback feature of the sample through a to-be-trained separation model to obtain a sample mapping feature, and splitting the sample mapping feature into a positive content prediction feature and a negative content prediction feature of a sample object; determining a first loss value between the positive content prediction feature and the sample trigger feature, and determining a second loss value between the negative content prediction feature and the sample trigger feature; determining the sum of the opposite number of the second loss value and the first loss value as a model loss value of the to-be-trained separation model; and correcting network parameters of the to-be-trained separation model according to the model loss value to obtain a content separation model. By implementing the embodiment of the invention, the recommendation accuracy of the multimedia data can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to a data processing method, device, equipment and medium. Background Art

[0002] With the development of data informatization, the amount of data is growing rapidly, and big data is showing a trend of diversification and decentralization. In the environment of large-scale data, the multimedia data recommended by various business platforms for objects may include multimedia data that the object is not interested in; the object can mark the multimedia data that he is not interested in (such as marking it as dislike) to reduce the recommendation of such multimedia data.

[0003] In current recommendation scenarios, multimedia data marked as disliked and multimedia data with the same data label as multimedia data marked as disliked can be blocked and not recommended. However, only a portion of the blocked multimedia data is not of interest to the subject, and there is still a small portion of content that the subject is interested in; for example, the multimedia data marked as disliked contains content A and content B, the subject is not interested in content A, but content B may be of interest to the subject. If all multimedia data marked as disliked by the subject are blocked, the multimedia data that the subject is actually interested in may be blocked, resulting in low accuracy in multimedia data recommendation. Summary of the invention

[0004] The embodiments of the present application provide a data processing method, apparatus, device, and medium, which can improve the accuracy of recommendation of multimedia data.

[0005] On the one hand, an embodiment of the present application provides a data processing method, including:

[0006] Acquire sample negative feedback features of sample negative feedback data in the multimedia sample data, and acquire sample trigger features of sample trigger data in the multimedia sample data;

[0007] Performing feature transformation on the negative feedback features of the sample through the separation model to be trained to obtain sample mapping features, and splitting the sample mapping features into positive content prediction features and negative content prediction features of the sample object;

[0008] Determine a first loss value between a positive content prediction feature and a sample trigger feature, and determine a second loss value between a negative content prediction feature and a sample trigger feature;

[0009] The sum of the inverse of the second loss value and the first loss value is determined as the model loss value of the separation model to be trained;

[0010] The network parameters of the separation model to be trained are modified according to the model loss value to obtain a content separation model; the content separation model is used to separate the positive content and the negative content in the negative feedback multimedia data.

[0011] An embodiment of the present application provides a data processing device, including:

[0012] A feature acquisition module, used to acquire sample negative feedback features of sample negative feedback data in the multimedia sample data, and acquire sample trigger features of sample trigger data in the multimedia sample data;

[0013] A feature splitting module is used to perform feature transformation on the negative feedback features of the sample through the separation model to be trained to obtain sample mapping features, and split the sample mapping features into positive content prediction features and negative content prediction features of the sample object;

[0014] A first determination module, used to determine a first loss value between a positive content prediction feature and a sample trigger feature, and to determine a second loss value between a negative content prediction feature and a sample trigger feature;

[0015] A second determination module is used to determine the sum of the inverse of the second loss value and the first loss value as the model loss value of the separation model to be trained;

[0016] The parameter correction module is used to correct the network parameters of the separation model to be trained according to the model loss value to obtain a content separation model; the content separation model is used to separate positive content and negative content in negative feedback multimedia data.

[0017] Wherein, the sample negative feedback data includes sample text data;

[0018] The feature acquisition module acquires the sample negative feedback features of the sample negative feedback data in the multimedia sample data, and is used to perform the following steps:

[0019] Acquire multimedia sample data of the sample object; the multimedia sample data includes sample negative feedback data and sample trigger data;

[0020] Performing text segmentation processing on sample text data in the multimedia sample data to obtain a negative feedback word set corresponding to the sample negative feedback data;

[0021] Performing vector conversion on each negative feedback word in the negative feedback word set to obtain the negative feedback word vector corresponding to each negative feedback word;

[0022] Perform semantic analysis on the negative feedback word vector to obtain the sample negative feedback features of the sample negative feedback data.

[0023] Among them, the feature splitting module performs feature transformation on the sample negative feedback feature through the separation model to be trained to obtain the sample mapping feature, and splits the sample mapping feature into the positive content prediction feature and the negative content prediction feature of the sample object to perform the following steps:

[0024] Perform feature transformation on the negative feedback features of the sample through the separation model to be trained to obtain the sample mapping features;

[0025] Obtaining a feature dimension of the sample mapping feature, and determining a center position of the sample mapping feature according to the feature dimension;

[0026] According to the center position, the sample mapping feature is split to obtain a first sequence sub-feature and a second sequence sub-feature of the sample mapping feature, and the first sequence sub-feature and the second sequence sub-feature have the same dimension;

[0027] The first sequence sub-features are determined as positive content prediction features of the sample object, and the second sequence sub-features are determined as negative content prediction features of the sample object.

[0028] The positive content prediction feature includes M-dimensional positive content prediction sub-features, and the sample trigger feature includes M-dimensional sample trigger sub-features; M is a positive integer;

[0029] The first determination module determines a first loss value between the positive content prediction feature and the sample trigger feature, and is used to perform the following steps:

[0030] Obtaining an error value between an i-th dimension positive content prediction sub-feature in the M-dimensional positive content prediction sub-feature and an i-th dimension sample trigger sub-feature in the M-dimensional sample trigger sub-feature;

[0031] A first loss value between the positive content prediction feature and the sample trigger feature is determined according to M error values ​​between the positive content prediction feature and the sample trigger feature.

[0032] The parameter correction module corrects the network parameters of the separation model to be trained according to the model loss value to obtain a content separation model for performing the following steps:

[0033] According to the model loss value, the gradient of the separation model to be trained is determined, and the network parameters of the separation model to be trained are modified according to the gradient;

[0034] If the model loss reaches the minimum value, the separation model to be trained including the corrected network parameters will be determined as the content separation model.

[0035] The device further comprises:

[0036] The forward feature acquisition module is used to perform the following steps:

[0037] Acquire negative feedback multimedia data and trigger multimedia data of the object;

[0038] Extracting features from negative feedback multimedia data to obtain negative feedback features of the negative feedback multimedia data, and extracting features from triggering multimedia data to obtain triggering features of the triggering multimedia data;

[0039] The negative feedback feature is transformed through the content separation model to obtain the negative feedback mapping feature, and the negative feedback mapping feature is split through the content separation model to obtain the negative feedback content feature and the positive feedback content feature of the object;

[0040] The attention weight of the negative feedback content feature is obtained, and the product of the negative feedback content feature and the attention weight is subtracted from the trigger feature to obtain the positive trigger content feature in the trigger feature.

[0041] Among them, the positive feature acquisition module obtains the attention weight of the negative feedback content feature to perform the following steps:

[0042] Calculate the similarity between the negative feedback content features and the trigger features to obtain an attention matrix, and normalize the attention matrix to obtain a normalized attention matrix;

[0043] The difference between the trigger feature and the negative feedback content feature is multiplied by the normalized attention matrix to obtain the attention weight.

[0044] The device further comprises:

[0045] A data acquisition module, used to acquire a candidate multimedia data set matching the positive trigger content feature from the service platform;

[0046] An indicator acquisition module, used to obtain an estimated feedback indicator value corresponding to each candidate multimedia data in the candidate multimedia data set;

[0047] A data sorting module is used to sort each candidate multimedia data in the candidate multimedia data set according to the feedback indicator estimation value to obtain a candidate multimedia data sequence of the object;

[0048] The data determination module is used to determine the multimedia recommendation data of the object according to the candidate multimedia data sequence.

[0049] On one hand, an embodiment of the present application provides a computer device, including a memory and a processor, wherein the memory is connected to the processor, the memory is used to store a computer program, and the processor is used to call the computer program so that the computer device executes the method provided in the above aspect of the embodiment of the present application.

[0050] On one hand, an embodiment of the present application provides a computer-readable storage medium, in which a computer program is stored. The computer program is suitable for being loaded and executed by a processor, so that a computer device with a processor executes the method provided in the above aspect of the embodiment of the present application.

[0051] According to one aspect of the present application, a computer program product is provided, which may include a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program, so that the computer device executes the method provided in the above aspect.

[0052] In the embodiment of the present application, sample negative feedback data and sample trigger data in multimedia sample data can be obtained, and feature extraction is performed on the sample negative feedback data and the sample trigger data respectively to obtain sample negative feedback features corresponding to the sample negative feedback data and sample trigger features corresponding to the sample trigger data. The sample negative feedback features are feature transformed by the separation model to be trained to obtain sample mapping features, and the sample mapping features are split into positive content prediction features and negative content prediction features of the sample object, and a first loss value between the positive content prediction features and the sample trigger features is determined, and a second loss value between the negative content prediction features and the trigger features is determined. The first loss value is subtracted from the second loss value to obtain a model loss value of the separation model to be trained, and the network parameters of the separation model to be trained are corrected according to the model loss value to obtain a content separation model. The separation model to be trained is trained by the sample negative feedback features and the sample trigger features to obtain a content separation model, and the content separation model can separate the positive content and the negative content in the sample negative feedback data, so that the data related to the negative content can be shielded, and the data related to the positive content can be recommended to the user, which can improve the recommendation accuracy of the multimedia data. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0054] Figure 1 It is a structural diagram of a network architecture provided by an embodiment of the present application;

[0055] Figure 2 It is a schematic diagram of a data processing scenario provided by an embodiment of the present application;

[0056] Figure 3This is a flow diagram of a data processing method provided in an embodiment of the present application. Figure 1 ;

[0057] Figure 4 It is a schematic diagram of a training process of a separation model to be trained provided in an embodiment of the present application;

[0058] Figure 5 This is a flow diagram of a data processing method provided in an embodiment of the present application. Figure 2 ;

[0059] Figure 6 This is a structural diagram of obtaining positive trigger content features provided by an embodiment of the present application;

[0060] Figure 7 is a structural schematic diagram of a data processing device provided in an embodiment of the present application;

[0061] Figure 8 It is a structural diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0063] It is understandable that in the specific implementation of the present application, it may involve interaction data between the business platform and the object (for example, clicks, favorites, comments, dislikes, etc. in the business platform), basic information of the object and other related information. When the embodiment of the present application is applied to a specific product or technology, it is necessary to obtain the permission or consent of the relevant agency or department, or the user himself, and the collection, use and processing of the relevant data need to comply with the relevant laws and standards of the relevant region.

[0064] For ease of understanding, the basic technical concepts involved in the embodiments of the present application are described below:

[0065] Multilayer Perceptron (MLP): is a feedforward artificial neural network model. The structure of MLP includes an input layer, at least one hidden layer, and an output layer. Each layer contains multiple neurons, and each neuron is connected to all neurons in the previous layer. MLP is trained through the back-propagation algorithm to minimize the error between the output layer and the target value.

[0066] Attention mechanism: refers to the process by which the nervous system processes specific stimuli by selectively focusing and concentrating attention when faced with complex perceptual information. In deep learning, the core idea of ​​the attention mechanism is to enable the model to selectively focus on different parts of the input sequence and assign different weights to each part of the input sequence to highlight information that is more critical to the task.

[0067] See also Figure 1 , Figure 1 1 is a schematic diagram of a network architecture provided by an embodiment of the present application; the network architecture may include a server 10d and a terminal cluster, the terminal cluster may include one or more terminal devices, and the number of terminal devices included in the terminal cluster is not limited. Figure 1 As shown, the terminal cluster may specifically include terminal device 10a, terminal device 10b, and terminal device 10c, etc.; all terminal devices in the terminal cluster (for example, including terminal device 10a, terminal device 10b, and terminal device 10c, etc.) may be connected to the server 10d through a network connection, so that each terminal device may exchange data with the server 10d through the network connection.

[0068] in, Figure 1 The terminal devices in the terminal cluster shown may include but are not limited to: smart phones, tablet computers, laptops, PDAs, desktop computers, wearable devices (such as smart watches, smart bracelets, etc.), smart voice interaction devices, smart home appliances (such as smart TVs, etc.), vehicle-mounted devices, aircraft and other electronic devices. This application does not limit the type of terminal devices.

[0069] Figure 1 The server 10d shown can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. This application does not limit the type of server.

[0070] In the embodiment of the present application, Figure 1 Each terminal device in the terminal cluster shown can run the service platform. When the service platform runs in each terminal device, it can be connected with Figure 1 The service platform running in each terminal device may be any service platform with a recommendation function, such as one or more of a news platform, a video platform, a game platform, a shopping platform, etc.

[0071] Figure 1 The server 10d shown can be a background server corresponding to the service platform running in each terminal device. If multiple service platforms are running in a terminal device (for example, terminal device 10a), then the multiple service platforms can correspond to different servers or the same server, and this application does not limit this. For example, if service platform A and service platform B are installed in terminal device 10a, then the server corresponding to service platform A and the server corresponding to service platform B can be the same server or different servers.

[0072] The recommendation function of the business platform can be realized by the recommendation system, such as recommending multimedia data of interest to the objects in the business platform through the recommendation system. The recommendation system is an information filtering system that analyzes the multimedia data in the business platform and the interactive data of the objects in the business platform (for example, the browsing history, purchase history, comment history, search history, like history, collection history, etc. of the objects), and uses various algorithms and neural network models to calculate the matching degree between the objects and the multimedia data, and screens out the multimedia data that may be of interest to the objects. The neural network models involved in the recommendation system may include but are not limited to: content separation model, recall network, click estimation network, rearrangement network, feature extraction network, etc. The content separation model provided in the embodiment of the present application can be used to strip positive feedback content features and negative feedback content features from the negative feedback multimedia data of the object. The positive feedback content features can be understood as the content of interest to the object in the negative feedback multimedia data, and the negative feedback content features can be understood as the content of no interest to the object in the negative feedback multimedia data. The positive feedback content features can be used as the basis for the recommendation system to recommend the content of interest to the object, and the negative feedback content features can be used as the basis for the recommendation system to shield the content that the object is not interested in.

[0073] The following description takes the news platform as an example. Figure 2 , Figure 2 Schematic diagram of a data processing scenario provided by an embodiment of the present application. Figure 2 As shown, the negative feedback news data and trigger news data of the object "Xiao Ming" are obtained. The negative feedback news data refers to the collection of news data marked as disliked (also understood as uninterested) by the object "Xiao Ming" in the news platform (for example, Figure 2 The triggered news data refers to the collection of news data on the news platform that the subject "Xiao Ming" clicks, browses, likes, etc. These news data can be considered as news data that the subject "Xiao Ming" likes or is interested in ( Figure 2 News data shown in 2).

[0074] Feature extraction is performed on the negative feedback news data and the trigger news data to obtain the negative feedback features corresponding to the negative feedback news data and the trigger features corresponding to the trigger news data. Feature extraction refers to the process of extracting key information from images, texts, or signals by a computer. For example, feature extraction is performed on the negative feedback news data (news data 1) to obtain the negative feedback features (e.g., feature x), and feature extraction is performed on the trigger news data (news data 2) to obtain the trigger features (e.g., feature y).

[0075] The negative feedback features are input into the content separation model, and the positive feedback content features and negative feedback content features in the negative feedback features are stripped by the content separation model. The positive feedback content features can refer to the content features that the object "Xiaoming" likes in the negative feedback news data. The negative feedback content features can refer to the content features that the object "Xiaoming" does not like in the negative feedback news data. For example, when the positive feedback content feature represents "Person A", "Person A" can be considered as the content that the object "Xiaoming" likes. When the negative feedback content feature represents "sports", "sports" can be considered as the content feature that the object "Xiaoming" does not like. The training process of the content separation model will be introduced in detail in the following content.

[0076] The attention weights of the negative feedback content features are calculated through the attention mechanism. The attention weights refer to the weights assigned to the negative content prediction features in the trigger features when processing the trigger features. The calculation of the attention weights of the negative feedback content features through the attention mechanism will be introduced in detail in the following content. The value range of the attention weights is [0, 1]. The trigger features are subtracted by the product of the negative feedback content features and the attention weights to obtain the positive trigger content features in the trigger features, as Figure 2 shown, to obtain the positive trigger content features in the trigger features. For example, the positive trigger content features can represent "Person A".

[0077] The news recommendation system will screen at least one news data related to "Person A" for the object "Xiaoming" from the news database according to the positive trigger content features (e.g., representing "Person A"). From the at least one related news data, the news data related to "Person A" and related to sports is blocked, and the news data related to "Person A" other than the news data related to "Person A" and related to sports is recommended to the object "Xiaoming". As Figure 2 shown, news data 3 "The sketch 'Happy XX' performed by Person A", news data 4 "The comedy 'You Are XXX' starring Person A", news data 5 "The movie and TV drama C starred by Person A", and news data 6 "The work D directed by Person A" can be recommended to the object "Xiaoming".

[0078] In the embodiment of the present application, negative feedback news data and trigger news data are obtained, feature extraction is performed on the negative feedback news data and trigger news data, and negative feedback features and trigger features corresponding to the negative feedback news data and trigger news data are obtained. The positive feedback content features and negative feedback content features in the negative feedback features are separated by a content separation model, and the attention weight of the negative feedback content features is calculated. The product of the negative feedback content features and the attention weight is subtracted from the trigger feature to remove the negative feedback content features in the trigger feature, so that only positive feedback content features are included in the trigger feature. The news recommendation system shields news data related to negative feedback content features, and only recommends news data with positive feedback content features to users, which can improve the recommendation accuracy of multimedia data.

[0079] See also Figure 3 , Figure 3 This is a flow diagram of a data processing method provided in an embodiment of the present application. Figure 1 It can be understood that the video data processing method is executed by a computer device, which may be a terminal device (e.g., Figure 1 Any terminal device in the terminal device set shown in FIG. 1 ), or a server (such as Figure 1 The server 10d is shown in the figure, and this application does not limit this. The video data processing method may include the following steps S101 to S105:

[0080] Step S101, obtaining sample negative feedback features of sample negative feedback data in multimedia sample data, and obtaining sample trigger features of sample trigger data in multimedia sample data;

[0081] Among them, a multimedia sample data may include sample negative feedback data and sample trigger data. At this time, the sample negative feedback data can be called negative sample data, and the sample trigger data can be called positive sample data. Multiple multimedia sample data can form a sample training set. The multimedia data may include but are not limited to: news, video, animation, audio, text and other data. This application does not limit the type of multimedia data. Sample negative feedback data refers to one or more multimedia data that the object does not like, and sample trigger data refers to one or more multimedia data that the object likes. For example, if the multimedia sample data is news data, the sample negative feedback data refers to one or more news data that the object does not like, and the sample trigger data refers to one or more news data that the object likes.

[0082] The sample negative feedback feature refers to the feature that can describe the basic characteristics and properties of the sample negative feedback data, and the sample trigger feature refers to the feature that can describe the basic characteristics and properties of the sample trigger data. Features (also called feature vectors) usually exist in the form of vectors, and the dimension of the feature can be understood as the dimension of the vector. For example, if feature X is a 4-dimensional vector, the feature dimension of feature X is 4.

[0083] For example, if the multimedia sample data is text data, the text information in the text data may include but is not limited to: keywords, themes, emotional words, character names, etc., and this application does not limit the text information. For another example, if the multimedia sample data is video data, the video information may include but is not limited to: video tags, video titles, video authors, etc., and this application does not limit the video information.

[0084] In practical applications, if the multimedia data is text data, the sample negative feedback features of the sample negative feedback data in the text sample data and the sample trigger features of the sample trigger data in the text sample data can be obtained by using a text feature extraction method based on statistics and a text feature extraction method based on a neural network model. The text feature extraction method based on statistics may include but is not limited to: term frequency (TF), inverse document frequency (IDF) and other methods, and the text feature extraction method based on a neural network model may include but is not limited to: bag-of-words model (BoW), N-gram model (N-gram model), word embedding model, etc. This application does not limit the text feature extraction method based on statistics and the text feature extraction method based on a neural network model.

[0085] In practical applications, if the multimedia data is video data, the sample negative feedback features of the sample negative feedback data in the video sample data and the sample trigger features of the sample trigger data in the text sample data can be obtained by using an image-based video feature extraction method and a deep learning-based video feature extraction method. The image-based video feature extraction method may include but is not limited to: image shape feature extraction, image texture feature extraction, image color feature extraction, etc., and the deep learning-based video feature extraction method may include but is not limited to: convolutional neural network (CNN), recurrent neural network (RNN), etc. This application does not limit the image-based video feature extraction method and the deep learning-based video feature extraction method.

[0086] This application takes multimedia sample data as text sample data as an example, and introduces how to obtain sample negative feedback features of sample negative feedback data in multimedia sample data and sample trigger features of sample trigger data in multimedia sample data through a word embedding model.

[0087] The specific process of obtaining sample negative feedback features of sample negative feedback data in multimedia sample data and sample trigger features of sample trigger data in multimedia sample data includes: obtaining multimedia sample data of sample objects; the multimedia sample data includes sample negative feedback data and sample trigger data; performing text segmentation processing on sample text data in the multimedia sample data to obtain a negative feedback word set corresponding to the sample negative feedback data; performing vector conversion on each negative feedback word in the negative feedback word set to obtain a negative feedback word vector corresponding to each negative feedback word; performing semantic analysis on the negative feedback word vector to obtain the sample negative feedback features of the sample negative feedback data.

[0088] The sample object refers to the user of the service platform, which may include but is not limited to: multimedia platforms (e.g., short video platforms, live broadcast platforms, video platforms), entertainment applications (e.g., game platforms), information platforms (e.g., news platforms), etc. This application does not limit the service platform. Sample negative feedback data refers to multimedia data that the sample object does not like, and sample trigger data refers to multimedia data that the sample object likes.

[0089] Text segmentation refers to splitting the words in a text to obtain a word set. For example, if a text A is "Xiao Ming likes to eat apples", the text segmentation of text A will obtain a word set of {"Xiao Ming", "like", "eat", "apple"}. Vector conversion refers to converting each word in a text into a word vector. Semantic analysis refers to analyzing each word vector according to the context semantics to determine the text features of a text.

[0090] For example, obtain the sample negative feedback data of sample object A (for example, Figure 2 News 1) and sample trigger data (e.g., Figure 2 News 2) shown in the figure, the sample negative feedback data (for example, Figure 2 News 1) and sample trigger data (e.g., Figure 2 The news 2 shown is subjected to text segmentation processing to obtain the word set corresponding to the news 1, and each word in the word set is vectorized to obtain the word vector corresponding to each word; the word vector is semantically analyzed to obtain the text features of the news 1.

[0091] It should be noted that the process of obtaining the sample trigger feature of the sample trigger data can refer to the process of obtaining the sample negative feedback feature of the sample negative feedback data, and this application will not elaborate on this.

[0092] Step S102, performing feature transformation on the sample negative feedback feature through the separation model to be trained to obtain a sample mapping feature, and splitting the sample mapping feature into a positive content prediction feature and a negative content prediction feature of the sample object;

[0093] Among them, the separation model to be trained refers to a neural network model in the training stage, and the separation model to be trained can be understood as a mapping network (for example, hidden layer, output layer). The mapping network can refer to a neural network model with a feature transformation function. The mapping network can include but is not limited to: MLP mapping network, self-organizing mapping network, projection mapping network, variational autoencoder, etc. This application does not limit the type of mapping network. The output layer in the mapping network refers to the last layer of the neural network structure of the separation model to be trained. Its main purpose is to receive features from the hidden layer and convert these features into the final output results.

[0094] The feature dimension of the output layer is twice the feature dimension of the trigger feature. The feature dimension of the output layer changes with the feature dimension of the trigger feature. This application does not limit the feature dimension of the output layer. Feature transformation refers to the process of mapping features from one vector space to another new vector space through a mapping network. For example, a 200-dimensional feature vector a is transformed into a 400-dimensional feature vector b through a mapping network.

[0095] Splitting refers to the ability to split a feature vector into multiple sub-feature vectors. Splitting may include but is not limited to: index splitting, length splitting, dimension splitting, etc. Index splitting may refer to splitting feature vectors by specifying indexes or index ranges, length splitting may refer to splitting feature vectors into multiple sub-feature vectors according to specified lengths, and dimension splitting may refer to splitting feature vectors according to dimensions in a multidimensional vector space. This application does not limit the type of splitting network.

[0096] Positive content prediction features may refer to the estimated positive content features in the sample mapping features, negative content prediction features may refer to the estimated negative content features in the sample mapping features, and positive content features may refer to the content features that the object likes (for example, Figure 2 A negative content feature may refer to a content feature that the subject does not like (e.g., Figure 2 "Motion" shown).

[0097] Among them, the specific process of performing feature transformation on the sample negative feedback feature through the separation model to be trained to obtain the sample mapping feature, and splitting the sample mapping feature into the positive content prediction feature and the negative content prediction feature of the sample object includes: performing feature transformation on the sample negative feedback feature through the separation model to be trained to obtain the sample mapping feature; obtaining the feature dimension of the sample mapping feature, and determining the center position of the sample mapping feature according to the feature dimension; splitting the sample mapping feature according to the center position to obtain the first sequence sub-feature and the second sequence sub-feature of the sample mapping feature, and the first sequence sub-feature and the second sequence sub-feature have the same dimension; determining the first sequence sub-feature as the positive content prediction feature of the sample object, and determining the second sequence sub-feature as the negative content prediction feature of the sample object.

[0098] For example, the separation model to be trained is an MLP mapping network. The MLP mapping network includes an input layer, three hidden layers, and an output layer. If the feature dimension of the sample trigger feature is 200 dimensions, the feature dimension of the output layer of the MLP mapping network is set to 400 dimensions. The sample negative feedback feature is input into the MLP mapping network. The input layer of the MLP mapping network receives the sample negative feedback feature. The sample negative feedback feature is input into the first hidden layer. The activation function of the first hidden layer is used to perform a linear transformation to obtain the first feature. The first feature is input into the second hidden layer. The activation function of the second hidden layer is used to perform a linear transformation to obtain the second feature. The second feature is input into the third hidden layer. The activation function of the third hidden layer is used to perform a linear transformation to obtain the third feature. The third feature is input into the output layer for feature dimension conversion to obtain a 400-dimensional sample mapping feature.

[0099] Further, the feature dimension of the sample mapping feature is obtained (for example, 400 dimensions), and the center position of the sample mapping feature is determined to be the position of the 200th dimension feature. According to the position of the 200th dimension feature, the sample mapping feature is split, and the first 200 dimensions of the sample mapping feature are used as the first sequence sub-features of the sample mapping feature, and the last 200 dimensions of the sample mapping feature are used as the second sequence sub-features of the sample mapping feature, to obtain the first sequence sub-features and the second sequence sub-features of the sample mapping feature, and the feature dimensions of the first sequence sub-features and the second sequence sub-features are both 200 dimensions. The first sequence sub-features are determined as the positive content prediction features of the sample object, and the second sequence sub-features are determined as the negative content prediction features of the sample object.

[0100] For another example, if the feature dimensions of the sample negative feedback feature and the sample trigger feature are 4-dimensional, the 4-dimensional sample negative feedback feature is transformed through the mapping network in the separation model to be trained to obtain an 8-dimensional sample mapping feature a. If the sample mapping feature a is [1, 2, 3, 4, 5, 6, 7, 8], the feature dimension of the sample mapping feature a is 8-dimensional. According to the feature dimension of the sample mapping feature, the center position of the sample mapping feature a is determined to be the position of the 4th dimension feature in the sample mapping feature. According to the position of the 4th dimension feature, the sample mapping feature is split into the first 4-dimensional sample mapping sub-features (which can be understood as the first sequence sub-features) and the second 4-dimensional sample mapping sub-features (which can be understood as the second sequence sub-features). The first sequence sub-features are [1, 2, 3, 4], and the second sequence sub-features are [5, 6, 7, 8]. The first sequence sub-features [1, 2, 3, 4] are determined as the positive content prediction features of the sample object, and the second sequence sub-features [5, 6, 7, 8] are determined as the negative content prediction features of the sample object.

[0101] Step S103, determining a first loss value between a positive content prediction feature and a sample trigger feature, and determining a second loss value between a negative content prediction feature and a sample trigger feature;

[0102] Among them, the first loss value refers to the function value of the loss function (for ease of understanding, it can also be called the error function) between the positive content prediction feature and the sample trigger feature at a specific value. For example, the value of the positive content prediction feature is a, and the value of the sample trigger feature is b. The first loss value is the function value of the loss function (error function) between the positive content prediction feature and the sample trigger feature when the value of the positive content prediction feature is a and the value of the sample trigger feature is b. The second loss value refers to the function value of the loss function (error function) between the negative content prediction feature and the sample trigger feature at a specific value. The loss function (error function) may include but is not limited to: Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), etc. This application does not limit the loss function.

[0103] The MSE loss function can refer to the average of the square of the error between data y and data x. The formula of the MSE loss function is shown in formula (1):

[0104]

[0105] The RMSE loss function can refer to the square root of MSE. The formula of the RMSE loss function is shown in formula (2):

[0106]

[0107] The MAE loss function can refer to the average of the absolute values ​​of the errors between data y and data x. The formula of the MAE loss function is shown in formula (3):

[0108]

[0109] Where n represents the number of data y or data x, y i represents the i-th data of data y, x i Represents the i-th data of data x.

[0110] This application takes the MSE loss function as an example to introduce the determination of the first loss value between the positive content prediction feature and the sample trigger feature. If the positive content prediction feature includes M-dimensional positive content prediction sub-features, and the sample trigger feature includes M-dimensional sample trigger sub-features; M is a positive integer; then the specific process of determining the first loss value between the positive content prediction feature and the sample trigger feature includes: obtaining the error value between the i-th dimension positive content prediction sub-feature in the M-dimensional positive content prediction sub-features and the i-th dimension sample trigger sub-feature in the M-dimensional sample trigger sub-features; and determining the first loss value between the positive content prediction feature and the sample trigger feature based on the M error values ​​between the positive content prediction feature and the sample trigger feature.

[0111] For example, if M is 10, the positive content prediction feature includes 10-dimensional positive content prediction sub-features, the sample trigger feature includes 10-dimensional sample trigger sub-features, and the error value between the i-th dimension positive content prediction sub-feature in the 10-dimensional positive content prediction sub-features and the i-th dimension sample trigger sub-feature in the 10-dimensional sample trigger sub-features is obtained, i=1, 2, 3, 4, 5, ..., 10. For example, the first dimension positive content prediction sub-feature y in the 10-dimensional positive content prediction sub-features is obtained. 1 , get the first dimension sample trigger sub-feature x in the 10-dimensional sample trigger sub-feature 1 , calculate the first dimension positive content prediction sub-feature y 1 and the first dimension sample trigger sub-feature x 1 The error value between 1 -x 1 ); Get the second dimension sample trigger sub-feature x in the 10-dimensional sample trigger sub-feature 1 , calculate the second dimension positive content prediction sub-feature y 1 and the first dimension sample trigger sub-feature x 1 The error value between 2 -x 2 ), ..., obtain the 10th dimension sample trigger sub-feature x in the 10-dimensional sample trigger sub-feature 1, calculate the 10th dimension positive content prediction sub-feature y 1 and the first dimension sample trigger sub-feature x 1 The error value between 10 -x 10 ).

[0112] According to the 10 error values ​​between the positive content prediction feature and the sample trigger feature, the first loss value between the positive content prediction feature and the sample trigger feature is determined. For example, if the loss function is the MSE loss function, the first loss value is MSE 1 =1 / 10((y 1 -x 1 ) 2 +(y 2 -x 2 ) 2+ (y 3 -x 3 ) 2 +(y 4 -x 4 ) 2 +(y 5 -x 5 ) 2 +(y 6 -x 6 ) 2 +(y 7 -x 7 ) 2 +(y 8 -x 8 ) 2+ (y 9 -x 9 ) 2 +(y 10 -x 10 ) 2 ), it can be understood that the process of determining the first loss value between the positive content prediction feature and the sample trigger feature and the process of determining the second loss value between the negative content prediction feature and the sample trigger feature are the same. If the i-th dimension negative content prediction sub-feature in the obtained 10-dimensional negative content prediction sub-feature is z i , then the second loss value is MSE 2 =1 / 10((z 1 -x 1 ) 2 +(z 2 -x 2 ) 2+ (z 3 -x 3 ) 2 +(z 4 -x 4 ) 2 +(z 5-x 5 ) 2 +(z 6 -x 6 ) 2 +(z 7 -x 7 ) 2 +(z 8 -x 8 ) 2+ (z 9 -x 9 ) 2 +(z 10 -x 10 ) 2 ).

[0113] It should be noted that the loss function between the positive content prediction feature and the sample trigger feature is the same as the loss function between the negative content prediction feature and the sample trigger feature. For example, if the loss function between the positive content prediction feature and the sample trigger feature is the MSE loss function, then the loss function between the negative content prediction feature and the sample trigger feature is also the MSE loss function. If the loss function between the positive content prediction feature and the sample trigger feature is the MAE loss function, then the loss function between the negative content prediction feature and the sample trigger feature is also the MAE loss function.

[0114] Step S104, determining the sum of the inverse of the second loss value and the first loss value as the model loss value of the separation model to be trained;

[0115] Specifically, the signs of two numbers that are opposite to each other are opposite. For example, if the second loss value is a, the opposite of the second loss value is -a. For another example, if the second loss value is 5, the opposite of the second loss value is -5. The sum of the opposite of the second loss value and the first loss value is determined as the model loss value of the separation model to be trained. That is, the model loss value of the separation model to be trained is determined by subtracting the second loss value from the first loss value. For example, if the first loss value is MSE 1 , the second loss value is MSE 2 , then the model loss value of the separation model to be trained MSE = MSE 1 -MSE 2 , for example, if the first loss value is MSE 1 , the second loss value is MSE 2 , then the model loss value of the separation model to be trained is MAE = MSE 1 -MSE 2 .

[0116] It is understandable that if the model loss value of the separation model to be trained is to be minimized, the first loss value needs to be minimized and the second loss value needs to be maximized, so that the model loss value of the separation model to be trained can be minimized. For example, taking the MSE loss function as an example, assuming that the minimum threshold of the first loss value is 20 and the maximum threshold of the second loss value is 50.

[0117] Among them, the change of the model loss value of the separation model to be trained can be shown in Table 1:

[0118] Table 1

[0119] <![CDATA[MSE 1 ]]> <![CDATA[MSE 2 ]]> <![CDATA[MSE=MSE 1 -MSE 2 ]]> 60 10 50 50 20 30 40 30 10 30 40 -10 20 50 -30

[0120] When the first loss value is MSE 1 =60, the second loss value is MSE 2 =10, the model loss value of the separation model to be trained is MSE=MSE 1 -MSE 2 =40-10=50, when the first loss value MSE 1 =50, the second loss value is MSE 2 =20, the model loss value of the separation model to be trained is MSE=MSE 1 -MSE 2 =50-20=30. When the first loss value is MSE 1 =40, the second loss value is MSE 2 =30, the model loss value of the separation model to be trained is MSE=MSE 1 -MSE 2 =40-30=10, when the first loss value is the minimum threshold MSE 1 =20, the second loss value is the maximum threshold MSE 2 =50, the model loss value of the separation model to be trained is MSE=MSE 1 -MSE 2 =-30. It can be seen that when the first loss value is the minimum threshold and the second loss value is the maximum threshold, the model loss value of the separation model to be trained is the minimum.

[0121] In one or more embodiments, the sample training set can be trained in batches, and the batch size can be set. The batch size can be set to 16, 32, 128, etc. The present application does not limit the setting of the batch size. For example, if the batch size is set to 32, the sample training set is divided into multiple batches, each batch includes 32 multimedia sample data, and for the multimedia sample data in a batch, the first loss value and the second loss value corresponding to each multimedia data can be calculated according to the above steps, and the first loss values ​​of all the multimedia sample data in the batch are averaged to obtain the first loss average value; the second loss values ​​of all the multimedia sample data in the batch are averaged to obtain the second loss average value, and the first loss average value is subtracted from the second loss average value to obtain the model loss value of the separation model to be trained.

[0122] Step S105, modifying the network parameters of the separation model to be trained according to the model loss value to obtain a content separation model; the content separation model is used to separate positive content and negative content in the negative feedback multimedia data.

[0123] Among them, the network parameters of the separation model to be trained are corrected according to the model loss value to obtain the specific process of the content separation model, including: determining the gradient of the separation model to be trained according to the model loss value, and correcting the network parameters of the separation model to be trained according to the gradient; if the model loss reaches the minimum value, the separation model to be trained containing the corrected network parameters is determined as the content separation model.

[0124] Gradient refers to a vector, which indicates that the model loss function changes fastest along this direction at a certain point (which can be understood as the maximum rate of change). The gradient can be used to find the direction that minimizes the model loss function. Gradient calculation refers to the calculation of the gradient of the model loss function to the model parameters using the back propagation algorithm. The back propagation algorithm is a calculation method based on the chain rule.

[0125] After obtaining the gradient, the network parameters of the separation model to be trained can be updated by the gradient descent method. Gradient descent method refers to an optimization algorithm based on gradient. Gradient descent method may include but is not limited to: Batch Gradient Descent (BGD), Stochastic Gradient Descent (SGD), Mini-Batch Gradient Descent (MBGD), etc. This application does not limit the gradient descent method. Batch gradient descent method refers to using the entire training set to calculate the gradient and update the network parameters of the model. Stochastic gradient descent method refers to using only one sample data or a small batch of sample data to calculate the gradient in each iteration process, and updating the network parameters of the model. Mini-batch gradient descent method refers to using a fixed number of sample data to calculate the gradient in each iteration process, and updating the network parameters of the model.

[0126] Specifically, according to the model loss value, the gradient of the model loss function with respect to the output layer in the separation model to be trained is calculated. Further, the gradient is back-propagated layer by layer to each layer of the neural network in the separation model to be trained using the chain rule, and the gradient of the network parameters of each layer of the neural network is calculated to obtain the gradient of the network parameters of each layer of the neural network in the separation model to be trained. According to the gradient, the network parameters in the separation model to be trained are updated through a gradient descent algorithm (for example, stochastic gradient descent method).

[0127] During the training process of the separation model to be trained, back-propagation (BP) can be performed based on the model loss value (which can be understood as the first loss value minus the second loss value) to calculate the gradient of the model loss function with respect to the network parameters of each layer of the neural network in the separation model to be trained. The network parameters of the separation model to be trained are iteratively adjusted according to the gradient, and the network parameters of the separation model to be trained can be iteratively updated by minimizing and optimizing the model loss function.

[0128] It is understandable that the training phase of the separation model to be trained may include multiple rounds of iterative training (epochs), each round of iterative training may traverse a sample training set once, and each time a batch of multimedia sample data may be obtained from the sample training set and input into the separation model to be trained to perform forward calculations to obtain sample prediction results. In the model training phase, when the number of training times of the separation model to be trained reaches a preset maximum number of iterations, or the model loss function value obtained from the previous iterative training is the same as the model loss function value obtained from the next iterative training, it can be considered that the model loss function at this time has reached the minimum value, the training is stopped, and the model parameters when the model loss function is the smallest are saved. At this time, the separation model to be trained can be determined as a content separation model that has completed training.

[0129] See also Figure 4 , Figure 4 Schematic diagram of the training process of a separation model to be trained provided in an embodiment of the present application. Figure 4 As shown, multimedia sample data is obtained, the multimedia sample data includes sample negative feedback data and sample trigger data, feature extraction is performed on the sample feedback data to obtain sample negative feedback features, and feature extraction is performed on the sample trigger data to obtain sample trigger features.

[0130] The sample negative feedback feature and the sample trigger feature are input into the separation model to be trained, and the sample negative feedback feature is transformed through the mapping network in the separation model to be trained to obtain the sample mapping feature, and the sample mapping feature is feature split to obtain the negative content prediction feature and the positive content prediction feature, and the feature dimensions of the negative content prediction feature and the positive content prediction feature are the same.

[0131] The first loss value between the positive content feature and the sample trigger feature is calculated by forward calculation, and the second loss value between the negative content prediction feature and the sample trigger feature is calculated by forward calculation. The model loss value of the separation model to be trained is obtained by subtracting the second loss value from the first loss value. According to the model loss value, the network parameters of the separation model to be trained are updated by back propagation. When the model loss value of the separation model to be trained is the smallest, the network parameters at this time are determined as the updated network parameters of the separation model to be trained, and the separation model to be trained containing the updated network parameters is determined as the content separation model.

[0132] In an embodiment of the present application, sample negative feedback data and sample trigger data in multimedia sample data are obtained, and feature extraction is performed on the sample negative feedback data and sample trigger data respectively to obtain sample negative feedback features and sample trigger features, and the sample negative feedback features and sample trigger features are input into the separation model to be trained, and the separation model to be trained is trained multiple times, and the network parameters of the separation model to be trained are continuously updated according to the loss value of the separation model to be trained, and when the model loss function reaches the minimum value, the separation model to be trained containing the updated network parameters is determined as the content separation model. The content separation model can be used to separate the positive content and the negative content in the negative feedback multimedia data, so that the data related to the negative content can be shielded, and the data related to the positive content can be recommended to the user, which can improve the recommendation accuracy of the multimedia data.

[0133] It is understandable that through Figure 3 and Figure 4 After the content separation model is trained in the corresponding embodiment, the content separation model can be applied to the recommendation system. The specific application process of the content separation model is described in detail below.

[0134] See also Figure 5 , Figure 5 This is a flow diagram of a data processing method provided in an embodiment of the present application. Figure 2 It can be understood that the video data processing method is executed by a computer device, which may be a terminal device (e.g., Figure 1 Any terminal device in the terminal device set shown in FIG. 1 ), or a server (such as Figure 1 The server 10d shown in the figure is not limited in this application. The video data processing method may include the following steps S201 to S205:

[0135] Step S201, acquiring negative feedback multimedia data and triggering multimedia data of an object, performing feature extraction on the negative feedback multimedia data to obtain negative feedback features of the negative feedback multimedia data, and performing feature extraction on the triggering multimedia data to obtain triggering features of the triggering multimedia data;

[0136] Among them, the process of acquiring the negative feedback multimedia data and trigger multimedia data of the object, performing feature extraction on the negative feedback multimedia data to obtain the negative feedback features of the negative feedback multimedia data, performing feature extraction on the trigger multimedia data to obtain the trigger features of the trigger multimedia data can refer to the sample negative feedback features of the sample negative feedback data in the multimedia sample data in step S101, and the process of acquiring the sample trigger features of the sample trigger data in the multimedia sample data, which will not be repeated here.

[0137] Step S202, performing feature transformation on the negative feedback feature through a content separation model to obtain a negative feedback mapping feature, and splitting the negative feedback mapping feature through the content separation model to obtain a negative feedback content feature and a positive feedback content feature of the object;

[0138] Among them, the content separation model refers to the separation model to be trained after multiple rounds of iterations. The negative feedback content features of the object can be understood as the positive content in the negative feedback multimedia data, and the positive feedback content features of the object can be understood as the negative content in the negative feedback multimedia data.

[0139] The process of transforming the negative feedback features through the content separation model to obtain the negative feedback mapping features can refer to the specific implementation process of transforming the sample negative feedback features through the separation model to be trained to obtain the sample mapping features in step S102, which will not be repeated in this application. By splitting the negative feedback mapping features through the content separation model, the negative feedback content features and positive feedback content features of the object can be obtained. The specific training process of the content separation model can be referred to steps S102 to S105, which will not be repeated here.

[0140] Step S203, obtaining the attention weight of the negative feedback content feature, subtracting the product of the negative feedback content feature and the attention weight from the trigger feature, and obtaining the positive trigger content feature in the trigger feature.

[0141] Among them, the positive trigger content feature may refer to the content feature that the user likes contained in the trigger feature, and the specific process of obtaining the attention weight of the negative feedback content feature includes: calculating the similarity between the negative feedback content feature and the trigger feature to obtain the attention matrix, normalizing the attention matrix to obtain the normalized attention matrix; performing a dot multiplication operation on the difference between the trigger feature and the negative feedback content feature and the normalized attention matrix to obtain the attention weight.

[0142] In practical applications, common methods for similarity calculation include but are not limited to: dot product calculation, cosine similarity, distance formula, MLP network, etc. This application does not limit the method for similarity calculation. Normalization refers to scaling the data so that it falls into a small specific interval, such as the interval [0, 1] or the interval [-1, 1].

[0143] Taking the dot product calculation as an example, the similarity calculation is performed on the negative feedback content features and the trigger features to obtain the attention matrix. If the negative feedback content features are feature X = [1, 2, 3, 4, 5] and the trigger features are feature Y = [6, 7, 8, 9, 10], then the dot product calculation is performed on the negative feedback content features and the trigger features, and the obtained attention matrix is ​​the matrix Normalized by the normalization function (softmax function), the normalized attention matrix is ​​obtained as matrix B.

[0144] The calculation formula of the normalization function (softmax function) is shown in formula (4):

[0145]

[0146] Where e represents the exponential function, e i represents the index of the i-th element in the attention matrix, represents the sum of the exponentials of all elements in the attention matrix, and N represents the number of elements in the attention matrix.

[0147] Subtract the negative feedback content feature (feature X) from the trigger feature (feature Y), and then perform a dot multiplication operation with the attention matrix (matrix B). The resulting attention weight is weight Z = (feature Y-feature X) × matrix B.

[0148] Calculate the product of the negative feedback content feature (feature X) and the attention weight (weight Z), and the obtained product of the negative feedback content feature and the attention weight is feature X×weight Z. Subtract the product of the negative feedback content feature (feature X) and the attention weight (weight Z) from the trigger feature (feature Y), and obtain the positive trigger content feature in the trigger feature as feature L=(feature Y-feature X)×weight Z.

[0149] Step S204, obtaining a candidate multimedia data set that matches the positive trigger content feature, and obtaining a feedback index estimate corresponding to each candidate multimedia data in the candidate multimedia data set;

[0150] Specifically, the candidate multimedia data set refers to a set containing at least one candidate multimedia data, and the candidate multimedia data that matches the positive triggering content feature refers to the candidate multimedia data related to the positive triggering content feature. For example, if the positive triggering content feature is a feature representation of "character A", then the candidate multimedia data that matches the positive triggering content feature is the candidate multimedia data related to "character A", such as film and television dramas starring "character A", skits starring "character A", etc.

[0151] The feedback indicator estimate refers to the predicted value of the feedback indicator, and the feedback indicator estimate may include but is not limited to: trigger indicator estimate (e.g., estimated click-through rate) and conversion indicator estimate (e.g., estimated conversion rate), etc., which are not limited in this application. The estimated click-through rate refers to the predicted probability of a user clicking on multimedia data, and the estimated conversion rate refers to the predicted benefit that the multimedia data can bring.

[0152] A candidate multimedia data set matching the positive triggering content feature is obtained from the business platform. For example, if the business platform is a video platform and the positive triggering content feature is a feature representation of "film and TV series D", then at least one video data related to "film and TV series D" is obtained from the video platform. At this time, the candidate multimedia data set can be understood as at least one video data here.

[0153] Obtain feedback indicator estimates corresponding to each candidate multimedia data in the candidate multimedia data set. For example, the feedback indicator estimates take the estimated click-through rate as an example. If the candidate multimedia data set includes 5 video data related to "film and television drama D", namely video is data 1, video is data 2, video is data 3, video is data 4, and video is data 5, then the 5 video data related to "film and television drama D" can be predicted by a click-through rate prediction model. The click-through rate prediction model refers to a neural network model for predicting the click rate of an input (for example, a click-through rate (CTR) prediction module), and the estimated click-through rates of the 5 video data are obtained. For example, the estimated click-through rates of the 5 video data are 80%, 50%, 90%, 60%, and 70%, respectively.

[0154] Step S205 , sorting each candidate multimedia data according to the feedback index estimation value to obtain a candidate multimedia data sequence of the object; and determining multimedia recommendation data of the object according to the candidate multimedia data sequence.

[0155] Among them, the candidate multimedia data sequence refers to a candidate multimedia data set with a sequence, and the multimedia recommendation data refers to multimedia data recommended to the user. According to the feedback index estimation value, for example, if the candidate multimedia data set includes video is data 1, video is data 2, video is data 3, video is data 4 and video is data 5, the feedback index estimation values ​​(for example, estimated click-through rates) corresponding to video is data 1, video is data 2, video is data 3, video is data 4 and video is data 5 are 80%, 50%, 90%, 60% and 70% respectively, the candidate multimedia data in the candidate multimedia data set are sorted, and the candidate multimedia data sequence of the object is obtained as candidate multimedia data sequence A {video is data 3, video is data 1, video is data 5, video is data 4, video is data 2}.

[0156] It is understandable that, by sorting each candidate multimedia data according to the feedback indicator estimate, the candidate multimedia data sequence of the object can be considered as a candidate multimedia data sequence obtained by considering only one feedback indicator estimate (for example, estimated click rate). In order to improve the user experience and ensure business benefits, it is necessary to comprehensively consider multiple feedback indicator estimates to sort the candidate multimedia data set to obtain an optimal candidate multimedia data sequence.

[0157] For example, if the estimated click-through rate and the estimated conversion rate are comprehensively considered to sort the candidate multimedia data set, the optimal candidate multimedia data sequence is candidate multimedia data sequence C {video is data 1, video is data 5, video is data 3, video is data 4, video is data 2}, then the candidate multimedia data sequence C is determined as the multimedia recommendation data of the object.

[0158] See also Figure 6 , Figure 6 Schematic diagram of a structure for obtaining positive trigger content features provided by an embodiment of the present application. Figure 6 As shown, multimedia data (e.g., news data) can be obtained from a business platform (e.g., a news platform), the multimedia data (e.g., news data) including negative feedback multimedia data and trigger multimedia data, feature extraction is performed on the negative feedback multimedia data and the trigger multimedia data through a feature extraction network to obtain negative feedback features of the negative feedback multimedia data and trigger features of the trigger data, the negative feedback features are input into a content separation model, the negative feedback features are transformed through a mapping network (e.g., an MLP mapping network) in the content separation model to obtain mapping features, and the mapping features are split into positive feedback content and negative feedback content features through a splitting network in the content separation model.

[0159] The attention weights of the trigger features and the negative feedback content features are calculated by an attention mechanism (e.g., an MLP attention mechanism) to obtain the attention weights of the negative feedback content features, and the product of the attention weights and the negative feedback content features is subtracted from the trigger features to obtain the positive trigger content features. The recommendation system can recommend multimedia data of interest to users based on the positive trigger content features. The recommendation system can include but is not limited to: a recall component, a click-through rate (CTR) estimation component, a rearrangement component, etc., which is not limited in this application.

[0160] The recall component can filter out candidate multimedia data sets that match the positive trigger content features from a massive multimedia database through various recall strategies (such as content-based recall strategies, collaborative filtering-based recall strategies, etc.). The click-through rate prediction component can predict the click probability of each candidate multimedia data in the candidate multimedia data set output by the recall component through machine learning algorithms (such as logistic regression algorithms, deep learning algorithms, etc.) to obtain the estimated click rate of each candidate multimedia data. Rearrangement can further adjust the order of each candidate multimedia data while ensuring the click rate and improving user experience and business revenue to obtain an optimal candidate multimedia data sequence, and then determine the multimedia recommendation data recommended to the user.

[0161] In the embodiment of the present application, the negative feedback features of the negative feedback multimedia data of the object and the trigger features of the trigger multimedia data can be obtained, and the negative feedback features are mapped and then split through the content separation model to obtain the negative feedback content features and positive feedback content features of the object, and the attention weight of the negative feedback content features is obtained. The product between the negative feedback content features and the attention weight is subtracted from the trigger feature to obtain the positive trigger content features in the trigger feature. Through the embodiment of the present application, the negative feedback features and the trigger features can be purified, and the negative feedback features only contain negative feedback content features, and the trigger features only contain positive feedback content features, so that the multimedia data recommended by the recommendation system is more accurate.

[0162] See also Figure 7 , Figure 7 Schematic diagram of a data processing device provided in an embodiment of the present application. Figure 7 As shown, the data processing device 1 may include: a feature acquisition module 101, a feature separation module 102, a first determination module 103, a second determination module 104, and a parameter correction module 105;

[0163] The feature acquisition module 101 is used to acquire the sample negative feedback feature of the sample negative feedback data in the multimedia sample data, and acquire the sample trigger feature of the sample trigger data in the multimedia sample data;

[0164] A feature splitting module 102 is used to perform feature transformation on the sample negative feedback feature through the separation model to be trained to obtain a sample mapping feature, and split the sample mapping feature into a positive content prediction feature and a negative content prediction feature of the sample object;

[0165] A first determination module 103, configured to determine a first loss value between a positive content prediction feature and a sample trigger feature, and to determine a second loss value between a negative content prediction feature and a sample trigger feature;

[0166] A second determination module 104 is used to determine the sum of the inverse of the second loss value and the first loss value as the model loss value of the separation model to be trained;

[0167] The parameter correction module 105 is used to correct the network parameters of the separation model to be trained according to the model loss value to obtain a content separation model; the content separation model is used to separate positive content and negative content in the negative feedback multimedia data.

[0168] In one or more embodiments, the sample negative feedback data includes sample text data;

[0169] In one or more embodiments, the feature acquisition module acquires sample negative feedback features of sample negative feedback data in the multimedia sample data, and is used to perform the following steps:

[0170] Acquire multimedia sample data of the sample object; the multimedia sample data includes sample negative feedback data and sample trigger data;

[0171] Performing text segmentation processing on sample text data in the multimedia sample data to obtain a negative feedback word set corresponding to the sample negative feedback data;

[0172] Performing vector conversion on each negative feedback word in the negative feedback word set to obtain the negative feedback word vector corresponding to each negative feedback word;

[0173] Perform semantic analysis on the negative feedback word vector to obtain the sample negative feedback features of the sample negative feedback data.

[0174] In one or more embodiments, the feature splitting module performs feature transformation on the sample negative feedback feature through the separation model to be trained to obtain the sample mapping feature, and splits the sample mapping feature into the positive content prediction feature and the negative content prediction feature of the sample object to perform the following steps:

[0175] Perform feature transformation on the negative feedback features of the sample through the separation model to be trained to obtain the sample mapping features;

[0176] Obtaining a feature dimension of the sample mapping feature, and determining a center position of the sample mapping feature according to the feature dimension;

[0177] According to the center position, the sample mapping feature is split to obtain a first sequence sub-feature and a second sequence sub-feature of the sample mapping feature, and the first sequence sub-feature and the second sequence sub-feature have the same dimension;

[0178] The first sequence sub-features are determined as positive content prediction features of the sample object, and the second sequence sub-features are determined as negative content prediction features of the sample object.

[0179] In one or more embodiments, the positive content prediction feature includes M-dimensional positive content prediction sub-features, and the sample trigger feature includes M-dimensional sample trigger sub-features; M is a positive integer;

[0180] In one or more embodiments, the first determination module determines a first loss value between the positive content prediction feature and the sample trigger feature, and is used to perform the following steps:

[0181] Obtaining an error value between an i-th dimension positive content prediction sub-feature in the M-dimensional positive content prediction sub-feature and an i-th dimension sample trigger sub-feature in the M-dimensional sample trigger sub-feature;

[0182] A first loss value between the positive content prediction feature and the sample trigger feature is determined according to M error values ​​between the positive content prediction feature and the sample trigger feature.

[0183] In one or more embodiments, the parameter modification module modifies the network parameters of the separation model to be trained according to the model loss value to obtain a content separation model for performing the following steps:

[0184] According to the model loss value, the gradient of the separation model to be trained is determined, and the network parameters of the separation model to be trained are modified according to the gradient;

[0185] If the model loss reaches the minimum value, the separation model to be trained including the corrected network parameters will be determined as the content separation model.

[0186] In one or more embodiments, the apparatus further comprises:

[0187] The forward feature acquisition module 106 is used to perform the following steps:

[0188] Acquire negative feedback multimedia data and trigger multimedia data of the object;

[0189] Extracting features from negative feedback multimedia data to obtain negative feedback features of the negative feedback multimedia data, and extracting features from triggering multimedia data to obtain triggering features of the triggering multimedia data;

[0190] The negative feedback feature is transformed through the content separation model to obtain the negative feedback mapping feature, and the negative feedback mapping feature is split through the content separation model to obtain the negative feedback content feature and the positive feedback content feature of the object;

[0191] The attention weight of the negative feedback content feature is obtained, and the product of the negative feedback content feature and the attention weight is subtracted from the trigger feature to obtain the positive trigger content feature in the trigger feature.

[0192] In one or more embodiments, the positive feature acquisition module acquires the attention weight of the negative feedback content feature to perform the following steps:

[0193] Calculate the similarity between the negative feedback content features and the trigger features to obtain an attention matrix, and normalize the attention matrix to obtain a normalized attention matrix;

[0194] The difference between the trigger feature and the negative feedback content feature is multiplied by the normalized attention matrix to obtain the attention weight.

[0195] In one or more embodiments, the apparatus further comprises:

[0196] The data acquisition module 107 is used to acquire a candidate multimedia data set matching the positive trigger content feature from the service platform;

[0197] The indicator acquisition module 108 is used to obtain the feedback indicator estimated value corresponding to each candidate multimedia data in the candidate multimedia data set;

[0198] The data sorting module 109 is used to sort each candidate multimedia data in the candidate multimedia data set according to the feedback indicator estimation value to obtain a candidate multimedia data sequence of the object;

[0199] The data determination module 110 is used to determine the multimedia recommendation data of the object according to the candidate multimedia data sequence.

[0200] According to an embodiment of the present application, the above Figure 3 and Figure 5 The relevant steps involved in the data processing method shown can be represented by Figure 7 The data processing device 1 shown in the figure is executed by each module. For example, Figure 3 Step S101 shown can be performed by Figure 7 The feature acquisition module 101 shown is used to execute, Figure 3 Step S102 shown can be performed by Figure 7 The feature splitting module 102 shown is used to perform, Figure 3 Step S103 shown can be performed by Figure 7 The first determination module 103 shown is used to execute, Figure 3 Step S104 shown can be performed by Figure 7 The second determination module 104 shown is used to execute, Figure 3 Step S105 shown can be performed by Figure 7 The parameter correction module 105 shown is used to perform the above operations.

[0201] According to one embodiment of the present application, Figure 7The modules in the data processing device 1 shown can be combined into one or more modules separately or in whole, or one (or more) of the modules can be further divided into at least two smaller units in terms of function, so as to achieve the same operation without affecting the realization of the technical effects of the embodiments of the present application. The above modules are divided based on logical functions. In practical applications, the functions of one module can also be realized by at least two units, or the functions of at least two modules can be realized by one module.

[0202] In the embodiment of the present application, negative feedback data and trigger data in multimedia data can be obtained, and feature extraction is performed on the negative feedback data and the trigger data respectively to obtain negative feedback features corresponding to the negative feedback data and trigger features corresponding to the trigger data. The negative feedback features are split by a content separation model to obtain positive feedback content features and negative feedback content features, and the attention weight of the negative feedback content features is calculated. The product of the negative feedback content features and the attention weight is subtracted from the trigger features to remove the negative feedback content features in the trigger features, so that only positive feedback content features are included in the trigger features. The recommendation system shields multimedia data related to negative feedback content features and only recommends multimedia data with positive feedback content features to users, thereby improving the recommendation accuracy of multimedia data.

[0203] See also Figure 8 , Figure 8 Schematic diagram of the structure of a computer device provided in an embodiment of the present application. Figure 8 As shown, the computer device 1000 may be a terminal device, for example, Figure 1 The terminal device 10a in the corresponding embodiment may also be a server, for example, Figure 1 The server 10d in the corresponding embodiment will not be limited here. For ease of understanding, this application takes a computer device as an example of a terminal device. The computer device 1000 may include: a processor 1001, a network interface 1004 and a memory 1005. In addition, the computer device 1000 may also include: a user interface 1003, and at least one communication bus 1002. Among them, the communication bus 1002 is used to realize the connection and communication between these components. Among them, the user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or it may be a non-volatile memory (non-volatile memory), such as at least one disk storage. The memory 1005 may optionally also be at least one storage device located away from the aforementioned processor 1001. As Figure 8As shown, the memory 1005 as a computer-readable storage medium may include an operating system, a network communication module, a user interface module, and a device control application program.

[0204] The network interface 1004 in the computer device 1000 can also provide a network communication function, and the optional user interface 1003 can also include a display screen (Display) and a keyboard (Keyboard). Figure 8 In the computer device 1000 shown, the network interface 1004 can provide a network communication function; the user interface 1003 is mainly used to provide an input interface for the user; and the processor 1001 can be used to call the device control application stored in the memory 1005 to achieve:

[0205] Acquire sample negative feedback features of sample negative feedback data in the multimedia sample data, and acquire sample trigger features of sample trigger data in the multimedia sample data;

[0206] Performing feature transformation on the negative feedback features of the sample through the separation model to be trained to obtain sample mapping features, and splitting the sample mapping features into positive content prediction features and negative content prediction features of the sample object;

[0207] Determine a first loss value between a positive content prediction feature and a sample trigger feature, and determine a second loss value between a negative content prediction feature and a sample trigger feature;

[0208] The sum of the inverse of the second loss value and the first loss value is determined as the model loss value of the separation model to be trained;

[0209] The network parameters of the separation model to be trained are modified according to the model loss value to obtain a content separation model; the content separation model is used to separate the positive content and the negative content in the negative feedback multimedia data.

[0210] It should be understood that the computer device 1000 described in the embodiment of the present application can execute the above Figure 3 , Figure 5 The description of the data processing method in any of the embodiments may also be performed as described above. Figure 7 The description of the data processing device 1 in the corresponding embodiment will not be repeated here. In addition, the description of the beneficial effects of the same method will not be repeated here either.

[0211] In addition, it should be pointed out here that: the embodiment of the present application also provides a computer-readable storage medium, and the computer-readable storage medium stores a computer program executed by the data processing device 1 mentioned above, and the computer program includes computer instructions. When the processor executes the computer instructions, it can execute the above-mentioned Figure 3 , Figure 5The description of the data processing method in any of the embodiments will not be repeated here. In addition, the description of the beneficial effects of the same method will not be repeated. Among them, the storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc. For technical details not disclosed in the computer-readable storage medium embodiment involved in this application, please refer to the description of the method embodiment of this application. As an example, program instructions can be deployed on a computer device for execution, or on multiple computer devices located at one location, or on multiple computer devices distributed at multiple locations and interconnected by a communication network. Multiple computer devices distributed at multiple locations and interconnected by a communication network can constitute a blockchain system.

[0212] In addition, it should be noted that: the embodiment of the present application also provides a computer program product, which may include a computer program, and the computer program may be stored in a computer-readable storage medium. The processor of the computer device reads the computer program from the computer-readable storage medium, and the processor may execute the computer program, so that the computer device performs the above Figure 3 , Figure 5 The description of the data processing method in any of the embodiments will not be repeated here. In addition, the description of the beneficial effects of the same method will not be repeated. For technical details not disclosed in the computer program product or computer program embodiment involved in the present application, please refer to the description of the method embodiment of the present application.

[0213] The terms "first", "second", etc. in the description, claims, and drawings of the embodiments of the present application are used to distinguish different media contents, rather than to describe a specific order. In addition, the term "comprising" and any variation thereof are intended to cover non-exclusive inclusions. For example, a process, method, device, product, or equipment comprising a series of steps or units is not limited to the listed steps or modules, but may optionally include steps or modules that are not listed, or may optionally include other step units inherent to these processes, methods, devices, products, or equipment.

[0214] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0215] The method and related apparatus provided by the embodiment of the present application are described with reference to the method flow chart and / or structural diagram provided by the embodiment of the present application. Specifically, each process and / or box in the method flow chart and / or structural diagram, as well as the combination of the processes and / or boxes in the flow chart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the process in the process. Figure 1 A process or multiple processes and / or structures Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the functions specified in the process. Figure 1 A process or multiple processes and / or structures Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A flow or multiple flows and / or structures illustrate the steps of the functions specified in one block or multiple blocks.

[0216] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0217] The above disclosure is only the preferred embodiment of the present application, which certainly cannot be used to limit the scope of rights of the present application. Therefore, equivalent changes made according to the claims of the present application are still within the scope covered by the present application.

Claims

1. A data processing method, characterized in that: include: Acquire sample negative feedback features of sample negative feedback data in the multimedia sample data, and acquire sample trigger features of sample trigger data in the multimedia sample data; Performing feature transformation on the sample negative feedback feature through the separation model to be trained to obtain a sample mapping feature, and splitting the sample mapping feature into a positive content prediction feature and a negative content prediction feature of the sample object; Determine a first loss value between the positive content prediction feature and the sample trigger feature, and determine a second loss value between the negative content prediction feature and the sample trigger feature; Determine the sum of the inverse of the second loss value and the first loss value as the model loss value of the separation model to be trained; The network parameters of the separation model to be trained are modified according to the model loss value to obtain a content separation model; the content separation model is used to separate positive content and negative content in negative feedback multimedia data.

2. The method according to claim 1, characterized in that The sample negative feedback data includes sample text data; The step of obtaining the sample negative feedback feature of the sample negative feedback data in the multimedia sample data includes: Acquire multimedia sample data of a sample object; the multimedia sample data includes sample negative feedback data and sample trigger data; Performing text segmentation processing on the sample text data in the multimedia sample data to obtain a negative feedback word set corresponding to the sample negative feedback data; Performing vector conversion on each negative feedback word in the negative feedback word set to obtain a negative feedback word vector corresponding to each negative feedback word; Perform semantic analysis on the negative feedback word vector to obtain sample negative feedback features of the sample negative feedback data.

3. The method according to claim 1, characterized in that The step of performing feature transformation on the sample negative feedback feature through the separation model to be trained to obtain a sample mapping feature, and splitting the sample mapping feature into a positive content prediction feature and a negative content prediction feature of the sample object includes: Perform feature transformation on the negative feedback features of the sample through the separation model to be trained to obtain the sample mapping features; Acquire a feature dimension of the sample mapping feature, and determine a center position of the sample mapping feature according to the feature dimension; Splitting the sample mapping feature according to the center position to obtain a first sequence sub-feature and a second sequence sub-feature of the sample mapping feature, wherein the first sequence sub-feature and the second sequence sub-feature have the same dimension; The first sequence sub-feature is determined as a positive content prediction feature of the sample object, and the second sequence sub-feature is determined as a negative content prediction feature of the sample object.

4. The method according to claim 1, characterized in that: The positive content prediction feature includes M-dimensional positive content prediction sub-features, and the sample trigger feature includes M-dimensional sample trigger sub-features; M is a positive integer; The determining a first loss value between the positive content prediction feature and the sample trigger feature includes: Obtaining an error value between an i-th dimension positive content prediction sub-feature in the M-dimensional positive content prediction sub-feature and an i-th dimension sample trigger sub-feature in the M-dimensional sample trigger sub-feature; A first loss value between the positive content prediction feature and the sample trigger feature is determined according to M error values ​​between the positive content prediction feature and the sample trigger feature.

5. The method according to claim 1, characterized in that: The step of modifying the network parameters of the separation model to be trained according to the model loss value to obtain a content separation model includes: Determine the gradient of the separation model to be trained according to the model loss value, and modify the network parameters of the separation model to be trained according to the gradient; If the model loss reaches the minimum value, the separation model to be trained including the corrected network parameters is determined as the content separation model.

6. The method according to claim 1, characterized in that The method further comprises: Acquire negative feedback multimedia data and trigger multimedia data of the object; Performing feature extraction on the negative feedback multimedia data to obtain negative feedback features of the negative feedback multimedia data, and performing feature extraction on the triggering multimedia data to obtain trigger features of the triggering multimedia data; Performing feature transformation on the negative feedback feature through the content separation model to obtain negative feedback mapping features, and splitting the negative feedback mapping features through the content separation model to obtain negative feedback content features and positive feedback content features of the object; The attention weight of the negative feedback content feature is obtained, and the product of the negative feedback content feature and the attention weight is subtracted from the trigger feature to obtain the positive trigger content feature in the trigger feature.

7. The method according to claim 6, characterized in that The obtaining of the attention weight of the negative feedback content feature includes: Calculating similarity between the negative feedback content feature and the trigger feature to obtain an attention matrix, and normalizing the attention matrix to obtain a normalized attention matrix; The difference between the trigger feature and the negative feedback content feature is multiplied by the normalized attention matrix to obtain an attention weight.

8. The method according to any one of claims 1 to 7, characterized in that: The method further comprises: Acquire a candidate multimedia data set matching the positive trigger content feature from a service platform; Obtaining a feedback indicator estimated value corresponding to each candidate multimedia data in the candidate multimedia data set; sorting each candidate multimedia data in the candidate multimedia data set according to the feedback indicator estimated value to obtain a candidate multimedia data sequence of the object; Determine the multimedia recommendation data of the object according to the candidate multimedia data sequence.

9. A data processing device, characterized in that: include: A feature acquisition module, used to acquire sample negative feedback features of sample negative feedback data in the multimedia sample data, and acquire sample trigger features of sample trigger data in the multimedia sample data; A feature splitting module, used to perform feature transformation on the sample negative feedback feature through the separation model to be trained to obtain a sample mapping feature, and split the sample mapping feature into a positive content prediction feature and a negative content prediction feature of the sample object; A first determination module, configured to determine a first loss value between the positive content prediction feature and the sample trigger feature, and determine a second loss value between the negative content prediction feature and the sample trigger feature; A second determination module, used to determine the sum of the inverse of the second loss value and the first loss value as the model loss value of the separation model to be trained; A parameter correction module is used to correct the network parameters of the separation model to be trained according to the model loss value to obtain a content separation model; the content separation model is used to separate positive content and negative content in negative feedback multimedia data.

10. A computer device, characterized in that: including memory and processor; The memory is connected to the processor, the memory is used to store a computer program, and the processor is used to call the computer program so that the computer device executes the method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded and executed by a processor, so that a computer device having the processor executes the method according to any one of claims 1 to 8.

Images