Data processing method, electronic device and storage medium

Abstract

The embodiment of the present application discloses a data processing method, an electronic device and a storage medium, which are applied to the field of data processing technology. The method includes: obtaining a first object attribute and a second object attribute of a recommended object, inputting the first object attribute into a recommendation model to obtain a first object feature vector, and inputting the second object attribute into the recommendation model to obtain a second object feature vector, obtaining a feature fusion matrix in the recommendation model, and decomposing the feature fusion matrix to obtain a first decomposition matrix for the first object attribute and a second decomposition matrix for the second object attribute, performing feature fusion on the first object feature vector and the second object feature vector according to the first decomposition matrix and the second decomposition matrix to obtain a fused object feature vector for the recommended object, and generating a prediction recommendation index for the recommended object according to the recommendation model and the fused object feature vector. By adopting the embodiment of the present application, the prediction accuracy for the recommended object (such as a video) can be improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a data processing method, an electronic device and a storage medium. Background Art

[0002] At present, in the recommendation scenario, the recommendation index (such as click-through rate) of the recommended object can be predicted, and then the recommended object can be ranked and recommended according to the recommendation index of the recommended object. When predicting, multiple object feature vectors can be obtained based on the multiple object attributes of the recommended object, and the multiple object feature vectors can be feature fused to obtain a fused object feature vector for the recommended object. Subsequently, the recommendation index can be predicted based on the fused object feature vector. The existing feature fusion method usually splices the multiple object feature vectors as a fused object feature vector. However, this fusion method makes the feature information interaction between the multiple object feature vectors shallow, resulting in low prediction accuracy. Therefore, how to improve the prediction accuracy for the recommended object has become a problem to be solved urgently. Summary of the invention

[0003] The embodiments of the present application provide a data processing method, an electronic device, and a storage medium, which can effectively improve the prediction accuracy for recommended objects.

[0004] On the one hand, an embodiment of the present application provides a data processing method, the method comprising:

[0005] Obtain a first object attribute and a second object attribute of a recommended object;

[0006] Inputting the first object attribute into the recommendation model to obtain a first object feature vector of the recommended object, and inputting the second object attribute into the recommendation model to obtain a second object feature vector of the recommended object; the recommendation model is trained based on the first sample object attribute, the second sample object attribute and the sample recommendation index for the sample object;

[0007] Acquire a feature fusion matrix in the recommendation model, and decompose the feature fusion matrix to obtain a first decomposition matrix for the first object attribute and a second decomposition matrix for the second object attribute;

[0008] According to a first decomposition matrix for the first object attribute and a second decomposition matrix for the second object attribute, feature fusion is performed on the first object feature vector and the second object feature vector to obtain a fused object feature vector for the recommended object;

[0009] Generate a prediction recommendation index for the recommended object based on the recommendation model and the fused object feature vector.

[0010] On the one hand, an embodiment of the present application provides a data processing device, the device comprising:

[0011] An acquisition module, used for acquiring a first object attribute and a second object attribute of a recommended object;

[0012] An input module, used to input the first object attribute into the recommendation model to obtain a first object feature vector of the recommended object, and input the second object attribute into the recommendation model to obtain a second object feature vector of the recommended object; the recommendation model is trained based on the first sample object attribute, the second sample object attribute of the sample object and the sample recommendation index for the sample object;

[0013] A decomposition module, used for obtaining a feature fusion matrix in the recommendation model, and decomposing the feature fusion matrix to obtain a first decomposition matrix for the first object attribute and a second decomposition matrix for the second object attribute;

[0014] A fusion module, configured to perform feature fusion on the first object feature vector and the second object feature vector according to a first decomposition matrix for the first object attribute and a second decomposition matrix for the second object attribute, so as to obtain a fused object feature vector for the recommended object;

[0015] The generation module is used to generate a prediction recommendation index for the recommended object according to the recommendation model and the fused object feature vector.

[0016] On the one hand, an embodiment of the present application provides an electronic device, which includes a processor and a memory, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute some or all of the steps in the above method.

[0017] On the one hand, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program, and the computer program includes program instructions, which, when executed by a processor, are used to execute some or all of the steps in the above method.

[0018] Accordingly, according to one aspect of the present application, a computer program product or computer program is provided, the computer program product or computer program comprising program instructions, the program instructions being stored in a computer-readable storage medium. A processor of a computer device reads the program instructions from the computer-readable storage medium, and the processor executes the program instructions, so that the computer device performs the data processing method provided above.

[0019] In the embodiment of the present application, the first object attribute and the second object attribute of the recommended object can be obtained, the first object attribute is input into the recommendation model to obtain the first object feature vector of the recommended object, and the second object attribute is input into the recommendation model to obtain the second object feature vector of the recommended object, and the feature fusion matrix in the recommendation model is obtained, and the feature fusion matrix is ​​decomposed to obtain the first decomposition matrix for the first object attribute and the second decomposition matrix for the second object attribute, and the first object feature vector and the second object feature vector are feature fused according to the first decomposition matrix for the first object attribute and the second decomposition matrix for the second object attribute to obtain the fused object feature vector for the recommended object, and the predicted recommendation index for the recommended object is generated according to the recommendation model and the fused object feature vector. By implementing the above-mentioned proposed method, by combining the feature fusion matrix to perform feature fusion on multiple object feature vectors, the interaction between features can be strengthened to improve the prediction accuracy of the recommendation index based on the fused object feature vector. In addition, when performing feature fusion, feature fusion can be achieved by decomposing the matrix for the feature fusion matrix, so that the amount of parameters used in feature fusion can be reduced, thereby reducing the complexity of the model at this stage and improving the prediction efficiency of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1 A schematic diagram of an application architecture provided for an embodiment of the present application;

[0022] Figure 2 A flowchart of a data processing method provided in an embodiment of the present application;

[0023] Figure 3 A flowchart of a data processing method provided in an embodiment of the present application;

[0024] Figure 4 A schematic diagram of a scenario for determining a fusion object feature vector provided in an embodiment of the present application;

[0025] Figure 5a A schematic diagram of an application scenario provided for an embodiment of the present application;

[0026] Figure 5b A schematic diagram of an application scenario provided for an embodiment of the present application;

[0027] Figure 6A schematic diagram of a scenario for determining a prediction recommendation index provided in an embodiment of the present application;

[0028] Figure 7 A schematic diagram of the structure of a data processing device provided in an embodiment of the present application;

[0029] Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0031] The data processing method proposed in the embodiment of the present application is implemented in an electronic device, which can be a server or a terminal. Among them, the server 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, CDN, and big data and artificial intelligence platforms. The terminal can be a smart phone, a tablet computer, a laptop computer, an intelligent voice interaction device, a smart home appliance, a car terminal, a desktop computer, a smart speaker, a smart watch, etc., but is not limited to this.

[0032] The embodiments of the present application relate to the field of machine learning technology. Machine Learning (ML) is a multi-disciplinary cross-disciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all fields of artificial intelligence. Machine learning and deep learning generally include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learning by teaching.

[0033] In some embodiments, see Figure 1 , Figure 1 The following is a schematic diagram of an application architecture provided in an embodiment of the present application. The data processing method proposed in the present application can be executed through the application architecture. Figure 1 As shown, Figure 1It may include a terminal device, an electronic device, and a recommendation model deployed in the electronic device. The electronic device may predict multiple recommended objects based on the recommendation model to obtain a predicted recommendation index for each recommended object. The electronic device may sort multiple recommended objects according to the predicted recommendation index for each recommended object, and send the sorted multiple recommended objects to the terminal device for display. Among them, the above-mentioned recommendation model may include a feature extraction layer, a feature fusion layer, and a feature prediction layer. The feature extraction layer may be used to extract features of the first object attribute of the recommended object input by the electronic device to obtain a first object feature vector of the recommended object, and may be used to extract features of the second object attribute of the recommended object input by the electronic device to obtain a second object feature vector of the recommended object; the feature fusion layer may be used to perform feature fusion on the first object feature vector and the second object feature vector to obtain a fused object feature vector for the recommended object; the feature prediction layer may be used to predict the fused object feature vector to generate a predicted recommendation index for the recommended object. Subsequently, the electronic device may push the recommended object in combination with the predicted recommendation index of the recommended object.

[0034] Understandably, Figure 1 It is only an exemplary representation of the possible application architecture of the technical solution of the present application, and does not limit the specific architecture of the technical solution of the present application, that is, the technical solution of the present application can also provide other forms of application architecture.

[0035] Optionally, in some embodiments, the electronic device may execute the data processing method to improve the prediction accuracy and prediction efficiency for the recommended object according to actual business needs. The technical solution of the present application can be applied to any recommendation scenario, that is, the electronic device can obtain multiple object attributes of the recommended object (such as the first object attribute and the second object attribute), and perform the feature fusion step included in the technical solution of the present application on the multiple object attributes to obtain a fused object feature vector for the recommended object, and perform prediction based on the recommendation model and the fused object feature vector to obtain a predicted recommendation index for the recommended object.

[0036] For example, the technical solution of the present application can be applied to the recommendation scenario of graphic data, in which case the recommended object is graphic data (such as video, etc.), the first object attribute can be the text attribute of the graphic data (such as video title), the second object attribute can be the image attribute of the graphic data (such as video cover frame), and the predicted recommendation index can be the predicted click rate for the graphic data. For another example, the technical solution of the present application can also be applied to the recommendation scenario of audio data, in which case the recommended object is audio data (such as songs in a playlist on a music software), the first object data can be the text attribute of the audio data (such as song title), the second object data can also be the text attribute of the audio data (such as song introduction), and the predicted recommendation index can be the predicted play rate for the audio data.

[0037] Optionally, the data involved in this application, such as the predicted recommendation indicators of the recommended objects, can be stored in a database, or can be stored in a blockchain, such as through a blockchain distributed system, which is not limited in this application.

[0038] It is understood that the above scenarios are only examples and do not constitute a limitation on the application scenarios of the technical solutions provided in the embodiments of the present application. The technical solutions of the present application can also be applied to other scenarios. For example, it is known to those skilled in the art that with the evolution of the system architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0039] The solution provided in the embodiments of the present application involves technologies such as machine learning of artificial intelligence, which is specifically described by the following embodiments:

[0040] Based on the above description, the present application embodiment proposes a data processing method, which can be executed by the electronic device mentioned above. Figure 2 , Figure 2 A flow chart of a data processing method provided in an embodiment of the present application. Figure 2 As shown, the process of the data processing method of the embodiment of the present application may include the following:

[0041] S201: Acquire a first object attribute and a second object attribute of a recommended object.

[0042] The recommended object can be any object that can be recommended, such as a video, a product, or news. The specific type of the recommended object is not limited here. The first object attribute and the second object attribute are used to represent the characteristics of different dimensions of the recommended object. The first object attribute and the second object attribute can be set by relevant business personnel according to the actual application scenario and are not limited here. For example, taking the recommended object as a video, the first object attribute can be the title information of the video, and the second object attribute can be the cover frame information of the video.

[0043] Optionally, the electronic device may also obtain multiple (more than two) object attributes of the recommended object, and obtain a fused object feature vector for the recommended object according to the multiple object attributes, for use in predicting a recommendation index of the recommended object.

[0044] S202: Input the first object attribute into the recommendation model to obtain a first object feature vector of the recommended object, and input the second object attribute into the recommendation model to obtain a second object feature vector of the recommended object.

[0045] In one possible implementation, the recommendation model may include a feature extraction layer, and the electronic device may input a first object attribute into the recommendation model, and the feature extraction layer processes the first object attribute to obtain a first object feature vector of the recommended object, and the first object feature vector may be understood as the feature of the first object attribute learned by the recommendation model, and the second object attribute may be input into the recommendation model, and the feature extraction layer processes the second object attribute to obtain a second object feature vector of the recommended object, and the second object feature vector may be understood as the feature of the second object attribute learned by the recommendation model. The number of elements in the first object feature vector may be the same as or different from the number of elements in the second object feature vector. That is, the first object feature vector is represented as f 1 ∈R 1*n , the second object feature vector is represented as f 2 ∈R 1*m , n and m may be the same or different.

[0046] S203: Acquire a feature fusion matrix in the recommendation model, and decompose the feature fusion matrix to obtain a first decomposition matrix for the first object attribute and a second decomposition matrix for the second object attribute.

[0047] In one possible implementation, the recommendation model may include a feature fusion layer, and the electronic device may obtain a feature fusion matrix in the feature fusion layer included in the recommendation model, and use the feature fusion matrix to perform feature fusion on the first object feature vector and the second object feature vector; the feature fusion matrix may be one or more. In some embodiments, the feature fusion matrix may be represented as W∈R n*m The dimension of the feature fusion matrix can be n*m, that is, the number of rows of the feature fusion matrix is ​​equal to the number of elements of the feature vector of the first object, and the number of columns of the feature fusion matrix is ​​equal to the number of elements of the feature vector of the second object.

[0048] In some embodiments, the electronic device may decompose the feature fusion matrix in the acquired recommendation model in the feature fusion layer to obtain a first decomposition matrix for the first object attribute and a second decomposition matrix for the second object attribute, and use the first decomposition matrix and the second decomposition matrix to perform feature fusion. This can achieve the purpose of deepening feature interaction based on the feature fusion matrix to improve the accuracy of model prediction. By decomposing the feature fusion matrix, the first decomposition matrix and the second decomposition matrix can also reduce the amount of parameters used, thereby improving the model prediction efficiency. Optionally, the electronic device may decompose the feature fusion matrix, that is, decompose the feature fusion matrix into a first matrix (U∈R n*k , k can be set by relevant business personnel based on experience) and the second matrix (V∈R k*m), and the first matrix is ​​used as the first decomposition matrix, and the second matrix is ​​used as the second decomposition matrix. By decomposing the feature fusion matrix (i.e., tensor decomposition method) to obtain the first decomposition matrix and the second decomposition matrix, and then performing feature fusion, the first decomposition matrix and the second decomposition matrix can be approximated to the feature fusion matrix, so as to reduce the number of parameters while deepening the interaction between features.

[0049] S204 , performing feature fusion on the first object feature vector and the second object feature vector according to the first decomposition matrix for the first object attribute and the second decomposition matrix for the second object attribute to obtain a fused object feature vector for the recommended object.

[0050] In a possible implementation, the electronic device may obtain a first fused feature matrix using the first decomposition matrix and the first object feature vector in the feature fusion layer, obtain a second fused feature matrix using the second decomposition matrix and the second object feature vector, and generate a fused object feature vector for the recommended object according to the first fused feature matrix and the second fused feature matrix. The number of elements in the fused object feature vector is the same as the number of the acquired feature fusion matrix.

[0051] In some embodiments, the electronic device may obtain the first fused feature matrix using the first decomposition matrix and the first object feature vector by transposing the first decomposition matrix to obtain the transposed matrix corresponding to the first decomposition matrix, transposing the first object feature vector to obtain the transposed vector corresponding to the first object feature vector, and taking the product of the transposed matrix corresponding to the first decomposition matrix and the transposed vector corresponding to the first object feature vector as the first fused feature matrix. And the electronic device may obtain the first fused feature matrix using the first decomposition matrix and the first object feature vector by transposing the second object feature vector to obtain the transposed vector corresponding to the second object feature vector, and taking the product of the second decomposition matrix and the transposed vector corresponding to the first object feature vector as the second fused feature matrix.

[0052] In some embodiments, the electronic device generates a fused object feature vector for a recommended object according to the first fused feature matrix and the second fused feature matrix by taking the product of the transposed matrix of the first fused feature matrix and the second fused feature matrix as fused feature elements, and obtaining the fused object feature vector according to the fused feature elements. That is:

[0053]

[0054] in, represents a first fusion feature matrix obtained based on the i-th feature fusion matrix in the feature fusion matrix acquired by the electronic device, represents the second fusion feature matrix obtained based on the i-th feature fusion matrix, logit i Represents the fused feature element obtained based on the i-th feature fusion matrix. Assume that there are N feature fusion matrices obtained, N is a positive integer of 1 or greater than 1, then i is a positive integer less than or equal to N, when N is 1, i is 1, when N is greater than 1, i is a positive integer from 1 to N. Therefore, a feature fusion matrix can obtain a fused feature element, and the electronic device can obtain a fusion object feature vector based on the fused feature element. When the number of feature fusion matrices is 1, the fused feature element is directly used as the fusion object feature vector, and when the number of feature fusion matrices is multiple, multiple fused feature elements are spliced ​​to obtain the fusion object feature vector.

[0055] S205: Generate a prediction recommendation index for the recommended object according to the recommendation model and the fused object feature vector.

[0056] In one possible implementation, the recommendation model may include a feature prediction layer, and the electronic device may predict the fused object feature vector in the feature prediction layer included in the recommendation model to generate a predicted recommendation index for the recommended object. Optionally, the feature prediction layer may be a fully connected layer, so the electronic device may input the fused object feature vector into the fully connected layer, and the fully connected layer maps the fused object feature vector and outputs a predicted recommendation index for the recommended object. Subsequently, the electronic device may sort the recommended objects according to the predicted recommendation indexes of the multiple recommended objects, and push them based on the sorted multiple recommended objects.

[0057] In the embodiment of the present application, the first object attribute and the second object attribute of the recommended object can be obtained, the first object attribute is input into the recommendation model to obtain the first object feature vector of the recommended object, and the second object attribute is input into the recommendation model to obtain the second object feature vector of the recommended object, and the feature fusion matrix in the recommendation model is obtained, and the feature fusion matrix is ​​decomposed to obtain the first decomposition matrix for the first object attribute and the second decomposition matrix for the second object attribute, and the first object feature vector and the second object feature vector are feature fused according to the first decomposition matrix for the first object attribute and the second decomposition matrix for the second object attribute to obtain the fused object feature vector for the recommended object, and the predicted recommendation index for the recommended object is generated according to the recommendation model and the fused object feature vector. By implementing the above-mentioned proposed method, by combining the feature fusion matrix to perform feature fusion on multiple object feature vectors, the interaction between features can be strengthened to improve the prediction accuracy of the recommendation index based on the fused object feature vector. In addition, when performing feature fusion, feature fusion can be achieved by decomposing the matrix for the feature fusion matrix, so that the amount of parameters used in feature fusion can be reduced, thereby reducing the complexity of the model at this stage and improving the prediction efficiency of the model.

[0058] See also Figure 3 , Figure 3 The following is a flow chart of a data processing method provided in an embodiment of the present application, which can be executed by the electronic device mentioned above. Figure 3 As shown, the process of the data processing method in the embodiment of the present application may include the following:

[0059] S301, obtaining a first object attribute and a second object attribute of a recommended object. The specific implementation of step S301 can refer to the relevant description of step S201, which will not be repeated here.

[0060] S302: Input the first object attribute into the recommendation model to obtain a first object feature vector of the recommended object, and input the second object attribute into the recommendation model to obtain a second object feature vector of the recommended object.

[0061] In one possible implementation, the feature extraction layer may include a neural network model, which may be used to encode object attributes to obtain corresponding object feature vectors. Therefore, the electronic device may input a first object attribute into the recommendation model, and the neural network model may encode the first object attribute to obtain a first object feature vector of the recommended object, and input a second object attribute into the recommendation model, and the neural network model may encode the second object attribute to obtain a second object feature vector of the recommended object.

[0062] Optionally, the neural network model for encoding the first object attribute and the neural network model for encoding the second object attribute may be the same neural network model or different neural network models. In some embodiments, when the object attribute is a text attribute, the neural network model may be used to semantically encode the text attribute to obtain a corresponding object feature vector, so the neural network model for encoding the text attribute may be a semantic model, such as a BERT (Bidirectional Encoder Representations from Transformer) model, an LSTM (Long Short-Time Memory) model, or a CNN (Convolutional Neural Networks) model. In some embodiments, when the object attribute is an image attribute, the neural network model may be used to encode the image attribute to obtain a corresponding object feature vector, so the neural network model for encoding the image attribute may be a ResNet (Residual Network) model, or other neural network models for encoding images. The specific type of the neural network model is not limited here. For example, if the recommended object is a video, the first object attribute may be the video title, and the second object attribute may be the video cover frame. Therefore, the neural network model for encoding the first object attribute may be a BERT model, and the neural network model for encoding the second object attribute may be a ResNet model.

[0063] Among them, a neural network model adapted to the object attribute type can be introduced into the recommendation model, that is, if the object attribute is a text attribute, the adapted neural network model can be a text semantic model (such as a BERT model), and if the object attribute is an image attribute, the adapted neural network model can be an image semantic model (such as a ResNet model), so that more features can be extracted from the object attributes, and it helps to improve the prediction accuracy when predicting the recommendation indicators later.

[0064] In one possible implementation, the recommendation model is obtained by training based on the first sample object attribute, the second sample object attribute and the sample recommendation index for the sample object. Therefore, the specific method for the electronic device to train the initial recommendation model to obtain the recommendation model may be: input the first sample object attribute of the sample object into the initial recommendation model to obtain the first sample object feature vector of the sample object, input the second sample object attribute of the sample object into the initial recommendation model to obtain the second sample object feature vector of the sample object, obtain the initial feature fusion matrix in the initial recommendation model, decompose the initial feature fusion matrix to obtain the first initial decomposition matrix for the first sample object attribute and the second initial decomposition matrix for the second sample object attribute, perform feature fusion on the first sample object feature vector and the second sample object feature vector according to the first initial decomposition matrix for the first sample object attribute and the second initial decomposition matrix for the second sample object attribute, obtain the fused sample object feature vector for the sample object, generate the sample prediction recommendation index for the sample object according to the initial recommendation model and the fused sample object feature vector, update the model parameters of the initial recommendation model using the sample recommendation index and the sample prediction recommendation index of the sample object, and obtain the recommendation model. The model parameters of the initial recommendation model may include an initial feature fusion matrix, so the recommendation model may include a feature fusion matrix obtained after updating the parameters of the initial feature fusion matrix. Optionally, the initial feature fusion matrix may be one or more, and a feature fusion matrix is ​​obtained after updating the parameters of an initial feature fusion matrix.

[0065] In some embodiments, the electronic device updates the model parameters of the initial recommendation model using the sample recommendation index and the sample prediction recommendation index of the sample object, and obtains the recommendation model by constructing a model loss function, and obtains the loss function value according to the model loss function, the sample recommendation index and the sample prediction recommendation index, and updates the model parameters of the initial recommendation model according to the loss function value until the model converges. In addition, the model training adopts a supervised learning training method, and the recommendation model can be trained by fitting the sample recommendation index of the sample object, without the need to additionally label the sample object and other related data, so it has flexibility and versatility, and when the first object attribute and / or the second object attribute is a text attribute, there is no restriction on the content size and context of the text attribute, so it has strong generalization ability and can be applied to a variety of prediction scenarios with text attributes.

[0066] Optionally, the electronic device may construct a model loss function based on regression, and the purpose of model training is to fit the sample recommendation index of the sample recommendation object. The model loss function may be a mean square error loss function (MSE, Mean Square Error). That is, the model loss function may be:

[0067]

[0068] Among them, n is the number of sample objects, y k is the sample prediction recommendation index for sample object k, y′ k is the sample recommendation index for sample object k.

[0069] Optionally, the model loss function constructed by the electronic device may also be a loss function in other regression models, such as a mean absolute value loss function (MAE, Mean Absolute Error) or a smoothed mean absolute error loss function (Huber).

[0070] For example, taking the recommendation object as a video document, the training process of the sample recommendation model can be:

[0071] (1) Construct a training data set, which may include training data of multiple sample recommendation objects (sample video documents), specifically, a sample video document, and sample title attributes obtained from the sample video document and key frame (such as cover frame) attributes extracted from the sample video document, the title attribute is used as the first object attribute, the cover frame attribute is used as the second object attribute, and includes a sample click-through rate for the sample video document, the sample click-through rate is the real click-through rate of the sample video document, as the training target value. Therefore, for a sample video document, the corresponding training data may be as follows:

[0072] Sample video document X = {"title attribute":"XXXX","cover frame attribute":"XXXX","sample click rate":0.75};

[0073] (2) Inputting the title attribute and the cover frame attribute of the sample video document into the sample recommendation model for feature extraction, and obtaining a first object feature vector corresponding to the title attribute and a second object feature vector corresponding to the cover frame attribute; wherein the feature extraction of the title attribute in the sample recommendation model may be to obtain a first object feature vector (f ) representing the semantic features of the title attribute by semantically encoding the title attribute through a BERT model. text ), and extracting features from the cover frame attributes in the sample recommendation model can be performed by encoding the cover frame attributes through a ResNet model to obtain a second object feature vector (f img ), that is, the formula representation can be as follows:

[0074] f text =BERT(text)

[0075] f img =ResNet(img)

[0076] Among them, text represents the title attribute, and img represents the video frame attribute;

[0077] Optionally, the BERT model and the ResNet model can be pre-trained models and deployed in the sample recommendation model, or can be models to be trained and trained together with the recommendation model to obtain trained models;

[0078] (3) using the feature fusion layer of the sample recommendation model to perform feature fusion on the first object feature vector and the second object feature vector to obtain a fused sample object feature vector, i.e., a multimodal feature vector;

[0079] (4) The sample predicted click rate for the sample video document is obtained according to the sample recommendation model and the fused sample object feature vector. The sample recommendation model is trained by the sample predicted click rate and the sample click rate to obtain a recommendation model. The purpose of the model is to fit the sample predicted click rate of the sample video document by regression. During model training, the implicit relationship between the fused semantic features of the title semantic features and the cover frame semantic features and the click rate can be modeled by regression, which can effectively capture the relationship between the fused semantic features and the click rate.

[0080] S303: Obtain a feature fusion matrix in the recommendation model.

[0081] Among them, the feature fusion matrix can be one or more, and the feature fusion matrix can be expressed as W [1:N] ∈R n*m , that is, there can be N feature fusion matrices, N is a positive integer, each feature fusion matrix is ​​a matrix with n rows and m columns, and the parameters of each feature fusion matrix can be the same or different. Optionally, the parameters of each feature fusion matrix can be obtained by training the sample recommendation model as described above, or can be fixed parameters set by relevant business personnel based on experience.

[0082] S304: If the number of acquired feature fusion matrices is N, decompose the N feature fusion matrices respectively to obtain a first decomposition matrix and a second decomposition matrix corresponding to each feature fusion matrix.

[0083] In a possible implementation, when the electronic device obtains N feature fusion matrices, the N feature fusion matrices are decomposed respectively to obtain a first decomposition matrix and a second decomposition matrix corresponding to each feature fusion matrix in the N feature fusion matrices; wherein one feature fusion matrix corresponds to one first decomposition matrix and one second decomposition matrix, the first decomposition matrix is ​​a matrix for the first object attribute, and the second decomposition matrix is ​​a matrix for the second object attribute. Therefore, the number of feature fusion matrices, the number of first decomposition matrices, and the number of second decomposition matrices are the same.

[0084] In some embodiments, the process and principle of decomposing each feature fusion matrix are the same, so any feature fusion matrix among the N feature fusion matrices is represented as the i-th feature fusion matrix, i is a positive integer less than or equal to N; the i-th feature fusion matrix is ​​decomposed so that the product of the first decomposition matrix and the second decomposition matrix corresponding to the i-th feature fusion matrix is ​​equal to the i-th feature fusion matrix, and the number of rows of the first decomposition matrix corresponding to the i-th feature fusion matrix is ​​equal to the number of elements in the first object feature vector; the number of columns of the second decomposition matrix corresponding to the i-th feature fusion matrix is ​​equal to the number of elements in the second object feature vector. That is, let the i-th feature fusion matrix be W i , the first decomposition matrix obtained by decomposing the i-th feature fusion matrix is ​​U i and the second decomposition matrix V i , so the relationship between the above three matrices is:

[0085] W i =U i *V i i=[1,2,...,N]

[0086] Among them, U i ∈R(n*k), that is, the first decomposition matrix is ​​a matrix with n rows and k columns, V i ∈R(k*m), that is, the second decomposition matrix is ​​a matrix with k rows and n columns; k can be set by relevant business personnel based on experience.

[0087] S305 . Perform feature fusion on the first object feature vector and the second object feature vector according to the first decomposition matrix and the second decomposition matrix corresponding to each feature fusion matrix to obtain a fused object feature vector for the recommended object.

[0088] In one possible implementation, the electronic device may perform feature fusion on the first object feature vector and the second object feature vector according to the first decomposition matrix and the second decomposition matrix corresponding to each feature fusion matrix, respectively, to obtain fusion feature elements corresponding to each feature fusion feature, and concatenate the fusion feature elements corresponding to each feature fusion matrix to obtain a fusion object feature vector. The number of fusion feature elements in the fusion object feature vector is the same as the number of feature fusion matrices, that is, when the number of feature fusion matrices is one, the fusion object feature vector is a vector composed of one fusion feature element; when the number of feature fusion matrices is N, the fusion object feature vector is a vector composed of N concatenated fusion feature elements.

[0089] In one possible implementation, the electronic device performs feature fusion on the first object feature vector and the second object feature vector respectively according to the first decomposition matrix and the second decomposition matrix corresponding to each feature fusion matrix, and the process and principle of obtaining the fused feature elements corresponding to each feature fusion matrix are the same. Therefore, taking the i-th feature fusion matrix as an example, the specific method of obtaining the fused feature elements corresponding to the i-th feature fusion matrix can be as follows: a first fused feature matrix is ​​generated according to the first decomposition matrix and the first object feature vector corresponding to the i-th feature fusion matrix, a second fused feature matrix is ​​generated according to the second decomposition matrix and the second object feature vector corresponding to the i-th feature fusion matrix, and a fused feature element corresponding to the i-th feature fusion matrix is ​​generated according to the first fused feature matrix and the second fused feature matrix.

[0090] In some embodiments, the electronic device generates the first fused feature matrix according to the first decomposition matrix corresponding to the i-th feature fusion matrix and the first object feature vector, which can be specifically: the transpose matrix (U i T ) and the transposed vector of the first object feature vector (f 1 T ) as the first fusion feature matrix, that is, the first fusion feature matrix is ​​U i T *f 1 T The electronic device generates a second fused feature matrix according to the second decomposition matrix and the second object feature vector corresponding to the i-th feature fusion matrix, specifically, by converting the second decomposition matrix (V i ) and the transposed vector of the second object feature vector (f 2 T ) as the second fusion feature matrix, that is, the first fusion feature matrix is ​​V i *f 2 T .

[0091] In some embodiments, the electronic device generates the fusion feature element corresponding to the i-th feature fusion matrix according to the first fusion feature matrix and the second fusion feature matrix, specifically, by taking the product of the transposed matrix of the first fusion feature matrix and the second fusion feature matrix as the corresponding fusion feature element, that is, the fusion feature element is:

[0092] logit i =(U i T *f 1 T ) T *(V i *f 2 T )i=[1,2,...,N]

[0093] Therefore, the fusion object feature vector is the vector concatenated by the fusion feature elements corresponding to each feature fusion matrix in the N feature fusion matrices. Therefore, the fusion object feature matrix can be logit = [logit 1 ,logit 2 ,...,logit i ,...,logit N ]. Optionally, the fusion object feature vector can also be the sum vector of the vector concatenated by the fusion feature elements corresponding to each feature fusion matrix and the bias vector, that is, logit = [logit 1 ,logit 2 ,...,logit i ,...,logit N ]+b. Wherein, b is a bias vector, which can be a vector set by relevant business personnel based on experience, or a vector obtained through training as a model parameter in a sample recommendation model.

[0094] For example, Figure 4 As shown, Figure 4 A schematic diagram of a scenario for determining a fusion object feature vector provided in an embodiment of the present application; wherein, the electronic device obtains N feature fusion matrices, and decomposes each feature fusion matrix into a corresponding first decomposition matrix and a second decomposition matrix, respectively, generates a first fusion feature matrix corresponding to each feature fusion matrix according to the first decomposition matrix and the first object feature vector corresponding to each feature fusion matrix, generates a second fusion feature matrix corresponding to each feature fusion matrix according to the second decomposition matrix and the second object feature vector corresponding to each feature fusion matrix, generates a fusion feature element corresponding to each feature fusion matrix according to the first fusion feature matrix and the second fusion feature matrix corresponding to each feature fusion matrix, and splices them to obtain a fusion object feature vector.

[0095] S306: Generate a prediction recommendation index for the recommended object according to the recommendation model and the fused object feature vector.

[0096] In a possible implementation, the electronic device may input the fused object feature vector into a fully connected layer in the recommendation model, and generate a predicted recommendation index for the recommended object according to the fused object feature vector via the fully connected layer.

[0097] Optionally, when the recommended object is graphic data, the first object data may be a text attribute of the graphic data, the second object attribute may be an image attribute of the graphic data, and the predicted recommendation indicator for the recommended object may be a predicted click-through rate for the graphic data. In some embodiments, if there are multiple graphic data, one graphic data corresponds to one predicted click-through rate, and the electronic device may sort the multiple graphic data according to the predicted click-through rate corresponding to each graphic data in the multiple graphic data, and send the sorted multiple graphic data to the recommendation client, so that the recommendation client outputs the sorted multiple graphic data, thereby achieving accurate push of the graphic data.

[0098] For example, Figure 5a-Figure 5b A schematic diagram of an application scenario provided in an embodiment of the present application, wherein a video search scenario applied to a mini program is taken as an example. Figure 5a In the process of rough sorting and / or fine sorting of video documents, the predicted click-through rate of video documents is a very important feature. Therefore, by obtaining the title attributes and cover frame attributes of the video documents to be sorted and inputting them into the recommendation model, the predicted click-through rate of each video document is generated. In the stage of sorting video documents, the predicted click-through rate can be combined to improve the accuracy of push and / or the quality of video documents. And optimize the user's search experience. In addition, when the recommendation model is combined with the title attribute and cover attribute of the video document for prediction, the predicted click-through rate of the video document with a cover and / or title that is more attractive to the user is higher and more accurate. According to the implicit relationship between the title attribute and the cover attribute and the click-through rate, the words with strong correlation between the title attribute and the click-through rate of the video document to be sorted can be captured, so that the predicted click-through rate of the video document with strong correlation can be higher. At the same time, the predicted click-through rate of the video document with inconsistent title and cover can be reduced to reduce the appearance of low-quality video documents (such as video documents with predicted click-through rate lower than the click-through rate threshold); Figure 5b There are multiple video files in the test. Figure 5bThe second and fourth video documents in the recommendation model will be more attractive to users to click on them, so the predicted click-through rate generated by the recommendation model will be higher, so that the two high-quality video documents (such as video documents with a predicted click-through rate equal to or higher than the click-through rate threshold) are more likely to be exposed; in addition, in this scenario, video documents with low exposure times are prone to be in the cold start stage, so high-quality video documents with low exposure times may be ranked low in the rough ranking / fine ranking stage and have few exposure opportunities. Therefore, the predicted click-through rate of high-quality video documents with low exposure is improved during prediction, and corresponding strategies (such as improving the weight adaptability of video documents with low exposure but high predicted click-through rate, etc.) are adjusted in the rough ranking / fine ranking stage to make the low-exposure high-quality video documents ranked as high as possible, thereby improving the problem that low-exposure high-quality video documents are often in the cold start stage.

[0099] For example, Figure 6 As shown, Figure 6 A schematic diagram of a scenario for determining a predictive recommendation indicator provided in an embodiment of the present application; wherein the recommended object is a video document, the electronic device obtains the title information of the video document as a title attribute, and extracts the cover frame (key frame) information from the video document as a cover frame attribute, the title attribute is input into the recommendation model, a first object feature vector is obtained through the BERT model of the feature extraction layer of the recommendation model, and the cover frame attribute is input into the recommendation model, a second object feature vector is obtained through the ResNet model of the feature extraction layer of the recommendation model, the first object feature vector and the second object feature vector are feature fused in the feature fusion layer to obtain a fused feature vector, and the fused feature vector is input into the feature prediction layer of the recommendation model, the fused feature vector is predicted through the fully connected layer in the feature prediction layer to obtain a predicted click-through rate for the video document.

[0100] In the embodiment of the present application, the first object attribute and the second object attribute of the recommended object can be obtained, the first object attribute is input into the recommendation model, the first object feature vector of the recommended object is obtained, and the second object attribute is input into the recommendation model, the second object feature vector of the recommended object is obtained, and the feature fusion matrix in the recommendation model is obtained. If the number of feature fusion matrices obtained is N, the N feature fusion matrices are decomposed respectively to obtain the first decomposition matrix for the first object attribute and the second decomposition matrix for the second object attribute. According to the first decomposition matrix for the first object attribute and the second decomposition matrix for the second object attribute, the first object feature vector and the second object feature vector are feature fused to obtain the fused object feature vector for the recommended object, and the predicted recommendation index for the recommended object is generated according to the recommendation model and the fused object feature vector. By implementing the above-mentioned method, by combining the feature fusion matrix to perform feature fusion on multiple object feature vectors, the interaction between features can be strengthened to improve the prediction accuracy of the recommendation index based on the fused object feature vector. In addition, when performing feature fusion, feature fusion can be achieved by decomposing the feature fusion matrix, so that the amount of parameters used in feature fusion can be reduced, thereby reducing the complexity of the model at this stage and improving the prediction efficiency of the model.

[0101] See also Figure 7 , Figure 7 This is a schematic diagram of the structure of a data processing device provided in this application. It should be noted that: Figure 7 The data processing device shown is used to execute the application Figure 2 and Figure 3 For the convenience of explanation, only the part related to the embodiment of the present application is shown, and the specific technical details are not disclosed. Figure 2 and Figure 3 The data processing device 700 may include: an acquisition module 701, an input module 702, a decomposition module 703, a fusion module 704, and a generation module 705. Among them:

[0102] An acquisition module 701 is used to acquire a first object attribute and a second object attribute of a recommended object;

[0103] An input module 702 is used to input the first object attribute into the recommendation model to obtain a first object feature vector of the recommended object, and input the second object attribute into the recommendation model to obtain a second object feature vector of the recommended object; the recommendation model is trained based on the first sample object attribute, the second sample object attribute of the sample object, and the sample recommendation index for the sample object;

[0104] A decomposition module 703 is used to obtain a feature fusion matrix in the recommendation model, and decompose the feature fusion matrix to obtain a first decomposition matrix for the first object attribute and a second decomposition matrix for the second object attribute;

[0105] A fusion module 704 is used to perform feature fusion on the first object feature vector and the second object feature vector according to the first decomposition matrix for the first object attribute and the second decomposition matrix for the second object attribute to obtain a fused object feature vector for the recommended object;

[0106] The generating module 705 is used to generate a prediction recommendation index for the recommended object according to the recommendation model and the fused object feature vector.

[0107] In a possible implementation, the number of the feature fusion matrices is N, where N is a positive integer;

[0108] When the decomposition module 703 is used to decompose the feature fusion matrix to obtain a first decomposition matrix for the first object attribute and a second decomposition matrix for the second object attribute, it is specifically used to:

[0109] The N feature fusion matrices are decomposed respectively to obtain a first decomposition matrix and a second decomposition matrix corresponding to each feature fusion matrix in the N feature fusion matrices; wherein one feature fusion matrix corresponds to one first decomposition matrix and one second decomposition matrix.

[0110] In a possible implementation, any one of the N feature fusion matrices is represented as the i-th feature fusion matrix, where i is a positive integer less than or equal to N; the product of the first decomposition matrix and the second decomposition matrix corresponding to the i-th feature fusion matrix is ​​equal to the i-th feature fusion matrix; the number of rows of the first decomposition matrix corresponding to the i-th feature fusion matrix is ​​equal to the number of elements in the first object feature vector; the number of columns of the second decomposition matrix corresponding to the i-th feature fusion matrix is ​​equal to the number of elements in the second object feature vector.

[0111] In a possible implementation, the fusion module 704 is used to perform feature fusion on the first object feature vector and the second object feature vector according to the first decomposition matrix for the first object attribute and the second decomposition matrix for the second object attribute to obtain a fused object feature vector for the recommended object, specifically for:

[0112] According to the first decomposition matrix and the second decomposition matrix corresponding to each feature fusion matrix, feature fusion is performed on the first object feature vector and the second object feature vector respectively to obtain fusion feature elements corresponding to each feature fusion matrix;

[0113] The fused feature elements corresponding to each feature fusion matrix are concatenated to obtain the fused object feature vector.

[0114] In a possible implementation manner, any one of the N feature fusion matrices is represented as an i-th feature fusion matrix, where i is a positive integer less than or equal to N;

[0115] The fusion module 704 is used to perform feature fusion on the first object feature vector and the second object feature vector according to the first decomposition matrix and the second decomposition matrix corresponding to each feature fusion matrix, respectively, to obtain the fusion feature element corresponding to each feature fusion matrix, specifically for:

[0116] Generate a first fused feature matrix according to the first decomposition matrix corresponding to the i-th feature fusion matrix and the first object feature vector;

[0117] Generate a second fused feature matrix according to the second decomposition matrix and the second object feature vector corresponding to the i-th feature fusion matrix;

[0118] A fused feature element corresponding to the i-th feature fusion matrix is ​​generated according to the first fused feature matrix and the second fused feature matrix.

[0119] In a possible implementation, the generating module 705 is further configured to:

[0120] Inputting a first sample object attribute of the sample object into the initial recommendation model to obtain a first sample object feature vector of the sample object, and inputting a second sample object attribute of the sample object into the initial recommendation model to obtain a second sample object feature vector of the sample object;

[0121] Acquire an initial feature fusion matrix in the initial recommendation model, and decompose the initial feature fusion matrix to obtain a first initial decomposition matrix for the first sample object attribute and a second initial decomposition matrix for the second sample object attribute;

[0122] According to a first initial decomposition matrix for the first sample object attribute and a second initial decomposition matrix for the second sample object attribute, feature fusion is performed on the first sample object feature vector and the second sample object feature vector to obtain a fused sample object feature vector for the sample object;

[0123] Generate a sample prediction recommendation index for the sample object based on the initial recommendation model and the fused sample object feature vector;

[0124] The model parameters of the initial recommendation model are updated using the sample recommendation index and the sample prediction recommendation index of the sample object to obtain a recommendation model; the recommendation model includes a feature fusion matrix obtained after updating the parameters of the initial feature fusion matrix.

[0125] In a possible implementation, the recommendation object is graphic data, the first object attribute is the text attribute of the graphic data, the second object attribute is the image attribute of the graphic data, and the predicted recommendation index is the predicted click rate for the graphic data.

[0126] In a possible implementation manner, the number of the above-mentioned graphic and text data is multiple, and one graphic and text data corresponds to one predicted click rate;

[0127] The generation module 705 is also used for:

[0128] sorting the plurality of graphic and text data according to the predicted click-through rate corresponding to each of the plurality of graphic and text data;

[0129] The sorted plurality of graphic and text data are sent to the recommendation client, so that the recommendation client outputs the sorted plurality of graphic and text data.

[0130] In an embodiment of the present application, an acquisition module acquires a first object attribute and a second object attribute of a recommended object; an input module inputs the first object attribute into a recommendation model to obtain a first object feature vector of the recommended object, and inputs the second object attribute into the recommendation model to obtain a second object feature vector of the recommended object; the recommendation model is trained based on the first sample object attribute, the second sample object attribute of the sample object, and a sample recommendation index for the sample object; a decomposition module acquires a feature fusion matrix in the recommendation model, and decomposes the feature fusion matrix to obtain a first decomposition matrix for the first object attribute and a second decomposition matrix for the second object attribute; a fusion module performs feature fusion on the first object feature vector and the second object feature vector according to the first decomposition matrix for the first object attribute and the second decomposition matrix for the second object attribute to obtain a fused object feature vector for the recommended object; and a generation module generates a predicted recommendation index for the recommended object according to the recommendation model and the fused object feature vector. By implementing the device proposed above, feature fusion is performed on feature vectors of multiple objects in combination with a feature fusion matrix, so that the interaction between features can be strengthened to improve the prediction accuracy of recommendation indicators based on the fused object feature vectors. In addition, when performing feature fusion, feature fusion can be achieved by decomposing the feature fusion matrix to reduce the amount of parameters used in feature fusion, thereby reducing the complexity of the model at this stage and improving the prediction efficiency of the model.

[0131] Each functional module in each embodiment of the present application can be integrated into one module, or each module can exist physically separately, or two or more modules can be integrated into one module. The above integrated modules can be implemented in the form of hardware or software functional modules, which is not limited in the present application.

[0132] See also Figure 8 , Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 8 As shown, the electronic device 800 includes: at least one processor 801 and a memory 802. Optionally, the electronic device may further include a network interface. The processor 801, the memory 802 and the network interface may exchange data, the network interface is controlled by the processor 801 to send and receive messages, the memory 802 is used to store a computer program, the computer program includes program instructions, and the processor 801 is used to execute the program instructions stored in the memory 802. The processor 801 is configured to call the program instructions to execute the above method.

[0133] Among them, the memory 802 may include a volatile memory (volatile memory), such as a random-access memory (RAM); the memory 802 may also include a non-volatile memory (non-volatile memory), such as a flash memory (flash memory), a solid-state drive (SSD), etc.; the memory 802 may also include a combination of the above types of memory.

[0134] The processor 801 may be a central processing unit (CPU). In one embodiment, the processor 801 may also be a graphics processing unit (GPU). The processor 801 may also be a combination of a CPU and a GPU.

[0135] In a possible implementation, the memory 802 is used to store program instructions, and the processor 801 may call the program instructions to perform the following steps:

[0136] Obtain a first object attribute and a second object attribute of a recommended object;

[0137] Inputting the first object attribute into the recommendation model to obtain a first object feature vector of the recommended object, and inputting the second object attribute into the recommendation model to obtain a second object feature vector of the recommended object; the recommendation model is trained based on the first sample object attribute, the second sample object attribute and the sample recommendation index for the sample object;

[0138] Acquire a feature fusion matrix in the recommendation model, and decompose the feature fusion matrix to obtain a first decomposition matrix for the first object attribute and a second decomposition matrix for the second object attribute;

[0139] According to a first decomposition matrix for the first object attribute and a second decomposition matrix for the second object attribute, feature fusion is performed on the first object feature vector and the second object feature vector to obtain a fused object feature vector for the recommended object;

[0140] Generate a prediction recommendation index for the recommended object based on the recommendation model and the fused object feature vector.

[0141] In a possible implementation, the number of the feature fusion matrices is N, where N is a positive integer;

[0142] When the processor 801 is used to decompose the feature fusion matrix to obtain a first decomposition matrix for the first object attribute and a second decomposition matrix for the second object attribute, it is specifically used to:

[0143] The N feature fusion matrices are decomposed respectively to obtain a first decomposition matrix and a second decomposition matrix corresponding to each feature fusion matrix in the N feature fusion matrices; wherein one feature fusion matrix corresponds to one first decomposition matrix and one second decomposition matrix.

[0144] In a possible implementation, any one of the N feature fusion matrices is represented as the i-th feature fusion matrix, where i is a positive integer less than or equal to N; the product of the first decomposition matrix and the second decomposition matrix corresponding to the i-th feature fusion matrix is ​​equal to the i-th feature fusion matrix; the number of rows of the first decomposition matrix corresponding to the i-th feature fusion matrix is ​​equal to the number of elements in the first object feature vector; the number of columns of the second decomposition matrix corresponding to the i-th feature fusion matrix is ​​equal to the number of elements in the second object feature vector.

[0145] In a possible implementation manner, when the processor 801 is used to perform feature fusion on the first object feature vector and the second object feature vector according to the first decomposition matrix for the first object attribute and the second decomposition matrix for the second object attribute to obtain the fused object feature vector for the recommended object, it is specifically used to:

[0146] According to the first decomposition matrix and the second decomposition matrix corresponding to each feature fusion matrix, feature fusion is performed on the first object feature vector and the second object feature vector respectively to obtain fusion feature elements corresponding to each feature fusion matrix;

[0147] The fused feature elements corresponding to each feature fusion matrix are concatenated to obtain the fused object feature vector.

[0148] In a possible implementation manner, any one of the N feature fusion matrices is represented as an i-th feature fusion matrix, where i is a positive integer less than or equal to N;

[0149] When the processor 801 is used to perform feature fusion on the first object feature vector and the second object feature vector respectively according to the first decomposition matrix and the second decomposition matrix corresponding to each feature fusion matrix to obtain the fusion feature element corresponding to each feature fusion matrix, it is specifically used to:

[0150] Generate a first fused feature matrix according to the first decomposition matrix corresponding to the i-th feature fusion matrix and the first object feature vector;

[0151] Generate a second fused feature matrix according to the second decomposition matrix and the second object feature vector corresponding to the i-th feature fusion matrix;

[0152] A fused feature element corresponding to the i-th feature fusion matrix is ​​generated according to the first fused feature matrix and the second fused feature matrix.

[0153] In a possible implementation, the processor 801 is further configured to:

[0154] Inputting a first sample object attribute of the sample object into the initial recommendation model to obtain a first sample object feature vector of the sample object, and inputting a second sample object attribute of the sample object into the initial recommendation model to obtain a second sample object feature vector of the sample object;

[0155] Acquire an initial feature fusion matrix in the initial recommendation model, and decompose the initial feature fusion matrix to obtain a first initial decomposition matrix for the first sample object attribute and a second initial decomposition matrix for the second sample object attribute;

[0156] According to a first initial decomposition matrix for the first sample object attribute and a second initial decomposition matrix for the second sample object attribute, feature fusion is performed on the first sample object feature vector and the second sample object feature vector to obtain a fused sample object feature vector for the sample object;

[0157] Generate a sample prediction recommendation index for the sample object based on the initial recommendation model and the fused sample object feature vector;

[0158] The model parameters of the initial recommendation model are updated using the sample recommendation index and the sample prediction recommendation index of the sample object to obtain a recommendation model; the recommendation model includes a feature fusion matrix obtained after updating the parameters of the initial feature fusion matrix.

[0159] In a possible implementation, the recommendation object is graphic data, the first object attribute is the text attribute of the graphic data, the second object attribute is the image attribute of the graphic data, and the predicted recommendation index is the predicted click rate for the graphic data.

[0160] In a possible implementation manner, the number of the above-mentioned graphic and text data is multiple, and one graphic and text data corresponds to one predicted click rate;

[0161] The processor 801 is further configured to:

[0162] sorting the plurality of graphic and text data according to the predicted click-through rate corresponding to each of the plurality of graphic and text data;

[0163] The sorted plurality of graphic and text data are sent to the recommendation client, so that the recommendation client outputs the sorted plurality of graphic and text data.

[0164] In a specific implementation, the above-described device, processor 801, memory 802, etc. can execute the implementation method described in the above-mentioned method embodiment, and can also execute the implementation method described in the embodiment of the present application, which will not be repeated here.

[0165] A computer (readable) storage medium is also provided in an embodiment of the present application, and the computer storage medium stores a computer program, and the computer program includes program instructions. When the program instructions are executed by the processor, the processor can execute some or all of the steps performed in the above method embodiment. Optionally, the computer storage medium can be volatile or non-volatile. The computer-readable storage medium can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function, etc.; the data storage area can store data created according to the use of the blockchain node, etc.

[0166] The term "multiple" as used herein refers to two or more than two. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the related objects are in an "or" relationship.

[0167] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the above-mentioned program can be stored in a computer storage medium, which can be a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the above-mentioned storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0168] The above disclosure is only part of the embodiments of the present application, which certainly cannot be used to limit the scope of rights of the present application. Ordinary technicians in this field can understand that all or part of the processes of implementing the above embodiments and 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: The method comprises: Acquire a first object attribute and a second object attribute of a recommended object; the first object attribute and the second object attribute are object attributes of different types; Inputting the first object attribute into a recommendation model to obtain a first object feature vector of the recommended object, and inputting the second object attribute into the recommendation model to obtain a second object feature vector of the recommended object; the recommendation model is trained based on the first sample object attribute and the second sample object attribute of the sample object and a sample recommendation index for the sample object; Acquire a feature fusion matrix in the recommendation model, and decompose the feature fusion matrix to obtain a first decomposition matrix for the first object attribute and a second decomposition matrix for the second object attribute; the dimension of the first decomposition matrix corresponds to the dimension of the first object attribute, and the dimension of the second decomposition matrix corresponds to the dimension of the second object attribute; Using the first decomposition matrix to perform a fusion operation on the first object feature vector to generate a first fused feature matrix, using the second decomposition matrix to perform a fusion operation on the second object feature vector to generate a second fused feature matrix, and performing a fusion operation on the first fused feature matrix and the second fused feature matrix to generate a fused object feature vector of the recommended object; A prediction recommendation index for the recommended object is generated according to the recommendation model and the fused object feature vector.

2. The method according to claim 1, characterized in that The number of the feature fusion matrices is N, where N is a positive integer; Decomposing the feature fusion matrix to obtain a first decomposition matrix for the first object attribute and a second decomposition matrix for the second object attribute includes: Decomposing N feature fusion matrices respectively to obtain a first decomposition matrix and a second decomposition matrix corresponding to each feature fusion matrix in the N feature fusion matrices; wherein one feature fusion matrix corresponds to one first decomposition matrix and one second decomposition matrix.

3. The method according to claim 2, characterized in that Any one of the N feature fusion matrices is represented as the i-th feature fusion matrix, where i is a positive integer less than or equal to N; the product of the first decomposition matrix and the second decomposition matrix corresponding to the i-th feature fusion matrix is ​​equal to the i-th feature fusion matrix; The number of rows of the first decomposition matrix corresponding to the i-th feature fusion matrix is ​​equal to the number of elements in the first object feature vector; the number of columns of the second decomposition matrix corresponding to the i-th feature fusion matrix is ​​equal to the number of elements in the second object feature vector.

4. The method according to claim 2, characterized in that: The step of using the first decomposition matrix to perform feature fusion on the first object feature vector to generate a first fused feature matrix, using the second decomposition matrix to perform feature fusion on the second object feature vector to generate a second fused feature matrix, and combining the first fused feature matrix and the second fused feature matrix to generate a fused object feature vector for the recommended object includes: Using the first decomposition matrix corresponding to each feature fusion matrix, perform feature fusion processing on the first object feature vector to generate a first fusion feature matrix corresponding to each feature fusion matrix, and using the second decomposition matrix corresponding to each feature fusion matrix to perform feature fusion processing on the second object feature vector to generate a second fusion feature matrix corresponding to each feature fusion matrix; Combining the first fused feature matrix and the second fused feature matrix corresponding to each feature fusion matrix respectively to obtain fused feature elements corresponding to each feature fusion matrix; The fusion feature elements corresponding to each feature fusion matrix are concatenated to obtain the fusion object feature vector.

5. The method according to claim 1, characterized in that The method further comprises: Inputting the first sample object attribute of the sample object into an initial recommendation model to obtain a first sample object feature vector of the sample object, and inputting the second sample object attribute of the sample object into the initial recommendation model to obtain a second sample object feature vector of the sample object; Acquire an initial feature fusion matrix in the initial recommendation model, and decompose the initial feature fusion matrix to obtain a first initial decomposition matrix for the first sample object attribute and a second initial decomposition matrix for the second sample object attribute; According to a first initial decomposition matrix for the first sample object attribute and a second initial decomposition matrix for the second sample object attribute, performing feature fusion on the first sample object feature vector and the second sample object feature vector to obtain a fused sample object feature vector for the sample object; Generating a sample prediction recommendation index for the sample object according to the initial recommendation model and the fused sample object feature vector; The model parameters of the initial recommendation model are updated using the sample recommendation index and the sample prediction recommendation index of the sample object to obtain the recommendation model; the recommendation model includes the feature fusion matrix obtained after the parameters of the initial feature fusion matrix are updated.

6. The method according to claim 1, characterized in that The recommendation object is graphic data, the first object attribute is a text attribute of the graphic data, the second object attribute is an image attribute of the graphic data, and the predicted recommendation index is a predicted click rate for the graphic data.

7. The method according to claim 6, characterized in that There are multiple graphic data, and each graphic data corresponds to one predicted click rate; The method further comprises: Sorting the plurality of graphic and text data according to the predicted click-through rate corresponding to each of the plurality of graphic and text data; The sorted plurality of graphic and text data are sent to a recommendation client, so that the recommendation client outputs the sorted plurality of graphic and text data.

8. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory is used to store a computer program, the computer program comprises program instructions, and the processor is configured to call the program instructions to execute the method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor is caused to perform the method according to any one of claims 1 to 7.

Images