Information recommendation method and device, electronic device and storage medium

Abstract

The disclosed embodiment provides an information recommendation method and device, an electronic device and a storage medium, wherein the information recommendation method comprises: encoding the user information of the target user to obtain a first feature vector, and encoding the content information of the recommended content to obtain a second feature vector; determining a relationship matrix; using a first recommendation model, obtaining a first output vector according to the first feature vector, the second feature vector and the relationship matrix; using a second recommendation model, obtaining a second output vector according to the first feature vector and the second feature vector; obtaining a first prediction result according to the first output vector and the second output vector. In this way, the interaction information between features of different categories is represented by the relationship matrix, so that the interaction information between features can be learned in a more detailed manner, the mining of the feature interaction information can be more comprehensive, and the result of the recommended information finally determined can be more accurate.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to an information recommendation method and device, an electronic device, and a computer-readable storage medium. Background Art

[0002] The Internet brings a large amount of information to users, satisfying their demand for information in the information age. However, it also makes it impossible for users to quickly obtain truly useful information from a large amount of information. Therefore, the efficiency of information use is reduced.

[0003] Recommendation systems can effectively help users quickly find interesting and high-quality information and improve user experience. In recommendation systems, recommendation algorithms are the core of recommendation systems. The key to recommendation algorithms is to mine data features and information between features, especially implicit information between features. However, in current related technologies, traditional recommendation algorithms such as those based on collaborative filtering or logistic regression, or deep learning sorting algorithms, do not fully consider the interactive information between features, especially ignoring the differences in interactive information between features of different categories, resulting in the determination of recommended information that cannot accurately reflect user needs, affecting user experience. Summary of the invention

[0004] In view of this, the present disclosure provides an information recommendation method and device, an electronic device, and a storage medium.

[0005] In a first aspect, the present application provides an information recommendation method, which may include:

[0006] Encoding the user information of the target user to obtain a first feature vector, and encoding the content information of the content to be recommended to obtain a second feature vector;

[0007] Determine a relationship matrix, wherein the relationship matrix includes: a first type of matrix and a second type of matrix, wherein the first type of matrix represents the relationship between the determined user information and the content to be recommended; and the second type of matrix represents the relationship between different types of content to be recommended;

[0008] Using a pre-trained first recommendation model, obtaining a first output vector according to the first eigenvector, the second eigenvector and the relationship matrix;

[0009] Using a pre-trained second recommendation model, obtaining a second output vector according to the first feature vector and the second feature vector;

[0010] A first prediction result is obtained according to the first output vector and the second output vector; wherein the first prediction result indicates the matching degree between the content attribute of the content to be recommended and the needs of the target user.

[0011] In some possible implementations, encoding the user information of the target user to obtain a first feature vector, and encoding the content information of the content to be recommended to obtain a second feature vector, includes:

[0012] Perform one-hot encoding on the user information to obtain a feature vector of the user information, and perform dimension reduction on the feature vector of the user information to obtain a first feature vector;

[0013] The content information is one-hot encoded to obtain a feature vector of the content information, and the feature vector of the content information is reduced in dimension to obtain a second feature vector.

[0014] In some possible implementations, using a pre-trained first recommendation model, obtaining a first output vector according to the first eigenvector, the second eigenvector, and the relationship matrix includes:

[0015] Multiplying the acquired first eigenvector by the relationship matrix to obtain first result information;

[0016] Based on the mapping relationship of the relationship matrix, multiply the first result information by the second eigenvector corresponding to the first eigenvector to obtain a first output vector;

[0017] and / or:

[0018] Multiplying the acquired second eigenvector by the relationship matrix to obtain first result information;

[0019] Based on the mapping relationship of the relationship matrix, the first result information is multiplied by the first eigenvector or the second eigenvector corresponding to the second eigenvector to obtain a first output vector.

[0020] In some possible implementations, the first recommendation model is a model based on a matrix factorization mechanism; and the second recommendation model is a deep neural network model including abandoned neurons.

[0021] In some possible implementations, using a pre-trained second recommendation model to obtain a second output vector according to the first feature vector and the second feature vector includes:

[0022] Selecting part of the first eigenvector and the second eigenvector based on the abandoned neurons;

[0023] The selected first feature vector and the second feature vector are input into a second recommendation model to obtain second recommendation information.

[0024] In some possible implementations, obtaining a first prediction result according to the first recommendation information and the second recommendation information includes:

[0025] The first recommendation information is combined with the second recommendation information using an activation function to obtain a first prediction result.

[0026] In some possible implementations, the content to be recommended includes at least one of the following:

[0027] Videos to be recommended;

[0028] Articles to be recommended;

[0029] Items to be recommended;

[0030] Recommended service.

[0031] In a second aspect, the present application provides an information recommendation device, the device comprising:

[0032] An information encoding module, used to encode the user information of the target user to obtain a first feature vector, and to encode the content information of the content to be recommended to obtain a second feature vector;

[0033] A determination module is used to determine a relationship matrix, wherein the relationship matrix includes: a first type of matrix and a second type of matrix, wherein the first type of matrix represents the relationship between the determined user information and the content to be recommended; and the second type of matrix represents the relationship between different types of content to be recommended;

[0034] A first obtaining module, used for obtaining a first output vector according to the first eigenvector, the second eigenvector and the relationship matrix by using a pre-trained first recommendation model;

[0035] A second obtaining module, configured to obtain a second output vector according to the first feature vector and the second feature vector by using a pre-trained second recommendation model;

[0036] The first prediction module is used to obtain a first prediction result according to the first output vector and the second output vector; wherein the first prediction result indicates the matching degree between the content attribute of the content to be recommended and the needs of the target user.

[0037] In some possible implementations, the information encoding module is specifically used to perform one-hot encoding on user information to obtain a feature vector of the user information, perform dimensionality reduction on the feature vector of the user information, and obtain a first feature vector; perform one-hot encoding on content information to obtain a feature vector of the content information, and perform dimensionality reduction on the feature vector of the content information to obtain a second feature vector.

[0038] In some possible implementations, the first obtaining module is specifically configured to multiply the obtained first eigenvector by the relationship matrix to obtain first result information; based on the mapping relationship of the relationship matrix, multiply the first result information by the second eigenvector corresponding to the first eigenvector to obtain a first output vector;

[0039] and / or:

[0040] The acquired second eigenvector is multiplied by the relationship matrix to obtain first result information; based on the mapping relationship of the relationship matrix, the first result information is multiplied by the first eigenvector or the second eigenvector corresponding to the second eigenvector to obtain a first output vector.

[0041] In some possible implementations, the first recommendation model is a model based on a matrix factorization mechanism; and the second recommendation model is a deep neural network model including abandoned neurons.

[0042] In some possible implementations, the second obtaining module is specifically configured to select part of the first feature vector and the second feature vector based on the abandoned neurons; input the selected first feature vector and the second feature vector into the second recommendation model to obtain the second recommendation information.

[0043] In some possible implementations, the first prediction module is specifically configured to combine the first recommendation information with the second recommendation information by using an activation function to obtain a first prediction result.

[0044] In some possible embodiments, the content to be recommended includes at least one of the following:

[0045] Videos to be recommended;

[0046] Articles to be recommended;

[0047] Items to be recommended;

[0048] Recommended service.

[0049] In a third aspect, the present application further provides an electronic device, including:

[0050] a memory for storing processor-executable instructions;

[0051] Processor; wherein the processor is configured to: implement the method of the first aspect and its possible implementation methods when executing executable instructions.

[0052] In a fourth aspect, the present application provides a computer-readable storage medium storing an executable program, wherein the executable program, when executed by a processor, implements the method of the first aspect and its possible implementation methods.

[0053] Compared with the prior art, the technical solution provided by the embodiment of the present application has the following beneficial effects:

[0054] In the present application, the user information of the target user is encoded to obtain a first feature vector, and the content information of the content to be recommended is encoded to obtain a second feature vector; and a relationship matrix is ​​determined for characterizing the relationship between the user information and the content to be recommended and the relationship between different types of content to be recommended, and a first output vector is obtained according to the first feature vector, the second feature vector and the relationship matrix using a pre-trained first recommendation model; a second output vector is obtained according to the first feature vector and the second feature vector using a pre-trained second recommendation model; and a first prediction result is obtained according to the first output vector and the second output vector. In this way, the interaction information between features of different categories is represented by the relationship matrix, so that the interaction information between features can be learned in a more detailed manner, and the mining of the feature interaction information can be more comprehensive, so that the final result of the recommendation information is more accurate and more in line with user needs.

[0055] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0057] Figure 1 It is a flowchart of an embodiment of an information recommendation method in an embodiment of the present application.

[0058] Figure 2 It is a flowchart of another information recommendation method in the embodiments of the present application.

[0059] Figure 3 It is a structural diagram of the first model in the embodiment of the present application.

[0060] Figure 4 It is a structural comparison diagram of a traditional network and a network with a Dropout mechanism in an embodiment of the present application.

[0061] Figure 5 It is a structural diagram of an information recommendation device in an embodiment of the present application.

[0062] Figure 6 It is a schematic diagram of the information recommendation model structure in an embodiment of the present application.

[0063] Figure 7 It is a schematic diagram of the structure of an electronic device in an embodiment of the present application. DETAILED DESCRIPTION

[0064] In the following description, reference is made to the drawings that form part of the present application and show specific aspects of the embodiments of the present application or specific aspects of the embodiments of the present application in an illustrative manner. It should be understood that the embodiments of the present application can be used in other aspects and may include structural or logical changes not depicted in the drawings. Therefore, the following detailed description should not be understood in a restrictive sense, and the scope of the present application is defined by the appended claims. For example, it should be understood that the disclosure in conjunction with the described method can be equally applicable to the corresponding device or apparatus for performing the method, and vice versa. For example, if one or more specific method steps are described, the corresponding device may include one or more units such as functional units to perform the one or more method steps described (for example, one unit performs one or more steps, or multiple units, each of which performs one or more of the multiple steps), even if such one or more units are not explicitly described or illustrated in the drawings. On the other hand, for example, if a specific device is described based on one or more units such as functional units, the corresponding method may include a step to perform the functionality of one or more units (for example, one step performs the functionality of one or more units, or multiple steps, each of which performs the functionality of one or more units in multiple units), even if such one or more steps are not explicitly described or illustrated in the drawings. Further, it should be understood that, unless explicitly stated otherwise, features of the various exemplary embodiments and / or aspects described herein may be combined with each other.

[0065] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0066] Figure 1 This is a flow chart of an example of an information recommendation method in an embodiment of the present application. The information recommendation method can be executed by an electronic device. The electronic device can be a program or application terminal installed on an application server. Various communication client applications can be installed on the terminal, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc. The terminal can be various electronic devices with a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, e-book readers, MP3 players, portable computers, desktop computers, etc.

[0067] See also Figure 1 As shown, the information recommendation method may include:

[0068] S101, encoding the user information of the target user to obtain a first feature vector, and encoding the content information of the content to be recommended to obtain a second feature vector;

[0069] S102, determining a relationship matrix, wherein the relationship matrix includes: a first type matrix and a second type matrix, wherein the first type matrix represents the relationship between the determined user information and the content to be recommended; and the second type matrix represents the relationship between different types of content to be recommended;

[0070] S103, using a pre-trained first recommendation model, obtaining a first output vector according to the first eigenvector, the second eigenvector and the relationship matrix;

[0071] S104, using a pre-trained second recommendation model to obtain a second output vector according to the first feature vector and the second feature vector;

[0072] S105, obtaining a first prediction result according to the first output vector and the second output vector; wherein the first prediction result indicates the matching degree between the content attribute of the content to be recommended and the needs of the target user.

[0073] In one embodiment, user information includes but is not limited to user basic attributes, social relationships, consumption capacity, behavioral characteristics, and psychological characteristics.

[0074] Exemplarily, the basic attributes of a user may include: age, gender, height, weight, ethnicity, education, occupation, geographic location and other basic attribute information.

[0075] Social relations may include: family information, educational information, occupational information, etc.

[0076] For example, family information may include whether one is married or unmarried, whether one has children, etc. Education information may include school, student status, etc. Career information may include employed, unemployed, company, position, etc.

[0077] Consumption capacity may include: monthly income, expenditure, debt status, etc.

[0078] Behavioral characteristics may include: interest preferences, historical behavior data, etc.

[0079] Psychological characteristics may include: price sensitivity, brand preference, difficulty in making choices, etc.

[0080] In one embodiment, the content to be recommended includes but is not limited to video content, article content, item content and / or service content.

[0081] Exemplarily, the video content may include: short videos, TV series, movies, variety shows, cartoons, animations, etc. The content information of the video content may be classification information of the video content.

[0082] For example, short videos can be about food, film reviews, science, life sharing, etc., and TV series can be about different themes such as ancient idol dramas, spy wars, political intrigues, the War of Resistance against Japanese Aggression, and fairy tales.

[0083] The content of the article can be educational examinations, literary fiction, science and technology IT, humanities and social sciences, business management inspiration and / or art, etc.;

[0084] The item content may be home appliances, household kitchenware, fresh food and / or clothing, etc. The content information of the item content may be basic information of the item content, such as the category, purpose, color, etc. of the home appliance;

[0085] The services may include food, hotels / homestays, cinemas, leisure and entertainment, etc.

[0086] In the embodiment of the present application, the user information and content information can be obtained from a big data system, such as a Hadoop system, etc. The big data system can store data of various types and channels, and can also provide various types of data for each system or channel. By obtaining the data stored in the big data system, the integrity and real-time nature of the data can be guaranteed, thereby improving the accuracy and completeness of the first prediction result.

[0087] In one embodiment, after obtaining user information and content information, such as Figure 2 As shown, the above step S101 may include:

[0088] S201, performing one-hot encoding on the user information to obtain a feature vector of the user information, and performing dimension reduction on the feature vector of the user information to obtain a first feature vector;

[0089] S202, performing one-hot encoding on the content information to obtain a feature vector of the content information, and performing dimension reduction on the feature vector of the content information to obtain a second feature vector.

[0090] As you can understand, one hot encoding is also called one-bit effective encoding, and its method is to represent categorical variables as binary vectors. Map categorical values ​​to integer values, and then each integer value is represented as a binary vector. Except for the integer index which is marked as 1, the others are all zero values. For example, user information can be gender, and gender features can include two values: male / female. For example, male is represented by 1 and female is represented by 2. The value after one hot encoding can be "01" or "10". For features with more value domain types, each category needs to be represented by 0 after one hot encoding, and only one integer value bit is represented by 1, which will cause the vector to be too sparse. For example, taking all the city names in China as an example, the result after one hot encoding can be: Beijing [1,0,0,...,0,0,0], Shanghai [0,1,0,...,0,0,0], Guangzhou [0,0,1,0,...,0,0,0]... In this case, not only is the obtained vector too sparse, but the feature dimension is also relatively high. Therefore, after one-hot encoding, the obtained feature vector can also be reduced in dimension.

[0091] In one embodiment, a low-dimensional real vector (Dristributed representation) method can be used for dimensionality reduction. Through training, each word is mapped to a shorter word vector, and all these word vectors constitute a vector space, and the dimension of the vector space can be set by itself during the training process to achieve dimensionality reduction of the vector. Alternatively, any existing method can be used to reduce the dimension of the feature vector, and the method used for reducing the dimension of the feature vector does not limit the solution to be protected by this application.

[0092] In one embodiment, after acquiring the user information and the content information, the method may further include: determining a relationship matrix based on the user information and the content information.

[0093] It is understandable that in the process of information recommendation, in addition to the feature information itself, the interaction results between different categories of feature information are also different. And, in addition to the interaction results between user information and content information, there are also interaction results between different categories of content information. For example, if the user is a woman or a pregnant woman, then the interaction information between the user information and beer and the interaction information between the user information and the parenting video must be different. At the same time, different content information such as the interaction information between beer and cups and the interaction information between beer and diapers must also be different. Therefore, in an embodiment of the present application, by respectively defining a first type of relationship matrix that represents the relationship between user information and the content to be recommended, and a second type of relationship matrix that represents the relationship between different categories of content to be recommended, different feature interaction information between different categories of features is learned.

[0094] In an embodiment of the present application, a relationship matrix is ​​defined to represent the interaction information between features in different categories, so that the interaction information between features can be learned in a more detailed manner, the mining of feature interaction information can be more comprehensive, the accuracy of information recommendation can be improved, user needs can be better met, and user experience can be improved.

[0095] In one embodiment, after obtaining the first eigenvector, the second eigenvector and the relationship matrix, the above step S103 may include:

[0096] Multiplying the acquired first eigenvector by the relationship matrix to obtain first result information;

[0097] Based on the mapping relationship of the relationship matrix, multiply the first result information by the second eigenvector corresponding to the first eigenvector to obtain a first output vector;

[0098] and / or:

[0099] Multiplying the acquired second eigenvector by the relationship matrix to obtain first result information;

[0100] Based on the mapping relationship of the relationship matrix, the first result information is multiplied by the first eigenvector or the second eigenvector corresponding to the second eigenvector to obtain a first output vector.

[0101] It can be understood that information recommendation is to select content information of the content to be recommended for the target user. Therefore, for user information, its associated characteristic information can only be content information. For content information, its associated characteristic information can be either user information or content information.

[0102] In one embodiment, the first recommendation model may be a model based on a matrix factorization mechanism.

[0103] Exemplarily, a matrix-based factorization mechanism model is used to learn low-order interaction information of features. For example, taking the field matrix factorization machine model (FmFM) as an example, Figure 3 is a schematic diagram of the structure of the first model in the embodiment of the present application, such as Figure 3 As shown, the method for calculating the first output vector may be:

[0104] First, based on the acquired features i and j, define a relationship matrix M F(i)，F(j) , for features i and j, the features are encoded to obtain feature vector v i and the eigenvector v j , the obtained feature vector v i Multiply the matrix to get the first result information v i，F(j) , v i，F(j) =v i ×M F(i)，F(j) , then vi，F(j) and the eigenvector v j Multiply them together to get the final result. Finally, the above FmFM model is as shown in expression (1):

[0105]

[0106] where w0, w i is the feature weight, x i , x j represents the eigenvalue, F(i), F(j) represent the feature categories of i and j. F(i)，F(j) The final result of feature interaction will be different. Therefore, the implicit information of different feature interactions can be more comprehensively mined to improve the CTR prediction results.

[0107] In one embodiment, the FmFM model can adjust the length of the output vector by changing the shape of the matrix for different categories.

[0108] Specifically, in practical applications, different category features have different value ranges, such as gender has two values, male and female, and color has seven values. The shape of the matrix is ​​set according to the variance of different category features. This can effectively reduce the size of the model and avoid overfitting problems.

[0109] For example, in actual applications, different category features have different value ranges. Based on the above calculation process, the following equation can be determined: i ×M F(i)，F(j) ×v j =v j ×M F(i)，F(j) ×v i . Such as gender characteristicsv i There are two values, male and female, namely v i is two-dimensional; color feature v j There are 20 values. j If it is 20-dimensional, then the last step of the calculation method on the right side of the equation only needs to calculate a 2*2 vector inner product, while the method on the left side of the equation is a 20*20 vector inner product.

[0110] In one embodiment, the second recommendation model is a deep neural network model including abandoned neurons. Based on this, step S104 may include:

[0111] Selecting a portion of the elements in the first eigenvector and a portion of the elements in the second eigenvector based on the abandoned neurons;

[0112] The first feature vector and the second feature vector after the abandoned neurons are selected are input into the second recommendation model to obtain a second output vector.

[0113] It can be understood that the deep neural network model (Deep Neural Networks, DNN) is used to learn high-order interactive information of features. By inputting the first feature vector and the second feature vector into a pre-trained second recommendation model, high-order feature interactive information can be effectively learned.

[0114] The abandoned neuron mechanism is to temporarily stop the network nodes from working with a certain probability. Taking the abandoned neuron mechanism as an example, Figure 4 This is a structural comparison diagram of a standard deep neural network and a deep neural network with a Dropout mechanism, see Figure 4 As shown:

[0115] In some related technologies, each node participates in the calculation of model training, while in the deep neural network with Dropout mechanism, a probability process r is added to each node. i Set the node value to 0 with a certain probability, where r i = Bernoulli(p), Bernoulli() function generates probability r vector, p is a hyperparameter. The calculation process of each hidden layer is: (l+1) =σ(W (l) a′ (l) +b (l) ), where a′ (l) =r (l) *a (l) , a represents the output value of the layer, l represents the layer, σ represents the activation function, W represents the weight of the layer, and b represents the bias term of the layer. By setting the number of hidden layers in the model, the final model result is shown in expression (2):

[0116] y DNN =σ(W |H|+1 ·(r |H| *a |H| )+b |H|+1 ) (2)

[0117] Among them, |H| is the number of hidden layers.

[0118] It is understandable that in the training process of the deep neural network model, if the model parameters are too many and the training samples are too few, the trained model is prone to overfitting. In the embodiment of the present application, by controlling some nodes not to participate in the calculation based on the abandoned neuron mechanism with probability, the model generalization is stronger, so that the model will not rely too much on certain local features. In this way, not only the training speed is faster under large-scale data sets, but also the overfitting problem can be avoided.

[0119] In one embodiment, after the first output vector and the second output vector are obtained, the first output vector and the second output vector are combined using an activation function to obtain a first prediction result.

[0120] Exemplarily, an activation function is a function that runs on a neuron of an artificial neural network and is responsible for mapping the input of the neuron to the output. The activation function includes but is not limited to at least one of the following:

[0121] Squeezing function (sigmoid), hyperbolic tangent function (Tanh), etc. Here, taking the activation function as sigmoid function as an example, the first prediction result is Then the first prediction result can be shown as expression (3):

[0122]

[0123] In this application, the low-order interaction information and high-order interaction information between features are learned respectively by the first recommendation model and the second recommendation model, and the low-order and high-order interactions of features are taken into account by combining them with the activation function. At the same time, the interaction information between features of different categories is further represented by setting a relationship matrix, so as to learn the interaction information between features in a more detailed manner, and to mine the feature interaction information more comprehensively, so that the final recommendation information is more accurate and more in line with user needs, thereby improving the user experience.

[0124] Figure 5 This is a schematic diagram of the structure of an information recommendation device in an embodiment of the present application, see Figure 5 As shown, the information recommendation device 500 may include:

[0125] An information encoding module 501 is used to encode the user information of the target user to obtain a first feature vector, and to encode the content information of the content to be recommended to obtain a second feature vector;

[0126] The determination module 502 is used to determine a relationship matrix, wherein the relationship matrix includes: a first type of matrix and a second type of matrix, wherein the first type of matrix represents the relationship between the determined user information and the content to be recommended; and the second type of matrix represents the relationship between different types of content to be recommended;

[0127] A first obtaining module 503 is used to obtain a first output vector according to the first eigenvector, the second eigenvector and the relationship matrix by using a pre-trained first recommendation model;

[0128] A second obtaining module 504 is used to obtain a second output vector according to the first feature vector and the second feature vector by using a pre-trained second recommendation model;

[0129] The first prediction module 505 is used to obtain a first prediction result according to the first output vector and the second output vector; wherein the first prediction result indicates the matching degree between the content attribute of the content to be recommended and the needs of the target user.

[0130] In some possible implementations, the information encoding module 501 is specifically used to perform one-hot encoding on the user information to obtain a feature vector of the user information, perform dimensionality reduction on the feature vector of the user information, and obtain a first feature vector; perform one-hot encoding on the content information to obtain a feature vector of the content information, and perform dimensionality reduction on the feature vector of the content information to obtain a second feature vector.

[0131] In some possible implementations, the first obtaining module 503 is specifically configured to multiply the obtained first eigenvector by the relationship matrix to obtain first result information; based on the mapping relationship of the relationship matrix, multiply the first result information by the second eigenvector corresponding to the first eigenvector to obtain a first output vector;

[0132] and / or:

[0133] The acquired second eigenvector is multiplied by the relationship matrix to obtain first result information; based on the mapping relationship of the relationship matrix, the first result information is multiplied by the first eigenvector or the second eigenvector corresponding to the second eigenvector to obtain a first output vector.

[0134] In some possible embodiments, the first recommendation model is a model based on a matrix factorization mechanism; and the second recommendation model is a deep neural network model including abandoned neurons.

[0135] In some possible embodiments, the second obtaining module 504 is specifically configured to select part of the first feature vector and the second feature vector based on the abandoned neurons; input the selected first feature vector and the second feature vector into the second recommendation model to obtain the second recommendation information.

[0136] In some possible embodiments, the first prediction module 505 is specifically configured to combine the first recommendation information with the second recommendation information by using an activation function to obtain a first prediction result.

[0137] In some possible embodiments, the content to be recommended includes at least one of the following:

[0138] Videos to be recommended;

[0139] Articles to be recommended;

[0140] Items to be recommended;

[0141] Recommended service.

[0142] It should be noted that the specific implementation process of the information encoding module 501, the determination module 502, the first acquisition module 503, the second acquisition module 504 and the first prediction module 505 can be referred to Figures 1 to 4 For the sake of brevity, the detailed description of the embodiments will not be repeated here.

[0143] In another embodiment of the present application, taking the first model as the FmFM model and the second model as the DNN model as an example, Figure 6 Schematic diagram of the information recommendation model structure in the embodiment of the present application. Figure 6 As shown:

[0144] The FmFM model shares the same input as the DNN model. The input of the model is a sparse feature vector. After the DenseEmbeddings layer, the sparse feature vector is converted into a dense vector, and then input into the FmFM module and the deep neural network module for feature learning. The output of the FmFM model is:

[0145]

[0146] The output of the deep neural network model is:

[0147] y DNN =σ(W |H|+1 ·(r |H| *a |H| )+b |H|+1 );

[0148] Finally, the results obtained from the two modules are combined using the sigmoid function to obtain the prediction result. The result can be expressed as:

[0149]

[0150] Based on the same inventive concept, an embodiment of the present application provides an electronic device, which may be consistent with the information recommendation method described in one or more of the above embodiments. Figure 7 This is a schematic diagram of the structure of an electronic device in an embodiment of the present application, see Figure 7 As shown, the electronic device 700 may adopt general computer hardware, including a processor 701 and a memory 702 .

[0151] In some possible implementations, at least one processor may constitute any physical device having a circuit that performs a logical operation on one or more inputs. For example, at least one processor may include one or more integrated circuits (ICs), including application specific integrated circuits (ASICs), microchips, microcontrollers, microprocessors, all or part of a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), a field programmable gate array (FPGA), or other circuits suitable for executing instructions or performing logical operations. The instructions executed by at least one processor may be preloaded, for example, into a memory integrated with or embedded in a controller, or may be stored in a separate memory. The memory may include a random access memory (RAM), a read-only memory (ROM), a hard disk, an optical disk, a magnetic medium, a flash memory, other permanent, fixed or volatile memory, or any other mechanism capable of storing instructions. In some embodiments, at least one processor may include more than one processor. Each processor may have a similar structure, or the processor may have different configurations that are electrically connected or disconnected from each other. For example, the processor may be a separate circuit or integrated in a single circuit. When more than one processor is used, the processor may be configured to operate independently or collaboratively. The processors may be coupled electrically, magnetically, optically, acoustically, mechanically, or by other means that allow them to interact.

[0152] Based on the same inventive concept, the present application provides a computer storage medium, which stores computer executable instructions. After the computer executable instructions are executed by a processor, the account registration method described in one or more of the above embodiments can be implemented.

[0153] Those skilled in the art will appreciate that the sequence numbers of the steps in the above embodiments do not imply a sequence of execution, and the execution sequence of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0154] The above is only an exemplary embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An information recommendation method, characterized in that: The method comprises: Encoding the user information of the target user to obtain a first feature vector, and encoding the content information of the content to be recommended to obtain a second feature vector; Determine a relationship matrix, wherein the relationship matrix includes: a first type of matrix and a second type of matrix, wherein the first type of matrix represents the relationship between the user information and the content to be recommended; the second type of matrix represents the relationship between different types of the content to be recommended; Using a pre-trained first recommendation model, obtaining a first output vector according to the first feature vector, the second feature vector and the relationship matrix; Using a pre-trained second recommendation model, obtaining a second output vector according to the first feature vector and the second feature vector; A first prediction result is obtained according to the first output vector and the second output vector; wherein the first prediction result indicates the matching degree between the content attribute of the to-be-recommended content and the needs of the target user.

2. The method according to claim 1, characterized in that: The step of encoding the user information of the target user to obtain a first feature vector and encoding the content information of the content to be recommended to obtain a second feature vector includes: Performing one-hot encoding on the user information to obtain a feature vector of the user information, and performing dimension reduction on the feature vector of the user information to obtain the first feature vector; The content information is one-hot encoded to obtain a feature vector of the content information, and the feature vector of the content information is reduced in dimension to obtain the second feature vector.

3. The method according to claim 2, characterized in that The using the pre-trained first recommendation model to obtain a first output vector according to the first feature vector, the second feature vector and the relationship matrix includes: Multiplying the acquired first eigenvector by the relationship matrix to obtain first result information; Based on the mapping relationship of the relationship matrix, multiply the first result information by the second eigenvector corresponding to the first eigenvector to obtain the first output vector; and / or: Multiplying the acquired second eigenvector by the relationship matrix to obtain first result information; Based on the mapping relationship of the relationship matrix, the first result information is multiplied by the first eigenvector corresponding to the second eigenvector to obtain the first output vector.

4. The method according to claim 3, characterized in that The first recommendation model is a model based on a matrix factorization mechanism; the second recommendation model is a deep neural network model including abandoned neurons.

5. The method according to claim 1, characterized in that The using the pre-trained second recommendation model to obtain a second output vector according to the first feature vector and the second feature vector includes: Selecting part of the first feature vector and the second feature vector based on abandoned neurons; The selected first feature vector and the second feature vector are input into the second recommendation model to obtain the second output vector.

6. The method according to claim 1, characterized in that The obtaining a first prediction result according to the first output vector and the second output vector includes: The first output vector and the second output vector are combined using an activation function to obtain the first prediction result.

7. The method according to claim 1, characterized in that The content to be recommended includes at least one of the following: Videos to be recommended; Articles to be recommended; Items to be recommended; Recommended service.

8. An information recommendation device, characterized in that: The device comprises: An information encoding module, used to encode the user information of the target user to obtain a first feature vector, and to encode the content information of the content to be recommended to obtain a second feature vector; A determination module is used to determine a relationship matrix, wherein the relationship matrix includes: a first type of matrix and a second type of matrix, wherein the first type of matrix represents the relationship between the user information and the content to be recommended; the second type of matrix represents the relationship between different types of the content to be recommended; A first obtaining module, configured to obtain a first output vector according to the first feature vector, the second feature vector and the relationship matrix by using a pre-trained first recommendation model; A second obtaining module, configured to obtain a second output vector according to the first feature vector and the second feature vector by using a pre-trained second recommendation model; The first prediction module is used to obtain a first prediction result according to the first output vector and the second output vector; wherein the first prediction result indicates the matching degree between the content attribute of the content to be recommended and the needs of the target user.

9. The device according to claim 8, characterized in that The information encoding module is specifically used to perform one-hot encoding on the user information to obtain a feature vector of the user information, and perform dimensionality reduction on the feature vector of the user information to obtain the first feature vector; perform one-hot encoding on the content information to obtain a feature vector of the content information, and perform dimensionality reduction on the feature vector of the content information to obtain the second feature vector.

10. The device according to claim 9, characterized in that The first obtaining module is specifically configured to multiply the first eigenvector obtained by the relationship matrix to obtain first result information; based on the mapping relationship of the relationship matrix, multiply the first result information by a second eigenvector corresponding to the first eigenvector to obtain the first output vector; and / or: The acquired second eigenvector is multiplied by the relationship matrix to obtain first result information; based on the mapping relationship of the relationship matrix, the first result information is multiplied by the first eigenvector corresponding to the second eigenvector to obtain the first output vector.

11. The device according to claim 10, characterized in that The first recommendation model is a model based on a matrix factorization mechanism; the second recommendation model is a deep neural network model including abandoned neurons.

12. The device according to claim 8, characterized in that The second acquisition module is specifically configured to select part of the first feature vector and the second feature vector based on abandoned neurons; input the selected first feature vector and the second feature vector into the second recommendation model to obtain the second output vector.

13. The device according to claim 8, characterized in that The first prediction module is specifically configured to combine the first output vector and the second output vector using an activation function to obtain the first prediction result.

14. The device according to claim 8, characterized in that The content to be recommended includes at least one of the following: Videos to be recommended; Articles to be recommended; Items to be recommended; Recommended service.

15. An electronic device, characterized in that: include: a memory for storing processor-executable instructions; Processor; wherein the processor is configured to: implement the method according to any one of claims 1 to 7 when executing the executable instructions.

16. A computer-readable storage medium, characterized in that: The readable storage medium stores an executable program, wherein the executable program implements the method according to any one of claims 1 to 7 when executed by a processor.

Images