A recommendation method and device

Abstract

The present application discloses a recommendation method and device, which performs low-order feature extraction on the user information of the target user and the object information of the candidate object to obtain a first feature vector, performs high-order feature extraction on the user information and the object information to obtain a connection vector, performs dimensionality reduction processing on the connection vector to obtain an activation vector, and finally performs residual fusion on the connection vector and the activation vector to obtain a second feature vector. Thus, based on the first feature vector and the second feature vector, the probability of the target user selecting the candidate object is predicted, and when the probability meets the set conditions, the candidate object is recommended to the target user. In this way, in the process of high-order feature extraction and dimensionality reduction processing of user information and object information, the residual fusion method is used to retain the loss of underlying information, successfully solve the degradation problem of the deep neural network algorithm, effectively prevent the gradient disappearance of the neural network, significantly improve the accuracy of the recommendation, and ensure the stability of the recommendation system.

Description

Technical Field

[0001] The present application relates to the field of recommendation technology, and in particular to a recommendation method and device. Background Art

[0002] With the rapid development of the Internet industry, online shopping and the use of the Internet for information dissemination have become very common. Various shopping platforms and information dissemination platforms have matured the technology of recommendation based on user behavior records. However, the current recommendation methods are highly dependent on historical data such as user's historical access records and purchase records. As a result, the quality of recommendations is seriously reduced in the case of cold start and data sparseness.

[0003] Application Contents

[0004] The present application discloses a recommendation method and device.

[0005] According to a first aspect of the present application, a recommendation method is provided, the method comprising: obtaining user information of a target user and object information of a candidate object; performing low-order feature extraction on the user information and the object information to obtain a first feature vector; performing high-order feature extraction on the user information and the object information to obtain a connection vector; performing dimensionality reduction processing on the connection vector to obtain an activation vector; performing residual fusion on the connection vector and the activation vector to obtain a second feature vector; predicting the probability of the target user selecting the candidate object based on the first feature vector and the second feature vector; and recommending the candidate object to the target user when the probability meets a set condition.

[0006] According to another embodiment of the present application, the extracting low-order features of the user information and the object information to obtain a first feature vector includes: performing matrix decomposition on the user information and the object information to obtain the first feature vector.

[0007] According to another embodiment of the present application, performing matrix decomposition on the user information and the object information to obtain the first eigenvector includes: performing matrix decomposition initialization on the user information and the object information, respectively, to obtain a user initial matrix vector and an object initial matrix vector respectively; performing global matrix decomposition on the user initial matrix vector and the object initial matrix vector to obtain the first eigenvector.

[0008] According to another embodiment of the present application, the performing high-order feature extraction on the user information and the object information to obtain the connection vector includes: utilizing a neural network algorithm to perform feature extraction on the user information and the object information to obtain the connection vector.

[0009] According to another embodiment of the present application, the use of a neural network algorithm to perform feature extraction on the user information and the object information to obtain the connection vector includes: performing feature extraction initialization on the user information and the object information respectively, and correspondingly obtaining a user initial perception vector and an object initial perception vector respectively; performing global matrix decomposition on the user initial perception vector and the object initial perception vector to obtain a shared feature vector; and performing full connection on the user initial perception vector, the object initial perception vector and the shared feature vector to obtain the connection vector.

[0010] According to another embodiment of the present application, performing dimensionality reduction processing on the connection vector to obtain the activation vector includes: performing nonlinear transformation on the connection vector to obtain the activation vector.

[0011] According to another embodiment of the present application, before performing nonlinear transformation on the connection vector to obtain the activation vector, the connection vector is subjected to dimensionality reduction processing to obtain the activation vector, and further includes at least one of the following: batch normalizing the connection vector; and processing the connection vector using an attention mechanism.

[0012] According to another embodiment of the present application, predicting the probability of the target user selecting the candidate object based on the first feature vector and the second feature vector includes: fully connecting the first feature vector and the second feature vector to obtain an output vector; and performing a nonlinear transformation on the output vector to obtain the probability.

[0013] According to another embodiment of the present application, the user information is the identity identifier of the target user; and / or the object information is the object identifier of the candidate object.

[0014] According to a second aspect of the present application, a recommendation device is disclosed, which includes: an acquisition module for acquiring user information of a target user and object information of a candidate object; a first feature extraction module for performing low-order feature extraction on the user information and the object information to obtain a first feature vector; a second feature extraction module for performing high-order feature extraction on the user information and the object information to obtain a connection vector; an activation module for performing dimensionality reduction processing on the connection vector to obtain an activation vector; a residual module for performing residual fusion on the connection vector and the activation vector to obtain a second feature vector; a prediction module for predicting the probability of the target user selecting the candidate object based on the first feature vector and the second feature vector; and a recommendation module for recommending the candidate object to the target user when the probability meets a set condition.

[0015] The above recommendation method performs low-order feature extraction on the user information of the target user and the object information of the candidate object to obtain a first feature vector, performs high-order feature extraction on the user information and object information to obtain a connection vector, performs dimensionality reduction processing on the connection vector to obtain an activation vector, and finally performs residual fusion on the connection vector and the activation vector to obtain a second feature vector. Thus, based on the first feature vector and the second feature vector, the probability of the target user selecting the candidate object is predicted, and when the probability meets the set conditions, the candidate object is recommended to the target user. In this way, in the process of high-order feature extraction and dimensionality reduction processing of user information and object information, the residual fusion method is used to retain the loss of underlying information, successfully solve the degradation problem of the deep neural network algorithm, effectively prevent the gradient disappearance of the neural network, significantly improve the accuracy of the recommendation, and ensure the stability of the recommendation system.

[0016] It should be understood that the teachings of the present application are not required to achieve all of the beneficial effects described above, but specific technical solutions can achieve specific technical effects, and other embodiments of the present application can also achieve beneficial effects not mentioned above. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] By reading the detailed description below with reference to the accompanying drawings, the above and other purposes, features and advantages of the exemplary embodiments of the present application will become easily understood. In the accompanying drawings, several embodiments of the present application are shown in an exemplary and non-limiting manner, wherein:

[0018] In the drawings, the same or corresponding reference numerals represent the same or corresponding parts.

[0019] Figure 1 A schematic diagram of the implementation process of the method recommended in the embodiment of the present application is shown;

[0020] Figure 2 A schematic diagram of the implementation process of a method recommended in another embodiment of the present application is shown;

[0021] Figure 3 A schematic diagram of the algorithm structure of a method recommended in another embodiment of the present application is shown;

[0022] Figure 4 A schematic diagram of the structure of the device recommended in the embodiment of the present application is shown. DETAILED DESCRIPTION

[0023] The principles and spirit of the present application will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present application, and are not intended to limit the scope of the present application in any way. On the contrary, these embodiments are provided to make the present application more thorough and complete, and to fully convey the scope of the present application to those skilled in the art.

[0024] The technical solution of the present application is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0026] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0027] Figure 1 The following is a schematic diagram showing the implementation process of the recommended method disclosed in this application.

[0028] refer to Figure 1 The recommendation method of the embodiment of the present application at least includes the following operation flow: operation 101, obtaining user information of the target user and object information of the candidate object; operation 102, performing low-order feature extraction on the user information and object information to obtain a first feature vector; operation 103, performing high-order feature extraction on the user information and object information to obtain a connection vector; operation 104, performing dimensionality reduction processing on the connection vector to obtain an activation vector; operation 105, performing residual fusion on the connection vector and the activation vector to obtain a second feature vector; operation 106, predicting the probability of the target user selecting the candidate object based on the first feature vector and the second feature vector; operation 107, when the probability meets the set conditions, recommending the candidate object to the target user.

[0029] In operation 101, user information of a target user and object information of a candidate object are obtained;

[0030] In this embodiment of the present application, the user information includes at least the user ID (Identity) of the target user, and the object information includes at least the product ID of the candidate product. Of course, the user information may also include the target user's historical click count and viewing time on the candidate product. The object information may also include product attribute information of the candidate product, such as product color, product category, and product applicable objects. Here, both the user information and the object information can be represented by a one-dimensional vector, for example: (0, 0, 1, 0...), and the length of the one-dimensional vector here can be 32.

[0031] In operation 102, low-order feature extraction is performed on the user information and the object information to obtain a first feature vector;

[0032] In this embodiment of the present application, performing low-order feature extraction on user information and object information refers to the features formed after a finite number of linear-nonlinear combination operations are performed on the user information and object information, for example, matrix decomposition is performed on the user information and object information.

[0033] In operation 103, high-order feature extraction is performed on the user information and the object information to obtain a connection vector;

[0034] In this embodiment of the present application, high-order feature extraction is performed on user information and object information, which means that the features formed after multiple linear-nonlinear combination operations are performed on the user information and object information, and the obtained features are highly abstract features. For example, a neural network algorithm can be used here to extract high-order features from user information and object information.

[0035] In operation 104, the connection vector is subjected to dimensionality reduction processing to obtain an activation vector;

[0036] In this embodiment of the present application, the connection vector is a feature vector formed after multiple linear-nonlinear combination operations are performed on user information and object information. The connection vector is a vector group formed by simply combining multiple vectors. If the connection vector is subjected to dimensionality reduction processing, each layer output is a linear function of the upper layer input, which has no use value and cannot ultimately obtain the probability of the target user selecting the candidate product. Therefore, it is necessary to perform dimensionality reduction operations on the connection vector. The connection vector can be nonlinearly transformed here, for example: using an activation function to perform nonlinear transformation on the connection vector. Common activation functions include: Sigmoid function, hyperbolic tangent activation function Tanh function and rectified linear unit ReLU function, etc.

[0037] In operation 105, residual fusion is performed on the connection vector and the activation vector to obtain a second feature vector.

[0038] In this embodiment of the present application, a deep neural network algorithm is used to extract high-order features from the features of user information and object information, abstract the data features layer by layer, obtain a connection vector, and perform dimensionality reduction processing on the connection vector to obtain an activation vector. However, a deeper neural network, with the support of a nonlinear activation function, has a larger hypothesis space. At this time, the input of the unit is directly fused with the output of the unit in the form of a skip-layer connection by using residual fusion, and then activated. When the feature is forward propagated, the input signal can be directly transmitted from any low layer to the high layer, which effectively solves the degradation problem of the deep neural network, prevents the network gradient from disappearing, and avoids the loss of underlying information.

[0039] In operation 106 , a probability of the target user selecting the candidate object is predicted based on the first feature vector and the second feature vector.

[0040] In this embodiment of the present application, the first feature vector and the second feature vector are connected and passed through a fully connected layer whose last activation function is sigmoid to output the predicted probability of the target user selecting the candidate object, where the probability is a value between 0 and 1.

[0041] In operation 107 , when the probability satisfies the set condition, the candidate object is recommended to the target user.

[0042] In this embodiment of the present application, when the probability is greater than the set threshold, the candidate object is recommended to the target user. For example, after the user information of the target user A and the object information of the candidate object 1 are processed through the above operations 101 to 107, the probability that the target user A selects the candidate object 1 is 0.76, and the set threshold is 0.65, then it is determined that the probability that the target user A selects the candidate object 1 meets the set condition, and the probability of recommending the candidate object 1 to the target user A is high.

[0043] Figure 2 A schematic diagram of the implementation process of another embodiment of the present application is shown. Figure 2 As shown, it at least includes the following operation processes:

[0044] Operation 201 : acquiring user information of a target user and object information of a candidate object.

[0045] Operation 202 : performing matrix decomposition on the user information and the object information to obtain a first feature vector.

[0046] In this embodiment of the present application, matrix decomposition can be performed on the user information and the object information to extract low-order features of the user information and the object information to obtain a first feature vector. Specifically, the user information and the object information can be initialized by matrix decomposition first, respectively, to obtain a user initial matrix vector and an object initial matrix vector, respectively, and then global matrix decomposition can be performed on the user initial matrix vector and the object initial matrix vector to obtain the first feature vector.

[0047] In this embodiment of the present application, the user information and the object information are respectively a one-dimensional vector, and the matrix decomposition initialization of the user information and the object information is a process of converting a sparse vector into a dense vector. The user information is used as the user ID (Identity) and the object ID, and after an embedding layer, it is converted into a user feature vector and an item feature vector, and then the user feature vector and the item feature vector are multiplied in position by the matrix decomposition formula to obtain the first feature vector.

[0048] Operation 203 : Using a neural network algorithm, feature extraction is performed on the user information and the object information to obtain a connection vector.

[0049] In this embodiment of the present application, a neural network algorithm is used to extract features from user information and object information to obtain a connection vector. Specifically, feature extraction initialization can be performed on user information and object information respectively, and corresponding user initial perception vectors and object initial perception vectors can be obtained respectively; global matrix decomposition can be performed on the user initial perception vector and the object initial perception vector to obtain a shared feature vector; and full connection can be performed on the user initial perception vector, the object initial perception vector, and the shared feature vector to obtain a connection vector.

[0050] Here, feature extraction is initialized for user information and object information respectively, and the same data processing method as the initialization method used for matrix decomposition initialization of user information and object information in operation 202 is used, and the user information and object information are initialized by random function method. The reason for using the same data processing method to initialize the user information twice in operation 202 and operation 203 is that the random function can learn user information from different interests, and provide more comprehensive and accurate user information for subsequent data processing. Similarly, the reason for initializing the object information twice is that the random function can learn object information from different interests, and provide more comprehensive and accurate object information for subsequent data processing.

[0051] In this embodiment of the present application, after each round of feature extraction in the neural network algorithm, some features will be screened and filtered, resulting in the final result being overly focused on a certain feature. In the present application, the shared feature vector, the user initial perception vector, and the object initial perception vector are simultaneously used as the input of the fully connected layer in the neural network algorithm, which can effectively alleviate the overfitting of the user vector and the product vector and increase the generalization ability of the model.

[0052] In operation 204, the connection vector is processed using an attention mechanism.

[0053] The attention model is introduced into the neural network algorithm. While improving the performance of the model, the attention mechanism increases the interpretability of the neural network structure. In the process of recommending candidate objects to the target user, the attention mechanism can assign attention weights to different features of user information to more effectively obtain the long-term and short-term interests of the target user. Obviously, not all interactions between the target user and the candidate object are related to the target user's preference for the candidate object, but the target user's interest will also change over time. Here, the attention mechanism is used to select the more critical information for the current task goal based on the characteristics of the target user and the information of the candidate object, increase the weight of the effective information, and effectively improve the accuracy of recommending candidate objects to the target user.

[0054] In operation 205 , batch normalization is performed on the concatenated vectors.

[0055] Batch normalization refers to the normalization of each batch of data in a neural network. In this embodiment of the present application, the goal of batch normalization is to normalize the features so that the output of each layer of the network is activated to obtain a zero mean state with a standard deviation of 1. Therefore, the opposite phenomenon of batch normalization is non-zero mean, which can be understood as a phenomenon that non-zero mean is that the data is not distributed around the value of 0, but most of the values ​​of the data are greater than 0 and less than 0, and the data will become very large or very small. Here, the connection vector is batch normalized, and the increasingly biased distribution is forced to be pulled back to a more standard distribution, so that the activation input value falls in an area where the nonlinear function is more sensitive to the input, so that a small change in the input will lead to a larger change in the loss function, which can make the gradient larger and avoid the problem of gradient vanishing. In this way, the convergence speed of the model can be effectively accelerated, computing power can be effectively saved, and the information processing speed for target users and candidate objects can be improved.

[0056] In operation 206, nonlinear transformation is performed on the connection vector to obtain an activation vector.

[0057] In this embodiment of the present application, the Relu function can be used to perform nonlinear transformation on the connection vector to obtain an activation vector. Without an activation function, each layer in the neural network is equivalent to matrix multiplication. Even if several layers are superimposed, it can only be matrix multiplication, and the final reference-valued parameters cannot be obtained, such as the probability of the target user selecting a candidate product. Here, the activation function is to reduce the dimension of the connection vector in the neural network model and increase the nonlinearity of the neural network model.

[0058] Of course, the Tanh activation function and the Sigmoid function can also be used here to perform nonlinear transformation on the connection vector to obtain the activation vector.

[0059] In operation 207, residual fusion is performed on the connection vector and the activation vector to obtain a second feature vector.

[0060] Operation 208 : predicting the probability of the target user selecting the candidate object based on the first feature vector and the second feature vector.

[0061] Operation 209 : when the probability meets the set condition, recommend the candidate object to the target user.

[0062] The specific implementation process of operations 201, 207 to 209 is as follows: Figure 1 The specific implementation processes of operations 101 and 105 to 107 in the illustrated embodiment are similar and will not be described in detail here.

[0063] Figure 3 A schematic diagram of the recommendation algorithm structure of another embodiment of the present application is shown.

[0064] like Figure 3 As shown in FIG. 1 , the traditional matrix decomposition and multi-layer perceptron are combined to extract low-dimensional and high-dimensional features of the target user's user information and the candidate object's object information. Specifically, Figure 3 The left half is the matrix decomposition, and the right half is the multi-layer perceptron.

[0065] The user information of the target user and the object information of the candidate object are obtained through the input layer. For example, the target user is user u and the candidate object is item i. The user information and object information are encoded to obtain User(u) and Item(i). User(u) and Item(i) are one-dimensional vectors respectively. For example, User(u) is a 32-bit one-dimensional vector and Item(i) is also a 32-bit one-dimensional vector. One-hot encoding can be used here.

[0066] Then, the user information and object information are initialized through the embedding layer. Specifically, User(u) is processed by random functions to obtain the matrix user vector MF user vector, and Item(i) is processed by random functions to obtain the matrix commodity vector MF item vector, and MF user vector and MF item vector are used as inputs of the matrix decomposition layer GMF Layer of the model layer. Through the matrix decomposition layer GMF Layer, the first eigenvector is obtained using the matrix decomposition formula.

[0067] Similarly, User(u) is processed by random functions to obtain the multi-layer perceptron user vector MLP uservector, and Item(i) is processed by random functions to obtain the multi-layer perceptron commodity vector MLP item vector. MLPuser vector and MLP item vector are used as the input of the model layer neural network algorithm. In addition, global matrix decomposition is performed on MLPuser vector and MLP item vector at the same time to obtain the shared global matrix decomposition layer vector ShareGMF Layer, which is also used as the input of the model layer neural network algorithm.

[0068] Therefore, in the process of extracting features of user information and object information using a multi-layer perceptron at the model layer, the Share GMF Layer, MLP user vector and MLP item vector are fully connected to obtain the connection vector Concat. This effectively solves the overfitting problem in the process of extracting features of user information and object information using a multi-layer perceptron.

[0069] Furthermore, a multi-layer RAN (Residual Attention Neural Network) processing is set here, including: Attention (attention mechanism), BN (Batch Normalization) and Relu (corrected linear unit) activation function. Among them, Attention effectively focuses on the effective features in the connection vector Concat to improve the accuracy of the model. BN can effectively accelerate the convergence speed of the model, improve computing power, thereby increasing data processing speed and significantly improving user experience. The Relu activation function here is just a simple activation function example, and other applicable activation functions can also be used here. For example: hyperbolic tangent activation function Tanh.

[0070] Furthermore, the connection vector Concat obtained by the multi-layer perceptron is residually fused with the vector obtained by the activation function Relu to obtain the second eigenvector.

[0071] Finally, the first eigenvector and the second eigenvector are concatentioned, and the fully connected layer with the last activation function as sigmoid is used in the output layer to output a recommendation value y_pred between 0 and 1. When the recommendation value y_pred is greater than the set threshold, the candidate object i is recommended to the target user u.

[0072] The above model verification method performs low-order feature extraction on the user information of the target user and the object information of the candidate object to obtain the first feature vector, performs high-order feature extraction on the user information and object information to obtain the connection vector, and performs dimensionality reduction processing on the connection vector to obtain the activation vector, and finally performs residual fusion on the connection vector and the activation vector to obtain the second feature vector. Therefore, based on the first feature vector and the second feature vector, the probability of the target user selecting the candidate object is predicted, and when the probability meets the set conditions, the candidate object is recommended to the target user. In this way, in the process of high-order feature extraction and dimensionality reduction processing of user information and object information, the residual fusion method is used to retain the loss of underlying information, successfully solve the degradation problem of the deep neural network algorithm, effectively prevent the gradient disappearance of the neural network, significantly improve the accuracy of the recommendation, and ensure the stability of the recommendation system.

[0073] Similarly, based on the above recommendation method, the present application also provides a computer-readable storage medium, which stores a program. When the program is executed by a processor, the processor performs at least the following operation steps: operation 101, obtaining user information of the target user and object information of the candidate object; operation 102, performing low-order feature extraction on the user information and object information to obtain a first feature vector; operation 103, performing high-order feature extraction on the user information and object information to obtain a connection vector; operation 104, performing dimensionality reduction processing on the connection vector to obtain an activation vector; operation 105, performing residual fusion on the connection vector and the activation vector to obtain a second feature vector; operation 106, predicting the probability of the target user selecting the candidate object based on the first feature vector and the second feature vector; operation 107, when the probability meets the set conditions, recommending the candidate object to the target user.

[0074] Furthermore, based on the above recommended method, the present application also provides a model verification device, such as Figure 4As shown, the device 40 includes: an acquisition module 401, which is used to acquire user information of a target user and object information of a candidate object; a first feature extraction module 402, which is used to perform low-order feature extraction on the user information and the object information to obtain a first feature vector; a second feature extraction module 403, which is used to perform high-order feature extraction on the user information and the object information to obtain a connection vector; an activation module 404, which is used to perform dimensionality reduction processing on the connection vector to obtain an activation vector; a residual module 405, which is used to perform residual fusion on the connection vector and the activation vector to obtain a second feature vector; a prediction module 406, which is used to predict the probability of the target user selecting the candidate object based on the first feature vector and the second feature vector; and a recommendation module 407, which is used to recommend the candidate object to the target user when the probability meets the set conditions.

[0075] According to another embodiment of the present application, the first feature extraction module 402 includes a matrix submodule, which is used to perform matrix decomposition on the user information and the object information to obtain a first feature vector.

[0076] According to another embodiment of the present application, the matrix submodule uses the following operation steps to perform matrix decomposition on user information and object information to obtain a first eigenvector: perform matrix decomposition initialization on the user information and object information respectively, and correspondingly obtain a user initial matrix vector and an object initial matrix vector respectively; perform global matrix decomposition on the user initial matrix vector and the object initial matrix vector to obtain a first eigenvector.

[0077] According to another embodiment of the present application, the second feature extraction module 403 includes: an extraction submodule, which is used to extract features from user information and object information using a neural network algorithm to obtain a connection vector.

[0078] According to another embodiment of the present application, the extraction submodule adopts the following operation steps to use a neural network algorithm to perform feature extraction on user information and object information to obtain a connection vector: perform feature extraction initialization on the user information and object information respectively, and correspondingly obtain a user initial perception vector and an object initial perception vector respectively; perform global matrix decomposition on the user initial perception vector and the object initial perception vector to obtain a shared feature vector; perform full connection on the user initial perception vector, the object initial perception vector and the shared feature vector to obtain a connection vector.

[0079] According to another embodiment of the present application, the activation module 404 includes a conversion submodule, which is used to perform nonlinear conversion on the connection vector to obtain an activation vector.

[0080] According to another embodiment of the present application, the activation module 404 also includes at least one of the following: a batch normalization submodule, which is used to batch normalize the connection vector before the conversion submodule performs nonlinear transformation on the connection vector to obtain the activation vector; and an attention submodule, which is used to process the connection vector using an attention mechanism before the conversion submodule performs nonlinear transformation on the connection vector to obtain the activation vector.

[0081] According to another embodiment of the present application, the prediction module 406 includes a connection submodule, which is used to fully connect the first feature vector and the second feature vector to obtain an output vector; and a probability submodule, which is used to perform nonlinear transformation on the output vector to obtain a probability.

[0082] According to another embodiment of the present application, the user information is the identity identifier of the target user; and / or the object information is the object identifier of the candidate object.

[0083] It should be noted that the above description of the embodiment of the recommended device is different from the above description of the embodiment of the recommended device. Figures 1 to 3 The description of the method embodiment shown is similar, with the same Figures 1 to 3 The method embodiments shown in the figure have similar beneficial effects, so they are not described in detail. For technical details not disclosed in the embodiments of the device recommended in this application, please refer to the aforementioned Figures 1 to 3 The description of the method embodiment shown in the figure can be understood, and in order to save space, it will not be repeated.

[0084] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0085] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0086] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0087] In addition, all functional units in the embodiments of the present application may be integrated into one processing unit, or each unit may be a separate unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0088] A person skilled in the art can understand that all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM), magnetic disks or optical disks, etc., various media that can store program codes.

[0089] Alternatively, if the above-mentioned integrated unit of the present application is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present application can essentially or in other words, the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods of each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0090] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which 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. A recommendation method, comprising: Obtain user information of the target user and object information of the candidate object; Performing low-order feature extraction on the user information and the object information to obtain a first feature vector; Performing high-order feature extraction on the user information and the object information to obtain a connection vector; Performing dimensionality reduction processing on the connection vector to obtain an activation vector; Perform residual fusion on the connection vector and the activation vector to obtain a second feature vector; Predicting, based on the first feature vector and the second feature vector, the probability of the target user selecting the candidate object; When the probability meets a set condition, recommending the candidate object to the target user; The extracting high-order features of the user information and the object information to obtain a connection vector includes: Using a neural network algorithm, feature extraction and initialization are performed on the user information and the object information, and accordingly, an initial user perception vector and an initial object perception vector are obtained respectively; Performing global matrix decomposition on the user initial perception vector and the object initial perception vector to obtain a shared feature vector; The user initial perception vector, the object initial perception vector and the shared feature vector are fully connected to obtain the connection vector.

2. According to the method of claim 1, the step of extracting low-order features from the user information and the object information to obtain a first feature vector comprises: Perform matrix decomposition on the user information and the object information to obtain the first feature vector.

3. The method according to claim 2, wherein the performing matrix decomposition on the user information and the object information to obtain the first feature vector comprises: Performing matrix decomposition initialization on the user information and the object information respectively, and correspondingly obtaining a user initial matrix vector and an object initial matrix vector respectively; Perform global matrix decomposition on the user initial matrix vector and the object initial matrix vector to obtain the first eigenvector.

4. According to the method of claim 1, the step of performing dimensionality reduction processing on the connection vector to obtain an activation vector comprises: The connection vector is nonlinearly transformed to obtain the activation vector.

5. The method according to claim 4, before performing nonlinear transformation on the connection vector to obtain the activation vector, performing dimensionality reduction processing on the connection vector to obtain the activation vector, further comprising at least one of the following: performing batch normalization on the connection vector; The connection vector is processed using an attention mechanism.

6. The method according to claim 1, wherein predicting the probability of the target user selecting the candidate object based on the first feature vector and the second feature vector comprises: Performing full connection on the first feature vector and the second feature vector to obtain an output vector; The output vector is nonlinearly transformed to obtain the probability.

7. The method according to any one of claims 1 to 6, The user information is the identity of the target user; and / or The object information is the object identifier of the candidate object.

8. A recommendation device, comprising: An acquisition module, used to acquire user information of a target user and object information of a candidate object; A first feature extraction module, used for performing low-order feature extraction on the user information and the object information to obtain a first feature vector; A second feature extraction module, used for performing high-order feature extraction on the user information and the object information to obtain a connection vector; The second feature extraction module is further used to use a neural network algorithm to perform feature extraction initialization on the user information and the object information, and correspondingly obtain a user initial perception vector and an object initial perception vector respectively; Performing global matrix decomposition on the user initial perception vector and the object initial perception vector to obtain a shared feature vector; Performing full connection on the user initial perception vector, the object initial perception vector and the shared feature vector to obtain the connection vector; An activation module, used for performing dimensionality reduction processing on the connection vector to obtain an activation vector; A residual module, used for performing residual fusion on the connection vector and the activation vector to obtain a second feature vector; A prediction module, configured to predict the probability of the target user selecting the candidate object according to the first feature vector and the second feature vector; The recommendation module is used to recommend the candidate object to the target user when the probability meets the set condition.

Images