CN104636486B - A kind of user characteristics abstracting method and draw-out device based on the conversion of non-negative alternating direction - Google Patents

Abstract

The invention proposes a user feature extraction method and extraction device based on non-negative alternating direction transformation. The extracting device includes a data receiving module, a data storage module and an execution module, wherein the data receiving module is connected to the data storage module, and the data receiving module is used to receive user behavior statistics data collected by the server, and collect the collected data from the server. The statistical data of user behavior is transmitted to the data storage module for storage, and the data storage module is connected with the execution module, and the execution module executes the user feature extraction instruction sent by the server, and stores the extracted user feature data into the data storage module. The present invention directly acts on the known data set in the user behavior statistical matrix, can process the extremely sparse user behavior statistical matrix with a large number of missing values, has fast convergence speed, high data restoration accuracy, and can solve problems in the big data processing environment User feature extraction problem.

Description

A user feature extraction method and extraction device based on non-negative alternating direction transformation

technical field

The invention relates to the technical field of computer big data processing, in particular to a method and device for extracting user features based on non-negative alternating direction transformation in an e-commerce system.

Background technique

Modern large-scale e-commerce systems have huge numbers of users and information. In this type of system, various objective behaviors of users, such as clicking, browsing, commenting, searching, etc., accumulate with the operating time of the system and aggregate into a huge user historical behavior data set. The data volume is at least in the TB level, which is a typical big data environment.

In a large-scale e-commerce system, a typical data description structure is a user behavior statistics matrix, in which each row corresponds to a user, and each column corresponds to an item; an item refers to any objective object in the system that may be operated by a user, such as news , pictures, commodities; each matrix element corresponds to the historical behavior data of a single user on a single item, which is composed of quantitative calculations using the user’s objective historical behavior data on the item, using mathematical statistical methods that conform to natural laws. In a large-scale e-commerce system, the number of users and items is very large, and the corresponding user behavior statistics matrix is also very large. At the same time, it is impossible for a user to operate all items, and it is impossible for an item to be operated by all users; generally speaking, the known data in the user behavior statistics matrix is far less than the unknown data, which is extremely sparse.

In the process of system operation, based on the known data in the user behavior statistics matrix, user characteristics are extracted from it, and user behavior can be effectively analyzed, and rules including user categories and behavior patterns can be excavated from them. In the process of extracting user features, maintaining the non-negativity of user features is a key, because non-negative user features are more in line with the natural law of positive user behavior data in e-commerce systems, and can better understand users. Behavioral representations. The existing non-negative feature extraction technology is mostly used in the field of computer vision. Its basic feature is that for a given graph or image, it is regarded as a full-rank matrix, and it is subjected to matrix factorization under the constraints of non-negative conditions, so that Extract the local object features of the graph or image. However, the problem of user feature extraction in e-commerce systems is quite different from the problem of non-negative object feature extraction in computer vision. This is because the graphics and images transformed by the non-negative object feature extraction in computer vision are full-rank matrices and do not have missing values. The non-negative matrix factorization of such matrices can be solved by conventional matrix iterative operations. However, for the non-negative user behavior extraction problem in e-commerce systems, the user behavior statistical matrix processed is usually extremely sparse, with a large number of missing values, which cannot be processed by traditional matrix factorization, and It needs to be processed with non-negative latent feature analysis that can operate on sparse matrices. However, the existing non-negative matrix latent feature analysis method has the disadvantages of slow convergence speed and low accuracy of data restoration.

Therefore, for the user behavior statistical matrix with a large number of missing values in large-scale e-commerce systems, how to perform non-negative latent feature analysis with fast convergence speed and high data restoration accuracy, so as to obtain user characteristics that can well describe the natural laws of user behavior , is a key issue that needs to be dealt with when analyzing the massive data generated by modern large-scale e-commerce systems.

Contents of the invention

In order to overcome the defects in the above-mentioned prior art, the object of the present invention is to provide a user feature extraction method and extraction device based on non-negative alternating direction transformation. The present invention directly acts on the known data set in the user behavior statistics matrix, It can handle extremely sparse user behavior statistical matrices with a large number of missing values, has fast convergence speed and high data restoration accuracy, and can solve the user feature extraction problem in the big data processing environment.

In order to achieve the above object of the present invention, the present invention provides a user feature extraction method based on non-negative alternating direction transformation, comprising the following steps:

S1. The server sends an instruction to the extraction device to extract user features;

S2. The extraction device receives instructions and initializes parameters. The initialization parameters include: feature space dimension f, dual learning rate η, Lagrangian enhancement factor λ, user feature matrix X, user training auxiliary matrices X_U, X_D and X_C, item features Matrix Y, project training auxiliary matrices Y_U, Y_D and Y_C, iteration control variable t, iteration upper limit n, convergence judgment threshold

S3. The extraction device constructs a cumulative absolute error ε(P, Q, X, Y), where P is the user feature constraint matrix, and Q is the item feature constraint matrix;

S4. The extraction device uses constraints to constrain the cumulative absolute error ε(P, Q, X, Y) to ensure the non-negativity of the parameters of the matrices P and Q during the training process;

S5. The extraction device constructs a unified loss function L (P, Q, X, Y, Γ, Κ), wherein Γ and Κ are dual parameters;

S6. The extraction device judges whether the iterative training control variable t has reached the upper limit n, if so, execute step S9, if not, execute step S7;

S7. The extraction device judges whether the unified loss function L converges on the known data set C in the user behavior statistics matrix with respect to P, Q, X, Y, Γ and K, if so, then execute step S9, if not, then execute Step S8;

S8. The extraction device performs iterative training on P, Q, X, Y, Γ and K on the known data in the known data set C in the user behavior statistics matrix, and then executes step S6;

S9. The extraction device outputs the user feature matrix X and item feature matrix Y obtained through iterative training, and stores them in the acquired feature storage unit in the data module.

In this method, the feature space dimension f in step S2 is the dimension of the feature space where the user feature is located, which determines the dimension of the feature vector, and is any positive integer in the set of positive real numbers.

The dual learning rate η is the learning rate for training the Lagrangian multipliers in the unified loss function, which is a floating-point number in the interval (0.0001, 0.05).

The Lagrangian enhancement factor λ is a factor that expresses the constraints in the unified loss function, and it is any decimal in the interval (0.01, 0.1).

The user feature matrix X is the feature to be extracted, which is a matrix of |A|×f, where A represents the set of users stored in the storage unit of the device. Each row of X corresponds to a user, and each row vector of X is a feature vector of a user. In the embodiment of the present invention, the initial value of each element in the user characteristic matrix X is set as a random number within the range of the open interval (0.4, 0.8).

User training auxiliary matrices X_U, X_D, and X_C are data structures used to assist iterative training of user features, all of which are |A|×f matrices.

The item feature matrix Y is the feature to be extracted, which is a |B|×f matrix, where B represents the item set stored in the storage unit of the device. Each row of Y corresponds to an item, and each row vector of Y is a feature vector of all users operating on an item.

Item training auxiliary matrices Y_U, Y_D, and Y_C are data structures used to assist iterative training of user features, all of which are |B|×f matrices.

The iteration control variable t is a variable that controls the feature training process, and the iteration control variable t is initialized to 0.

The iteration upper limit n is a variable that controls the iteration upper limit of the training process, and is any positive integer in the set of positive real numbers.

Convergence Decision Threshold It is the threshold parameter for judging whether the iterative training has converged.

This method directly acts on the known data set in the user behavior statistical matrix, and can handle the extremely sparse user behavior statistical matrix with a large number of missing values. User feature extraction problem.

Preferably, the formula for calculating the absolute error described in step S3 is:

s.t.P=X,P≥0,

Q=Y, Q≥0.

Among them, C represents the known data set in the user behavior statistics matrix; r _u,i represents the element value of row u and column i in the user behavior statistics matrix, representing the historical behavior statistics data of user u on item i; x _{u, k} represent the u-th row and k-th column element of the user feature matrix X; y _{i, k} represent the i-th row and k-th column element of the item feature matrix Y; P is the user feature constraint matrix, and Q is the item feature constraint matrix .

Step S3 constructs the cumulative absolute error ε(P, Q, X, Y) to fully express the error and non-negativity constraints, and at the same time, provides conditions for the introduction of dual parameters.

Step S4 uses constraint conditions to constrain the cumulative absolute error ε(P, Q, X, Y) to ensure the non-negativity of relevant model parameters during the training process.

The construction of the unified loss function in step S5 is to unify the loss function and related constraints by using the Lagrange multiplier method, so that the constraints of the constraints are satisfied during the training process.

Preferably, step S4 includes the following steps:

S4-1. For each element p _u,k in P, if it is not equal to the corresponding element x _u,k in X, then let p _u,k =x _u,k ;

S4-2. For each element q _i,k in Q, if it is not equal to the corresponding element y _i,k in Y, then let q _i,k =y _i,k ;

S4-3. For each element p _u,k in P, if it is less than 0, set p _u,k =0;

S4-4. For each element q _i,k in Q, if it is less than 0, set q _i,k =0.

Among them, p _{u, k} represent elements in row u and column k in the user feature constraint matrix P, and q _{i, k} represent elements in row i and column k in the item feature constraint matrix.

Preferably, the calculation formula of the loss function described in step S5 is:

Wherein Γ and Κ are dual parameters, γ _{u, k} represent elements in row u and column k in Γ, and κ _{i, k} represent elements in row i and column k in K, and this formula adopts the reduced Lagrang The Augmented Lagrangian method, the Augmented Lagrangian method is based on the Lagrangian multiplier method, adding a statute item corresponding to the restriction condition, and the statute item is ρ is the reduction parameter of the reduced Lagrangian multiplier method, which refers to the matrix X and is a constant in the calculation.

Preferably, the iterative training in step S8 includes the following steps:

S8-1. Determine the iterative training target, that is, all parameters P, Q, X, Y, Γ and Κ, so that it meets the unified loss function L relative to P, Q, X, Y, Γ and Κ in the user behavior statistics matrix The smallest known data set C, expressed as a formula:

τ is the reduction parameter of the reduced Lagrange multiplier method, It is a specification item, this parameter refers to the matrix Y, and it is a constant in the calculation.

S8-2. use non-negative direction alternating transformation, carry out sequential training to the single element in P, Q, X, Y, Γ and Κ, training rule is expressed as formula:

for k=1～f,

S8-3, for each element in P, Q, X, Y, Γ and Κ, perform it according to the following formula

ρ _u =λ|C(u)|,τ _i =λ|C(i)|;

training updates, Among them, C(u) and C(i) respectively represent the subsets associated with user u and item i in the known data set C, and τ in τ _i is the reduction parameter of the reduced Lagrangian multiplier method, the The parameter refers to the matrix Y, and i corresponds to the ith row of Y.

The present invention also proposes an extraction device based on a user feature extraction method based on non-negative alternating direction transformation, including a data receiving module, a data storage module and an execution module, wherein the data receiving module is connected to the data storage module, and the data receiving module uses To receive the user behavior statistics data collected by the server, and transfer the received user behavior statistics data collected by the server to the data storage module for storage, the data storage module is connected to the execution module, and the execution module executes the user characteristic sent by the server extracting instructions, and storing the extracted user feature data into the data storage module.

The data receiving module is used to collect user behavior statistics data of the server, the execution module is used to execute user feature extraction instructions, and the data storage module stores the server user behavior statistics data collected by the data receiving module and the user feature data extracted by the execution module. The device can directly act on the known data set in the user behavior statistical matrix, can process the extremely sparse user behavior statistical matrix with a large number of missing values, and can solve the user feature extraction problem in the big data processing environment.

Further, the data storage module includes an acquisition feature storage unit and a statistical data storage unit, the acquisition feature storage unit is connected to the execution module, and is used to store user characteristic data extracted by the execution module; the statistical data storage unit and The data receiving module is connected to store user behavior statistical data delivered by the data receiving module.

The user behavior statistics data of the server collected by the data receiving module and the user characteristic data extracted by the execution module are stored in units, so that it is more convenient, accurate and faster to retrieve data.

Further, the execution module includes an initialization unit, a training unit and an output unit,

The initialization unit initializes the parameters that the user feature extraction process depends on, and the initialization parameters include: feature space dimension f, dual learning rate η, Lagrangian enhancement factor λ, user feature matrix X, user training auxiliary matrix X_U, X_D and X_C, project characteristic matrix Y, project training auxiliary matrix Y_U, Y_D and Y_C, iteration control variable t, iteration upper limit n, convergence judgment threshold

The input end of the training unit is connected with the initialization unit and the data storage module respectively, according to the parameters initialized by the initialization unit and the user behavior statistics data in the data storage module to construct user characteristic data, including user characteristic matrix X and item characteristic matrix Y, training unit First construct the cumulative absolute error ε(P,Q,X,Y), where P is the user feature constraint matrix, Q is the item feature constraint matrix, and then construct a unified loss function L(P,Q,X,Y,Γ,K) , where Γ and Κ are dual parameters, and then iteratively train P, Q, X, Y, Γ and Κ until the unified loss function L(P, Q, X, Y, Γ, K) is relative to P, Q , X, Y, Γ and Κ converge on the known data set C in the user behavior statistics matrix, or the iteration control variable t is equal to the iteration upper limit n;

The input end of the output unit is connected to the output end of the training unit, the output end of the output unit is connected to the data storage module, and the output unit outputs and stores the user characteristic data constructed by the training unit into the data storage module.

The execution module is divided into three units, so that when extracting user characteristic data, initialization of parameters, extraction and storage of data can be more accurate and quicker.

The beneficial effects of the present invention are: the method aims to directly act on the known data set in the user behavior statistical matrix through non-negative alternating direction transformation, and has the following advantages:

1. Able to handle extremely sparse user behavior statistics matrix with a large number of missing values;

2. The convergence speed is fast, and the data restoration accuracy is high, which can solve the user feature extraction problem in the big data processing environment.

Additional aspects and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.

Description of drawings

The above and/or additional aspects and advantages of the present invention will become apparent and comprehensible from the description of the embodiments in conjunction with the following drawings, wherein:

Fig. 1 is a structural representation of the present invention;

Fig. 2 is a schematic flow chart of the present invention;

Fig. 3 is before and after applying the embodiment of the present invention, the comparison diagram of the convergence speed of extracting user features;

Fig. 4 is a comparison chart of data restoration accuracy for extracting user features before and after applying the embodiment of the present invention.

detailed description

Embodiments of the present invention are described in detail below, examples of which are shown in the drawings, wherein the same or similar reference numerals designate the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the figures are exemplary only for explaining the present invention and should not be construed as limiting the present invention.

In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "connection" should be understood in a broad sense, for example, it can be mechanical connection or electrical connection, or two The internal communication of each element may be directly connected or indirectly connected through an intermediary. Those skilled in the art can understand the specific meanings of the above terms according to specific situations.

As shown in Figure 1, the present invention provides a kind of user feature extraction method based on non-negative alternating direction transformation, comprising the following steps:

S1. The server sends an instruction to the extraction device to extract user features;

S2. The extraction device receives instructions and initializes parameters. The initialization parameters include: feature space dimension f, dual learning rate η, Lagrangian enhancement factor λ, user feature matrix X, user training auxiliary matrices X_U, X_D and X_C, item features Matrix Y, project training auxiliary matrices Y_U, Y_D and Y_C, iteration control variable t, iteration upper limit n, convergence judgment threshold

Among them, the feature space dimension f is the dimension of the feature space where the user feature is located, which determines the dimension of the feature vector, and is any positive integer in the set of positive real numbers, such as 25.

The dual learning rate η is the learning rate for training the Lagrangian multipliers in the unified loss function, which is a floating-point number in the interval (0.0001, 0.05), such as 0.001.

The Lagrangian enhancement factor λ is a factor that expresses the constraints in the unified loss function, which is any decimal in the interval (0.01, 0.1), such as 0.05.

The user feature matrix X is the feature to be extracted, which is a matrix of |A|×f, where A represents the set of users stored in the storage unit of the device. Each row of X corresponds to a user, and each row vector of X is a feature vector of a user. In the embodiment of the present invention, the initial value of each element in the user feature matrix X is set to a random number within the open interval (0.4, 0.8), such as 0.47.

User training auxiliary matrices X_U, X_D, and X_C are data structures used to assist iterative training of user features, all of which are |A|×f matrices, where X_U is used to cache the initial value of X matrix during training, and X_D is used to Cache the target value of the X matrix during the training process, and X_C is used to cache the update value of the X matrix during the training process. In the embodiment of the present invention, the initialization of each element in X_U, X_D and X_C is 0.

The item feature matrix Y is the feature to be extracted, which is a |B|×f matrix, where B represents the item set stored in the storage unit of the device. Each row of Y corresponds to an item, and each row vector of Y is a feature vector of all users operating on an item.

Project training auxiliary matrices Y_U, Y_D, and Y_C are data structures used to assist iterative training of user features, all of which are |B|×f matrices, where Y_U is used to cache the initial value of the Y matrix during the training process, and Y_D is used to Cache the target value of the Y matrix during the training process, and Y_C is used to cache the update value of the Y matrix during the training process. In the embodiment of the present invention, the initialization of each element in Y_U, Y_D and Y_C is 0.

The iteration control variable t is a variable that controls the feature training process, and the iteration control variable t is initialized to 0.

The iteration upper limit n is a variable that controls the iteration upper limit of the training process, which is any positive integer in the positive real number set, such as 500.

Convergence Decision Threshold It is the threshold parameter for judging whether the iterative training has converged. In the embodiment of the present invention, it is set to any decimal within the open interval (0,0.001), such as 0.0001.

S3. The extraction device constructs a cumulative absolute error ε(P, Q, X, Y), where P is the user feature constraint matrix, and Q is the item feature constraint matrix;

The formula for calculating the absolute error described in this step is:

s.t.P=X,P≥0,

Q=Y, Q≥0.

Among them, C represents the known data set in the user behavior statistics matrix; r _u,i represents the element value of row u and column i in the user behavior statistics matrix, representing the historical behavior statistics data of user u on item i; x _{u, k} represent the u-th row and k-th column element of the user feature matrix X; y _{i, k} represent the i-th row and k-th column element of the item feature matrix Y; P is the user feature constraint matrix, and Q is the item feature constraint matrix .

S4. The extraction device constrains the cumulative absolute error ε(P, Q, X, Y) using constraint conditions, where P=X, P≥0, Q=Y, and Q≥0 are all constraint conditions.

S5. The extraction device constructs a unified loss function L (P, Q, X, Y, Γ, Κ), wherein Γ and Κ are dual parameters;

The calculation formula of the loss function described in this step is:

S6. The extraction device judges whether the iterative training control variable t has reached the upper limit n,

If the upper limit has been reached, then step S9 is performed: the user feature matrix X and the item feature matrix Y output obtained by the iterative training are output by the extraction device, stored in the acquisition feature storage unit in the data module, and the extraction of the user features is completed;

If the upper limit is not reached, step S7 is executed;

In this step, the extraction device first accumulates 1 on the iteration control variable t, and then judges whether the iteration control variable t is greater than the iteration upper limit n.

S7. The extraction device judges whether the unified loss function L converges on the known data set C in the user behavior statistics matrix with respect to P, Q, X, Y, Γ and Κ,

If so, then perform step S9: the extraction device will output the user feature matrix X and the item feature matrix Y obtained through iterative training, and store them in the acquisition feature storage unit in the data module to complete the extraction of user features;

If not, execute step S8;

In this step, the basis for the user feature extraction device to judge whether the unified loss function L has converged with respect to the user feature matrix X, item feature matrix Y, and user features on the known data set C in the user behavior statistics matrix is that the current round of iteration Before the training starts, the value of the unified loss function L is compared with the value of the unified loss function L before the start of the last round of iterative training. Whether the absolute value of the difference is less than the convergence judgment threshold If it is less than , it is judged to have converged, and vice versa.

S8. The extraction device performs iterative training on P, Q, X, Y, Γ and K on the known data in the known data set C in the user behavior statistics matrix, and then performs step S6, and so on, until step S9 is completed The extraction device outputs the user feature matrix X and item feature matrix Y obtained through iterative training, and stores them in the acquired feature storage unit in the data module to complete the extraction of user features.

As a preferred solution of this embodiment, step S4 includes the following steps:

S4-1. For each element p _u,k in P, if it is not equal to the corresponding element x _u,k in X, then let p _u,k =x _u,k ;

S4-2. For each element q _i,k in Q, if it is not equal to the corresponding element y _i,k in Y, then let q _i,k =y _i,k ;

S4-3. For each element p _u,k in P, if it is less than 0, set p _u,k =0;

S4-4. For each element q _i,k in Q, if it is less than 0, set q _i,k =0.

Among them, p _u,k represents the u-th row and k-th column element in the user characteristic constraint matrix P; x _u,k represents the u-th row and k-th column element in the user characteristic matrix X; q _i,k represents the item characteristic constraint matrix The element in row i and column k in Q; y _i,k represents the element in row i and column k of item feature matrix Y.

The iterative training in step S8 comprises the following steps:

S8-1. Determine the iterative training target, that is, all parameters P, Q, X, Y, Γ and Κ, solve the parameters P, Q, X, Y, Γ and Κ, so that it satisfies the unified loss function L relative to P , Q, X, Y, Γ and Κ are the smallest on the known data set C in the user behavior statistics matrix, expressed as a formula:

S8-2. use non-negative direction alternating transformation, carry out sequential training to the single element in P, Q, X, Y, Γ and Κ, training rule is expressed as formula:

for k=1～f,

Among them, t and t+1 represent the t-th round and the t+1-th round of iterations, respectively.

S8-3, for each element in P, Q, X, Y, Γ and Κ, perform it according to the following formula

ρ _u =λ|C(u)|,τ _i =λ|C(i)|;

training updates,

Among them, C(u) and C(i) represent the subsets associated with user u and item i in the known data set C, respectively.

According to this method, P, Q, X, Y, Γ and K are iteratively trained, and then step S6 is repeated, and so on, until the extraction of user features is completed.

The present invention also proposes an extraction device based on a user feature extraction method based on non-negative alternating direction transformation, as shown in Figure 2, including a data receiving module, a data storage module and an execution module, wherein the data receiving module and the data storage module connection, the data receiving module is used to receive the user behavior statistical data collected by the server, and transfer the received user behavior statistical data collected by the server to the data storage module for storage, the data storage module is connected to the execution module, and the execution module executes The server sends an instruction to extract user features, and stores the extracted user feature data into the data storage module.

The data receiving module is used to collect user behavior statistics data of the server, the execution module is used to execute user feature extraction instructions, and the data storage module stores the server user behavior statistics data collected by the data receiving module and the user feature data extracted by the execution module. The device can directly act on the known data set in the user behavior statistical matrix, can process the extremely sparse user behavior statistical matrix with a large number of missing values, and can solve the user feature extraction problem in the big data processing environment.

In this embodiment, the data storage module includes an acquisition feature storage unit and a statistical data storage unit, and the acquisition feature storage unit is connected to the execution module for storing user feature data extracted by the execution module; the statistical data storage The unit is connected with the data receiving module, and is used for storing user behavior statistics data delivered by the data receiving module.

The user behavior statistics data of the server collected by the data receiving module and the user characteristic data extracted by the execution module are stored in units, so that it is more convenient, accurate and faster to retrieve data.

As a preferred solution of this embodiment, the execution module includes an initialization unit, a training unit and an output unit.

The initialization unit initializes the parameters that the user feature extraction process depends on, and the initialization parameters include: feature space dimension f, dual learning rate η, Lagrangian enhancement factor λ, user feature matrix X, user training auxiliary matrix X_U, X_D and X_C, project characteristic matrix Y, project training auxiliary matrix Y_U, Y_D and Y_C, iteration control variable t, iteration upper limit n, convergence judgment threshold Among them, the user feature matrix X, user training auxiliary matrices X_U, X_D and X_C are matrixes of |A| rows and |f| columns established according to the current user set A and the current feature space dimension f; the user feature matrix X The initial value of each element is a random number within the interval (0.2,0.6), and the initial value of each element in the user training auxiliary matrices X_U, X_D and X_C is 0. Item feature matrix Y, item training auxiliary matrix Y_U, Y_D, and Y_C are matrices of |B| rows and |f| columns established according to the current item set B and the current feature space dimension f; each item feature matrix X The initial value of the element is a random number within the interval (0.2,0.6), and the initial value of each element in the item training auxiliary matrix Y_U, Y_D and Y_C is 0.

The input end of the training unit is connected with the initialization unit and the data storage module respectively, and constructs user characteristic data according to the parameters initialized by the initialization unit and the user behavior statistics data in the data storage module, including user characteristic matrix X and item characteristic matrix Y, among X Each row vector in Y corresponds to a user's non-negative behavior feature; each row vector in Y corresponds to the known non-negative operation feature of all users for an item. Training and constructing user feature data further includes that the training unit first constructs the cumulative absolute error ε(P, Q, X, Y), where P is the user feature constraint matrix, Q is the item feature constraint matrix, and then uses the enhanced Lagrangian multiplier method to construct a unified loss function L(P, Q, X, Y, Γ, K), where Γ and Κ are dual parameters, and solve the related parameters P, Q, X, Y, Γ and Κ, so that the global loss The function is minimized on the known data set C in the user behavior statistics matrix, and then uses non-negative alternating direction transformation to iteratively train P, Q, X, Y, Γ and Κ until the unified loss function L(P, Q, X, Y, Γ, K) converge on the known data set C in the user behavior statistics matrix with respect to P, Q, X, Y, Γ and Κ, or the iteration control variable t is equal to the iteration upper limit n;

The input of the output unit is connected with the output of the training unit, and the output of the output unit is connected with the data storage module, and the output unit will output the user characteristic data constructed by the training unit, including the user characteristic matrix X and the item characteristic matrix Y, and output and stored in the data storage module.

In the specific implementation, the example analysis uses the number of training iterations as an index to measure the convergence speed of user feature extraction. The fewer the number of training iterations, the faster the convergence speed of user feature extraction; the mean absolute error MAE is used as the data for user feature extraction The indicator of restoration accuracy, the lower the mean absolute error MAE, the higher the accuracy of data restoration for user feature extraction.

FIG. 3 is a comparison of the convergence speed of extracting user features before and after applying this embodiment. After applying the embodiment of the present invention, when user features are extracted under non-negative constraints, the number of iterations is significantly reduced, and the convergence speed is significantly improved.

FIG. 4 is a comparison of data restoration accuracy of extracted user features before and after applying this embodiment. After applying the embodiment of the present invention, when user features are extracted under non-negative constraints, the mean absolute error MAE is significantly reduced, and the accuracy of data restoration is significantly improved.

In the description of this specification, descriptions with reference to the terms "one embodiment", "some embodiments", "example", "specific examples", or "some examples" mean that specific features described in connection with the embodiment or example , structure, material or feature is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

Although the embodiments of the present invention have been shown and described, those skilled in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the invention is defined by the claims and their equivalents.

Claims (8)

1. A user feature extraction method based on non-negative alternating direction transformation, characterized in that, comprising the following steps: S1. The server sends an instruction to the extraction device to extract user features; S2. The extraction device receives instructions and sets initialization parameters. The initialization parameters include: feature space dimension f, dual learning rate η, Lagrangian enhancement factor λ, user feature matrix X, user training auxiliary matrices X_U, X_D and X_C, items Feature matrix Y, project training auxiliary matrix Y_U, Y_D and Y_C, iteration control variable t, iteration upper limit n, convergence judgment threshold S3. The extraction device constructs a cumulative absolute error ε(P, Q, X, Y), where P is the user feature constraint matrix, and Q is the item feature constraint matrix; S4. The extraction device uses constraints to constrain the cumulative absolute error ε(P, Q, X, Y) to ensure the non-negativity of the parameters of the matrices P and Q during the training process; S5. The extraction device constructs a unified loss function L(P, Q, X, Y, Γ, K), where Γ and K are dual parameters; S6. The extraction device judges whether the iterative training control variable t has reached the upper limit n, if so, execute step S9, if not, execute step S7; S7. The extraction device judges whether the unified loss function L converges on the known data set C in the user behavior statistics matrix with respect to P, Q, X, Y, Γ and K, if so, execute step S9, if not, execute Step S8; S8. The extraction device performs iterative training on P, Q, X, Y, Γ and K on the known data in the known data set C in the user behavior statistics matrix, and then executes step S6; S9. The extraction device outputs the user feature matrix X and item feature matrix Y obtained through iterative training, and stores them in the acquired feature storage unit in the data module. 2. the user feature extraction method based on non-negative alternating direction transformation according to claim 1, is characterized in that, the computing formula of absolute error described in step S3 is: <mrow><mi>&amp;epsiv;</mi><mrow><mo>(</mo><mi>P</mi><mo>,</mo><mi>Q</mi><mo>,</mo><mi>X</mi><mo>,</mo><mi>Y</mi><mo>)</mo></mrow><mo>=</mo><munder><mo>&amp;Sigma;</mo><mrow><mo>(</mo><mi>u</mi><mo>,</mo><mi>i</mi><mo>)</mo><mo>&amp;Element;</mo><mi>C</mi></mrow></munder><msup><mrow><mo>(</mo><msub><mi>r</mi><mrow><mi>u</mi><mo>,</mo><mi>i</mi></mrow></msub><mo>-</mo><munderover><mo>&amp;Sigma;</mo><mrow><mi>k</mi><mo>=</mo><mn>1</mn></mrow><mi>f</mi></munderover><msub><mi>x</mi><mrow><mi>u</mi><mo>,</mo><mi>k</mi></mrow></msub><msub><mi>y</mi><mrow><mi>i</mi><mo>,</mo><mi>k</mi></mrow></msub><mo>)</mo></mrow><mn>2</mn></msup><mo>,</mo></mrow> s.t.P=X,P≥0, Q=Y,Q≥0. Among them, C represents the known data set in the user behavior statistics matrix; r _u,i represents the element value of row u and column i in the user behavior statistics matrix, representing the historical behavior statistics data of user u on item i; x _{u, k} represent the u-th row and k-th column element of the user feature matrix X; y _{i, k} represent the i-th row and k-th column element of the item feature matrix Y; P is the user feature constraint matrix, and Q is the item feature constraint matrix . 3. the user feature extraction method based on non-negative alternating direction transformation according to claim 1, is characterized in that, step S4 comprises the following steps: S4-1. For each element p _u,k in P, if it is not equal to the corresponding element x _u,k in X, then let p _u,k =x _u,k ; S4-2. For each element q _i,k in Q, if it is not equal to the corresponding element y _i,k in Y, then let q _i,k =y _i,k ; S4-3. For each element p _u,k in P, if it is less than 0, set p _u,k =0; S4-4. For each element q _i,k in Q, if it is less than 0, set q _i,k =0. 4. The user feature extraction method based on non-negative alternating direction transformation according to claim 1, wherein the unified loss function calculation formula described in step S5 is: <mrow><mtable><mtr><mtd><mrow><mi>L</mi><mrow><mo>(</mo><mi>P</mi><mo>,</mo><mi>Q</mi><mo>,</mo><mi>X</mi><mo>,</mo><mi>Y</mi><mo>,</mo><mi>&amp;Gamma;</mi><mo>,</mo><mi>K</mi><mo>)</mo></mrow><mo>=</mo><mfrac><mn>1</mn><mn>2</mn></mfrac><munder><mo>&amp;Sigma;</mo><mrow><mo>(</mo><mi>u</mi><mo>,</mo><mi>i</mi><mo>)</mo><mo>&amp;Element;</mo><mi>C</mi></mrow></munder><msup><mrow><mo>(</mo><msub><mi>r</mi><mrow><mi>u</mi><mo>,</mo><mi>i</mi></mrow></msub><mo>-</mo><munderover><mo>&amp;Sigma;</mo><mrow><mi>k</mi><mo>=</mo><mn>1</mn></mrow><mi>f</mi></munderover><msub><mi>x</mi><mrow><mi>u</mi><mo>,</mo><mi>k</mi></mrow></msub><msub><mi>y</mi><mrow><mi>i</mi><mo>,</mo><mi>k</mi></mrow></msub><mo>)</mo></mrow><mn>2</mn></msup><mo>+</mo><munder><mo>&amp;Sigma;</mo><mrow><mo>(</mo><mi>u</mi><mo>,</mo><mi>k</mi><mo>)</mo></mrow></munder><msub><mi>&amp;gamma;</mi><mrow><mi>u</mi><mo>,</mo><mi>k</mi></mrow></msub><mrow><mo>(</mo><msub><mi>x</mi><mrow><mi>u</mi><mo>,</mo><mi>k</mi></mrow></msub><mo>-</mo><msub><mi>p</mi><mrow><mi>u</mi><mo>,</mo><mi>k</mi></mrow></msub><mo>)</mo></mrow></mrow></mtd></mtr><mtr><mtd><mrow><mo>+</mo><munder><mo>&amp;Sigma;</mo><mrow><mo>(</mo><mi>i</mi><mo>,</mo><mi>k</mi><mo>)</mo></mrow></munder><msub><mi>&amp;kappa;</mi><mrow><mi>i</mi><mo>,</mo><mi>k</mi></mrow></msub><mrow><mo>(</mo><msub><mi>y</mi><mrow><mi>i</mi><mo>,</mo><mi>k</mi></mrow></msub><mo>-</mo><msub><mi>q</mi><mrow><mi>i</mi><mo>,</mo><mi>k</mi></mrow></msub><mo>)</mo></mrow><mo>+</mo><mfrac><mi>&amp;rho;</mi><mn>2</mn></mfrac><munder><mo>&amp;Sigma;</mo><mrow><mo>(</mo><mi>u</mi><mo>,</mo><mi>k</mi><mo>)</mo></mrow></munder><msup><mrow><mo>(</mo><msub><mi>x</mi><mrow><mi>u</mi><mo>,</mo><mi>k</mi></mrow></msub><mo>-</mo><msub><mi>p</mi><mrow><mi>u</mi><mo>,</mo><mi>k</mi></mrow></msub><mo>)</mo></mrow><mn>2</mn></msup><mo>+</mo><mfrac><mi>&amp;tau;</mi><mn>2</mn></mfrac><munder><mo>&amp;Sigma;</mo><mrow><mo>(</mo><mi>i</mi><mo>,</mo><mi>k</mi><mo>)</mo></mrow></munder><msup><mrow><mo>(</mo><msub><mi>y</mi><mrow><mi>i</mi><mo>,</mo><mi>k</mi></mrow></msub><mo>-</mo><msub><mi>q</mi><mrow><mi>i</mi><mo>,</mo><mi>k</mi></mrow></msub><mo>)</mo></mrow><mn>2</mn></msup></mrow></mtd></mtr></mtable><mo>,</mo></mrow> Among them, Γ and K are dual parameters, r _{u, i} represent the element value of row u and column i in the user behavior statistics matrix, representing the historical behavior statistics data of user u on item i; x _{u, k} represent user characteristics The u-th row and k-th column element of matrix X; y _i,k represents the i-th row and k-th column element of the item feature matrix Y; γ _u,k represents the u-th row and k-th column element in Γ, κ _{i, k} represents the i-th row and the k-th column element in K. This formula uses the reduced Lagrangian multiplier method, which adds the corresponding The statute of limitations, the statute is τ is the reduction parameter of the reduced Lagrange multiplier method, p _u,k represents the u-th row and k-th column element in the user characteristic constraint matrix P; q _i,k represents the i-th row, the item characteristic constraint matrix Q The k column elements, ρ and τ are the reduction parameters of the reduction Lagrangian multiplier method. 5. the user feature extraction method based on non-negative alternating direction transformation according to claim 1, is characterized in that, iterative training in step S8 comprises the following steps: S8-1. Determine the iterative training target, that is, all parameters P, Q, X, Y, Γ and K, so that it satisfies the unified loss function L relative to P, Q, X, Y, Γ and K in the user behavior statistics matrix The smallest known data set C, expressed as a formula: Among them, r _{u, i} represent the element value of row u and column i in the user behavior statistics matrix, representing the historical behavior statistics data of user u on item i; x _{u, k} represents the uth row of user feature matrix X, k-th column element; y _{i, k} means the i-th row and k-th column element of the item feature matrix Y; γ _{u, k} means the u-th row and k-th column element in Γ; κ _{i, k} means the i-th row in K , the k-th column element, the formula uses the reduced Lagrange multiplier method, the reduced Lagrangian multiplier method is based on the Lagrange multiplier method, adding the corresponding restriction items, the reduced Item is τ is the reduction parameter of the reduced Lagrange multiplier method, p _u,k represents the u-th row and k-th column element in the user characteristic constraint matrix P; q _i,k represents the i-th row, the item characteristic constraint matrix Q K column elements, ρ and τ are the reduction parameters of the reduction Lagrangian multiplier method; S8-2. Use non-negative direction alternating transformation to carry out sequential training to the single element in P, Q, X, Y, Γ and K, and the training rule is expressed as a formula: for k=1～f, <mfenced open = "{" close = ""><mtable><mtr><mtd><mrow><msubsup><mi>X</mi><mrow><mo>,</mo><mi>k</mi></mrow><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msubsup><mo>:</mo><munder><mrow><mi>arg</mi><mi>min</mi></mrow><msub><mi>X</mi><mrow><mo>.</mo><mo>,</mo><mi>k</mi></mrow></msub></munder><mi>L</mi><mrow><mo>(</mo><msup><mi>P</mi><mi>t</mi></msup><mo>,</mo><msup><mi>Q</mi><mi>t</mi></msup><mo>,</mo><msubsup><mi>X</mi><mrow><mo>,</mo><mn>1</mn><mo>~</mo><mi>k</mi><mo>-</mo><mn>1</mn></mrow><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msubsup><mo>,</mo><msub><mi>X</mi><mrow><mo>,</mo><mi>k</mi></mrow></msub><mo>,</mo><msubsup><mi>X</mi><mrow><mo>,</mo><mi>k</mi><mo>+</mo><mn>1</mn><mo>~</mo><mi>f</mi></mrow><mi>t</mi></msubsup><mo>,</mo><msubsup><mi>Y</mi><mrow><mo>,</mo><mn>1</mn><mo>~</mo><mi>k</mi><mo>-</mo><mn>1</mn></mrow><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msubsup><mo>,</mo><msubsup><mi>Y</mi><mrow><mo>,</mo><mi>k</mi><mo>~</mo><mi>f</mi></mrow><mi>t</mi></msubsup><mo>,</mo><msup><mi>&amp;Gamma;</mi><mi>t</mi></msup><mo>,</mo><msup><mi>K</mi><mi>t</mi></msup><mo>)</mo></mrow></mrow></mtd></mtr><mtr><mtd><mrow><msubsup><mi>Y</mi><mrow><mo>,</mo><mi>k</mi></mrow><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msubsup><mo>:</mo><munder><mrow><mi>arg</mi><mi>min</mi></mrow><msub><mi>Y</mi><mrow><mo>.</mo><mo>,</mo><mi>k</mi></mrow></msub></munder><mi>L</mi><mrow><mo>(</mo><msup><mi>P</mi><mi>t</mi></msup><mo>,</mo><msup><mi>Q</mi><mi>t</mi></msup><mo>,</mo><msubsup><mi>X</mi><mrow><mo>,</mo><mn>1</mn><mo>~</mo><mi>k</mi><mo>-</mo><mn>1</mn></mrow><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msubsup><mo>,</mo><msubsup><mi>X</mi><mrow><mo>,</mo><mi>k</mi><mo>~</mo><mi>f</mi></mrow><mi>t</mi></msubsup><mo>,</mo><msubsup><mi>Y</mi><mrow><mo>,</mo><mn>1</mn><mo>~</mo><mi>k</mi><mo>-</mo><mn>1</mn></mrow><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msubsup><mo>,</mo><msub><mi>Y</mi><mrow><mo>,</mo><mi>k</mi></mrow></msub><mo>,</mo><msubsup><mi>Y</mi><mrow><mo>,</mo><mi>k</mi><mo>+</mo><mn>1</mn><mo>~</mo><mi>f</mi></mrow><mi>t</mi></msubsup><mo>,</mo><msup><mi>&amp;Gamma;</mi><mi>t</mi></msup><mo>,</mo><msup><mi>K</mi><mi>t</mi></msup><mo>)</mo></mrow></mrow></mtd></mtr></mtable></mfenced> <mrow><mfenced open = "{" close = ""><mtable><mtr><mtd><mrow><msup><mi>P</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>=</mo><munder><mi>argmin</mi><mi>P</mi></munder><mi>L</mi><mrow><mo>(</mo><mi>P</mi><mo>,</mo><msup><mi>Q</mi><mi>t</mi></msup><mo>,</mo><msup><mi>X</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>,</mo><msup><mi>Y</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>,</mo><msup><mi>&amp;Gamma;</mi><mi>t</mi></msup><mo>,</mo><msup><mi>K</mi><mi>t</mi></msup><mo>)</mo></mrow></mrow></mtd></mtr><mtr><mtd><mrow><msup><mi>Q</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>=</mo><munder><mi>argmin</mi><mi>Q</mi></munder><mi>L</mi><mrow><mo>(</mo><msup><mi>P</mi><mi>t</mi></msup><mo>,</mo><mi>Q</mi><mo>,</mo><msup><mi>X</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>,</mo><msup><mi>Y</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>,</mo><msup><mi>&amp;Gamma;</mi><mi>t</mi></msup><mo>,</mo><msup><mi>K</mi><mi>t</mi></msup><mo>)</mo></mrow></mrow></mtd></mtr><mtr><mtd><mrow><msup><mi>&amp;Gamma;</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>=</mo><msup><mi>&amp;Gamma;</mi><mi>t</mi></msup><mo>+</mo><mi>&amp;eta;</mi><msub><mo>&amp;dtri;</mo><mi>&amp;Gamma;</mi></msub><mi>L</mi><mrow><mo>(</mo><msup><mi>P</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>,</mo><msup><mi>Q</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>,</mo><msup><mi>X</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>,</mo><msup><mi>Y</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>,</mo><mi>&amp;Gamma;</mi><mo>,</mo><msup><mi>K</mi><mi>t</mi></msup><mo>)</mo></mrow></mrow></mtd></mtr><mtr><mtd><mrow><msup><mi>K</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>=</mo><msup><mi>K</mi><mi>t</mi></msup><mo>+</mo><mi>&amp;eta;</mi><msub><mo>&amp;dtri;</mo><mi>K</mi></msub><mi>L</mi><mrow><mo>(</mo><msup><mi>P</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>,</mo><msup><mi>Q</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>,</mo><msup><mi>X</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>,</mo><msup><mi>Y</mi><mrow><mi>t</mi><mo>+</mo><mn>1</mn></mrow></msup><mo>,</mo><msup><mi>&amp;Gamma;</mi><mi>t</mi></msup><mo>,</mo><mi>K</mi><mo>)</mo></mrow></mrow></mtd></mtr></mtable></mfenced><mo>;</mo></mrow> S8-3. For each element in P, Q, X, Y, Γ and K, perform its calculation according to the following formula ρ _u =λ|C(u)|,τ _i =λ|C(i)|; training updates, Among them, C(u) and C(i) represent the subsets associated with user u and item i in the known data set C, respectively. 6. An extraction device according to any one of claims 1-5 based on a user feature extraction method based on non-negative alternating direction transformation, characterized in that it includes a data receiving module, a data storage module and an execution module, wherein, The data receiving module is connected to the data storage module, the data receiving module is used to receive the user behavior statistics data collected by the server, and transfer the received user behavior statistics data collected by the server to the data storage module for storage, The data storage module is connected with the execution module, and the execution module executes the user feature extraction instruction sent by the server, and stores the extracted user feature data into the data storage module. 7. The extracting device based on the user feature extraction method of non-negative alternating direction transformation according to claim 6, wherein the data storage module includes an acquisition feature storage unit and a statistical data storage unit, The acquisition feature storage unit is connected to the execution module, and is used to store the user feature data extracted by the execution module; The statistical data storage unit is connected with the data receiving module, and is used for storing user behavior statistical data delivered by the data receiving module. 8. The extracting device based on the user feature extraction method of non-negative alternating direction exchange according to claim 6, wherein the execution module includes an initialization unit, a training unit and an output unit, The initialization unit initializes the parameters that the user feature extraction process depends on, and the initialization parameters include: feature space dimension f, dual learning rate η, Lagrangian enhancement factor λ, user feature matrix X, user training auxiliary matrix X_U, X_D and X_C, project characteristic matrix Y, project training auxiliary matrix Y_U, Y_D and Y_C, iteration control variable t, iteration upper limit n, convergence judgment threshold The input end of the training unit is connected with the initialization unit and the data storage module respectively, according to the parameters initialized by the initialization unit and the user behavior statistics data in the data storage module to construct user characteristic data, including user characteristic matrix X and item characteristic matrix Y, training unit First construct the cumulative absolute error ε(P,Q,X,Y), where P is the user feature constraint matrix, Q is the item feature constraint matrix, and then construct a unified loss function L(P,Q,X,Y,Γ,K) , where Γ and K are dual parameters, and then iteratively train P, Q, X, Y, Γ and K until the unified loss function L(P, Q, X, Y, Γ, K) is relative to P, Q , X, Y, Γ and K converge on the known data set C in the user behavior statistics matrix, or the iteration control variable t is equal to the iteration upper limit n; The input end of the output unit is connected to the output end of the training unit, the output end of the output unit is connected to the data storage module, and the output unit outputs and stores the user characteristic data constructed by the training unit into the data storage module.