Multi-interest recommendation method and device

Abstract

The application provides a multi-interest recommendation method and device, and belongs to the field of artificial intelligence. The multi-interest recommendation method comprises the following steps: extracting a user's historical multi-interest expression by using a Kalman capsule network according to a user behavior sequence; extracting a user's possible interest expression by using the Kalman capsule network according to the user behavior sequence and a target item; and estimating the click rate of the user on the target item by using a first attention model according to the historical multi-interest expression of the user and the possible interest expression of the user. The technical scheme of the application solves the problem of high iteration complexity of a dynamic routing algorithm when multi-interest recommendation is performed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of business, in particular to a multi-interest recommendation method and device. BACKGROUND

[0002] Personalized recommendation is to recommend materials that meet user interests from a large number of candidate materials according to user behavior records and material characteristics. Thus, the correct thing is distributed to the right person, avoiding inefficient filtering of information by users.

[0003] Multi-interest recommendation: User behavior is usually implicit information, which implies the user's click selection logic. If the user behavior sequence is directly input into the neural network for learning, part of the user selection logic can indeed be learned, but the learning efficiency is low and it is not interpretable. By modeling user behavior, the user's multiple interest representations are abstracted. For neural networks, this part of content can be better learned, and the interpretability is stronger.

[0004] Capsule network: The neuron of the ordinary neural network is a scalar, and the information it can represent is limited. The neuron of the capsule network is a vector, which can represent more information. Capsule network is first used in the field of image recognition, and the purpose is to represent both the local features of a certain region of the image and the position information of the region through a neuron. The capsule vector of a capsule network can represent the existence probability of the feature by the modulus, and the attitude information of the feature by the direction, so as to achieve the expected purpose.

[0005] Kalman filter: Kalman filter is a dynamic estimation algorithm that estimates the hidden space state by fusing the prediction from the empirical model and the correction from the external measurement. It is usually used in communication, navigation, control and other fields. Since Kalman filter only depends on the last iteration result, it has the advantages of fast calculation speed and fast iteration convergence speed. Moreover, the formula of Kalman filter is simple and flexible, and can be modeled in any space.

[0006] The prior art changes the vector and weight fusion mode in the dynamic routing algorithm of the capsule network when making multi-interest recommendation, replaces the original network simple multiplication mode with a gated unit form. The dynamic routing algorithm itself is a method of gradually converging multiple times of recursion, and the complexity is further increased after adding the gated unit. On the other hand, the dynamic routing multiple iteration mode accelerates the convergence of parameters. The gated unit is in the loop, and although multiple loops, the parameter environment is the same, and a small amount of training samples cause a large number of parameters in the gated unit to converge to a local minimum, which cannot learn the content from a large number of subsequent samples, thereby losing the significance of setting a large number of parameters by the gated unit; in addition, when predicting the click rate of the user to the target item, the user interest vector and the target item vector are simply fused, the target item and the user behavior are not effectively modeled, and the possible interest representation of the user is not extracted from the target item. SUMMARY

[0007] The technical problem to be solved by the present application is to provide a multi-interest recommendation method and device, which solves the problem of high iteration complexity of the dynamic routing algorithm when making multi-interest recommendation.

[0008] To solve the above technical problems, the embodiments of the present application provide the technical solutions as follows:

[0009] On the one hand, a multi-interest recommendation method is provided, comprising:

[0010] extracting a user's historical multi-interest expression by using a Kalman capsule network according to a user behavior sequence;

[0011] extracting a user's possible interest expression by using a Kalman capsule network according to the user behavior sequence and a target item;

[0012] estimating a click rate of the user to the target item by using a first attention model according to the user's historical multi-interest expression and the user's possible interest expression.

[0013] In some embodiments, the extracting a user's historical multi-interest expression by using a Kalman capsule network according to a user behavior sequence comprises:

[0014] calculating an interest vector by using each Kalman filter respectively to obtain m interest vectors, m is an integer greater than 1, and the calculating an interest vector by using each Kalman filter comprises:

[0015] representing each behavior in the user behavior sequence as a vector to form a behavior vector matrix;

[0016] mapping the behavior vector matrix from a low-level space to a high-level space by using a weight matrix w;

[0017] The Kalman filter is used to calculate the weight of each behavior vector;

[0018] The result of the weighted sum of the plurality of behavior vectors using the weight updated by the Kalman filter is taken as the interest vector.

[0019] In some embodiments, the use of the Kalman filter to calculate the weight of each behavior vector comprises:

[0020] The weight of each behavior vector is randomly initialized as c i , which is configured in the form of a diagonal matrix c, and the weight values of each behavior vector are sequentially arranged on the main diagonal line. The weight matrix c is a parameter tracked by the Kalman filter, and the process equation of the Kalman filter is defined as follows:

[0021] c k = Ac k-1 + w k-1

[0022] The measurement equation of the Kalman filter is defined as follows:

[0023] z k = Hc k + v k

[0024] where c k c k-1 is a system state matrix, i.e., a parameter tracked by the Kalman filter, A is a state transition matrix, H is a state observation matrix, z k is an observation, w k-1 is process noise, and v k is measurement noise. The process equation is defined as a process of transferring the weight matrix from the previous time c k-1 to the current time c k , and the result is the prior estimate at the current time. The measurement equation is defined as a process of calculating the observation z k using the prior estimate c k of the current weight matrix. Assuming that the weight value distribution conforms to a standard Gaussian distribution, the observation is defined as the mean and variance of the current prior estimate c k , and the standard value is the mean and variance of the standard Gaussian distribution.

[0025] In some embodiments, the parameter update process of the Kalman filter is as follows:

[0026]

[0027]

[0028]

[0029]

[0030]

[0031] wherein P k is the state estimation covariance matrix, z is the observation standard value, Q is the process noise covariance, R is the measurement noise covariance, A, H are parameter learning matrices, P, Q, R are hyperparameter matrices.

[0032] In some embodiments, the extracting the possible interest expression of the user from the user behavior sequence and the target item using the Kalman capsule network comprises:

[0033] placing the target item at the back of the user behavior sequence, and inputting the target item and the user behavior sequence into a Kalman filter for weight calculation, and outputting a vector after weighted average and nonlinear activation as the possible interest expression of the user.

[0034] In some embodiments, the predicting the click rate of the user on the target item from the historical multi-interest expression of the user and the possible interest expression of the user using a first attention model comprises:

[0035] modeling the historical multi-interest expression vector of the user and the possible interest expression vector of the user using the first attention model.

[0036] fitting a plurality of vectors output by the first attention model into a value using a logistic regression, and outputting the value as the click probability of the user on the target item after sigmod function activation.

[0037] In some embodiments, the loss function of the first attention model is a cross-entropy loss function.

[0038] The embodiments of the present application also provide a multi-interest recommendation device, comprising:

[0039] a first extraction module configured to extract a historical multi-interest expression of a user from a user behavior sequence using a Kalman capsule network;

[0040] a second extraction module configured to extract a possible interest expression of the user from the user behavior sequence and a target item using the Kalman capsule network;

[0041] a processing module configured to predict a click rate of the user on the target item from the historical multi-interest expression of the user and the possible interest expression of the user using a first attention model.

[0042] In some embodiments, the first extraction module is configured to calculate an interest vector from the user behavior sequence using m groups of Kalman filters to obtain m interest vectors, m is an integer greater than 1, and each Kalman filter is configured to calculate an interest vector, which comprises:

[0043] each behavior in the user behavior sequence is represented by a vector, and a behavior vector matrix is formed;

[0044] the behavior vector matrix is mapped from a low-level space to a high-level space by using a weight matrix w;

[0045] the weight of each behavior vector is calculated by using a Kalman filter;

[0046] a plurality of behavior vectors are weighted and summed by using the weight updated by the Kalman filter, and the result is taken as an interest vector.

[0047] In some embodiments, the first extraction module is specifically configured to randomly initialize the weight of each behavior vector as c i , which is configured in the form of a diagonal matrix c, and the weight values of each behavior vector are sequentially arranged on the main diagonal line, and the weight matrix c is a Kalman filter tracking parameter, and the Kalman filter process equation is defined as follows:

[0048] c k = Ac k-1 + w k-1

[0049] The Kalman filter measurement equation is defined as follows:

[0050] z k = Hc k + v k

[0051] wherein c k c k-1 is a system state matrix, that is, a Kalman filter tracking parameter, A is a state transition matrix, H is a state observation matrix, z k is an observation, w k-1 is a process noise, and v k is a measurement noise, wherein the process equation is defined as a process of transferring the weight matrix from the previous time c k-1 to the current time c k , and the result is the prior estimation at the current time; the measurement equation is defined as a process of calculating the observation z k by using the prior estimation c k of the current weight matrix. Assuming that the weight value distribution conforms to a standard Gaussian distribution, the observation is defined as the mean and variance of the current prior estimation c k , and the standard value is the mean and variance of the standard Gaussian distribution.

[0052] In some embodiments, the Kalman filter parameter updating process is as follows:

[0053]

[0054]

[0055]

[0056]

[0057]

[0058] wherein, P k is a state estimation covariance matrix, z is an observation standard value, Q is a process noise covariance, R is a measurement noise covariance, A, H are parameter learning matrices, P, Q, R are hyperparameter matrices.

[0059] In some embodiments, the second extraction module is configured to place the target item at the back of the user behavior sequence and input the target item and the user behavior sequence into a Kalman filter to calculate a weight, and output a vector after weighted average and nonlinear activation of the user as a possible interest expression of the user.

[0060] In some embodiments, the processing module is configured to model the historical multi-interest expression vector of the user and the possible interest expression vector of the user by using the first attention model, and fit a plurality of vectors output by the first attention model into a value by using a logistic regression, and output the value after sigmod function activation as a click probability of the target item by the user.

[0061] In some embodiments, the loss function of the first attention model is a cross-entropy loss function.

[0062] The embodiments of the present application also provide a multi-interest recommendation device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor; and the processor implements the multi-interest recommendation method as described above when executing the program.

[0063] In some embodiments, the processor is configured to extract a historical multi-interest expression of a user by using a Kalman capsule network according to a user behavior sequence, extract a possible interest expression of the user by using the Kalman capsule network according to the user behavior sequence and a target item, and estimate a click rate of the target item by the user by using a first attention model according to the historical multi-interest expression of the user and the possible interest expression of the user.

[0064] In some embodiments, the processor is configured to calculate an interest vector by using m groups of Kalman filters respectively according to the user behavior sequence to obtain m interest vectors, m is an integer greater than 1, and the calculation of the interest vector by using each Kalman filter comprises:

[0065] each behavior in the user behavior sequence is represented by a vector to form a behavior vector matrix.

[0066] mapping the behavior vector matrix from the low-level space to the high-level space using a weight matrix w;

[0067] calculating the weight of each behavior vector using a Kalman filter;

[0068] summing the plurality of behavior vectors using the weight updated by the Kalman filter to obtain a result as an interest vector.

[0069] In some embodiments, the processor is configured to randomly initialize the weight of each behavior vector as c i , which is configured in the form of a diagonal matrix c, and the weight values of each behavior vector are sequentially arranged on the main diagonal line, and the weight matrix c is a Kalman filter tracking parameter, and the Kalman filter process equation is defined as follows:

[0070] c k = Ac k-1 + w k-1

[0071] The Kalman filter measurement equation is defined as follows:

[0072] z k = Hc k + v k

[0073] where c k c k-1 is a system state matrix, i.e., a Kalman filter tracking parameter, A is a state transition matrix, H is a state observation matrix, z k is an observation, w k-1 is a process noise, and v k is a measurement noise, wherein the process equation is defined as a process of transferring the weight matrix from the previous time c k-1 to the current time c k , and the result is the prior estimate at the current time; the measurement equation is defined as a process of calculating the observation z k using the prior estimate c k of the current weight matrix. Assuming that the weight value distribution conforms to a standard Gaussian distribution, the observation is defined as the mean and variance of the current prior estimate c k , and the standard value is the mean and variance of the standard Gaussian distribution.

[0074] In some embodiments, the Kalman filter parameter updating process is as follows:

[0075]

[0076]

[0077]

[0078]

[0079]

[0080] wherein, P k is a state estimation covariance matrix, z is an observation standard value, Q is a process noise covariance, R is a measurement noise covariance, A, H are parameter learning matrices, P, Q, R are hyperparameter matrices.

[0081] In some embodiments, the processor is configured to place the target item at the back of the user behavior sequence, and input the target item and the user behavior sequence into a Kalman filter together for weight calculation, and output a vector after weighted average and nonlinear activation as a possible interest expression of the user.

[0082] In some embodiments, the processor is configured to model the historical multi-interest expression vector of the user and the possible interest expression vector of the user by using the first attention model; and fit a plurality of vectors output by the first attention model into a value by using a logistic regression, and after activation by a sigmod function, the value is taken as a click probability of the target item by the user.

[0083] In some embodiments, the loss function of the first attention model is a cross-entropy loss function.

[0084] The embodiment of the present application also provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the steps in the multi-interest recommendation method.

[0085] The embodiment of the present application has the following beneficial effects:

[0086] In the above scheme, by introducing a Kalman filter into a capsule network, the problem of high iteration complexity of a dynamic routing algorithm is solved, and the problem of effective recommendation for a cold start or a user with less behavior is solved; by jointly modeling a target item and a user behavior sequence, a possible interest expression of a user is estimated, attention calculation is performed on the possible interest expression and an existing historical interest vector, and therefore the relationship between the target item and the user behavior sequence is modeled in the interest vector dimension, and the accuracy of a recommendation system and user experience are improved. BRIEF DESCRIPTION OF DRAWINGS

[0087] Figure 1 FIG. 1 is a flowchart of a multi-interest recommendation method according to an embodiment of the present application;

[0088] Figure 2 FIG. 2 is a structural diagram of a multi-interest recommendation device according to an embodiment of the present application;

[0089] Figure 3 FIG. 3 is a component diagram of the multi-interest recommendation device according to an embodiment of the present application. DETAILED DESCRIPTION

[0090] To make the technical problems, technical solutions and advantages to be solved by the embodiments of the present application more clear, the following will be described in detail with reference to the drawings and specific embodiments.

[0091] The embodiments of the present application provide a multi-interest recommendation method and device, which solve the problem of high iteration complexity of a dynamic routing algorithm when multi-interest recommendation is performed.

[0092] The embodiments of the present application provide a multi-interest recommendation method, as shown in the accompanying drawings, comprising: Figure 1

[0093] Step 101: extracting a historical multi-interest expression of a user according to a user behavior sequence by using a Kalman capsule network;

[0094] Step 102: extracting a possible interest expression of the user according to the user behavior sequence and a target item by using the Kalman capsule network;

[0095] Step 103: estimating a click rate of the target item by the user by using a first attention model according to the historical multi-interest expression of the user and the possible interest expression of the user.

[0096] In the embodiments, by introducing a Kalman filter into the capsule network, the problem of high iteration complexity of the dynamic routing algorithm is solved, and the problem of effective recommendation for cold start or users with less behavior is solved; by jointly modeling the target item and the user behavior sequence, the possible interest expression of the user is estimated, and attention calculation is performed on the existing historical interest vector, so that the relationship between the target item and the user behavior sequence is modeled in the interest vector dimension, and the accuracy of the recommendation system and the user experience are improved.

[0097] In some embodiments, the extracting the historical multi-interest expression of the user according to the user behavior sequence by using the Kalman capsule network comprises:

[0098] m groups of Kalman filters are used to calculate interest vectors respectively by using the user behavior sequence, m is an integer greater than 1, and the calculation of the interest vector by each Kalman filter comprises:

[0099] each behavior in the user behavior sequence is represented by a vector to form a behavior vector matrix;

[0100] the behavior vector matrix is mapped from a low-level space to a high-level space by using a weight matrix w;

[0101] the weight of each behavior vector is calculated by using the Kalman filter;

[0102] ​The interest vector is obtained by weighted sum of the plurality of behavior vectors using the updated weights of the Kalman filter.

[0103] In some embodiments, the step of calculating the weight of each behavior vector using the Kalman filter comprises:

[0104] Randomly initializing the weight of each behavior vector as c i , which is configured in the form of a diagonal matrix c, and the weight values of each behavior vector are sequentially arranged on the main diagonal line. The weight matrix c is a parameter tracked by the Kalman filter, and the process equation of the Kalman filter is defined as follows:

[0105] c k = Ac k-1 + w k-1

[0106] The measurement equation of the Kalman filter is defined as follows:

[0107] z k = Hc k + v k

[0108] , where c k c k-1 is a system state matrix, i.e., a parameter tracked by the Kalman filter, A is a state transition matrix, H is a state observation matrix, z k is an observation, w k-1 is a process noise, and v k is a measurement noise. The process equation is defined as a process of transferring the weight matrix from the previous time c k-1 to the current time c k , and the result is the prior estimate at the current time. The measurement equation is defined as a process of calculating the observation z k using the prior estimate c k of the current weight matrix. Assuming that the weight value distribution conforms to a standard Gaussian distribution, the observation is defined as the mean and variance of the current prior estimate c k , and the standard value is the mean and variance of the standard Gaussian distribution.

[0109] In some embodiments, the parameter updating process of the Kalman filter is as follows:

[0110]

[0111]

[0112]

[0113]

[0114]

[0115] wherein, P k is a state estimation covariance matrix, z is an observation standard value, Q is a process noise covariance, R is a measurement noise covariance, A, H are parameter learning matrices, P, Q, R are hyperparameter matrices.

[0116] In some embodiments, the extracting the possible interest expression of the user by using the Kalman capsule network according to the user behavior sequence and the target item comprises:

[0117] placing the target item at the back of the user behavior sequence, and inputting the target item and the user behavior sequence into a Kalman filter for weight calculation, and taking an output vector after weighted average and nonlinear activation as the possible interest expression of the user.

[0118] In some embodiments, the predicting the click rate of the user on the target item by using a first attention model according to the historical multi-interest expression of the user and the possible interest expression of the user comprises:

[0119] modeling the historical multi-interest expression vector of the user and the possible interest expression vector of the user by using the first attention model;

[0120] fitting a plurality of vectors output by the first attention model into a value by using a logistic regression, and taking a value after sigmod function activation as the click probability of the user on the target item.

[0121] In some embodiments, the loss function of the first attention model is a cross-entropy loss function.

[0122] The embodiment proposes a multi-interest recommendation method based on a Kalman capsule network, specifically, a dynamic routing algorithm for calculating a capsule vector fusion weight in an original capsule network is replaced by a Kalman filter method for state estimation, and a multi-iteration dynamic routing algorithm is replaced by a single-estimation Kalman filter to reduce complexity, wherein a process matrix of the Kalman filter is obtained through training and can be regarded as modeling of an input sequence. A user behavior sequence and a target item are jointly input into a Kalman capsule network to extract a possible interest vector of the user, a user historical behavior sequence and the target item are used to estimate a possible interest expression of the user, an attention mechanism is modeled with the historical interest expression of the user, and a click probability of the user on the target item is predicted from an interest level. The input of the unified attention mechanism is an interest expression vector, and the problem of non-computability caused by different domains of vectors is avoided.

[0123] The embodiment of the application further provides a multi-interest recommendation device, as shown in Figure 2 the device comprises:

[0124] The first extraction module 11 is configured to extract historical multi-interest expressions of a user according to a user behavior sequence by using a Kalman capsule network.

[0125] The second extraction module 12 is configured to extract possible interest expressions of the user according to the user behavior sequence and a target item by using the Kalman capsule network.

[0126] The processing module 13 is configured to predict a click rate of the user on the target item according to the historical multi-interest expressions of the user and the possible interest expressions of the user by using a first attention model.

[0127] In the embodiment, the Kalman filter is introduced into the capsule network to solve the problem of high iteration complexity of a dynamic routing algorithm and the problem of effective recommendation for a cold start or a user with less behaviors. The target item and the user behavior sequence are jointly modeled to predict possible interest expressions of the user, and attention calculation is performed on the existing historical interest vector, so as to model the relationship between the target item and the user behavior sequence in the interest vector dimension, and improve the accuracy of the recommendation system and the user experience.

[0128] In some embodiments, the first extraction module 11 is configured to calculate interest vectors by using m groups of Kalman filters respectively on the user behavior sequence to obtain m interest vectors, m is an integer greater than 1, and the calculation of the interest vector by using each Kalman filter includes:

[0129] Each behavior in the user behavior sequence is represented by a vector to form a behavior vector matrix;

[0130] The behavior vector matrix is mapped from a low-level space to a high-level space by using a weight matrix w;

[0131] The weight of each behavior vector is calculated by using a Kalman filter;

[0132] The result obtained by weighted summation of a plurality of behavior vectors by using the weight updated by the Kalman filter is taken as the interest vector.

[0133] In some embodiments, the first extraction module 11 is specifically configured to randomly initialize the weight of each behavior vector as c i , which is constructed in the form of a diagonal matrix c, and the weight values of each behavior vector are sequentially arranged on the main diagonal line. The weight matrix c is a tracking parameter of the Kalman filter, and the process equation of the Kalman filter is defined as follows:

[0134] c k = Ac k-1 +w k-1

[0135] The measurement equation of the Kalman filter is defined as follows:

[0136] zk = Hc k + v k

[0137] where c k c k-1 is the system state matrix, i.e., the Kalman filter tracks parameters, A is the state transition matrix, H is the state observation matrix, z k is the observation, w k-1 is the process noise, v k is the measurement noise, where the process equation defines the process of transferring the weight matrix from the last time c k-1 to the current time c k , and the result is the prior estimate at the current time; the measurement equation defines the process of calculating the observation z k using the prior estimate c k of the current weight matrix. Assuming that the weight value distribution conforms to the standard Gaussian distribution, the observation is defined as the mean and variance of the current prior estimate c k , and the standard value is the mean and variance of the standard Gaussian distribution.

[0138] In some embodiments, the Kalman filter parameter updating process is as follows:

[0139]

[0140]

[0141]

[0142]

[0143]

[0144] where P k is the state estimation covariance matrix, z is the observation standard value, Q is the process noise covariance, R is the measurement noise covariance, A and H are parameter learning matrices, and P, Q and R are hyperparameter matrices.

[0145] In some embodiments, the second extraction module 12 is configured to place the target item at the back of the user behavior sequence, and input the target item and the user behavior sequence to the Kalman filter for weight calculation. The output vector after weighted average and nonlinear activation is used as the possible interest expression of the user.

[0146] In some embodiments, the processing module 13 is configured to model the historical multi-interest expression vector of the user and the possible interest expression vector of the user by using the first attention model; and fit the vectors output by the first attention model into a value by using a logistic regression, which is activated by a sigmod function and used as the click probability of the user on the target item.

[0147] In some embodiments, the loss function of the first attention model is a cross-entropy loss function.

[0148] The embodiments of the present application also provide a multi-interest recommendation device, which comprises a memory 21, a processor 22, and a computer program stored in the memory 21 and executable on the processor 22; and the processor 22 implements the multi-interest recommendation method as described above when executing the program. Figure 3 As shown in the figure, the device comprises a memory 21, a processor 22, and a computer program stored in the memory 21 and executable on the processor 22; and the processor 22 implements the multi-interest recommendation method as described above when executing the program.

[0149] In some embodiments, the processor 22 is configured to extract the historical multi-interest expression of the user by using the Kalman capsule network according to the user behavior sequence; extract the possible interest expression of the user by using the Kalman capsule network according to the user behavior sequence and the target item; and estimate the click rate of the user on the target item by using the first attention model according to the historical multi-interest expression of the user and the possible interest expression of the user.

[0150] In the embodiments, the Kalman filter is introduced into the capsule network to solve the problem of high iteration complexity of the dynamic routing algorithm and the problem of effective recommendation for cold start or users with less behaviors; the target item and the user behavior sequence are jointly modeled to estimate the possible interest expression of the user, and the attention calculation is performed on the existing historical interest vector, so as to model the relationship between the target item and the user behavior sequence in the interest vector dimension, and improve the accuracy of the recommendation system and the user experience.

[0151] In some embodiments, the processor 22 is configured to calculate the interest vector by using the user behavior sequence by using m groups of Kalman filters respectively, to obtain m interest vectors, m is an integer greater than 1, and the calculation of the interest vector by using each Kalman filter comprises:

[0152] Each behavior in the user behavior sequence is represented by a vector to form a behavior vector matrix;

[0153] The behavior vector matrix is mapped from a low-level space to a high-level space by using a weight matrix w;

[0154] The weight of each behavior vector is calculated by using the Kalman filter;

[0155] The result obtained by weighted summing a plurality of the behavior vectors by using the weight updated by the Kalman filter is used as the interest vector.

[0156] In some embodiments, the processor 22 is configured to randomly initialize the weight of each behavior vector as c i , which is configured as a diagonal matrix c, whose diagonal elements are the weight values of each behavior vector in turn, and the weight matrix c is the Kalman filter tracking parameter. The Kalman filter process equation is defined as follows:

[0157] c k = Ac k-1 + w k-1

[0158] The Kalman filter measurement equation is defined as follows:

[0159] z k = Hc k + v k

[0160] Where c k c k-1 is the system state matrix, i.e. the Kalman filter tracking parameter, A is the state transition matrix, H is the state observation matrix, z k is the observation, w k-1 is the process noise, and v k is the measurement noise. The process equation is defined as the process of transferring the weight matrix from the previous time c k-1 to the current time c k , and the result is the prior estimate at the current time; the measurement equation is defined as the process of calculating the observation z k using the prior estimate c k of the current weight matrix. Assuming that the weight value distribution conforms to the standard Gaussian distribution, the observation is defined as the mean and variance of the current prior estimate c k , and the standard value is the mean and variance of the standard Gaussian distribution.

[0161] In some embodiments, the Kalman filter parameter update process is as follows:

[0162]

[0163]

[0164]

[0165]

[0166]

[0167] Where P kFor state estimation covariance matrix, z is the standard value of observation, Q is the process noise covariance, R is the measurement noise covariance, A, H are parameter learning matrices, P, Q, R are hyperparameter matrices.

[0168] In some embodiments, the processor 22 is configured to place the target item at the back of the user behavior sequence, and input the target item and the user behavior sequence into a Kalman filter to calculate a weight, and output a vector after weighted average and nonlinear activation as a possible interest expression of the user.

[0169] In some embodiments, the processor 22 is configured to model the historical multi-interest expression vector of the user and the possible interest expression vector of the user by using the first attention model; and fit the vectors output by the first attention model into a value by using a logistic regression, and after sigmod function activation, output the value as a click probability of the user on the target item.

[0170] In some embodiments, the loss function of the first attention model is a cross-entropy loss function.

[0171] The embodiments of the present application also provide a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the multi-interest recommendation method.

[0172] The computer readable medium includes permanent and non-permanent, removable and non-removable media, and can be implemented by any method or technology to store information. The information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage, or any other non-transmission medium that can be used to store information that can be accessed by a computing device to be detected. According to the definition in this paper, the computer readable medium does not include transitory computer readable media, such as modulated data signals and carriers.

[0173] The above is the preferred embodiment of the present application. It should be noted that for those skilled in the art, without departing from the principles of the present application, some improvements and refinements can be made, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A multi-interest recommendation method, characterized by, The method comprises the following steps: According to the user behavior sequence, the user's historical multi-interest expression is extracted by using the Kalman capsule network; According to the user behavior sequence and the target item, the user's possible interest expression is extracted by using the Kalman capsule network; According to the user's historical multi-interest expression and the user's possible interest expression, the click rate of the user to the target item is estimated by using the first attention model; The method according to the user behavior sequence, the user's historical multi-interest expression is extracted by using the Kalman capsule network, comprising: An m-group Kalman filter is used to calculate an interest vector respectively by using the user behavior sequence, m is an integer greater than 1, and each Kalman filter is used to calculate an interest vector, comprising: Each behavior in the user behavior sequence is represented by a vector to form a behavior vector matrix; The behavior vector matrix is mapped from a low-level space to a high-level space by using a weight matrix w; The weight of each behavior vector is calculated by using the Kalman filter; The result obtained by weighted sum of a plurality of behavior vectors by using the weight updated by the Kalman filter is used as an interest vector.

2. The multiple interest recommendation method of claim 1, wherein, The weight of each behavior vector calculated by using the Kalman filter comprises: The weight of each behavior vector is initialized randomly as c i It is configured into a diagonal matrix c, and the weight values of each behavior vector are sequentially arranged on the main diagonal line. The weight matrix c is the tracking parameter of the Kalman filter. The process equation of the Kalman filter is defined as follows: c k = Ac k-1 +w k-1 The measurement equation of the Kalman filter is defined as follows: z k = Hc k + v k where c k c k-1 is the system state matrix, i.e., the Kalman filter tracks parameters, A is the state transition matrix, H is the state observation matrix, z k is the observation, w k-1 is the process noise, v k is the measurement noise, where the process equation defines the process of transferring the weight matrix from the last time c k-1 to the current time c k , and the result is the prior estimation at the current time; the measurement equation defines the process of calculating the observation z k using the prior estimation c k of the current weight matrix, assuming that the weight value distribution conforms to the standard Gaussian distribution, the observation is defined as the mean and variance of the current prior estimation c k , and the standard value is the mean and variance of the standard Gaussian distribution.

3. The multiple interest recommendation method of claim 1, wherein, The method according to the user behavior sequence and the target item, the user's possible interest expression is extracted by using the Kalman capsule network, comprising: The target item is placed behind the user behavior sequence and input to the Kalman filter for weight calculation, and the output vector after weighted average and nonlinear activation is used as the user's possible interest expression.

4. The multiple interest recommendation method of claim 1, wherein, According to the user's historical multi-interest expression and the user's possible interest expression, the click rate of the user to the target item is estimated by using the first attention model, comprising: The first attention model is used to model the user's historical multi-interest expression vector and the user's possible interest expression vector; The plurality of vectors output by the first attention model are fitted into a value by using a logistic regression, and the click probability of the user to the target item is obtained after the sigmod function activation.

5. The multiple interest recommendation method of claim 1, wherein, The loss function of the first attention model is a cross-entropy loss function.

6. A multiple interest recommendation apparatus, characterized by, The method comprises the following steps: The first extraction module is used to extract the user's historical multi-interest expression by using the Kalman capsule network according to the user behavior sequence; The second extraction module is used to extract the user's possible interest expression by using the Kalman capsule network according to the user behavior sequence and the target item; The processing module is used to estimate the click rate of the user to the target item by using the first attention model according to the user's historical multi-interest expression and the user's possible interest expression. The first extraction module is specifically configured to calculate interest vectors by using m groups of Kalman filters respectively on the user behavior sequence, to obtain m interest vectors, m being an integer greater than 1, and the calculation of the interest vector by using each Kalman filter comprising: representing each behavior in the user behavior sequence by a vector to form a behavior vector matrix; mapping the behavior vector matrix from a low-level space to a high-level space by using a weight matrix w; calculating the weight of each behavior vector by using a Kalman filter; and obtaining the interest vector by weighted summing of a plurality of the behavior vectors by using the weight updated by the Kalman filter. 7.A multi-interest recommendation apparatus, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor; characterized in that, The processor implements the multi-interest recommendation method of any one of claims 1-5 when executing the program.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps in the multi-interest recommendation method of any one of claims 1-5.

Images