Personalized search recommendation method and system based on sparse variational auto-encoder
Through a personalized search recommendation method based on sparse variational autoencoder, a user interest preference model is built, combined with evolutionary algorithms and matrix decomposition, the user-generated content mining problem is solved, and the accuracy and efficiency of personalized recommendations are improved.
Patent Information
- Application Number
- CN202411054508.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-02
- Publication Date
- 2024-11-29
AI Technical Summary
The existing technology is difficult to deeply explore and utilize user-generated content, accurately capture users' potential needs and interests and preferences, and predict user selection tendencies and behavioral patterns, resulting in poor personalized search and recommendation results.
Using a personalized search recommendation method based on sparse variational autoencoder, by collecting multi-source heterogeneous user-generated data, constructing user positive and negative interest preference models, combining evolutionary algorithms and matrix decomposition methods, designing user evaluation agent models and distribution estimates Sampling probability model to achieve personalized recommendations.
It improves the accuracy and efficiency of personalized recommendations, reduces the burden of user evaluation, and improves the effectiveness of search and recommendation.
Smart Images

Figure CN119046528A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a personalized search recommendation method and system, and in particular to a personalized search recommendation method and system based on a sparse variational autoencoder. Background Art
[0002] With the rapid development of the information age, Internet services such as social media and e-commerce platforms have flourished rapidly. People have left a variety of data information on the Internet, and the ways to obtain information have become more diverse. However, traditional search methods and technologies can no longer meet the personalized needs and personalized interest preferences of users, so the importance of personalized search and recommendation systems is becoming increasingly apparent. Personalized search and recommendation systems in big data environments cover a large amount of user-generated content, such as user ratings, interactive behaviors, text reviews, product category labels, attribute short descriptions, time series, pictures or videos, social networks and other data. These user-generated data provide a rich source of data information for personalized search and recommendation methods. How to deeply mine and make full use of user-generated content, efficiently obtain useful information, accurately capture users' potential needs and interest preferences, predict users' selection tendencies and behavioral patterns, and recommend items or content that users may be interested in to current users, these practical tasks are still full of huge challenges. Summary of the invention
[0003] Purpose of the invention: The first purpose of the present invention is to provide a personalized search recommendation method based on sparse variational autoencoder that can deeply mine and utilize user-generated content to predict users' selection tendencies and behavioral patterns.
[0004] The second object of the present invention is to provide a personalized search recommendation system based on sparse variational autoencoder.
[0005] Technical solution: The present invention discloses a personalized search recommendation method based on sparse variational autoencoder, which includes the following steps:
[0006] Collect user-generated content in the network environment to form multi-source heterogeneous user-generated data; the user-generated content includes user rating data, project category labels, user text comments and user historical interaction data;
[0007] Collect user-generated content in the network environment to form multi-source heterogeneous user-generated data;
[0008] Vectorize multi-source heterogeneous user-generated data;
[0009] Based on multi-source heterogeneous user-generated data, filter out the project groups that the current user likes as the dominant population P a , select the project group that the current customers do not like as the disadvantaged population Pd ;
[0010] In the framework of evolutionary algorithm, the vectorized representation of the project category label in the multi-source heterogeneous user-generated data and the vectorized representation of the user text comments of the project are concatenated as the genotype representation of the evolutionary individual of the project. a , construct and train a user active interest preference model based on sparse variational autoencoder, according to the disadvantaged population P d , build and train a user negative preference model based on sparse variational autoencoders, use the positive interest preference model based on sparse variational autoencoders and the negative interest preference model based on sparse variational autoencoders to jointly build a user preference model based on double sparse variational autoencoders, and obtain model parameters containing excellent solution feature distribution information and model parameters containing characteristic distribution information of inferior solutions
[0011] In the framework of evolutionary algorithm, according to the user interest preference model based on double sparse variational autoencoder, the user evaluation proxy model based on user interest preference is designed by joint matrix decomposition method. u (x);
[0012] In the framework of interactive distribution estimation algorithm, the dominant population P a As the initial population M (t = 0), t represents the population algebra, through the user's active interest preference model based on the sparse variational autoencoder, the Gaussian probability density function is obtained, and the distribution estimation sampling probability model G based on the user's interest preference is constructed based on the Gaussian probability density function t (X');
[0013] Use the distribution estimation sampling probability model G based on user interest preferences t (X'), generate k new evolutionary individuals N containing user preference information t , according to the similarity criterion, calculate the newly evolved individual N t The similarity between the feasible solution items in the search space to be searched is used to set the screening conditions and screen the feasible solution items that meet the similarity conditions to form the recommended collection s u ;
[0014] Using the user evaluation proxy model based on user interest preferences u (x) Treat recommended collections u The feasible solution project individuals in the prediction score are used as the fitness value of the new evolutionary individual, and the fitness value is used instead of the user's real evaluation to participate in the interactive evolution;
[0015] According to the elite selection strategy, select the collection to be recommended uThe top N project individuals in the list submit the top N project recommendation list to the current user for evaluation. If the user feedback is not ideal, the user's latest evaluation data content is added, and the top N is combined with the original population to update the population M(t)→M(t+1);
[0016] Utilizing the model management mechanism, the user interest preference model based on the double sparse variational autoencoder is dynamically updated according to the newly added user-generated content.
[0017] Furthermore, the specific steps for vectorizing the multi-source heterogeneous user-generated data are as follows: for the item category labels, Multi-hot encoding is used to represent them as binary vectors; for user text comments, the natural language Bert model and feature extraction technology are used to vectorize them.
[0018] Furthermore, based on the user historical interaction data, the similarity between users is calculated through Pearson correlation and the user social network relationship is established.
[0019] Furthermore, the dominant population P a and the inferior population P d The screening steps are as follows: record the user rating as R = [r ui ], represents the real score of user u on item i, and set the threshold α to represent the like, and select items with real scores higher than the threshold α to form the dominant population P a , let the threshold β represent dislike, and select the projects with real scores lower than the threshold β to form the disadvantaged population P d .
[0020] Furthermore, the steps for constructing the user active interest preference model based on sparse variational autoencoder are as follows:
[0021] Splice the dominant population P of the current user a The vectorized representation of the project category label and the vectorized representation of the user text comment in where X∈P a , establish a training dataset for user active interest preference model based on sparse variational autoencoder
[0022] The mean of each sample in the training data set is calculated through the encoder network of the sparse variational autoencoder and variance
[0023]
[0024] From the above, we can get the variational distribution
[0025]
[0026]
[0027] in, is the input sample The latent variables mapped to the latent space are obtained from the variational distribution using the reparameterization technique The multivariate random variable p(Z Pa ), ε is a random noise drawn from a standard Gaussian distribution N(0, I), and They are The corresponding mean and variance;
[0028] The trained user active interest preference model based on sparse variational autoencoder effectively obtains the latent variables containing user preferences. Hidden variables As a decoder network Input, reconstruction and original samples Reconstruct samples as similar as possible
[0029]
[0030]
[0031]
[0032] and They are Corresponding to the mean and variance of the Gaussian distribution.
[0033] Ignore the decoder network output expression Item, the output Item as reconstruction sample
[0034]
[0035] Joint Encoder Network and decoder network Get Input The ELBO lower bound on the log-likelihood is:
[0036]
[0037] Among them, KL is the Kullback-Leibler divergence distance, which indicates the distribution With standard Gaussian distribution The difference between
[0038] Using the reparameterization technique, the log-likelihood lower bound ELBO is used to minimize the negativity of the loss function and introduce a sparsity penalty term Get the loss function L of the sparse variational autoencoder SVAE for:
[0039]
[0040] in, η is a hyperparameter used to balance L VAE and The relative importance of the two losses; Sparse variational autoencoders introduce sparsity penalty terms
[0041]
[0042] Where p is the expected proportion of hidden layer neuron activation; q is the actual average activation of hidden layer neurons;
[0043] In the iterative training process of the user active interest preference model based on the sparse variational autoencoder, the difference between p and q is calculated by KL divergence as the penalty term Achieve sparsity.
[0044] Furthermore, the steps for constructing the user negative interest preference model based on sparse variational autoencoder are as follows:
[0045] According to the inferior population P d , build and train a user negative interest preference model based on sparse variational autoencoder; the disadvantaged population contains the current user's negative preference information, and integrates the category labels of the current user's negative preference items and the user's text comments on these items to obtain Establishing a training dataset for user passive interest preference model based on sparse variational autoencoder
[0046] The mean of each sample is calculated through the encoder network of the sparse variational autoencoder and variance
[0047]
[0048]
[0049] From the above, we can get the variational distribution
[0050]
[0051]
[0052] in, is the input sample The latent variables mapped to the latent space are obtained from the variational distribution using the reparameterization technique A multivariate random variable obtained by sampling and obeying a standard Gaussian distribution ε is random noise drawn from the standard Gaussian distribution N(0,I), and They are The corresponding mean and variance;
[0053] The trained user negative interest preference model based on sparse variational autoencoder effectively obtains the latent variables containing user negative preferences. Hidden variables As a decoder network Input, reconstruction and original samples Reconstruct samples as similar as possible
[0054]
[0055]
[0056]
[0057] and They are The mean and variance of the corresponding Gaussian distribution;
[0058] The decoder network output expression is usually ignored Item, the output Item as reconstruction sample
[0059]
[0060] Joint Encoder Network and decoder network Get Input The ELBO lower bound on the log-likelihood is:
[0061]
[0062] Among them, KL represents the distribution With standard Gaussian distribution The difference between
[0063] Using the reparameterization technique, the log-likelihood lower bound ELBO is used to minimize the negativity of the loss function and introduce a sparsity penalty term Get the loss function L of the sparse variational autoencoderSVAE for:
[0064]
[0065] in, is a hyperparameter used to balance L VAE and The relative importance of the two losses; Sparse variational autoencoders introduce sparsity penalty terms
[0066]
[0067] Where p is the expected proportion of hidden layer neuron activation; q is the actual average activation of hidden layer neurons;
[0068] In the iterative training process of the user negative interest preference model based on the sparse variational autoencoder, the difference between p and q is calculated by KL divergence as the penalty term Achieve sparsity.
[0069] Furthermore, the user evaluation proxy model based on user interest preference u The construction steps of (x) are as follows: vectorize the individual items in the dominant population As the hidden representation of the project's positive preference, it is sent to the trained user's positive interest preference model based on sparse variational autoencoder to reconstruct the user's positive preference hidden representation
[0070]
[0071] Vectorize the individual items in the disadvantaged population As the hidden representation of the negative preference of the item, it is sent to the trained user negative interest preference model based on sparse variational autoencoder to reconstruct the hidden representation of the user negative preference
[0072]
[0073] Using the user social network relationship and combining the matrix decomposition method, we can obtain the current user's predicted score for the item individual x. and The calculation formula is as follows:
[0074]
[0075]
[0076] Among them, b uis the user's bias, which represents the rating tendency of other users in the current user's social network relationship to the item individual x; b i is the bias of the project individual, indicating the basic popularity of the project; μ All is the average of all item ratings;
[0077] Therefore, according to the user interest preference model based on the dual sparse variational autoencoder and the matrix decomposition method, a user evaluation proxy model based on user interest preference is designed. u (x):
[0078]
[0079] Among them, f u (x) is the predicted rating of item x by user u; and They respectively represent the predicted scores of the user's positive interest preference model and the user's negative interest preference model based on the sparse variational autoencoder for the item individual x; w is the adjustment coefficient.
[0080] Furthermore, the steps for constructing the distribution estimation sampling probability model based on user interest preferences are as follows:
[0081] Through the trained user interest preference model based on double sparse variational autoencoders, the activation probability r of the reconstruction layer neural unit of the sparse variational autoencoder in the user's active interest preference model is obtained. i :
[0082]
[0083] Among them, the state h of the jth neuron in the hidden layer of the decoder in the user active interest preference model based on the sparse variational autoencoder is j :
[0084] h j =tanh(W lj z l +b j )
[0085] Thus, the reconstructed representation of the decision variables of the individual items in the initial population M (t = 0) is obtained: And according to the n1-dimensional variables of the evolving individuals Calculate the mean of the individual reconstructed representations of the items and the covariance matrix elements Get the mean of the item category attribute features containing user preference information and the covariance matrix C tBased on the above mean and covariance matrix, the Gaussian probability density function is parameterized and a Gaussian probability model is constructed as a distribution estimation sampling probability model G based on user interest preferences. t (X'):
[0086]
[0087] Furthermore, the steps for estimating the fitness value of the project individuals in the updated population are as follows: According to the principle of survival of the fittest, the top n% of individuals with higher fitness values of the project individuals are selected to form the optimal subpopulation S t , calculate S t The maximum likelihood estimate of
[0088]
[0089] By storing the optimal subpopulation S of each generation t Get Archive A t , and combined with the current population optimal solution S t , estimate the covariance matrix C of the new population t+1 :
[0090]
[0091] Among them, A t =S t-1 ∪S t-2 …∪S t-l , S t-i represents the optimal subpopulation selected by the ti-th generation population, l refers to the set archive length, and represents archive A t The length of storage is the optimal subpopulation set selected by the population in the last l generations, and l is a non-negative integer;
[0092] Import archive A t , which weakens the dependence of the estimated covariance matrix on the current population, and then uses the estimated and C t+1 Update the distribution estimation sampling probability model G based on user interest preferences t (X')→G t+1 (X'), thereby guiding the search direction of the evolutionary optimization algorithm.
[0093] Based on the same inventive concept, the present invention also discloses a personalized search recommendation system based on a sparse variational autoencoder, comprising:
[0094] A user-generated content collection module is used to collect multi-source heterogeneous user historical interaction data in a network environment to form multi-source heterogeneous user-generated data;
[0095] The data preprocessing module is used to vectorize the multi-source heterogeneous user-generated data and establish the social network relationship information of users based on the similarity between users;
[0096] The dominant group and the disadvantaged group construction module sets up screening conditions based on multi-source heterogeneous user-generated data, and screens the project groups that the current user likes as the dominant population P a , select the project group that the current customers do not like as the disadvantaged population P d ;
[0097] The user interest preference model building module, under the framework of evolutionary algorithm, concatenates the vectorized representation of the project category label in the multi-source heterogeneous user-generated data with the vectorized representation of the user text comments of the project as the genotype representation of the project evolution individual, and selects the dominant population P according to the dominant population P. a , construct and train a user active interest preference model based on sparse variational autoencoder, according to the disadvantaged population P d , build and train a user negative preference model based on sparse variational autoencoders, use the positive interest preference model based on sparse variational autoencoders and the negative interest preference model based on sparse variational autoencoders to jointly build a user preference model based on double sparse variational autoencoders, and obtain model parameters containing excellent solution feature distribution information and model parameters containing characteristic distribution information of inferior solutions
[0098] The user evaluation proxy model establishment module designs a user evaluation proxy model based on user interest preference according to the trained user interest preference model based on dual sparse variational autoencoders and the combined matrix decomposition method. u (x):
[0099]
[0100] In the formula, f u (x) is the user's predicted rating for the item; and They represent the predicted scores of the user's positive preference model and the user's negative preference model for the item individual x; w is the adjustment coefficient;
[0101] The distribution estimation sampling probability module based on user interest preference obtains the Gaussian probability density function through the user active preference model in the trained user interest preference model based on double sparse variational autoencoder, thereby constructing the Gaussian probability model G t (X'), as a distribution estimation sampling probability model based on user interest preferences:
[0102]
[0103] In the formula, is the mean of the individual reconstruction representations of the items; C t The covariance matrix of the reconstructed representation for the individual items.
[0104] The project recommendation set building module uses the distribution estimation sampling probability model based on user interest preferences to sample and generate K excellent new evolutionary individuals N containing user preference information t ; According to the similarity criterion, calculate the similarity between the newly evolved individuals and the feasible solution project individuals in the search space, select the feasible solution project individuals with higher similarity, and form the recommended set s u ;
[0105] The evolutionary individual fitness estimation module uses the user evaluation proxy model based on user interest preferences f u (x), estimate the set of recommendations s u The predicted scores of individual items in the project are used as the fitness values of the evolving individuals, replacing the real evaluation of users, participating in the interactive evolutionary optimization process, and according to the model parameters containing the distribution information of the excellent solution features and model parameters containing characteristic distribution information of inferior solutions Guide the search direction of the evolutionary optimization algorithm;
[0106] The project recommendation list building module selects s according to the elite selection strategy u The top N project individuals in the list submit the top N project recommendation list to the current user for evaluation. If the user feedback is not ideal, the user's latest evaluation data content is added, and the top N is combined with the original population to update the population M(t)→M(t+1);
[0107] The model management module uses the model management mechanism to dynamically update the user interest preference model based on the double sparse variational autoencoder according to the newly added user-generated content.
[0108] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: the present invention utilizes a sparse variational autoencoder to construct a user interest preference model based on a double sparse variational autoencoder, which is conducive to accurately describing the user's personalized interest preferences; the present invention designs a distribution estimation sampling probability model based on the user's interest preference based on the user's active interest preference model based on the sparse variational autoencoder, and designs an interactive evolutionary optimization operator and strategy in combination with an interactive distribution estimation algorithm, which is conducive to improving the accuracy of personalized recommendations; the present invention combines a matrix decomposition model with the constructed user interest preference model to establish a user evaluation proxy model based on user interest preferences, estimates the fitness value of the evolved individual, and replaces the user's real evaluation in the interactive personalized search process, which can reduce the user's evaluation burden and psychological fatigue, and is conducive to improving the efficiency and effect of personalized search and recommendation. BRIEF DESCRIPTION OF THE DRAWINGS
[0109] Figure 1 is a flow chart of the method of the present invention;
[0110] Figure 2 This is a schematic diagram of the structure of the user interest preference model based on double sparse variational autoencoders;
[0111] Figure 3 It is a schematic diagram of the structure of the system of the present invention. DETAILED DESCRIPTION
[0112] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.
[0113] Example 1
[0114] The personalized search recommendation method based on sparse variational autoencoder described in the present invention is as follows: Figure 1 As shown, the following steps are included:
[0115] Collect user-generated content in the network environment to form multi-source heterogeneous user-generated data; user-generated content includes user rating data, project category labels, user text comments and user historical interaction data, such as user ratings, likes, attention, tag collection, browsing time, forwarding and sharing, purchases, user comments, project category labels, project attribute descriptions, images, life shows, social networks, geographic locations and other multi-source heterogeneous data information.
[0116] Vectorize the multi-source heterogeneous user-generated data; for the project category label, Multi-hot encoding is used to represent it as a binary vector. Specifically, if a project contains n1 category labels, the project is represented as If the item contains the category label c i1 , then c i1 =1, and if the item does not contain the category label c i1 , then c i1 = 0; For user text comments, the natural language Bert model and feature extraction technology are used to perform vector representation. Specifically, the text comment of user u on project i is represented as a vector Where n2 is the length of the text comment vector; based on the user historical interaction data, the similarity between users is calculated and the user social network relationship is established through Pearson correlation.
[0117] Based on multi-source heterogeneous user-generated data, filter out the project groups that the current user likes as the dominant population P a , select the project group that the current customers do not like as the disadvantaged population P d ; Preferably, the user rating is recorded as R = [rui ], represents the real score of user u on item i, and set the threshold α to represent the like, and select items with real scores higher than the threshold α to form the dominant population P a , let the threshold β represent dislike, and select the projects with real scores lower than the threshold β to form the disadvantaged population P d .
[0118] In the framework of evolutionary algorithms, the genotype representation of the project evolution individuals is designed, and the vectorized representation of the project category label and the vectorized representation of the text comments obtained for the project are concatenated as follows:
[0119] As a genotype representation of the individuals evolving in this project.
[0120] According to the dominant population P a , build and train a user active interest preference model based on sparse variational autoencoder, and splice the dominant population P of the current user a The vectorized representation of the project category label and the vectorized representation of the user text comment in where X∈P a , establish a training dataset for user active interest preference model based on sparse variational autoencoder
[0121] The encoder network of the sparse variational autoencoder calculates the mean of the samples in each training data set and variance
[0122]
[0123]
[0124] From the above, we can get the variational distribution
[0125]
[0126]
[0127] in, is the input sample X Pa The latent variables mapped to the latent space are obtained from the variational distribution using the reparameterization technique A multivariate random variable obtained by sampling and obeying a standard Gaussian distribution ε is random noise drawn from the standard Gaussian distribution Ν(0,I), and They are The corresponding mean and variance;
[0128] The trained user active interest preference model based on sparse variational autoencoder effectively obtains the latent variables containing user preferences. Hidden variables As a decoder network Input, reconstruction and original samples Reconstruct samples as similar as possible
[0129]
[0130]
[0131]
[0132] and They are Corresponding to the mean and variance of the Gaussian distribution.
[0133] Ignore the decoder network output expression Item, the output Item as reconstruction sample
[0134]
[0135] Joint Encoder Network and decoder network Get Input The log-likelihood ELBO (Evidence Lower Bound Objective):
[0136]
[0137] Among them, KL is the Kullback-Leibler divergence distance, which indicates the distribution With standard Gaussian distribution The difference between
[0138] Using the reparameterization technique, the log-likelihood lower bound ELBO is used to minimize the negativity of the loss function and introduce a sparsity penalty term Get the loss function L of the sparse variational autoencoder SVAE for:
[0139]
[0140] in, η is a hyperparameter used to balance L VAE and The relative importance of the two losses; Sparse variational autoencoders introduce sparsity penalty terms
[0141]
[0142] Where p is the expected proportion of hidden layer neuron activation; q is the actual average activation of hidden layer neurons;
[0143] In the iterative training process of the user active interest preference model based on the sparse variational autoencoder, the difference between p and q is calculated by KL divergence as the penalty term Achieve sparsity.
[0144] According to the inferior population P d , build and train a user negative preference model based on sparse variational autoencoder. The steps for building a user negative interest preference model based on sparse variational autoencoder are as follows:
[0145] According to the inferior population P d , build and train a user negative interest preference model based on sparse variational autoencoder; the disadvantaged population contains the current user's negative preference information, and integrates the category labels of the current user's negative preference items and the user's text comments on these items to obtain Establishing a training dataset for user passive interest preference model based on sparse variational autoencoder
[0146] The mean of each sample is calculated through the encoder network of the sparse variational autoencoder and variance
[0147]
[0148]
[0149] From the above, we can get the variational distribution
[0150]
[0151]
[0152] in, is the input sample The latent variables mapped to the latent space are obtained from the variational distribution using the reparameterization technique A multivariate random variable obtained by sampling and obeying a standard Gaussian distribution ε is random noise drawn from the standard Gaussian distribution Ν(0,I), and They are The corresponding mean and variance;
[0153] The trained user negative interest preference model based on sparse variational autoencoder effectively obtains the latent variables containing user negative preferences. Hidden variables As a decoder network Input, reconstruction and original samples Reconstruct samples as similar as possible
[0154]
[0155]
[0156]
[0157] and They are Corresponding to the mean and variance of the Gaussian distribution.
[0158] The decoder network output expression is usually ignored Item, the output Item as reconstruction sample
[0159]
[0160] Joint Encoder Network and decoder network Get Input The ELBO lower bound on the log-likelihood is:
[0161]
[0162] Among them, KL represents the distribution With standard Gaussian distribution The difference between
[0163] Using the reparameterization technique, the log-likelihood lower bound ELBO is used to minimize the negativity of the loss function and introduce a sparsity penalty term Get the loss function L of the sparse variational autoencoder SVAE for:
[0164]
[0165] in, is a hyperparameter used to balance L VAE and The relative importance of the two losses; Sparse variational autoencoders introduce sparsity penalty terms
[0166]
[0167] Where p is the expected proportion of hidden layer neuron activation; q is the actual average activation of hidden layer neurons;
[0168] In the iterative training process of the user negative interest preference model based on the sparse variational autoencoder, the difference between p and q is calculated by KL divergence as the penalty term Achieve sparsity.
[0169] like Figure 2 As shown, a positive interest preference model based on a sparse variational autoencoder and a negative interest preference model based on a sparse variational autoencoder are used to jointly construct a user preference model based on a double sparse variational autoencoder, and the model parameters containing excellent solution feature distribution information are obtained. and model parameters containing characteristic distribution information of inferior solutions Domain knowledge information is introduced into the subsequent interactive evolutionary optimization algorithms and strategies based on user interest preferences, thereby guiding the personalized evolutionary search process.
[0170] In the framework of evolutionary algorithm, according to the user interest preference model based on double sparse variational autoencoder, the joint matrix is divided into vector representations of individual items in the dominant population. As the hidden representation of the project's positive preference, it is sent to the trained user's positive interest preference model based on sparse variational autoencoder to reconstruct the user's positive preference hidden representation
[0171]
[0172] Vectorize the individual items in the disadvantaged population As the hidden representation of the negative preference of the item, it is sent to the trained user negative interest preference model based on sparse variational autoencoder to reconstruct the hidden representation of the user negative preference
[0173]
[0174] Using the user social network relationship and combining the matrix decomposition method, we can obtain the current user's predicted score for the item individual x. and The calculation formula is as follows:
[0175]
[0176]
[0177] Among them, b u is the user's bias, which represents the rating tendency of other users in the current user's network social relationship towards the item individual x; bi is the bias of the project individual, indicating the basic popularity of the project; μ All is the average of all item ratings;
[0178] Therefore, according to the user interest preference model based on the dual sparse variational autoencoder and the matrix decomposition method, a user evaluation proxy model based on user interest preference is designed. u (x):
[0179]
[0180] Among them, f u (x) is the predicted rating of item x by user u; and They respectively represent the predicted scores of the user's positive interest preference model and the user's negative interest preference model based on the sparse variational autoencoder for the item individual x; w is the adjustment coefficient.
[0181] In the framework of interactive distribution estimation algorithm, the dominant population P a As the initial population M (t = 0), t represents the population algebra, through the trained user interest preference model based on double sparse variational autoencoder, the activation probability r of the reconstruction layer neural unit of the sparse variational autoencoder in the user's active interest preference model is obtained i :
[0182]
[0183] Among them, the state h of the jth neuron in the hidden layer of the decoder in the user active interest preference model based on the sparse variational autoencoder is j :
[0184] h j =tanh(W lj z l +b j )
[0185] Thus, the reconstructed representation of the decision variables of the individual items in the initial population M (t = 0) is obtained: And according to the n1-dimensional variables of the evolving individuals Calculate the mean of the individual reconstructed representations of the items and the covariance matrix elements Get the mean of the item category attribute features containing user preference information and the covariance matrix C t Based on the above mean and covariance matrix, the Gaussian probability density function is parameterized and a Gaussian probability model is constructed as a distribution estimation sampling probability model G based on user interest preferences. t (X'):
[0186]
[0187] in, refers to the mean of the individual reconstruction representations of the items, C t Refers to the covariance matrix of the individual reconstructed representations of the items;
[0188] Use the distribution estimation sampling probability model G based on user interest preferences t (X'), generate k new evolutionary individuals N containing user preference information t , according to the similarity criterion, calculate the newly evolved individual N t The similarity between the feasible solution items in the search space to be searched is used to set the screening conditions and screen the feasible solution items that meet the similarity conditions to form the recommended collection s u .
[0189] Using the user evaluation proxy model based on user interest preferences u (x) Treat recommended collections u The feasible solution project individuals in the population are predicted and scored, and the predicted scores are used as the fitness values of the new evolutionary individuals. The fitness values are used to replace the real user evaluations to participate in the interactive evolution. The steps for estimating the fitness values of the project individuals in the updated population are as follows: According to the principle of survival of the fittest, the top n% of individuals with higher fitness values of the project individuals are selected to form the optimal subpopulation S t , calculate S t The maximum likelihood estimate of
[0190]
[0191] By storing the optimal subpopulation S of each generation t Get Archive A t , and combined with the current population optimal solution S t , estimate the covariance matrix C of the new population t+1 :
[0192]
[0193] Among them, A t =S t-1 ∪S t-2 …∪S t-l , S t-i represents the optimal subpopulation selected by the ti-th generation population, l refers to the set archive length, and represents archive A t The length of storage is the optimal subpopulation set selected by the population in the last l generations, and l is a non-negative integer;
[0194] Import archive A t, which weakens the dependence of the estimated covariance matrix on the current population, and then uses the estimated and C t+1 Update the distribution estimation sampling probability model G based on user interest preferences t (X')→G t+1 (X'), thereby guiding the search direction of the evolutionary optimization algorithm.
[0195] According to the elite selection strategy, select the collection to be recommended u The outstanding TopN project individuals in the population submit the TopN project recommendation list to the current user for evaluation. If the user feedback is not ideal, the user’s recent evaluation data content is added, and the TopN is combined with the original population to update the population M(t)→M(t+1).
[0196] Utilizing the model management mechanism, the user interest preference model based on the double sparse variational autoencoder is dynamically updated according to the newly added user-generated content.
[0197] Example 2
[0198] The personalized search recommendation system based on sparse variational autoencoder described in the present invention is as follows: Figure 3 As shown, including,
[0199] A user-generated content collection module is used to collect multi-source heterogeneous user historical interaction data in a network environment to form multi-source heterogeneous user-generated data, including but not limited to: user rating data, project category labels, user text comments, social network relationships and other information;
[0200] The data preprocessing module is used to vectorize the data generated by multi-source heterogeneous users and establish the social network relationship information of users based on the similarities between users; the Bert text vectorization model and feature extraction technology based on natural language processing technology are used to vectorize user text comments. Through Multi-hot encoding, the project category labels are vectorized Establishing social network relationship information of users based on similarities between users;
[0201] The dominant group and the disadvantaged group construction module sets up screening conditions based on multi-source heterogeneous user-generated data, and screens the project groups that the current user likes as the dominant population P a , select the project group that the current customers do not like as the disadvantaged population P d ;
[0202] The user interest preference model building module, under the framework of evolutionary algorithm, concatenates the vectorized representation of the project category label in the multi-source heterogeneous user-generated data with the vectorized representation of the user text comments of the project as the genotype representation of the project evolution individual, and selects the dominant population P according to the dominant population P. a , construct and train a user active interest preference model based on sparse variational autoencoder, according to the disadvantaged population P d , build and train a user negative preference model based on sparse variational autoencoders, use the positive interest preference model based on sparse variational autoencoders and the negative interest preference model based on sparse variational autoencoders to jointly build a user preference model based on double sparse variational autoencoders, and obtain model parameters containing excellent solution feature distribution information and model parameters containing characteristic distribution information of inferior solutions
[0203] The user evaluation proxy model establishment module designs a user evaluation proxy model based on user interest preference according to the trained user interest preference model based on dual sparse variational autoencoders and the combined matrix decomposition method. u (x):
[0204]
[0205] In the formula, f u (x) is the user's predicted rating for the item, and They represent the predicted scores of the user's positive preference model and the user's negative preference model for the item individual x, respectively, and w is the adjustment coefficient;
[0206] The distribution estimation sampling probability module based on user interest preference obtains the Gaussian probability density function through the user active preference model in the trained user interest preference model based on double sparse variational autoencoder, thereby constructing the Gaussian probability model G t (X'), as a distribution estimation sampling probability model based on user interest preferences:
[0207]
[0208] In the formula, is the mean of the individual reconstruction representations of the items; C t The covariance matrix of the reconstructed representation for the individual items.
[0209] The project recommendation set building module uses the distribution estimation sampling probability model based on user interest preferences to sample and generate K excellent new evolutionary individuals N containing user preference information t; According to the similarity criterion, calculate the similarity between the newly evolved individuals and the feasible solution project individuals in the search space, select the feasible solution project individuals with higher similarity, and form the recommended set s u ;
[0210] The evolutionary individual fitness estimation module uses the user evaluation proxy model based on user interest preferences f u (x), estimate the set of recommendations s u The predicted scores of individual items in the project are used as the fitness values of the evolving individuals, replacing the real evaluation of users, participating in the interactive evolutionary optimization process, and according to the model parameters containing the distribution information of the excellent solution features and model parameters containing characteristic distribution information of inferior solutions Guide the search direction of the evolutionary optimization algorithm;
[0211] The project recommendation list building module selects s according to the elite selection strategy u The top N project individuals in the list submit the top N project recommendation list to the current user for evaluation. If the user feedback is not ideal, the user's latest evaluation data content is added, and the top N is combined with the original population to update the population M(t)→M(t+1);
[0212] The model management module uses the model management mechanism to dynamically update the user interest preference model based on the double sparse variational autoencoder according to the newly added user-generated content.
Claims
1. A personalized search recommendation method based on sparse variational autoencoder, characterized by: The following steps are involved: Collect user-generated content in the network environment to form multi-source heterogeneous user-generated data; User-generated content includes user rating data, item category labels, user text comments, and user historical interaction data; Vectorize multi-source heterogeneous user-generated data; Based on multi-source heterogeneous user-generated data, filter out the project groups that the current user likes as the dominant population P a , select the project group that the current customers do not like as the disadvantaged population P d ; The vectorized representation of the project category label in the multi-source heterogeneous user-generated data is concatenated with the vectorized representation of the user text comments of the project as the genotype representation of the evolutionary individual of the project. a , construct and train a user active interest preference model based on sparse variational autoencoder, according to the disadvantaged population P d , construct and train a user negative preference model based on a sparse variational autoencoder, use the positive interest preference model based on a sparse variational autoencoder and the negative interest preference model based on a sparse variational autoencoder to jointly construct a user preference model based on a double sparse variational autoencoder, and obtain the model parameter γ containing the distribution information of the excellent solution features Pa = {W Pa ,b Pa } and the model parameter γ containing the characteristic distribution information of the inferior solution Pd = {W Pd ,b Pd }; According to the user interest preference model based on dual sparse variational autoencoders, the user evaluation proxy model based on user interest preference is designed by joint matrix decomposition method. u (x); The dominant population P a As the initial population M (t = 0), t represents the population algebra, through the user's active interest preference model based on the sparse variational autoencoder, the Gaussian probability density function is obtained, and the distribution estimation sampling probability model G based on the user's interest preference is constructed based on the Gaussian probability density function t (X'); Use the distribution estimation sampling probability model G based on user interest preferences t (X'), generate k new evolutionary individuals N containing user preference information t , according to the similarity criterion, calculate the newly evolved individual N t The similarity between the feasible solution items in the search space to be searched is used to set the screening conditions and screen the feasible solution items that meet the similarity conditions to form the recommended collection s u ; Using the user evaluation proxy model based on user interest preferences u (x) Treat recommended collections u The feasible solution project individuals in the prediction score are used as the fitness value of the new evolutionary individual, and the fitness value is used instead of the user's real evaluation to participate in the interactive evolution; According to the elite selection strategy, select the collection to be recommended u The top N project individuals in the list submit the top N project recommendation list to the current user for evaluation. If the user feedback is not ideal, the user's latest evaluation data content is added, and the top N is combined with the original population to update the population M(t)→M(t+1); Utilizing the model management mechanism, the user interest preference model based on the double sparse variational autoencoder is dynamically updated according to the newly added user-generated content.
2. The personalized search recommendation method based on sparse variational autoencoder according to claim 1, characterized in that: The specific steps for vectorizing multi-source heterogeneous user-generated data are as follows: for item category labels, Multi-hot encoding is used to represent them as binary vectors; for user text comments, natural language Bert model and feature extraction technology are used to vectorize them.
3. The personalized search recommendation method based on sparse variational autoencoder according to claim 1, characterized in that: Based on the user's historical interaction data, the similarity between users is calculated through Pearson correlation and the user social network relationship is established.
4. The personalized search recommendation method based on sparse variational autoencoder according to claim 1, characterized in that: The dominant population P a and the inferior population P d The screening steps are as follows: record the user rating as R = [r ui ], represents the real score of user u on item i, and set the threshold α to represent the like, and select items with real scores higher than the threshold α to form the dominant population P a , let the threshold β represent dislike, and select the projects with real scores lower than the threshold β to form the disadvantaged population P d .
5. The personalized search recommendation method based on sparse variational autoencoder according to claim 1, characterized in that: The steps for constructing a user active interest preference model based on sparse variational autoencoder are as follows: Splice the dominant population P of the current user a The vectorized representation of the project category label and the vectorized representation of the user text comment in X Pa , where X∈P a , establish a training dataset for user active interest preference model based on sparse variational autoencoder The mean of each sample in the training data set is calculated through the encoder network of the sparse variational autoencoder and variance From the above, we can get the variational distribution in, is the input sample The latent variables mapped to the latent space are obtained from the variational distribution using the reparameterization technique A multivariate random variable obtained by sampling and obeying a standard Gaussian distribution ε is random noise drawn from the standard Gaussian distribution N(0,I), and They are The corresponding mean and variance; The trained user active interest preference model based on sparse variational autoencoder effectively obtains the latent variables containing user preferences. The latent variable Z Pa As a decoder network Input, reconstruction and original samples Reconstruct samples as similar as possible and They are The mean and variance of the corresponding Gaussian distribution; Ignore the decoder network output expression Item, the output Item as reconstruction sample Joint Encoder Network and decoder network Get Input The ELBO lower bound on the log-likelihood is: Among them, KL is the Kullback-Leibler divergence distance, which represents the distribution q φ (Z Pa |X Pa ) and the standard Gaussian distribution The difference between Using the reparameterization technique, the log-likelihood lower bound ELBO is used to minimize the negativity of the loss function and introduce a sparsity penalty term Get the loss function L of the sparse variational autoencoder SVAE for: in, η is a hyperparameter used to balance L VAE and the relative importance of the two losses; Sparse variational autoencoder introduces sparsity penalty term Where p is the expected proportion of hidden layer neuron activation; q is the actual average activation of hidden layer neurons; In the iterative training process of the user active interest preference model based on the sparse variational autoencoder, the difference between p and q is calculated by KL divergence as the penalty term Achieve sparsity.
6. The personalized search recommendation method based on sparse variational autoencoder according to claim 1, characterized in that: The steps for constructing the user negative interest preference model based on sparse variational autoencoder are as follows: According to the inferior population P d , build and train a user negative interest preference model based on sparse variational autoencoder; the disadvantaged population contains the current user's negative preference information, and integrates the category labels of the current user's negative preference items and the user's text comments on these items to obtain Establishing a training dataset for user passive interest preference model based on sparse variational autoencoder The mean of each sample is calculated through the encoder network of the sparse variational autoencoder and variance From the above, we can get the variational distribution in, is the input sample The latent variables mapped to the latent space are obtained from the variational distribution using the reparameterization technique A multivariate random variable obtained by sampling and obeying a standard Gaussian distribution ε is random noise drawn from the standard Gaussian distribution N(0,I), and They are The corresponding mean and variance; The trained user negative interest preference model based on sparse variational autoencoder effectively obtains the latent variables containing user negative preferences. Hidden variables As a decoder network Input, reconstruction and original samples Reconstruct samples as similar as possible and They are The mean and variance of the corresponding Gaussian distribution; The decoder network output expression is usually ignored Item, the output Item as reconstruction sample Joint Encoder Network and decoder network Get Input The ELBO lower bound on the log-likelihood is: Among them, KL represents the distribution With standard Gaussian distribution The difference between Using the reparameterization technique, the log-likelihood lower bound ELBO is used to minimize the negativity of the loss function and introduce a sparsity penalty term Get the loss function L of the sparse variational autoencoder SVAE for: in, is a hyperparameter used to balance L VAE and the relative importance of the two losses; Sparse variational autoencoder introduces sparsity penalty term Where p is the expected proportion of hidden layer neuron activation; q is the actual average activation of hidden layer neurons; In the iterative training process of the user negative interest preference model based on the sparse variational autoencoder, the difference between p and q is calculated by KL divergence as the penalty term Achieve sparsity.
7. The personalized search recommendation method based on sparse variational autoencoder according to claim 3, characterized in that:The user evaluation proxy model based on user interest preference u The construction steps of (x) are as follows: vectorize the individual items in the dominant population As the hidden representation of the project's positive preference, it is sent to the trained user's positive interest preference model based on sparse variational autoencoder to reconstruct the user's positive preference hidden representation Vectorize the individual items in the disadvantaged population As the hidden representation of the negative preference of the item, it is sent to the trained user negative interest preference model based on sparse variational autoencoder to reconstruct the hidden representation of the user negative preference Using the user social network relationship and combining the matrix decomposition method, we can obtain the current user's predicted score for the item individual x. and The calculation formula is as follows: Among them, b u is the user's bias, which represents the rating tendency of other users in the current user's social network relationship to the item individual x; b i is the bias of the project individual, indicating the basic popularity of the project; μ All is the average of all item ratings; Therefore, according to the user interest preference model based on the dual sparse variational autoencoder and the matrix decomposition method, a user evaluation proxy model based on user interest preference is designed. u (x): Among them, f u (x) is the predicted rating of item x by user u; and They respectively represent the predicted scores of the user's positive interest preference model and the user's negative interest preference model based on the sparse variational autoencoder for the item individual x; w is the adjustment coefficient.
8. The personalized search recommendation method based on sparse variational autoencoder according to claim 1, characterized in that: The steps for constructing the distribution estimation sampling probability model based on user interest preferences are as follows: Through the trained user interest preference model based on double sparse variational autoencoders, the activation probability r of the reconstruction layer neural unit of the sparse variational autoencoder in the user's active interest preference model is obtained. i : Among them, the state h of the jth neuron in the hidden layer of the decoder in the user active interest preference model based on the sparse variational autoencoder is j : h j =tanh(W lj z l +b j ) Thus, the reconstructed representation of the decision variables of the individual items in the initial population M (t = 0) is obtained: And according to the n1-dimensional variables of the evolving individuals Calculate the mean of the individual reconstructed representations of the items and the covariance matrix elements Get the mean of the item category attribute features containing user preference information and the covariance matrix C t Based on the above mean and covariance matrix, the Gaussian probability density function is parameterized and a Gaussian probability model is constructed as a distribution estimation sampling probability model G based on user interest preferences. t (X'):
9. The personalized search recommendation method based on sparse variational autoencoder according to claim 1, characterized in that: The steps for estimating the fitness value of the project individuals in the updated population are as follows: According to the principle of survival of the fittest, select the top n% of individuals with higher fitness values of the project individuals to form the optimal subpopulation S t , calculate S t The maximum likelihood estimate of By storing the optimal subpopulation S of each generation t Get Archive A t , and combined with the current population optimal solution S t , estimate the covariance matrix C of the new population t+1 : Among them, A t =S t-1 ∪S t-2 …∪S t-l , S t-i represents the optimal subpopulation selected by the ti-th generation population, l refers to the set archive length, and represents archive A t The length of the storage is the optimal subpopulation set selected by the population in the last l generations, and l is a non-negative integer; Import archive A t , which weakens the dependence of the estimated covariance matrix on the current population, and then uses the estimated and C t+1 Update the distribution estimation sampling probability model G based on user interest preferences t (X')→G t+1 (X'), thereby guiding the search direction of the evolutionary optimization algorithm.
10. A personalized search recommendation system based on sparse variational autoencoder, characterized by: include, A user-generated content collection module is used to collect multi-source heterogeneous user historical interaction data in a network environment to form multi-source heterogeneous user-generated data; The data preprocessing module is used to vectorize the multi-source heterogeneous user-generated data and establish the social network relationship information of users based on the similarity between users; The dominant group and the disadvantaged group construction module sets up screening conditions based on multi-source heterogeneous user-generated data, and screens the project groups that the current user likes as the dominant population P a , select the project group that the current customers do not like as the disadvantaged population P d ; The user interest preference model building module, under the framework of evolutionary algorithm, concatenates the vectorized representation of the project category label in the multi-source heterogeneous user-generated data with the vectorized representation of the user text comments of the project as the genotype representation of the project evolution individual, and selects the dominant population P according to the dominant population P. a , construct and train a user active interest preference model based on sparse variational autoencoder, according to the disadvantaged population P d , build and train a user negative preference model based on sparse variational autoencoders, use the positive interest preference model based on sparse variational autoencoders and the negative interest preference model based on sparse variational autoencoders to jointly build a user preference model based on double sparse variational autoencoders, and obtain model parameters containing excellent solution feature distribution information and model parameters containing characteristic distribution information of inferior solutions The user evaluation proxy model establishment module designs a user evaluation proxy model based on user interest preference according to the trained user interest preference model based on dual sparse variational autoencoders and the combined matrix decomposition method. u (x): In the formula, f u (x) is the user's predicted rating for the item; and They represent the predicted scores of the user's positive preference model and the user's negative preference model for the item individual x; w is the adjustment coefficient; The distribution estimation sampling probability module based on user interest preference obtains the Gaussian probability density function through the user active preference model in the trained user interest preference model based on double sparse variational autoencoder, thereby constructing the Gaussian probability model G t (X'), as a distribution estimation sampling probability model based on user interest preferences: In the formula, is the mean of the individual reconstruction representations of the items; C t Reconstruct the covariance matrix of the representation for the individual items; The project recommendation set building module uses the distribution estimation sampling probability model based on user interest preferences to sample and generate K excellent new evolutionary individuals N containing user preference information t ; According to the similarity criterion, the similarity between the newly evolved individuals and the feasible solution project individuals in the search space is calculated, and the feasible solution project individuals with higher similarity are selected to form the recommended set s u ; The evolutionary individual fitness estimation module uses the user evaluation proxy model based on user interest preferences f u (x), estimate the set of recommendations s u The predicted scores of individual items in the project are used as the fitness values of the evolving individuals, replacing the real evaluation of users, participating in the interactive evolutionary optimization process, and according to the model parameters containing the distribution information of the excellent solution features and model parameters containing characteristic distribution information of inferior solutions Guide the search direction of the evolutionary optimization algorithm; The project recommendation list building module selects s according to the elite selection strategy u The top N project individuals in the list submit the top N project recommendation list to the current user for evaluation. If the user feedback is not ideal, the user's latest evaluation data content is added, and the top N is combined with the original population to update the population M(t)→M(t+1); The model management module uses the model management mechanism to dynamically update the user interest preference model based on the double sparse variational autoencoder according to the newly added user-generated content.
Citation Information
Cited By
Commodity auction intelligent operation management method and system based on artificial intelligence
CN119323460A
