Explainable and efficient autoencoder system and method for implicit recommendation

Abstract

The present invention provides an explainable and efficient autoencoder system and method for implicit recommendation, comprising: step S1: establishing an interaction matrix based on the interaction records of all users and all items; step S2: preprocessing the interaction matrix to obtain a preprocessed interaction matrix; step S3: using the preprocessed interaction matrix to train a linear encoder model to obtain a trained linear encoder model; step S4: using the trained linear encoder model to perform personalized recommendations based on the interaction records of users and items.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation systems, and in particular, to an explainable and efficient autoencoder system and method for implicit recommendation, and more particularly, to an explainable and efficient autoencoder model for implicit recommendation. Background Art

[0002] In implicit recommendation, how to utilize the interaction records between users and items to improve the accuracy of recommendations is the main research focus.

[0003] Latent factorization model singular value decomposition (SVD), which learns left singular vectors and right singular vector matrices to represent users and items. Matrix factorization (MF) is one of the most popular collaborative filtering methods, which can map users and items to a low-dimensional embedding space. In addition, neural network matrix factorization (NNMF) enhances previous matrix factorization methods by inputting features into a neural network to model more complex interactions.

[0004] The classic autoencoder model AutoRec takes the rating vector of the user or item as the input of the encoder, and the decoder reconstructs it. Similarly, variational autoencoders (VAE) have also been applied to recommendations. User-based AutoRec and item-based AutoRec are complementary. The joint collaboration of the two AutoRecs can simultaneously learn the relationship between users and the relationship between items.

[0005] GCN, a rising star in graph network models, has advantages over other collaborative filtering models. Advanced graph methods regard users and items as nodes in a bipartite graph and use the natural advantages of the graph to constrain the relationship between users and items. In order to explicitly capture the high-order interactions in the user-item graph, the collaborative signal is embedded into the message propagation process. LightGCN, a simplified graph model, simplifies the model by removing nonlinear operators and achieves better performance, which shows that linear models can perform CF tasks better than nonlinear models when only node IDs or rating matrices are available. Therefore, the development of CF graph models shows that collaborative filtering is essentially a bipartite graph link prediction task, in which capturing the internal structural relationship between users and items is very important. However, more complex graph models not only make training difficult, but also weaken the robustness and generalization ability of the model.

[0006] Patent document CN112347362A (application number: 202011283015.5) discloses a personalized recommendation method based on a graph autoencoder. It uses the interaction between users and items to construct an adjacency matrix and normalizes it. It uses a graph convolutional network to perform convolution operations to obtain the hidden layer representation of the node; uses user comment text and item description text as the source of node information to obtain the initial feature vector of each node, and then uses a graph attention network to aggregate neighbor node features and update node information; uses the attribute features of users and items to construct a fully connected network to calculate the hidden layer features; splices the hidden layer features to obtain new node information, constructs a fully connected network, encodes, and uses a bilinear decoder to reconstruct the user's rating of the item as a predicted rating, and uses Top-N recommendation to generate a list of recommended items based on the predicted rating.

[0007] The present invention relates to the collaborative filtering problem in recommendation, and how to improve the quality of recommendation when there are only interaction records between users and items. The application of existing autoencoders in implicit recommendation relies too much on complex nonlinear networks. The present invention mainly simplifies the autoencoder into a more interpretable and efficient recommendation model, and improves the induction and generalization capabilities of recommendation.

[0008] The automatic encoder of the present invention is simplified to a linear encoder with only two parameters, and the input is dropout. The decoder is regarded as a weighted matrix decomposition, which theoretically unifies the above three models and further introduces a joint linear autoencoder combined with hinge-loss to enhance the recommendation effect. Summary of the invention

[0009] In view of the defects in the prior art, the object of the present invention is to provide an explainable and efficient autoencoder system and method for implicit recommendation.

[0010] According to the present invention, an explainable and efficient autoencoder system for implicit recommendation is provided, comprising:

[0011] Module M1: Establish an interaction matrix based on the interaction records of all users and all items;

[0012] Module M2: preprocessing the interaction matrix to obtain a preprocessed interaction matrix;

[0013] Module M3: Use the preprocessed interaction matrix to perform batch parallel training on the linear encoder model to obtain a trained linear encoder model;

[0014] Module M4: Personalized recommendation based on the interaction records between users and items using the trained linear encoder model;

[0015] The linear encoder model realizes parallel training and achieves the recommendation effect by preprocessing the input in the encoder and decomposing the weight matrix applied to the decoder.

[0016] Preferably, the step S2 adopts: preprocessing the batch-sized interaction records input into the encoder through L1 normalization and dropout with a preset drop probability to obtain preprocessed interaction records.

[0017] Preferably, the linear encoder model comprises an encoder and a decoder;

[0018] The encoder aggregates neighbor representations of users or items based on interaction records of users or items to obtain representations of target users or items;

[0019] The decoder obtains the predicted interaction record by using the dot product of the representation of the target user or item and the item or user parameter representation in the decoder.

[0020] Preferably, in the encoder module, the matrix expression of the encoder module is:

[0021] Z=RW1

[0022] Among them, Z represents the potential representation of the target user; W1 represents the parameters of the encoder; R represents the interaction record matrix of users or items; the dimension of R is m*n, m represents the number of users, and n represents the number of items; the dimension of W1 is n*d, d represents the dimension of the hidden layer.

[0023] Preferably, in the decoder module, the matrix expression of the decoder module is:

[0024]

[0025] in, Represents the predicted user interaction record; W2 represents the weight parameter of the decoder, the dimension is d*n, d represents the dimension of the hidden layer; n represents the number of items.

[0026] Preferably, in the encoder module, the interaction record r of user u u The potential representation of user u obtained by the encoder (1*n) is as follows:

[0027] z u =r u1 W1[1,:]+r u2 W1[1,:]+…+r un W1[n,:]

[0028] Among them, W1[i,:] represents the i-th row of W1, and the middle represents the dimension of 1*d.

[0029] Preferably, in the decoder module, the latent representation zu of user u is decoded by the decoder to obtain the final interaction prediction of user u as follows:

[0030]

[0031] Where W2[:,i] represents the i-th row and column of W2; the superscript T represents transposition;

[0032] Preferably, a weighted regression loss is used to assign a confidence C(m*n) to each prediction in the interaction matrix:

[0033]

[0034] Where F represents the Frobenius norm and L is the distance loss between the decoder and the original input R.

[0035] Preferably, a joint linear autoencoder is designed by simultaneously reconstructing the rows and columns of the interaction matrix and averaging and adding them;

[0036]

[0037] in, represents the user-based reconstruction matrix, Represents the item-based reconstruction matrix.

[0038] According to the present invention, an explainable and efficient autoencoder method for implicit recommendation is provided, comprising:

[0039] Step S1: Establish an interaction matrix based on the interaction records of all users and all items;

[0040] Step S2: preprocessing the interaction matrix to obtain a preprocessed interaction matrix;

[0041] Step S3: using the preprocessed interaction matrix to perform batch parallel training on the linear encoder model to obtain a trained linear encoder model;

[0042] Step S4: Perform personalized recommendations based on the interaction records between users and items using the trained linear encoder model;

[0043] The linear encoder model realizes parallel training and achieves the recommendation effect by preprocessing the input in the encoder and decomposing the weight matrix applied to the decoder.

[0044] Compared with the prior art, the present invention has the following beneficial effects: based on permutation invariance, the linear encoder can learn the structural representation of users and items on the interaction graph, and then predict the link relationship, which can approximate the graph neural network, and ensure that the input dropout will not be over-smoothed, and finally be equivalent to an infinitely multi-layer graph network, which has a powerful graph representation learning ability. The joint linear encoder is combined with the hinge-loss function. It can also enhance the recommendation ranking. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0046] Figure 1 Schematic diagram of the linear autoencoder model.

[0047] Figure 2 Schematic diagram of the joint linear autoencoder model. DETAILED DESCRIPTION

[0048] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0049] The present invention simplifies the autoencoder into a more interpretable linear model with fewer parameters and uses it for implicit recommendation. Specifically: In order to optimize the decoder more effectively, the decoder is regarded as a weighted matrix decomposition. In order to simultaneously capture the correlation between users and items, a joint linear autoencoder framework is introduced to learn user preferences by reconstructing the rows and columns of the user-item interaction matrix. At the same time, the present invention adopts the hinge-loss objective function to improve the top-K recommendation. The key idea of ​​the present invention is that linear autoencoders have the potential to approximate singular value decomposition through neural networks, so the important conclusion of the joint linear autoencoder framework is that the final matrix reconstruction benefits from simultaneously learning the left and right eigenvectors of the matrix. In addition, the permutation invariance of the linear autoencoder, that is, independent of the input order of the user or item ID, indicates that the linear autoencoder can also learn structural representations and provide induction and generalization capabilities for recommendations.

[0050] Example 1

[0051] According to the present invention, an explainable and efficient autoencoder method for implicit recommendation is provided. Figures 1 to 2 As shown, including:

[0052] Step S1: Establish an interaction matrix based on the interaction records of all users and all items;

[0053] Step S2: preprocessing the interaction matrix to obtain a preprocessed interaction matrix;

[0054] Step S3: using the preprocessed interaction matrix to perform batch parallel training on the linear encoder model to obtain a trained linear encoder model;

[0055] Step S4: Perform personalized recommendations based on the interaction records between users and items using the trained linear encoder model;

[0056] The linear encoder model realizes parallel training and achieves the recommendation effect by preprocessing the input in the encoder and decomposing the weight matrix applied to the decoder.

[0057] Specifically, the step S2 adopts: preprocessing the batch-sized interaction records input into the encoder through L1 normalization and dropout with a preset drop probability to obtain the preprocessed interaction records.

[0058] Specifically, the linear encoder model includes an encoder and a decoder;

[0059] The encoder input is the interaction record of the user or item, that is, the row or column of the original interaction matrix R, which contains only one parameter, representing all items or users, and is used to aggregate the representations of the neighbors of the user or item (the items or users that interact with it) to output the representation of the target user or item. Specifically, the representations of the items that the user interacts with are added together to approximate the representation of the user, because only the value of 1 is considered in a row of interaction records, and the corresponding rows selected in the encoder parameters are added to obtain the output of the encoder, and vice versa.

[0060] The decoder input is the representation of the target user or item, contains only one parameter, which represents the structural representation of all items or users, and the input dot product, and outputs the predicted interaction record.

[0061] Specifically, in the encoder module, the matrix expression of the encoder module is:

[0062] Assume that the user interaction records are used as input, that is, the rows of the interaction matrix R are used as input. The same is true for the case where the item interaction records are used as input. W1 is the weight parameter of the encoder, z is the target user's potential representation, and the matrix expression of the encoder module is:

[0063] Z=RW1

[0064] The dimension of R is m*n, where m is the number of users and n is the number of items. The dimension of W1 is n*d, where d is the dimension of the hidden layer.

[0065] Specifically, in the decoder module, the matrix expression of the decoder module is:

[0066]

[0067] in, Represents the predicted user interaction record; W2 represents the weight parameter of the decoder, the dimension is d*n, d represents the dimension of the hidden layer; n represents the number of items.

[0068] Specifically, in the encoder module, the interaction record r of user u is u The potential representation of user u obtained by the encoder (1*n) is as follows:

[0069] z u =r u1 W1[1,:]+r u2 W1[1,:]+…+r un W1[n,:]

[0070] Among them, W1[i,:] represents the i-th row of W1, and the middle represents the dimension of 1*d.

[0071] Specifically, in the decoder module, the latent representation z of user u is u After decoding by the decoder, the final interaction prediction of user u is as follows:

[0072]

[0073] Where W2[:,i] represents the i-th row and column of W2; the superscript T represents transposition;

[0074] Weight matrix decomposition: For implicit recommendations, observed interactions are sparse, while the scale of unobserved interactions is much larger, which means that users will have potential interest in those unobserved items. Here, weighted regression loss (WMF) is used, which assigns a confidence C(m*n) to each prediction in the interaction matrix. Generally speaking, the distance loss between the final output ZW2 of the encoder and the initial input R is the model loss. However, in implicit recommendations, there are many interactions without interactions, that is, there are many 0s in the interaction matrix R, but it does not mean that 0 means that the user is not interested. At this time, it cannot be treated the same as the case of value 1. Here, the interaction with value 1 can be assigned a confidence of 1, and the interaction with value 0 can be assigned a confidence value less than l, indicating that the user may have a certain degree of interest in this uninteracted item:

[0075]

[0076] Where F represents the Frobenius norm and L is the distance loss between the decoder and the original input R.

[0077] Node dropout: The training method is to randomly drop an element in each user rating vector, which has been L1 normalized, where those non-zero elements are rescaled after being dropped. Dropout helps the encoder simulate "sampling" and thus generalize well to unobserved ratings.

[0078] Batch training: The mini-batch training process is convenient because only mini-batch rows of the rating matrix and weight matrix are selected at randomly permuted indices without loading the entire matrix into the model. After normalizing the interaction matrix rows, a dropout operation with appropriate retention probability is required to avoid overfitting. The high efficiency of training mainly depends on two aspects: 1) The whole process is calculated in parallel. 2) Only two parameters W1 and W2 need to be trained, as well as the sparsity of the interaction matrix.

[0079] Simplify the autoencoder into a more interpretable linear model with fewer parameters for implicit recommendation, which has the advantages of permutation invariance and correlation with other recommendation models SVD and graph neural network methods to prove the interpretability of this model;

[0080] Permutation invariance: The result of the linear autoencoder is independent of the input order of the user interaction records, which means that the structural representation of the items in the interaction graph is independent of the input order of the user interaction records, that is, the two representations W1 and W2 of the items are independent of the user ID encoding, and the structural representation of the items can be learned (the relationship between the items is independent of the user-specific ID encoding, and the ID can be exchanged at will without affecting the correlation between the items, that is, the structured representation)

[0081] Approximate SVD: Through appropriate parameter setting, W1 can be regarded as a solution of the right singular value vector of R.

[0082] Approximate Graph Neural Network Methods:

[0083] In graph neural network recommendation without auxiliary information, the user's preference prediction for unobserved items can be transformed into a link prediction problem in a bipartite graph. The simplified framework of CF's graph method is defined as follows:

[0084] 1) Aggregation and propagation The target node aggregates the features of neighbor nodes and updates its own representation.

[0085] 2) Combine each user or item in each layer by combining the node representations in all layers, such as concatenation or summation.

[0086] 3) Model prediction performs a dot product operation on the final representation of the user and item to obtain the prediction score.

[0087] By applying dropout to the input of the linear encoder, the linear autoencoder can be equivalent to an infinitely multi-layered graph network.

[0088] Specifically, a joint linear autoencoder can be designed, that is, the rows and columns of the interaction matrix are reconstructed at the same time and averaged (both dimensions are m*n). Two cases of row input and column input are given here. When the two cases are considered together, the reconstructed matrix is ​​the output of the final decoder. Because the distance loss between the reconstructed matrix and the original input matrix R is the model loss, the reconstructed matrix is ​​the final predicted matrix, similar to the interaction matrix R above:

[0089]

[0090] in represents the user-based reconstruction matrix, Represents the item-based reconstruction matrix.

[0091] According to the present invention, an explainable and efficient autoencoder system for implicit recommendation is provided, comprising:

[0092] Module M1: Establish an interaction matrix based on the interaction records of all users and all items;

[0093] Module M2: preprocessing the interaction matrix to obtain a preprocessed interaction matrix;

[0094] Module M3: Use the preprocessed interaction matrix to perform batch parallel training on the linear encoder model to obtain a trained linear encoder model;

[0095] Module M4: Personalized recommendation based on the interaction records between users and items using the trained linear encoder model;

[0096] The linear encoder model realizes parallel training and achieves the recommendation effect by preprocessing the input in the encoder and decomposing the weight matrix applied to the decoder.

[0097] Specifically, the module M2 adopts: preprocessing the batch-sized interaction records input into the encoder through L1 normalization and dropout with a preset dropout probability to obtain preprocessed interaction records.

[0098] Specifically, the linear encoder model includes an encoder and a decoder;

[0099] The encoder input is the interaction record of the user or item, that is, the row or column of the original interaction matrix R, which contains only one parameter, representing all items or users, and is used to aggregate the representations of the neighbors of the user or item (the items or users that interact with it) to output the representation of the target user or item. Specifically, the representations of the items that the user interacts with are added together to approximate the representation of the user, because only the value of 1 is considered in a row of interaction records, and the corresponding row selected in W1 is added to obtain the output of the encoder, and vice versa.

[0100] The decoder input is the representation of the target user or item, contains only one parameter, which represents the structural representation of all items or users, and the input dot product, and outputs the predicted interaction record.

[0101] Specifically, in the encoder module, the matrix expression of the encoder module is:

[0102] Assume that the user interaction records are used as input, that is, the rows of the interaction matrix R are used as input. The same is true for the case where the item interaction records are used as input. W1 is the weight parameter of the encoder, Z is the potential representation of the target user, and the matrix expression of the encoder module is:

[0103] Z=RW1

[0104] The dimension of R is m*n, where m is the number of users and n is the number of items. The dimension of W1 is n*d, where d is the dimension of the hidden layer.

[0105] Specifically, in the decoder module, the matrix expression of the decoder module is:

[0106]

[0107] in, Represents the predicted user interaction record; W2 represents the weight parameter of the decoder, the dimension is d*n, d represents the dimension of the hidden layer; n represents the number of items.

[0108] Specifically, in the encoder module, the interaction record r of user u is u The potential representation of user u obtained by the encoder (1*n) is as follows:

[0109] z u =r u1 W1[1,:]+r u2 W1[1,:]+…+r un W1[n,:]

[0110] Among them, W1[i,:] represents the i-th row of W1, and the middle represents the dimension of 1*d.

[0111] Specifically, in the decoder module, the latent representation z of user u isu After decoding by the decoder, the final interaction prediction of user u is as follows:

[0112]

[0113] Where W2[:,i] represents the i-th row and column of W2; the superscript T represents transposition;

[0114] Weight matrix decomposition: For implicit recommendations, observed interactions are sparse, while the scale of unobserved interactions is much larger, which means that users will have potential interest in those unobserved items. Here, weighted regression loss (WMF) is used, which assigns a confidence C(m*n) to each prediction in the interaction matrix. Generally speaking, the distance loss between the final output ZW2 of the encoder and the initial input R is the model loss. However, in implicit recommendations, there are many interactions without interactions, that is, there are many 0s in the interaction matrix R, but it does not mean that 0 means that the user is not interested. At this time, it cannot be treated the same as the case of value 1. Here, the interaction with value 1 can be assigned a confidence of 1, and the interaction with value 0 can be assigned a confidence value less than 1, indicating that the user may have a certain degree of interest in this uninteracted item:

[0115]

[0116] Where F represents the Frobenius norm and L is the distance loss between the decoder and the original input R.

[0117] Node dropout: The training method is to randomly drop an element in each user rating vector, which has been L1 normalized, where those non-zero elements are rescaled after being dropped. Dropout helps the encoder simulate "sampling" and thus generalize well to unobserved ratings.

[0118] Batch training: The mini-batch training process is convenient because only mini-batch rows of the rating matrix and weight matrix are selected at randomly permuted indices without loading the entire matrix into the model. After normalizing the interaction matrix rows, a dropout operation with appropriate retention probability is required to avoid overfitting. The high efficiency of training mainly depends on two aspects: 1) The whole process is calculated in parallel. 2) Only two parameters W1 and W2 need to be trained, as well as the sparsity of the interaction matrix.

[0119] Simplify the autoencoder into a more interpretable linear model with fewer parameters for implicit recommendation, which has the advantages of permutation invariance and correlation with other recommendation models SVD and graph neural network methods to prove the interpretability of this model;

[0120] Permutation invariance: The result of the linear autoencoder is independent of the input order of the user interaction records, which means that the structural representation of the items in the interaction graph is independent of the input order of the user interaction records, that is, the two representations W1 and W2 of the items are independent of the user ID encoding, and the structural representation of the items can be learned (the relationship between the items is independent of the user-specific ID encoding, and the ID can be exchanged at will without affecting the correlation between the items, that is, the structured representation)

[0121] Approximate SVD: Through appropriate parameter setting, W1 can be regarded as a solution of the right singular value vector of R.

[0122] Approximate Graph Neural Network Methods:

[0123] In graph neural network recommendation without auxiliary information, the user's preference prediction for unobserved items can be transformed into a link prediction problem in a bipartite graph. The simplified framework of CF's graph method is defined as follows:

[0124] 1) Aggregation and propagation The target node aggregates the features of neighbor nodes and updates its own representation.

[0125] 2) Combine each user or item in each layer by combining the node representations in all layers, such as concatenation or summation.

[0126] 3) Model prediction performs a dot product operation on the final representation of the user and item to obtain the prediction score.

[0127] By applying dropout to the input of the linear encoder, the linear autoencoder can be equivalent to an infinitely multi-layered graph network.

[0128] Specifically, a joint linear autoencoder can be designed, that is, the rows and columns of the interaction matrix are reconstructed at the same time and averaged (both dimensions are m*n). Two cases of row input and column input are given here. When the two cases are considered together, the reconstructed matrix is ​​the output of the final decoder. Because the distance loss between the reconstructed matrix and the original input matrix R is the model loss, the reconstructed matrix is ​​the final predicted matrix, similar to the interaction matrix R above:

[0129]

[0130] in represents the user-based reconstruction matrix, Represents the item-based reconstruction matrix.

[0131] Example 2

[0132] Embodiment 2 is a preferred embodiment of Embodiment 1

[0133] The present invention provides a recommended model of an efficient linear encoder. The present invention is described in more detail below through preferred examples.

[0134] Taking the ML1M movie recommendation dataset as an example, the dataset has 6072 users, 3062 items, and a total of 574026 interaction records. 80% of the dataset is used as a training set, 20% as a test set, and 10% of the test set is used as a validation set for parameter adjustment.

[0135] For the linear encoder model, we will do the following steps in sequence:

[0136] 1) Input data: interaction matrix R

[0137] 2) Model training

[0138] a. Initialize model parameters: Randomly initialize the parameters in the model.

[0139] b. Use the training set to train the model: Usually we randomly extract 256 samples from the training set to form a batch, then perform L1 normalization, use dropout with an appropriate dropout probability, and then input the samples into the neural network for training. We use a learning rate of 0.001 and select the best hyperparameter based on its performance on the validation set. We train each model for 100 epochs, and the actual results show that the models basically converge within 25 epochs.

[0140] c. Select hyperparameters: Train the model by setting different hyperparameters, such as weight matrix, drop probability, and weight parameter dimension. Finally, we use the validation set to compare the performance of different models and select the model with the smallest prediction error, that is, the highest prediction accuracy, as the final model.

[0141] 3) Prediction

[0142] Use the best model trained and selected in (2) to make predictions on the test set.

[0143] For the joint linear encoder model, we will perform the following steps in sequence:

[0144] 1) Input data: interaction matrix R

[0145] 2) Model training

[0146] a. Initialize model parameters: Randomly initialize the parameters in the model.

[0147] b. Use the training set to train the model: Usually we randomly extract 1280 samples from the training set to form a row batch and a column batch, then perform L1 normalization on each, use dropout with an appropriate dropout probability, and then input two linear autoencoders for training. The loss function consists of the following: the average loss of the row reconstruction loss and the column reconstruction loss, and the pairwise hinge-loss loss composed of the positive and negative samples selected from the row and column batch intersection. We use a learning rate of 0.001 and select the best hyperparameter based on its performance on the validation set. We train each model for 100 epochs. The actual situation shows that the model basically converges within 25 epochs.

[0148] c. Select hyperparameters: Train the model by setting different hyperparameters, such as weight matrix, drop probability, weight parameter dimension, and hinge-loss edge parameter. Finally, we use the validation set to compare the performance of different models and select the model with the smallest prediction error, that is, the highest prediction accuracy, as the final model.

[0149] Those skilled in the art know that, in addition to implementing the system, device and its various modules provided by the present invention in a purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers and embedded microcontrollers by logically programming the method steps. Therefore, the system, device and its various modules provided by the present invention can be considered as a hardware component, and the modules included therein for implementing various programs can also be considered as structures within the hardware component; the modules for implementing various functions can also be considered as both software programs for implementing the method and structures within the hardware component.

[0150] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. An explainable and efficient autoencoder system for implicit recommendation, characterized in that: include: Module M1: Establish an interaction matrix based on the interaction records of all users and all items; Module M2: preprocessing the interaction matrix to obtain a preprocessed interaction matrix; Module M3: Use the preprocessed interaction matrix to perform batch parallel training on the linear encoder model to obtain a trained linear encoder model; Module M4: Personalized recommendation based on the interaction records between users and items using the trained linear encoder model; The linear encoder model realizes parallel training and achieves the recommendation effect by preprocessing the input in the encoder and decomposing the weight matrix applied to the decoder; In the encoder module, the matrix expression of the encoder module is: in, Indicates the potential representation of target users; represents the parameters of the encoder; R represents the interaction record matrix of users or items; the dimension of R is m*n, where m represents the number of users and n represents the number of items; The dimension is n*d, where d represents the dimension of the hidden layer; In the decoder module, the matrix expression of the decoder module is: in, Represents the predicted user interaction record; represents the weight parameter of the decoder, the dimension is d*n, d represents the dimension of the hidden layer; n represents the number of items; In the encoder module, the user Interaction records The user obtained by the encoder The potential representation of is as follows: in, express No. Row, the middle indicates the dimension is 1*d; In the decoder module, the user Potential representation of After decoding by the decoder, the final user The interactive predictions are as follows: in, express No. Row and column; superscript T indicates transpose; A weighted regression loss is used to assign a confidence score C(m*n) to each prediction in the interaction matrix: Where F represents the Frobenius norm and L is the distance loss between the decoder and the original input R.

2. The explainable and efficient autoencoder system for implicit recommendation according to claim 1, characterized in that: Step S2 adopts: preprocessing the batch size interaction records input into the encoder through L1 normalization and dropout with a preset drop probability to obtain preprocessed interaction records.

3. The explainable and efficient autoencoder system for implicit recommendation according to claim 1, characterized in that: The linear encoder model includes an encoder and a decoder; The encoder aggregates neighbor representations of users or items based on interaction records of users or items to obtain representations of target users or items; The decoder obtains the predicted interaction record by using the dot product of the representation of the target user or item and the item or user parameter representation in the decoder.

4. The explainable and efficient autoencoder system for implicit recommendation according to claim 1, characterized in that: A joint linear autoencoder is designed by reconstructing the rows and columns of the interaction matrix simultaneously and averaging them. in, represents the user-based reconstruction matrix, Represents the item-based reconstruction matrix.

5. An explainable and efficient autoencoder method for implicit recommendation, characterized in that: include: Step S1: Establish an interaction matrix based on the interaction records of all users and all items; Step S2: preprocessing the interaction matrix to obtain a preprocessed interaction matrix; Step S3: using the preprocessed interaction matrix to perform batch parallel training on the linear encoder model to obtain a trained linear encoder model; Step S4: Perform personalized recommendations based on the interaction records between users and items using the trained linear encoder model; The linear encoder model realizes parallel training and achieves the recommendation effect by preprocessing the input in the encoder and decomposing the weight matrix applied to the decoder; In the encoder module, the matrix expression of the encoder module is: in, Indicates the potential representation of target users; represents the parameters of the encoder; R represents the interaction record matrix of users or items; the dimension of R is m*n, where m represents the number of users and n represents the number of items; The dimension is n*d, where d represents the dimension of the hidden layer; In the decoder module, the matrix expression of the decoder module is: in, Represents the predicted user interaction record; represents the weight parameter of the decoder, the dimension is d*n, d represents the dimension of the hidden layer; n represents the number of items; In the encoder module, the user Interaction records The user obtained by the encoder The potential representation of is as follows: in, express No. Row, the middle indicates the dimension is 1*d; In the decoder module, the user Potential representation of After decoding by the decoder, the final user The interactive predictions are as follows: in, express No. Row and column; superscript T indicates transpose; A weighted regression loss is used to assign a confidence score C(m*n) to each prediction in the interaction matrix: Where F represents the Frobenius norm and L is the distance loss between the decoder and the original input R.

Images