A user representation learning recommendation method and system based on fusion confidence

Abstract

The present invention belongs to the technical field of text processing, and provides a user representation learning recommendation method and system based on fusion confidence. The method comprises: collecting the first type of comment texts of the same user on multiple items and the second type of comment texts of different users on the same item, inputting the pre-trained word vector model, and outputting the first embedding vector and the second embedding vector; constructing a confidence matrix, which is used as the correction coefficient of each comment text; correcting the first embedding vector and the second embedding vector according to the constructed confidence matrix, and then performing feature interactive learning after convolution and pooling to obtain a user representation learning recommendation model; inputting the comment text to be processed into the user representation learning recommendation model to obtain user preference recommendation information. The present invention measures the relationship between the score outliers and the misleading comments through the confidence matrix, reduces the influence of the misleading comments on the model, and improves the accuracy of the model recommendation.

Description

Technical Field

[0001] The present invention relates to the technical field of text processing, and in particular to a user representation learning recommendation method and system based on fusion confidence. Background Art

[0002] In recent years, recommendation systems have been widely used in many fields such as news, e-commerce, and movies. Due to information overload, it is becoming increasingly difficult for consumers to choose suitable products on online platforms. Therefore, recommendation systems, which play the role of information filtering in many online platforms, use users' historical behavior data to understand users' hobbies and interests and predict their preferences.

[0003] Most existing recommendation methods based on reviews still have some problems. For example, when most of them use user reviews to construct user features, they do not remove the review noise, and cannot filter effective user reviews to construct user features. Most recommendation models lack the influence of the confidence rate of reviews on the recommendation model, which affects the recommendation performance. In addition, there are problems of sparse user ratings and information overload in the recommendation systems of existing online platforms.

[0004] Therefore, it is necessary to provide a user representation learning recommendation method and system based on fusion confidence to solve the above problems. Summary of the invention

[0005] The present invention aims to provide a user representation learning recommendation method and system based on fusion confidence, so as to solve the technical problems in the prior art that the user features constructed by the existing methods have comment noise, the user preference information has large errors due to sparse user ratings, information overload and comment noise, and thus cause low model prediction accuracy. The technical problem to be solved by the present invention is achieved by the following technical solution.

[0006] The first aspect of the present invention proposes a user representation learning recommendation method based on fusion confidence, comprising:

[0007] The first category of comment texts on multiple items by the same user and the second category of comment texts on the same item by different users are collected; the collected first category of comment texts and the second category of comment texts are input into a pre-trained word vector model to output a first embedding vector and a second embedding vector; based on a confidence function, a confidence matrix is ​​constructed according to all the user's ratings on different items and all the ratings obtained for the same item, and is used as a correction coefficient for each comment text; the first embedding vector and the second embedding vector are corrected according to the constructed confidence matrix, and then after convolution and pooling, feature interaction learning is performed to obtain a user representation learning recommendation model; the comment text to be processed is input into the user representation learning recommendation model to obtain user preference recommendation information.

[0008] The second aspect of the present invention proposes a user representation learning recommendation system based on fusion confidence, which adopts the user representation learning recommendation method based on fusion confidence described in the first aspect of the present invention. The user representation learning recommendation system includes: a data collection module, which collects the first-class comment texts of the same user on multiple items and the second-class comment texts of different users on the same item; inputs the collected first-class comment texts and second-class comment texts into a pre-trained word vector model, and outputs a first embedding vector and a second embedding vector;

[0009] Based on the confidence function, a confidence matrix is ​​constructed according to all user ratings of different items and all ratings of the same item, and is used as a correction coefficient for each comment text; the first embedding vector and the second embedding vector are corrected according to the constructed confidence matrix, and then convolution and pooling are performed to perform feature interaction learning to obtain a user representation learning recommendation model; the comment text to be processed is input into the user representation learning recommendation model to obtain user preference recommendation information.

[0010] The third aspect of the present invention provides an electronic device, comprising: one or more processors; a storage device for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the user representation learning recommendation method based on fusion confidence described in the first aspect of the present invention.

[0011] A fourth aspect of the present invention provides a computer-readable medium having a computer program stored thereon, and when the computer program is executed by a processor, the method for user representation learning and recommendation based on fusion confidence described in the first aspect of the present invention is implemented.

[0012] The embodiments of the present invention include the following advantages:

[0013] Compared with the prior art, the present invention utilizes user reviews and item reviews to learn the effectiveness of user preferences and the interactivity between users and items, and has achieved significant and stable improvements. It measures the relationship between rating outliers and misleading reviews by constructing a confidence matrix, and corrects the review text. It further utilizes the interactivity of review information to construct potential factors of interaction between users and items, so as to reduce the impact of misleading reviews on the model and thereby improve the accuracy of model recommendations.

[0014] In addition, the loss function is constructed using the maximum a posteriori estimation theory, and the batch gradient descent algorithm is introduced to optimize the loss function to optimize the model training process and obtain a more accurate user representation learning recommendation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a flowchart of an example of a user representation learning recommendation method based on fusion confidence of the present invention;

[0016] Figure 2 It is a partial principle schematic diagram of the user representation learning recommendation method based on fusion confidence of the present invention;

[0017] Figure 3 It is a structural block diagram of the user representation learning recommendation system based on fusion confidence of the present invention;

[0018] Figure 4 is a schematic structural diagram of an electronic device embodiment according to the present invention;

[0019] Figure 5 It is a schematic diagram of the structure of a computer-readable medium embodiment according to the present invention. DETAILED DESCRIPTION

[0020] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0021] In view of the above problems, the present invention proposes a user representation learning recommendation method based on fusion confidence. The present invention takes into account the credibility factor of user comments to remove the noise of basic data, and can more accurately construct user dynamic preference features. Specifically, through comment representation learning and historical ratings, an improved confidence-aware recommendation model (i.e., user representation learning recommendation model) is proposed. In addition, in order to reduce the problem of sparse rating data, a new constraint method based on user-item interactivity is introduced, and the potential factors of interaction between the same items and different items are constructed using single review information (multiple users on the same item or project). At the same time, considering the negative impact of misleading comments on the model, a confidence matrix is ​​introduced to eliminate the negative impact of the above misleading comments on the model, thereby further improving the accuracy of recommendations.

[0022] It should be noted that the effectiveness of the user preferences learned by the present invention using user reviews and item reviews is consistent with the interactivity between users and items, and has achieved significant and stable improvements compared to existing methods. In fact, in terms of user preferences, the ratings are consistent with the reviews, but there are also misleading reviews (such as false positive reviews and false negative reviews). To this end, the potential factors of interaction between users and items are constructed by utilizing the interactivity of review information. In addition, the confidence matrix is ​​used to measure the relationship between rating outliers and misleading reviews to further improve the accuracy of the model and reduce the impact of misleading reviews on the model. In addition, the maximum a posteriori estimation theory is used to construct the loss function. Finally, the batch gradient descent algorithm is introduced to optimize the loss function, so that a more accurate user representation learning recommendation model can be obtained. Example

[0023] Refer to the following Figures 1 to 3 , the contents of the present invention will be described in detail.

[0024] Figure 1 It is a flowchart of an example of the user representation learning recommendation method based on fusion confidence of the present invention. Figure 2 It is a schematic diagram of an application example of the user representation learning recommendation method based on fusion confidence of the present invention.

[0025] like Figure 1 As shown, in step S101, first-category review texts of the same user on multiple items and second-category review texts of different users on the same item are collected.

[0026] Specifically, for example, from the first category of comments texts (corresponding to Figure 2 The multiple items include similar items and different items. Specifically, they include various physical items such as shoes, clothes, books, and electrical appliances. They also include movies and music.

[0027] It should be noted that in the present invention, the items include various physical items such as shoes, clothes, and electrical appliances, as well as visual artworks such as movies and short videos, and auditory artworks. The above is only used as an optional example and cannot be understood as a limitation of the present invention.

[0028] Furthermore, for example, the second type of review texts of different users on the same item j are collected from public data on the Internet, that is, review texts of at least two or more users on the same item j.

[0029] Preferably, both the first category review text and the second category review text contain the following information: rating information (e.g., number of stars), whether it is a negative review, a neutral review, or a positive review, and whether there are multiple reviews for the same item (multiple reviews contain contradictory review texts, such as negative reviews and positive reviews).

[0030] Optionally, a single comment text information is extracted through a convolutional neural network (CNN) to obtain the first category comment text and the second category comment text.

[0031] By collecting the first category of comment texts and the second category of comment texts, we can obtain sample text data corresponding to outliers in the ratings and text sample data corresponding to misleading reviews.

[0032] It should be noted that the above is only described as an optional example and should not be understood as a limitation to the present invention.

[0033] Next, in step S102, the collected first category comment texts and second category comment texts are input into a pre-trained word vector model, and a first embedding vector and a second embedding vector are output.

[0034] Specifically, the first category of comment texts and the second category of comment texts collected in step S101 are input into a pre-trained word vector model (eg, word2vec), and a first embedding vector and a second embedding vector are output.

[0035] The original text information is converted into an embedding vector for input into the subsequent model (i.e., user representation learning recommendation model).

[0036] For example, the following vector D is used to represent the embedding vector:

[0037] ;

[0038] Where D represents the word vector set of user i’s comment text on item j, including the first embedding vector and the second embedding vector , Indicates The word vector of .

[0039] It should be noted that the above is only described as an optional example and should not be understood as a limitation to the present invention.

[0040] Next, in step S103, based on the confidence function, a confidence matrix is ​​constructed according to all ratings given by users to different items and all ratings obtained for the same item, and is used as a correction coefficient for each review text.

[0041] Specifically, the latent factors are regarded as the prior constraints on the interaction between the user and the item latent factors.

[0042] A confidence matrix is ​​constructed through the relationship between rating outliers and misleading reviews.

[0043] Based on the confidence function, a confidence matrix is ​​constructed according to all the ratings given by users to different items and all the ratings obtained for the same item.

[0044] The confidence matrix is ​​calculated using the following expression:

[0045] ;

[0046] in, represents a confidence matrix, which is used to represent the contribution of different comment texts to the construction of the user representation learning recommendation model, so as to be used as a correction coefficient for each comment text, and each element in the confidence matrix represents the inverse of the offline probability; Indicates user For items Ratings, j represents items, limited to users Purchased items with ratings, Indicates user For items Rating; Indicates user The score of item j, where Represents the user, and Not equal; Indicates user The mean of all ratings, represents the mean of all the ratings obtained by item j.

[0047] It should be noted that in this example, the calculated confidence matrix is ​​used as a relative adjustment of the regularization of each comment text, and its element values ​​are scaled between 0 and 1. The real comment has a larger weight value to learn the interactive information of the potential factors of the real comment; while the misleading comment has a smaller weight value to learn the interactive information of the potential factors of the misleading comment. When a user gives a high score to an item, the score actually deviates from the average score of the user's historical high score behavior and the average score of the item's history. Therefore, in this case, one or more outliers will appear. Therefore, in this case, misleading comments are given a lower weight. That is, the calculated confidence matrix is ​​used to measure the relationship between the outliers of the score and the misleading comments, thereby improving the performance and robustness of the recommendation algorithm.

[0048] Next, in step S104, the first embedding vector and the second embedding vector are corrected according to the constructed confidence matrix, and then convolution and pooling are performed to perform feature interaction learning to obtain a user representation learning recommendation model.

[0049] A CNN algorithm is used to construct a user representation learning recommendation model, which includes a weighted word vector layer (or embedding layer), a convolution layer, and a pooling layer.

[0050] A training dataset is established using the first category of review texts (specifically, the review texts of user i on items) and the second category of review texts (specifically, the review texts obtained by item j) annotated with abnormal rating labels (for example, two abnormal labels that are higher than a specified value or lower than a specified value) and / or misleading evaluation labels (two labels, misleading positive rating labels and misleading negative rating labels).

[0051] The training dataset is used to train the user representation learning recommendation model.

[0052] In a specific embodiment, the collected first category comment texts and second category comment texts are input into a pre-trained word vector model, and a first embedding vector and a second embedding vector are output to serve as inputs of the weighted word vector layer.

[0053] Before the convolutional layer, the first embedding vector [D']_i and the second embedding vector [D']_j are corrected according to the constructed confidence matrix, and then the corrected first embedding vector and the second embedding vector are input into the weighted word vector layer of the user representation learning recommendation model.

[0054] Specifically, the following expression is used to perform correction calculation at the weighted word vector layer to obtain the input of the convolution layer:

[0055] H= + ;

[0056] Among them, H represents the input of the convolutional layer; represents the vector of the first embedding vector corrected by the confidence matrix; represents the vector of the second embedding vector corrected by the confidence matrix;

[0057] H 单 = × ;

[0058] in, represents the confidence matrix; represents the first embedding vector;

[0059] H 多 = × ;

[0060] in, represents the confidence matrix; Denotes the second embedding vector.

[0061] More specifically, the following matrix is ​​used to represent the input matrix of the convolutional layer:

[0062] ;

[0063] Wherein, H represents the input of the convolution layer, for example, multiple input matrices; represents the sth word vector, represents the parameter value of the sth word vector in the confidence matrix, represents the s+1th word vector, Represents the parameter value of the s+1th word vector in the confidence matrix, represents the s+2th word vector, Represents the parameter value of the s+2th word vector in the confidence matrix. Specifically, the corresponding parameter value in the confidence matrix is ​​used as the word vector to obtain the weight to correct the authenticity of the comment word vector, which is used as the output of the embedding layer and as the input of the convolution layer.

[0064] In the convolution layer, convolution processing is performed to obtain the output result of the convolution layer.

[0065] Use the following expression to get the output of the convolutional layer.

[0066] ;

[0067] in, The output result of the convolution vector representing the review text of item j; represents the context feature vector of the review text obtained by extracting item j through the jth shared weight; t represents the sliding window parameter of the review text obtained by item j, which is used to define the number of word vectors of the review text obtained by item j; represent The offset of represents the input matrix of the convolutional layer of the score obtained by item j, H represents the embedding layer, and i, j, and t are all positive integers.

[0068] Furthermore, the total output result of the following convolutional layer is obtained.

[0069] ;

[0070] in, Represents the convolutional layer vector of the review text obtained for item j, The first element in the convolutional vector representing the review text for item j; Represents the Lth element in the convolutional layer vector of the comment text obtained for item j; L represents the number of convolutional vectors obtained after the comment text of item j passes through the convolutional layer, t is the sliding window parameter of the comment text of item j, and both L and t are positive integers.

[0071] It should be noted that, in this example, the comment text of user i on the item is input into the convolution layer, and the process of obtaining the convolution vector after passing through the convolution layer is roughly the same as the process of obtaining the convolution vector after the comment text obtained by the above-mentioned item j passes through the convolution layer. Therefore, the description of the same part is omitted.

[0072] Next, pooling is performed in the pooling layer to obtain the pooling result.

[0073] ;

[0074] Among them, d represents selecting all the maximum values ​​from the L convolution vectors of the convolution layer. Represents the first convolution vector of the comment text obtained by item j after the convolution layer, It represents the Lth convolution vector of the comment text obtained by item j after the convolution layer, t is the sliding window parameter of the comment text of item j, and L, j, and t are all positive integers.

[0075] Furthermore, the first type of comment text (corresponding to Figure 2 After convolution and pooling, the first maximum pooling vector d_i is obtained, and the second type of comment text (corresponding to Figure 2 After learning, the second largest pooling vector d_j is obtained.

[0076] The interactive features are learned and predicted. In the interactive feature learning process, the Hadamard product is used for learning. Specifically, the following expression is used:

[0077] ;

[0078] Among them, d i Represents the first maximum pooling vector corresponding to the first type of comment text; d j Represents the second largest pooling vector corresponding to the second type of comment text, d i1 Represents the first maximum pooling vector of user i’s comment text on the item, d in represents the nth largest pooled vector of the user's comment text on the item, and n represents the number of the largest pooled vectors obtained by the user i's comment text on the item in the pooling layer; d j1 Represents the first maximum pooled vector of the review text obtained for item j; d jm represents the mth maximum pooled vector of the comment text obtained by item j, and m represents the number of maximum pooled vectors obtained by the comment text obtained by item j in the pooling layer.

[0079] In the output layer, the existing nonlinear mapping is used to obtain the potential factors of the comments. Specifically, the following expression is used:

[0080] ;

[0081] in, It means using CNN to learn the interactive features of the latent factors of user reviews and item reviews, i.e., the latent factors of reviews; represents the first adjustable parameter, a weight parameter used to represent the interaction characteristics of potential factors, 0.5 to 0.9; represents the second adjustable parameter, which is used to adjust the weight parameter of the interactive features representing the potential factors, ranging from 0.3 to 0.7; d represents the interactive feature vector obtained by multiplying the first maximum pooling vector corresponding to the first type of comment text and the second maximum pooling vector corresponding to the second type of comment text; represents a first offset, used to correct the first offset of the interaction feature; Represents the second offset, which is used to correct the second offset of the interactive feature.

[0082] It should be noted that in this example, sufficient data must be used to create a confidence matrix, and the complete Yelp dataset can only guarantee that each user has at least four ratings. The Yelp dataset is equivalent to a certain review platform, which specifically contains user information, product information, user ratings of products, and comments. Therefore, some users (those who contribute less than 20 ratings) are deleted from the Yelp dataset to meet the sufficiency requirement. Among them, the experiment partitions each dataset and arranges it in chronological order, with 80% used for training and 20% for testing.

[0083] In an optional implementation, model optimization is performed during the model training process, a loss function is constructed through the maximum a posteriori estimation theory, and a mini-batch gradient descent algorithm is introduced to optimize the loss function.

[0084] And the limiting conditions for stopping model iteration, for example, when the loss value between the predicted score of user i for the item and the actual score of user i for the item is in the range of 0.4 to 0.6 (preferably, when the loss value is less than 0.5), stop training and obtain a trained user representation learning recommendation model.

[0085] It should be noted that the above is only described as an optional example and should not be understood as a limitation to the present invention.

[0086] Next, in step S105, the comment text to be processed is input into the user representation learning recommendation model to obtain user preference recommendation information.

[0087] In a specific implementation, when the user representation learning recommendation model is used for prediction, the comment text of the current user i, the comment text of the item j, the confidence matrix constructed by the score of the item by user i and the score obtained by the item j are input, and the prediction score of the item j by user i is output, and the user preference recommendation information is obtained according to the output prediction score of the item j by user i. The prediction score has a corresponding association relationship with the user preference recommendation information. The user preference recommendation information includes at least two of the following information: item type (e.g., comedy movies), the number of items of the same type (e.g., two movies), and the name of the item (e.g., movie a, movie b).

[0088] In another specific embodiment, when the user representation learning recommendation model is used for prediction, the comment text of the current user i, the comment text of the item j, the confidence matrix constructed by the rating of the item by the user i and the rating of the item j are input, and the user preference recommendation information is output. In the process of training the user representation learning recommendation model, the corresponding association relationship between the comment text of the current user i, the comment text of the item j and the user preference recommendation information is learned.

[0089] It should be noted that the above is only described as an optional example and should not be understood as a limitation to the present invention.

[0090] In addition, the drawings are only schematic illustrations of the processes included in the method according to an exemplary embodiment of the present invention, and are not intended to be limiting. It is easy to understand that the processes shown in the drawings do not indicate or limit the time sequence of these processes. In addition, it is also easy to understand that these processes can be performed synchronously or asynchronously, for example, in multiple modules.

[0091] Compared with the prior art, the present invention utilizes user reviews and item reviews to learn the effectiveness of user preferences and the interactivity between users and items, and has achieved significant and stable improvements. It measures the relationship between rating outliers and misleading reviews by constructing a confidence matrix, and corrects the review text. It further utilizes the interactivity of review information to construct potential factors of interaction between users and items, so as to reduce the impact of misleading reviews on the model and thereby improve the accuracy of model recommendations.

[0092] In addition, the loss function is constructed using the maximum a posteriori estimation theory, and the batch gradient descent algorithm is introduced to optimize the loss function to optimize the model training process and obtain a more accurate user representation learning recommendation model. Example

[0093] The following is a system embodiment of the present invention, which can be used to implement the method embodiment of the present invention. For details not disclosed in the system embodiment of the present invention, please refer to the method embodiment of the present invention.

[0094] Figure 3 4 is a schematic diagram of the structure of an example of a user representation learning recommendation system based on fusion confidence according to the present invention.

[0095] Reference Figure 3 The second aspect of the present disclosure provides a user representation learning recommendation system 300 based on fusion confidence, which adopts the user representation learning recommendation method based on fusion confidence described in the first aspect of the present invention.

[0096] Specifically, the user representation learning recommendation system 300 includes a text collection module 310 , a vector conversion module 320 , a construction module 330 , a correction processing module 340 , and a prediction module 350 .

[0097] In a specific embodiment, the text collection module 310 is used to collect the first type of comment texts of the same user on multiple items and the second type of comment texts of different users on the same item. The vector conversion module 320 is used to input the collected first type of comment texts and second type of comment texts into the pre-trained word vector model, and output the first embedding vector and the second embedding vector. The construction module 330 constructs a confidence matrix based on the confidence function according to all the user's ratings of different items and all the ratings obtained for the same item, and uses it as the correction coefficient for each comment text. The correction processing module 340 corrects the first embedding vector and the second embedding vector according to the constructed confidence matrix, and then performs feature interaction learning after convolution and pooling to obtain a user representation learning recommendation model. The prediction module 350 inputs the comment text to be processed into the user representation learning recommendation model to obtain user preference recommendation information.

[0098] According to an optional implementation, based on the confidence function, a confidence matrix is ​​constructed according to all ratings given by users to different items and all ratings obtained for the same item.

[0099] The confidence matrix is ​​calculated using the following expression:

[0100] ;

[0101] in, represents a confidence matrix, which is used to represent the contribution of different comment texts to the construction of the user representation learning recommendation model, so as to be used as a correction coefficient for each comment text, and each element in the confidence matrix represents the inverse of the offline probability; Indicates user For items Ratings, j represents items, limited to users Purchased items with ratings, Indicates user For items Rating; Indicates user The score of item j, where Represents the user, and Not equal; Indicates user The mean of all ratings, represents the mean of all the ratings obtained by item j.

[0102] A CNN algorithm is used to construct a user representation learning recommendation model, which includes a weighted word vector layer, a convolution layer, and a pooling layer.

[0103] The collected first category comment texts and second category comment texts are input into a pre-trained word vector model, and a first embedding vector and a second embedding vector are output as inputs of the weighted word vector layer.

[0104] Before inputting the convolutional layer, the first embedding vector and the second embedding vector are corrected according to the constructed confidence matrix.

[0105] According to an optional implementation, the following expression is used to perform a correction calculation to obtain the input of the convolution layer:

[0106] H= + ;

[0107] Among them, H represents the input of the convolutional layer; represents the vector of the first embedding vector corrected by the confidence matrix; Represents the vector of the second embedding vector corrected by the confidence matrix.

[0108] H 单 = × ;

[0109] in, represents the confidence matrix; represents the first embedding vector;

[0110] H 多 = × ;

[0111] in, represents the confidence matrix; Denotes the second embedding vector.

[0112] Furthermore, after convolution and pooling, the first maximum pooling vector of the first type of comment text is obtained as d_i; after convolution and pooling, the second maximum pooling vector of the second type of comment text is obtained as d_j;

[0113] The Hadamard product method is used to make the first maximum pooling d_i and the second maximum pooling d_j for feature interactive learning to obtain the potential factors of the comments.

[0114] In the output layer, the existing nonlinear mapping is used to obtain the potential factors of the comments. Specifically, the following expression is used:

[0115] ;

[0116] in, It means using CNN to learn the interactive features of the latent factors of user reviews and item reviews, i.e., the latent factors of reviews; represents the first adjustable parameter, a weight parameter used to represent the interaction characteristics of potential factors, 0.5 to 0.9; represents the second adjustable parameter, which is used to adjust the weight parameter of the interactive features representing the potential factors, ranging from 0.3 to 0.7; d represents the interactive feature vector obtained by multiplying the first maximum pooling vector corresponding to the first type of comment text and the second maximum pooling vector corresponding to the second type of comment text; represents a first offset, used to correct the first offset of the interaction feature; Represents the second offset, which is used to correct the second offset of the interactive feature.

[0117] According to an optional implementation, the review text to be processed includes the first category review text of the current user i, the second category review text for the same item j, and a confidence matrix constructed by the rating of the item by user i and the rating obtained for item j.

[0118] The comment text to be processed is input into the user representation learning recommendation model, and the prediction score of the current user i for the item j is output. The user preference recommendation information is obtained according to the output prediction score of the current user i for the item j, and the user preference recommendation information includes the item type and the number of items of the same category.

[0119] It should be noted that the content of the user representation learning recommendation method based on fusion confidence adopted by the user representation learning recommendation system based on fusion confidence in this embodiment is the same as Figure 1 The contents of the user representation learning recommendation method based on fusion confidence are roughly the same, so the description of the same parts is omitted.

[0120] Compared with the prior art, the present invention utilizes user reviews and item reviews to learn the effectiveness of user preferences and the interactivity between users and items, and has achieved significant and stable improvements. It measures the relationship between rating outliers and misleading reviews by constructing a confidence matrix, and corrects the review text. It further utilizes the interactivity of review information to construct potential factors of interaction between users and items, so as to reduce the impact of misleading reviews on the model and thereby improve the accuracy of model recommendations.

[0121] In addition, the maximum a posteriori estimation theory is used to construct the loss function, and the batch gradient descent algorithm is introduced to optimize the loss function to optimize the model training process and obtain a more accurate user representation learning recommendation model. Figure 4 As shown, the electronic device is presented in the form of a general computing device. The processor may be one or more and work in coordination. The present invention does not exclude distributed processing, that is, the processor may be dispersed in different physical devices. The electronic device of the present invention is not limited to a single entity, but may also be the sum of multiple physical devices.

[0122] The memory stores a computer executable program, which is usually a machine-readable code. The computer-readable program can be executed by the processor to enable the electronic device to perform the method of the present invention, or at least part of the steps in the method.

[0123] The memory includes a volatile memory, such as a random access memory unit (RAM) and / or a cache memory unit, and may also be a non-volatile memory, such as a read-only memory unit (ROM).

[0124] Optionally, in this embodiment, the electronic device further includes an I / O interface, which is used for the electronic device to exchange data with an external device. The I / O interface can represent one or more of several types of bus structures, including a storage unit bus or a storage unit controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of a variety of bus structures.

[0125] It should be understood that Figure 4 The electronic device shown is only an example of the present invention, and the electronic device of the present invention may also include elements or components not shown in the above examples. For example, some electronic devices also include display units such as display screens, and some electronic devices also include human-computer interaction elements such as buttons, keyboards, etc. As long as the electronic device can execute the computer-readable program in the memory to implement the method of the present invention or at least part of the steps of the method, it can be considered as an electronic device covered by the present invention.

[0126] Through the above description of the implementation, it is easy for those skilled in the art to understand that the example implementation described here can be implemented by software, or by combining software with necessary hardware. Figure 5 As shown, the technical solution according to the embodiment of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes a number of commands to enable a computing device (which can be a personal computer, a server, or a network device, etc.) to execute the above method according to the embodiment of the present invention.

[0127] The software product may be any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, a system, device or component of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination thereof. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0128] The computer readable storage medium may include a data signal propagated in a baseband or as part of a carrier wave, wherein a readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. The readable storage medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with a command execution system, device, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the above.

[0129] Program code for performing the operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0130] The computer-readable medium carries one or more programs (eg, computer-executable programs). When the one or more programs are executed by a device, the computer-readable medium implements the method of the present disclosure.

[0131] Those skilled in the art will appreciate that the above modules can be distributed in the device according to the description of the embodiment, or can be changed accordingly and only used in one or more devices different from the embodiment. The modules of the above embodiments can be combined into one module, or further divided into multiple sub-modules.

[0132] Through the description of the above embodiments, it is easy for those skilled in the art to understand that the exemplary embodiments described here can be implemented by software, or by software combined with necessary hardware. Therefore, the technical solution according to the embodiment of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes several commands to enable a computing device (which can be a personal computer, a server, a mobile terminal, or a network device, etc.) to execute the method according to the embodiment of the present invention.

[0133] The exemplary embodiments of the present invention are specifically shown and described above. It should be understood that the present invention is not limited to the detailed structure, configuration or implementation method described herein; on the contrary, the present invention is intended to cover various modifications and equivalent configurations included in the spirit and scope of the appended claims.

Claims

1. A user representation learning recommendation method based on fusion confidence, characterized in that: include: Collect the first-class review texts of the same user on multiple items and the second-class review texts of different users on the same item; Input the collected first-category review texts and second-category review texts into the pre-trained word vector model, and output a first embedding vector and a second embedding vector; Based on the confidence function, a confidence matrix is ​​constructed according to all the ratings of different items and all the ratings of the same item by the user, and used as the correction coefficient for each review text; After correcting the first embedding vector and the second embedding vector according to the constructed confidence matrix, and then performing feature interactive learning after convolution and pooling, a user representation learning recommendation model is obtained; The comment text to be processed is input into the user representation learning recommendation model to obtain user preference recommendation information, wherein: The confidence matrix is ​​calculated using the following expression; Among them, Q ij represents a confidence matrix, which is used to represent the contribution of different comment texts to the construction of the user representation learning recommendation model, so as to be used as a correction coefficient for each comment text. Each element in the confidence matrix represents the inverse of the offline probability; R ij represents the rating of user i on item j, where j represents the item, limited to items purchased by user i and with ratings, R ip represents the rating of item p by user i; R qj represents the rating of user q on item j, where q represents the user and q is not equal to i; represents the mean of all ratings of user i, represents the mean of all the scores obtained by item j; A CNN algorithm is used to construct a user representation learning recommendation model, which includes a weighted word vector layer, a convolution layer, and a pooling layer; the collected first-category comment texts and second-category comment texts are input into a pre-trained word vector model, and a first embedding vector and a second embedding vector are output as inputs of the weighted word vector layer; before inputting the convolution layer, the first embedding vector and the second embedding vector are corrected according to the constructed confidence matrix; Use the following expression to perform the correction calculation to get the input of the convolution layer: H=D i +D h Among them, H represents the input of the convolutional layer; D i represents the vector of the first embedded vector corrected by the confidence matrix; D h represents the vector of the second embedding vector corrected by the confidence matrix; H 单 =Q ij ×D i Among them, Q ij represents the confidence matrix; D i represents the first embedding vector; H 多 =Q ij ×D j Among them, Q ij represents the confidence matrix; D j represents the second embedding vector; After convolution and pooling, the first maximum pooling vector of the first type of comment text is obtained as d i ; After convolution and pooling, the second largest pooling vector of the second type of comment text is d h ; Use the Hadamard product method to make the first maximum pooling d i and the second maximum pooling is d j Conduct feature interaction learning to obtain latent factors of comments; In the output layer, the existing nonlinear mapping is used to obtain the potential factors of the comments. Specifically, the following expression is used: in, It represents the interactive features of the latent factors of user comments and item comments learned by CNN, that is, the latent factors of comments; W2 represents the first adjustable parameter, which is the weight parameter of the interactive features of the latent factors, ranging from 0.5 to 0.9; w1 represents the second adjustable parameter, which is used to adjust the weight parameter of the interactive features of the latent factors, ranging from 0.3 to 0.7; d represents the interactive feature vector obtained by multiplying the first maximum pooling vector corresponding to the first type of comment text with the second maximum pooling vector corresponding to the second type of comment text; b1 represents the first offset, which is used to correct the first offset of the interactive feature; b2 represents the second offset, which is used to correct the second offset of the interactive feature.

2. The user representation learning recommendation method based on fusion confidence according to claim 1 is characterized in that: Further including: The review text to be processed includes the first type of review text of the current user i, the second type of review text for the same item j, and the confidence matrix constructed by the rating of the item by the user i and the rating obtained for the item j; The comment text to be processed is input into the user representation learning recommendation model, and the prediction score of the current user i for the item j is output. The user preference recommendation information is obtained according to the output prediction score of the current user i for the item j, and the user preference recommendation information includes the item type and the number of items of the same category.

3. A user representation learning recommendation system based on fusion confidence, characterized in that: The user representation learning recommendation method based on fusion confidence as described in claim 1 is adopted, and the user representation learning recommendation system comprises: A text collection module, used to collect the first-class review texts of the same user on multiple items and the second-class review texts of different users on the same item; A vector conversion module, used for inputting the collected first-category review text and second-category review text into a pre-trained word vector model, and outputting a first embedding vector and a second embedding vector; The construction module builds a confidence matrix based on the confidence function, according to all the ratings of different items by users and all the ratings of the same item, and uses it as the correction coefficient for each review text; A correction processing module, which corrects the first embedding vector and the second embedding vector according to the constructed confidence matrix, and then performs feature interactive learning after convolution and pooling to obtain a user representation learning recommendation model; The prediction module inputs the comment text to be processed into the user representation learning recommendation model to obtain user preference recommendation information.

Images