CN108920665B - Recommendation scoring method and device based on network structure and review text - Google Patents

Abstract

Embodiments of the present invention provide a method and device for recommendation scoring based on a network structure and comment text, the method includes: determining a target user in a plurality of sample users, and determining a target commodity in a plurality of sample commodities; The first type of feature matrix for the user, and the second type of feature matrix for the target product; the third type of feature matrix for the target user and the fourth type of feature matrix for the target product are obtained; the first type of feature matrix for the target user is obtained The matrix and the third-type feature matrix, as well as the second-type feature matrix and the fourth-type feature matrix for the target product, are input into the recommendation network model to obtain the predicted score value of the target user for the target product. Therefore, the interaction information between the user and the product, including the comment text information and the rating information, can be fully considered, and the user's purchase expectation of the product can be predicted more accurately.

Description

Recommendation scoring method and device based on network structure and review text

technical field

The present invention relates to the technical field of machine learning, in particular to a method and device for recommendation scoring based on network structure and review text.

Background technique

With the development of e-commerce, Internet companies can provide a large number of commodities for users to choose from, and it is difficult for users to make choices in the face of a large number of commodities. Currently, recommender systems are mainly used to help users make choices. Among them, the recommendation system is a system obtained by using the deep learning method to train the neural network model by using the user's purchase information of the product.

When using the recommendation system to recommend products for users, for each candidate product, the user's purchase information of different products and the purchased information of the candidate product are input into the recommendation system, and then the user's rating value for the candidate product is obtained, The rating value represents the user's purchase expectation of the product. The recommender system recommends candidate products with higher ratings to users.

It can be seen from the above that when recommending products to users, the scoring value of candidate products plays an important role.

However, the current recommendation system is only obtained by training users' purchase information of commodities. The user's purchase information of the product does not fully represent the interaction information between the user and the product. Only using the user's purchase information for the product to train the neural network model, the obtained recommendation system does not accurately score the candidate products, resulting in ineffectiveness. Help users make choices about products.

SUMMARY OF THE INVENTION

The purpose of the embodiments of the present invention is to provide a recommendation scoring method and device based on a network structure and comment text, so as to improve the accuracy of predicting the user's scoring of a commodity. The specific technical solutions are as follows:

In order to improve the accuracy of predicting a user's rating of a product, an embodiment of the present invention provides a recommendation rating method based on a network structure and comment text, the method comprising:

Determine the target user among the multiple sample users, and determine the target product among the multiple sample products;

Obtain a first-type feature matrix for the target user and a second-type feature matrix for the target product; both the first-type feature matrix and the second-type feature matrix are based on the multiple sample users Determined by the comment text information of the plurality of sample commodities;

Obtain a third-type feature matrix for the target user and a fourth-type feature matrix for the target product; both the third-type feature matrix and the fourth-type feature matrix are based on the multiple sample users The purchase network structure information of the plurality of sample commodities is determined;

Input the first-type feature matrix and the third-type feature matrix for the target user, as well as the second-type feature matrix and the fourth-type feature matrix for the target product into the recommendation network model, and obtain the target user's opinion on the The predicted score value of the target product;

The recommendation network model is a model obtained by training according to a training set, and the training set includes: a first-type feature matrix for each sample user, a second-type feature matrix for each sample commodity, and a second-type feature matrix for each sample The user's third-type feature matrix, the fourth-type feature matrix for each sample item, and the real rating values of multiple sample users for multiple sample items.

Optionally, the first-type feature matrix for the target user and the second-type feature matrix for the target product are determined by the following steps:

Acquiring a plurality of comment text information of the plurality of sample users on the plurality of sample commodities;

Perform word segmentation processing on the plurality of comment text information to obtain a word vector of each comment text information;

determining a first-type feature matrix for the target user according to word vectors of multiple comment text information of the target user;

A second type feature matrix for the target product is determined according to word vectors of multiple comment text information of the target product.

Optionally, the purchase network structure information is network structure information of a heterogeneous information network based on the purchase information of the plurality of sample users for the plurality of sample commodities, and the nodes of the heterogeneous information network include the plurality of samples. a sample user and the plurality of sample products, the edge of the heterogeneous information network is used to connect the sample user and the sample product, and the edge of the heterogeneous information network is used to indicate that the sample user has a real rating value for the sample product;

The third type of feature matrix for the target user and the fourth type of feature matrix for the target product are determined by the following steps:

Based on the connection relationship between sample users and sample commodities in the heterogeneous information network, a random walk is performed with the target user as the starting node, a plurality of first random walk sequences are determined, and the target commodity is used as the starting node to perform a random walk. Random walk, determining a plurality of second random walk sequences; the first random walk sequence includes a first preset number of nodes; the second random walk sequence includes a second preset number of nodes;

determining a third type of feature matrix for the target user according to a plurality of the first random walk sequences;

A fourth type of feature matrix for the target commodity is determined according to a plurality of the second random walk sequences.

Optionally, the recommendation network model includes: a first neural network model, a second neural network model, a third neural network model, and a fourth neural network model;

The first type of feature matrix and the third type of feature matrix for the target user, and the second type of feature matrix and the fourth type of feature matrix for the target product are input into the recommendation network model, to obtain the target user pair. The predicted score value of the target product, including:

Inputting the first type of feature matrix for the target user into the first neural network model to obtain a first feature vector;

Inputting the second type of feature matrix for the target product into the second neural network model to obtain a second feature vector;

Inputting the third type of feature matrix for the target user into the third neural network model to obtain a third feature vector;

Inputting the fourth type of feature matrix for the target product into the fourth neural network model to obtain a fourth feature vector;

Fusing the first feature vector, the second feature vector, the third feature vector and the fourth feature vector according to a multi-view machine learning algorithm to obtain a fusion feature vector;

The fusion feature vector is processed by an activation function to obtain the predicted rating value of the target product by the target user.

Optionally, the recommendation network model is obtained by training in the following steps:

Obtain a preset neural network model and the training set;

Input the first-type feature matrix for multiple sample users, the second-type feature matrix for multiple sample products, the third-type feature matrix for multiple sample users, and the fourth-type feature matrix for multiple sample products. The neural network model described above is used to obtain the predicted rating values of multiple sample users for multiple sample commodities;

Determine the loss value according to the obtained predicted score value and the real score value included in the training set;

Determine whether the neural network model converges according to the loss value;

If not, adjust the parameter values in the neural network model, and return the first-type feature matrix for multiple sample users, the second-type feature matrix for multiple sample commodities, and the multiple sample users. The third type of feature matrix and the fourth type of feature matrix for a plurality of sample commodities are input into the neural network model, and the steps of obtaining the predicted score values of a plurality of sample users for a plurality of sample commodities;

If so, the current neural network model is determined as the recommendation network model.

In order to achieve the above object, an embodiment of the present invention further provides a recommendation scoring device based on a network structure and comment text, the device comprising:

a determination module, used for determining a target user among the multiple sample users, and determining the target product among the multiple sample products;

The first obtaining module is used to obtain a first-type feature matrix for the target user and a second-type feature matrix for the target commodity; both the first-type feature matrix and the second-type feature matrix are Determined according to the textual information of comments of the plurality of sample users on the plurality of sample commodities;

The second acquiring module is configured to acquire the third type of feature matrix for the target user and the fourth type of feature matrix for the target product; both the third type of feature matrix and the fourth type of feature matrix are Determined according to the purchase network structure information of the plurality of sample users for the plurality of sample commodities;

The prediction module is used for inputting the first-type feature matrix and the third-type feature matrix for the target user, and the second-type feature matrix and the fourth-type feature matrix for the target product into the recommendation network model, to obtain the The predicted rating value of the target product by the target user.

Optionally, the first obtaining module is specifically used for:

Acquiring a plurality of comment text information of the plurality of sample users on the plurality of sample commodities;

Perform word segmentation processing on the plurality of comment text information to obtain a word vector of each comment text information;

determining a first-type feature matrix for the target user according to word vectors of multiple comment text information of the target user;

A second type feature matrix for the target product is determined according to word vectors of multiple comment text information of the target product.

Optionally, the purchase network structure information is network structure information of a heterogeneous information network based on the purchase information of the plurality of sample users for the plurality of sample commodities, and the nodes of the heterogeneous information network include the plurality of samples. a sample user and the plurality of sample products, the edge of the heterogeneous information network is used to connect the sample user and the sample product, and the edge of the heterogeneous information network is used to indicate that the sample user has a real rating value for the sample product;

The second acquisition module is specifically used for:

Based on the connection relationship between sample users and sample commodities in the heterogeneous information network, a random walk is performed with the target user as the starting node, a plurality of first random walk sequences are determined, and the target commodity is used as the starting node to perform a random walk. Random walk, determining a plurality of second random walk sequences; the first random walk sequence includes a first preset number of nodes; the second random walk sequence includes a second preset number of nodes;

determining a third type of feature matrix for the target user according to a plurality of the first random walk sequences;

A fourth type of feature matrix for the target commodity is determined according to a plurality of the second random walk sequences.

Optionally, the recommendation network model includes: a first neural network model, a second neural network model, a third neural network model, and a fourth neural network model;

The prediction module is specifically used for:

Inputting the first type of feature matrix for the target user into the first neural network model to obtain a first feature vector;

Inputting the second type of feature matrix for the target product into the second neural network model to obtain a second feature vector;

Inputting the third type of feature matrix for the target user into the third neural network model to obtain a third feature vector;

Inputting the fourth type of feature matrix for the target product into the fourth neural network model to obtain a fourth feature vector;

Fusing the first feature vector, the second feature vector, the third feature vector and the fourth feature vector according to a multi-view machine learning algorithm to obtain a fusion feature vector;

The fusion feature vector is processed by an activation function to obtain the predicted rating value of the target product by the target user.

Optionally, the device further includes:

a training module for training the recommendation network model;

The training module is specifically used for:

Obtain a preset neural network model and the training set;

Input the first-type feature matrix for multiple sample users, the second-type feature matrix for multiple sample products, the third-type feature matrix for multiple sample users, and the fourth-type feature matrix for multiple sample products. The neural network model described above is used to obtain the predicted rating values of multiple sample users for multiple sample commodities;

Determine the loss value according to the obtained predicted score value and the real score value included in the training set;

Determine whether the neural network model converges according to the loss value;

If not, adjust the parameter values in the neural network model, and return the first-type feature matrix for multiple sample users, the second-type feature matrix for multiple sample commodities, and the multiple sample users. The third type of feature matrix and the fourth type of feature matrix for a plurality of sample commodities are input into the neural network model, and the steps of obtaining the predicted score values of a plurality of sample users for a plurality of sample commodities;

If so, the current neural network model is determined as the recommendation network model.

In order to achieve the above object, an embodiment of the present invention also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus; wherein, the processor, the communication interface, and the memory communicate with each other through the communication bus;

memory for storing computer programs;

The processor is configured to implement any of the above method steps when executing the program stored in the memory.

To achieve the above object, an embodiment of the present invention further provides a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, any one of the above method steps is implemented.

In the embodiment of the present invention, based on the first-type feature matrix and the third-type feature matrix for each sample user, the second-type feature matrix and the fourth-type feature matrix for each sample commodity, and the multiple sample user-to-many The real rating value of each sample product is trained to obtain the recommendation network model. Input the first-type feature matrix and third-type feature matrix for the target user, as well as the second-type feature matrix and the fourth-type feature matrix for the target product into the trained recommendation network model, and obtain the target user's prediction score for the target product value.

Among them, the first type of feature matrix and the second type of feature matrix are both determined according to the textual information of multiple sample users' comments on multiple sample products, and the third type of feature matrix and the fourth type of feature matrix are based on multiple sample users. The purchase network structure information of multiple sample commodities is determined. According to the first type of feature matrix, the second type of feature matrix, the third type of feature matrix and the fourth type of feature matrix, the predicted score value of the target user for the target product is determined, and the interaction information between the user and the product is fully considered. The review text information and rating information between the user and the product can be more accurately predicted, and the user's purchase expectation of the product can be predicted, and the accuracy of the product rating can be improved.

Of course, it is not necessary for any product or method of the present invention to achieve all of the advantages described above at the same time.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative efforts.

Fig. 1 is a kind of flow chart of the recommendation scoring method based on network structure and comment text provided by the embodiment of the present invention;

2 is a schematic diagram of a heterogeneous information network provided by an embodiment of the present invention;

3 is an exemplary schematic diagram of a recommendation scoring method based on a network structure and comment text provided by an embodiment of the present invention;

4 is a schematic structural diagram of a recommendation scoring device based on a network structure and comment text provided by an embodiment of the present invention;

FIG. 5 is a schematic structural diagram of an electronic device provided by an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

In order to solve the problem of inaccurate product scoring, an embodiment of the present invention provides a recommendation scoring method based on a network structure and comment text. Referring to FIG. 1, FIG. 1 is a network structure and comment text-based method provided by an embodiment of the present invention. A flowchart of the recommendation scoring method of , the method includes the following steps:

Step S101: Determine a target user among the plurality of sample users, and determine a target commodity among the plurality of sample commodities.

In this embodiment of the present invention, a large number of sample users and a large number of sample commodities may be acquired in advance. Among them, some sample users have rating information for some sample products, and some sample users have comment text information for some sample products. For example, the sample users are users of a shopping platform, the sample products are the products provided by the shopping platform, these sample users have purchased some sample products of the shopping platform and rated them, and some of these sample users Purchased sample items are reviewed.

Based on these information data, for a sample user and a sample product, wherein the sample user does not have rating information and comment text information for the sample product, what the embodiment of the present invention needs to achieve is to predict the predicted score of the sample user for the sample product The predicted score value can also represent the purchase expectation of the sample user for the sample product.

That is, in step S101, a target user may be determined from a plurality of sample users first, and a target commodity may be determined from a plurality of sample commodities, wherein the target user has no rating information and comment text information on the target commodity.

Step S102: Obtain a first-type feature matrix for the target user and a second-type feature matrix for the target commodity. The first type of feature matrix and the second type of feature matrix are both determined according to the textual information of comments of multiple sample users on multiple sample commodities.

In the embodiment of the present invention, the text information of the comments made by the target user on multiple sample commodities can be converted into a matrix representation, and the first-type feature matrix for the target user can be obtained. The text information of comments made by multiple sample users on the target product can be converted into a matrix representation to obtain the second type of feature matrix for the target product.

In an implementation manner of the present invention, the first-type feature matrix for the target user and the second-type feature matrix for the target commodity may be determined through the following steps.

Step 11: Acquire multiple comment text information of multiple sample users on multiple sample commodities.

In this step, all comment text information made by multiple sample users on multiple sample products can be obtained.

Step 12: Perform word segmentation on a plurality of comment text information to obtain a word vector of each comment text information.

In this step, word segmentation processing may be performed on all the comment text information obtained in step 11 to obtain a word vector of each comment text information.

In an implementation manner, for each comment text information, during word segmentation processing, all punctuation marks in the comment text information can be deleted, and then the comment text information can be split into the form of words. Each word of the comment text information is correspondingly converted into a digital form, and a word vector corresponding to the comment text information containing a string of numbers is obtained.

Step 13: Determine a first-type feature matrix for the target user according to word vectors of multiple comment text information of the target user.

Wherein, the multiple comment text information of the target user is: the sum of the comment text information made by the target user on multiple sample commodities. For example, if the target user has purchased sample product I ₁ , sample product I ₂ , sample product I ₃ and sample product I ₄ , and has commented on sample product I ₁ , sample product I ₂ , and sample product I ₄ , the target user will The comment text information made by the user on the sample product I ₁ , the sample product I ₂ , and the sample product I ₄ is determined as the comment text information of the target user.

In an implementation manner, after the target user is determined, word vectors corresponding to multiple comment text information of the target user can be obtained, and these word vectors can be combined to obtain the first-type feature matrix for the target user.

Step 14: Determine a second-type feature matrix for the target product according to word vectors of multiple comment text information of the target product.

Wherein, the multiple comment text information of the target product is: the sum of the comment text information made by multiple different sample users on the target product. For example, the target product has been purchased by user U ₁ , user U ₂ , user U ₃ and user U ₄ , and user U ₁ , user U ₂ , and user U ₄ have commented on the target product, then user U ₁ , user U 2 , and user U 4 have commented on the target product. The comment text information made by U ₂ and the user U ₄ on the target commodity is determined as the comment text information of the target commodity.

In an implementation manner, after the target product is determined, word vectors of multiple comment text information of the target product can be obtained, and these word vectors can be combined to obtain the second-type feature matrix for the target product.

In this embodiment of the present invention, the execution order of step 13 and step 14 is not limited.

Certainly, the multiple comment text information of each sample user or sample product may be converted into a matrix using the method flow of the above steps 11-14 provided in the embodiment of the present invention, which is not limited.

In an implementation manner of the present invention, when it is necessary to determine the predicted rating value of the target user for the target product, the first type of feature matrix for the target user can be determined according to the method flow of the above steps 11-14, and the target product to improve the accuracy of the determined first-type feature matrix for the target user and the second-type feature matrix for the target commodity, thereby improving the accuracy of the determined prediction score value.

In another implementation manner of the present invention, a first-type feature matrix for each sample user and a second-type feature matrix for each sample commodity may be pre-calculated, and a plurality of first-type feature matrices obtained by calculation and multiple second-type feature matrices are pre-stored in the database as data in the training set.

When it is necessary to determine the predicted score value of the target user for the target product, the first type feature matrix for the target user and the second type feature matrix for the target product are directly obtained from the database to improve the calculation efficiency.

Step S103: Acquire a third-type feature matrix for the target user and a fourth-type feature matrix for the target commodity. Both the third type of feature matrix and the fourth type of feature matrix are determined according to the purchase network structure information of multiple sample users for multiple sample commodities.

In the embodiment of the present invention, the purchase network structure information indicates which sample users have purchased which sample commodities and have scored the purchased commodities. The purchase network structure information can be represented in the form of a heterogeneous information network, see FIG. 2 , FIG. 2 is a schematic diagram of a heterogeneous information network provided by an embodiment of the present invention, wherein the heterogeneous information network includes nodes and edges, the nodes include sample users and sample commodities, and the edge connecting the sample users and the sample commodities indicates that the sample users are paired with each other. This sample product has published ratings.

In an implementation manner of the present invention, the third-type feature matrix for the target user and the fourth-type feature matrix for the target commodity may be determined through the following steps.

Step 21: Based on the connection relationship between the sample users and the sample commodities in the heterogeneous information network, take the target user as the starting node to perform a random walk, determine a plurality of first random walk sequences, and take the target commodity as the starting node to perform random walks. walk, and determine a plurality of second random walk sequences. Wherein, the first random walk sequence includes a first preset number of nodes; the second random walk sequence includes a second preset number of nodes.

In practical applications, when determining the first random walk sequence and the second random walk sequence, since there are many sample users and many sample products, one sample user can score multiple sample products, so the target user is the starting point. There will be many first random walk sequences of nodes. Similarly, there will be many second random walk sequences.

2 , assuming that the first preset number is 5 and the target user is user U ₃ , the user U ₃ is used as the starting node, and a random walk is performed along the edge connecting the user and the product to obtain 2 The first random walk sequence includes: {U ₃ , I ₂ , U ₂ , I ₁ , U ₁ }, {U ₃ , I ₃ , U ₄ , I ₄ , U ₅ }.

Similarly, it is also possible to determine a plurality of second random walk sequences obtained by performing random walks with the target commodity as a starting node.

In an implementation of the present invention, each sample user and each sample commodity can be numbered in advance, and then each first random walk sequence and second random walk sequence can be converted to include a string of numbers vector.

For example, in the heterogeneous information network shown in FIG. 2 , user U ₁ , user U ₂ , user U ₃ , user U ₄ , and user U ₅ may be numbered as 01, 02, 03, 04, and 05, respectively, and the commodity I ₁ , commodity I ₂ , commodity I ₃ , and commodity I ₄ are respectively numbered 11, 12, 13, and 14, then for the first random walk sequence U ₃ , I ₂ , U ₂ , I ₁ , and U ₁ , we can Represented as a vector containing {03, 12, 02, 11, 01}.

Step 22: Determine a third type of feature matrix for the target user according to a plurality of first random walk sequences.

After a plurality of first random walk sequences are obtained, a third type of feature matrix for the target user can be obtained by combining the plurality of first random walk sequences.

In an implementation manner of the present invention, the multiple first random walk sequences of the target user are all vector representations including a string of numbers. Then the third type of feature matrix for the target user can be obtained by combining these vectors.

The description will still be made with reference to the example in step 21 . Multiple first random walk sequences for user U ₃ , including: {U ₃ , I ₂ , U ₂ , I ₁ , U ₁ }, {U ₃ , I ₃ , U ₄ , I ₄ , U ₅ }. Then you can connect multiple first random walk sequences to get a long sequence: {U ₃ , I ₂ , U ₂ , I ₁ , U ₁ , U ₃ , I ₃ , U ₄ , I ₄ , U ₅ } .

In this embodiment of the present invention, both the user and the commodity may be represented by a vector, and the vector may contain a certain number of numerical values. For example, user U ₁ , user U ₂ , user U ₃ , user U ₄ , user U ₅ can be represented as {0.1, 0.2, 0.3}, {0.1, 0.3, 0.1}, {0.2, 0.1, 0.3}, {0.2, 0.3, 0.1}, {0.3, 0.1, 0.2}, commodity I ₁ , commodity I ₂ , commodity I ₃ , commodity I ₄ can be expressed as {0.7, 0.8, 0.9}, {0.7, 0.9, 0.8, respectively }, {0.8, 0.7, 0.9}, {0.8, 0.9, 0.7}, then the long sequence {U ₃ , I ₂ , U ₂ , _{I 1 ,} U ₁ , U ₃ , I ₃ , U ₄ , I ₄ , U ₅ } can be transformed into the following matrices:

Of course, the above-mentioned embodiment is only an example. In practical applications, any method for converting multiple sequences into matrices can be applied in the embodiments of the present invention.

Step 23: Determine the fourth type of feature matrix for the target commodity according to the plurality of second random walk sequences.

After multiple second random walk sequences are acquired, a fourth type of feature matrix for the target user can be obtained by combining the multiple second random walk sequences.

In an implementation manner of the present invention, the multiple second random walk sequences of the target user are all vector representations including a string of numbers. Then the fourth type of feature matrix for the target user can be obtained by combining these vectors. For details, refer to the example in step 22, which is not repeated here.

In this embodiment of the present invention, the execution order of step 22 and step 23 is not limited.

In an implementation manner of the present invention, when it is necessary to determine the predicted score value of the target user for the target product, the third type of feature matrix for the target user can be determined according to the method flow of the above steps 21-23, and the target product to improve the accuracy of the determined third-type feature matrix for the target user and the fourth-type feature matrix for the target commodity, thereby improving the accuracy of the determined prediction score value.

In another implementation manner of the present invention, the third-type feature matrix for each sample user and the fourth-type feature matrix for each sample commodity may be pre-calculated, and the calculated third-type feature matrices and a plurality of fourth-type feature matrices are pre-stored in the database as data in the training set.

When it is necessary to determine the predicted score value of the target user for the target product, the third-type feature matrix for the target user and the fourth-type feature matrix for the target product are directly obtained from the database to improve the calculation efficiency.

In this embodiment of the present invention, the execution order of step S103 and step S102 is not limited.

Step S104: Input the first-type feature matrix and the third-type feature matrix for the target user, and the second-type feature matrix and the fourth-type feature matrix for the target product into the recommendation network model, and obtain the target user's prediction score for the target product value.

Among them, the recommendation network model is a model obtained by training according to the training set, and the training set includes: a first-type feature matrix for each sample user, a second-type feature matrix for each sample commodity, and a third-type feature matrix for each sample user. The class feature matrix, the fourth class feature matrix for each sample item, and the real rating values of multiple sample users for multiple sample items.

After determining the first-type feature matrix and the third-type feature matrix for the target user, and the second-type feature matrix and the fourth-type feature matrix for the target product, these feature matrices can be input into the recommendation network model, and the target can be obtained. The user's predicted rating value for the target item.

Referring to FIG. 3 , in this embodiment of the present invention, the recommendation network model may include a first neural network model, a second neural network model, a third neural network model, and a fourth neural network model. In the step of inputting the feature matrix and the third-type feature matrix, and the second-type feature matrix and the fourth-type feature matrix for the target product into the recommendation network model, the following refinement steps may be included:

Input the first type feature matrix for the target user into the first neural network model to obtain the first feature vector; input the second type feature matrix for the target product into the second neural network model to obtain the second feature vector; The third-type feature matrix of The feature vector, the third feature vector, and the fourth feature vector are fused according to a multi-view machine learning algorithm to obtain a fused feature vector.

Wherein, the first neural network model, the second neural network model, the third neural network model and the fourth neural network model may be a recurrent neural network model, a convolutional neural network model, a recurrent convolutional neural network model, a deep neural network model, etc. . This embodiment of the present invention does not limit this.

Each neural network model can output a feature vector. In order to comprehensively consider the user's comment text information on the product and the user's rating information on the product, in this embodiment of the present invention, the above-mentioned first feature vector, second feature vector, The third feature vector and the fourth feature vector are fused.

In an implementation manner of the present invention, the fusion can be performed according to a multi-view machine learning algorithm, and the numerical values in each feature vector can interact with each other during fusion using this algorithm, thereby better integrating the four feature vectors. Fusion of multiple vectors based on a multi-view machine learning algorithm belongs to the category of the prior art, and details are not described here.

After the first eigenvector, the second eigenvector, the third eigenvector and the fourth eigenvector are fused, a fused eigenvector can be obtained, and the fused eigenvector is processed by the activation function to obtain the target user's predicted rating value for the target product . In an implementation manner of the present invention, a linear rectification function (Rectified Linear Unit, ReLU) can be used to process the fusion feature vector, a sigmoid activation function can also be used to process the fusion feature vector, and other activation functions can also be used to process the fusion feature vector. The activation function may be selected and determined according to actual requirements, which is not limited in this embodiment of the present invention.

It can be seen that in the embodiment of the present invention, based on the first-type feature matrix and the third-type feature matrix for each sample user, the second-type feature matrix and the fourth-type feature matrix for each sample product, and a plurality of sample users For the real rating values of multiple sample products, the recommendation network model is obtained by training. Input the first-type feature matrix and third-type feature matrix for the target user, as well as the second-type feature matrix and the fourth-type feature matrix for the target product into the trained recommendation network model, and obtain the target user's prediction score for the target product value.

Among them, the first type of feature matrix and the second type of feature matrix are both determined according to the textual information of multiple sample users' comments on multiple sample products, and the third type of feature matrix and the fourth type of feature matrix are based on multiple sample users. The purchase network structure information of multiple sample commodities is determined. According to the first type of feature matrix, the second type of feature matrix, the third type of feature matrix and the fourth type of feature matrix, the predicted score value of the target user for the target product is determined, and the interaction information between the user and the product is fully considered. The review text information and rating information between the user and the product can be more accurately predicted, and the user's purchase expectation of the product can be predicted, and the accuracy of the product rating can be improved.

The predicted score value determined in the embodiment of the present invention represents the target user's purchase expectation for the target product, and a threshold may be preset. If the predicted score value is greater than the threshold value, it indicates that the target user has a high purchase expectation for the target product, and can be set as The target product is recommended to the target user.

In the embodiment of the present invention, the recommendation network model can be obtained by training in the following steps.

Step 31: Acquire a preset neural network model and the training set.

The recommended network model used to predict the rating value is obtained by training according to the training set, wherein the training set includes the first-type feature matrix for each sample user, the second-type feature matrix for each sample product, and the second-type feature matrix for each sample product. The user's third-type feature matrix, the fourth-type feature matrix for each sample item, and the real rating values of multiple sample users for multiple sample items.

In the above embodiments, the method for determining the feature matrix included in the training set has been introduced, and will not be repeated here.

Step 32: Combine the first-type feature matrix for multiple sample users, the second-type feature matrix for multiple sample products, the third-type feature matrix for multiple sample users, and the fourth-type feature matrix for multiple sample products. The matrix is input to the neural network model, and the predicted rating values of multiple sample users for multiple sample products are obtained.

The steps of inputting the above four types of feature matrices into the neural network model to obtain the predicted score value may refer to the description of step S104 in the above-mentioned embodiment of the recommendation scoring method.

In this embodiment of the present invention, the real rating value of a sample user for the sample product, the first-type feature matrix and the third-type feature matrix of the sample user, and the second-type feature matrix and the fourth-type feature matrix of the sample product Matrix as a set of training set data.

In an implementation manner of the present invention, a group of training set data may be input each time to obtain a prediction score value corresponding to the group of training set data.

In a preferred implementation manner of the present invention, multiple sets of training set data may also be input each time to obtain a predicted score value corresponding to each set of training set data.

For example, if the training set data includes 10,000 real score data, and each real score data corresponds to a sample user and sample product, then 100 real score data can be selected for training in each round of training, that is, these 100 real score data can be selected for training. The rating data and its corresponding feature matrix of sample users and sample products are input into the neural network model, and 100 predicted rating values are obtained.

Step 33: Determine the loss value according to the obtained predicted score value and the real score value included in the training set.

In the embodiments of the present application, including but not limited to, using the mean squared error (Mean Squared Error, MSE) formula as the loss function to obtain the loss value.

Step 34: Determine whether the neural network model has converged according to the obtained loss value. If so, go to step 35. If not, step 36.

Step 35: Determine the current neural network model as the recommendation network model.

Step 36: Adjust the parameter values in the neural network model, and return to step 32.

Based on the same inventive concept, according to the above-mentioned embodiment of the recommendation scoring method based on network structure and review text, an embodiment of the present invention also provides a recommendation scoring device based on network structure and review text, referring to FIG. 4 , which may include the following modules:

A determination module 401, configured to determine a target user in a plurality of sample users, and determine a target commodity in the plurality of sample commodities;

The first acquisition module 402 is configured to acquire a first-type feature matrix for target users and a second-type feature matrix for target commodities; both the first-type feature matrix and the second-type feature matrix are based on multiple sample user-to-many ratios. The comment text information of each sample product is determined;

The second obtaining module 403 is configured to obtain a third-type feature matrix for the target user and a fourth-type feature matrix for the target product; both the third-type feature matrix and the fourth-type feature matrix are based on multiple sample users to The purchase network structure information of each sample commodity is determined;

The prediction module 404 is configured to input the first-type feature matrix and the third-type feature matrix for the target user, and the second-type feature matrix and the fourth-type feature matrix for the target product into the recommendation network model, so as to obtain the target user's preference for the target product. The predicted score value of ;

The network structure and review text-based recommendation scoring device provided by the embodiment of the present invention compares the first-type feature matrix and the third-type feature matrix for the target user, and the second-type feature matrix and the fourth-type feature matrix for the target product. Input the trained recommendation network model to obtain the predicted rating value of the target user for the target product, wherein the first-type feature matrix and the second-type feature matrix are determined according to the textual information of the comments of multiple sample users on multiple sample products , the third type of feature matrix and the fourth type of feature matrix are determined according to the purchase network structure information of multiple sample users for multiple sample commodities. Therefore, the interaction information between the user and the product, including the comment text information and the rating information, can be fully considered, and the user's purchase expectation of the product can be predicted more accurately.

In an embodiment of the present invention, the first obtaining module 402 can be specifically used for:

Acquiring a plurality of comment text information of the plurality of sample users on the plurality of sample commodities;

Perform word segmentation processing on the plurality of comment text information to obtain a word vector of each comment text information;

determining a first-type feature matrix for the target user according to word vectors of multiple comment text information of the target user;

A second type feature matrix for the target product is determined according to word vectors of multiple comment text information of the target product.

In one embodiment of the present invention, the purchase network structure information is the network structure information of a heterogeneous information network based on the purchase information of a plurality of sample users for a plurality of sample commodities, and the nodes of the heterogeneous information network include a plurality of sample users and a plurality of samples. A sample product, the edge of the heterogeneous information network is used to connect the sample user and the sample product, and the edge of the heterogeneous information network is used to indicate that the sample user has a real rating value for the sample product.

The second obtaining module 403 can be specifically used for:

Based on the connection relationship between sample users and sample commodities in the heterogeneous information network, a random walk is performed with the target user as the starting node, a plurality of first random walk sequences are determined, and the random walk is performed with the target commodity as the starting node. determining a plurality of second random walk sequences; the first random walk sequence includes a first preset number of nodes; the second random walk sequence includes a second preset number of nodes;

determining a third type of feature matrix for the target user according to a plurality of first random walk sequences;

A fourth type of feature matrix for the target item is determined according to the plurality of second random walk sequences.

In an embodiment of the present invention, the recommendation network model includes: a first neural network model, a second neural network model, a third neural network model, and a fourth neural network model;

The prediction module 404 can be specifically used for:

Input the first type feature matrix for the target user into the first neural network model to obtain the first feature vector; input the second type feature matrix for the target product into the second neural network model to obtain the second feature vector; The third-type feature matrix of The feature vector, the third feature vector and the fourth feature vector are fused according to the multi-view machine learning algorithm to obtain a fused feature vector; the fused feature vector is processed by an activation function to obtain the target user's predicted rating value for the target product.

In an embodiment of the present invention, on the basis of the apparatus embodiment shown in FIG. 3 , a training module may also be included for training a recommendation network model, specifically for:

Obtain preset neural network models and training sets;

Input the first-type feature matrix for multiple sample users, the second-type feature matrix for multiple sample products, the third-type feature matrix for multiple sample users, and the fourth-type feature matrix for multiple sample products into the neural network. Network model to obtain the predicted rating values of multiple sample users for multiple sample commodities;

Determine the loss value according to the obtained predicted score value and the real score value included in the training set;

Determine whether the neural network model converges according to the loss value;

If not, adjust the parameter values in the neural network model, and return the first-type feature matrix for multiple sample users, the second-type feature matrix for multiple sample products, and the third-type feature matrix for multiple sample users. Matrix, input the neural network model for the fourth type feature matrix of multiple sample commodities, and obtain the steps of obtaining the predicted score values of multiple sample users for multiple sample commodities;

If so, the current neural network model is determined as the recommendation network model.

Based on the same inventive concept, according to the above embodiments of the recommendation scoring method based on network structure and comment text, an embodiment of the present invention further provides an electronic device, as shown in FIG. 5 , including a processor 501 , a communication interface 502 , and a memory 503 and a communication bus 504, wherein the processor 501, the communication interface 502, and the memory 503 complete the communication with each other through the communication bus 504,

a memory 503 for storing computer programs;

The processor 501 is configured to implement the above-mentioned embodiment of the recommendation scoring method based on the network structure and the comment text shown in FIG. 1 when executing the program stored in the memory 503 . Among them, recommendation scoring methods based on network structure and review text include:

Determine the target user among the multiple sample users, and determine the target product among the multiple sample products;

Obtain a first-type feature matrix for the target user and a second-type feature matrix for the target product; both the first-type feature matrix and the second-type feature matrix are based on the multiple sample users Determined by the comment text information of the plurality of sample commodities;

Obtain a third-type feature matrix for the target user and a fourth-type feature matrix for the target product; both the third-type feature matrix and the fourth-type feature matrix are based on the multiple sample users The purchase network structure information of the plurality of sample commodities is determined;

Input the first-type feature matrix and the third-type feature matrix for the target user, as well as the second-type feature matrix and the fourth-type feature matrix for the target product into the recommendation network model, and obtain the target user's opinion on the The predicted score value of the target product;

The recommendation network model is a model obtained by training according to a training set, and the training set includes: a first-type feature matrix for each sample user, a second-type feature matrix for each sample commodity, and a second-type feature matrix for each sample The user's third-type feature matrix, the fourth-type feature matrix for each sample item, and the real rating values of multiple sample users for multiple sample items.

In the embodiment of the present invention, the first-type feature matrix and the third-type feature matrix for the target user, and the second-type feature matrix and the fourth-type feature matrix for the target product are input into the trained recommendation network model to obtain the target The user's predicted rating value for the target product, wherein the first type of feature matrix and the second type of feature matrix are determined according to the textual information of the comments of multiple sample users on multiple sample products, the third type of feature matrix and the fourth type of feature matrix are determined. The feature matrix is determined according to the purchase network structure information of multiple sample users for multiple sample commodities. Therefore, the interaction information between the user and the product, including the comment text information and the rating information, can be fully considered, and the user's purchase expectation of the product can be predicted more accurately.

The communication bus 504 mentioned in the above electronic device may be a peripheral component interconnect standard (Peripheral Component Interconnect, PCI) bus or an Extended Industry Standard Architecture (Extended Industry Standard Architecture, EISA) bus or the like. The communication bus 504 can be divided into an address bus, a data bus, a control bus, and the like. For ease of presentation, only one thick line is used in FIG. 5, but it does not mean that there is only one bus or one type of bus.

The communication interface 502 is used for communication between the above-mentioned electronic device and other devices.

The memory 503 may include random access memory (Random Access Memory, RAM), or may include non-volatile memory (Non-Volatile Memory, NVM), such as at least one disk storage. Optionally, the memory 503 may also be at least one storage device located away from the aforementioned processor.

The above-mentioned processor 501 may be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU), a network processor (Network Processor, NP), etc.; it may also be a digital signal processor (Digital Signal Processing, DSP), an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), Field-Programmable Gate Array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

Based on the same inventive concept, according to the above-mentioned embodiments of the recommendation scoring method based on network structure and comment text, in another embodiment provided by the present invention, a computer-readable storage medium is also provided, in which the computer-readable storage medium stores A computer program is stored, and when the computer program is executed by the processor, the above-mentioned steps of the recommendation scoring method based on the network structure and the review text shown in FIG. 1 are implemented.

It should be noted that, in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any relationship between these entities or operations. any such actual relationship or sequence exists. Moreover, the terms "comprising", "comprising" or any other variation thereof are intended to encompass a non-exclusive inclusion such that a process, method, article or device that includes a list of elements includes not only those elements, but also includes not explicitly listed or other elements inherent to such a process, method, article or apparatus. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in a process, method, article or apparatus that includes the element.

Each embodiment in this specification is described in a related manner, and the same and similar parts between the various embodiments may be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the apparatus embodiment, the electronic device embodiment and the storage medium embodiment, since they are basically similar to the method embodiment, the description is relatively simple. For the relevant details, please refer to the embodiment of the recommendation scoring method based on network structure and comment text part of the description.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims (6)

1. A recommendation scoring method based on a network structure and comment texts is characterized by comprising the following steps:

determining a target user in the plurality of sample users and determining a target commodity in the plurality of sample commodities;

acquiring a first class feature matrix for the target user and a second class feature matrix for the target commodity; the first type feature matrix and the second type feature matrix are determined according to comment text information of the sample users on the sample commodities;

acquiring a third type feature matrix for the target user and a fourth type feature matrix for the target commodity; the third type feature matrix and the fourth type feature matrix are determined according to the purchasing network structure information of the plurality of sample users on the plurality of sample commodities;

inputting the first class feature matrix and the third class feature matrix for the target user and the second class feature matrix and the fourth class feature matrix for the target commodity into a recommendation network model to obtain a prediction score value of the target user on the target commodity;

the recommended network model is a model obtained by training according to a training set, wherein the training set comprises: the first class feature matrix aiming at each sample user, the second class feature matrix aiming at each sample commodity, the third class feature matrix aiming at each sample user, the fourth class feature matrix aiming at each sample commodity and the real score values of a plurality of sample users on a plurality of sample commodities;

the first class characteristic matrix aiming at the target user and the second class characteristic matrix aiming at the target commodity are determined by the following steps:

acquiring a plurality of comment text messages of the plurality of sample users on the plurality of sample commodities;

performing word segmentation processing on the plurality of comment text messages to obtain a word vector of each comment text message;

determining a first class feature matrix aiming at the target user according to the word vectors of the comment text information of the target user;

determining a second class of feature matrix aiming at the target commodity according to the word vectors of the comment text messages of the target commodity;

the purchasing network structure information is network structure information of a heterogeneous information network based on purchasing information of the plurality of sample users on the plurality of sample commodities, nodes of the heterogeneous information network comprise the plurality of sample users and the plurality of sample commodities, edges of the heterogeneous information network are used for connecting the sample users and the sample commodities, and edges of the heterogeneous information network are used for indicating that the sample users have real scoring values on the sample commodities;

the third type feature matrix for the target user and the fourth type feature matrix for the target commodity are determined by the following steps:

based on the connection relation between the sample user and the sample commodity in the heterogeneous information network, randomly walking by taking the target user as an initial node to determine a plurality of first random walking sequences, randomly walking by taking the target commodity as the initial node to determine a plurality of second random walking sequences; the first random walk sequence comprises a first preset number of nodes; the second random walk sequence comprises a second preset number of nodes;

determining a third type of feature matrix aiming at the target user according to the plurality of first random walk sequences;

determining a fourth type feature matrix aiming at the target commodity according to the plurality of second random walk sequences;

the recommended network model includes: a first neural network model, a second neural network model, a third neural network model, and a fourth neural network model;

the step of inputting the first class feature matrix and the third class feature matrix for the target user and the second class feature matrix and the fourth class feature matrix for the target commodity into a recommendation network model to obtain the predicted score value of the target user on the target commodity comprises:

inputting a first class feature matrix aiming at the target user into the first neural network model to obtain a first feature vector;

inputting a second type of feature matrix aiming at the target commodity into the second neural network model to obtain a second feature vector;

inputting a third type of feature matrix aiming at the target user into the third neural network model to obtain a third feature vector;

inputting a fourth type of feature matrix aiming at the target commodity into the fourth neural network model to obtain a fourth feature vector;

fusing the first feature vector, the second feature vector, the third feature vector and the fourth feature vector according to a multi-view machine learning algorithm to obtain a fused feature vector;

and processing the fusion feature vector through an activation function to obtain the prediction score value of the target user on the target commodity.

2. The method of claim 1, wherein the recommended network model is obtained by training using the following steps:

acquiring a preset neural network model and the training set;

inputting a first class feature matrix aiming at a plurality of sample users, a second class feature matrix aiming at a plurality of sample commodities, a third class feature matrix aiming at a plurality of sample users and a fourth class feature matrix aiming at a plurality of sample commodities into the neural network model to obtain the prediction score values of the plurality of sample users on the plurality of sample commodities;

determining a loss value according to the obtained prediction score value and a real score value included in the training set;

determining whether the neural network model converges according to the loss value;

if not, adjusting parameter values in the neural network model, and returning to the step of inputting the first class feature matrix aiming at the plurality of sample users, the second class feature matrix aiming at the plurality of sample commodities, the third class feature matrix aiming at the plurality of sample users and the fourth class feature matrix aiming at the plurality of sample commodities into the neural network model to obtain the predicted score values of the plurality of sample users on the plurality of sample commodities;

and if so, determining the current neural network model as the recommended network model.

3. The method of claim 1, further comprising:

judging whether the predicted score value of the target user on the target commodity reaches a threshold value; and if so, recommending the target commodity to the target user.

4. A recommendation scoring apparatus based on a network structure and a comment text, the apparatus comprising:

the determining module is used for determining a target user in the plurality of sample users and determining a target commodity in the plurality of sample commodities;

the first acquisition module is used for acquiring a first class feature matrix aiming at the target user and a second class feature matrix aiming at the target commodity; the first type feature matrix and the second type feature matrix are determined according to comment text information of the sample users on the sample commodities;

the second acquisition module is used for acquiring a third type feature matrix aiming at the target user and a fourth type feature matrix aiming at the target commodity; the third type feature matrix and the fourth type feature matrix are determined according to the purchasing network structure information of the plurality of sample users on the plurality of sample commodities;

the prediction module is used for inputting the first class characteristic matrix and the third class characteristic matrix aiming at the target user and the second class characteristic matrix and the fourth class characteristic matrix aiming at the target commodity into a recommendation network model to obtain a prediction score value of the target user on the target commodity;

the first obtaining module is specifically configured to:

acquiring a plurality of comment text messages of the plurality of sample users on the plurality of sample commodities;

performing word segmentation processing on the plurality of comment text messages to obtain a word vector of each comment text message;

determining a first class feature matrix aiming at the target user according to the word vectors of the comment text information of the target user;

determining a second class of feature matrix aiming at the target commodity according to the word vectors of the comment text messages of the target commodity;

the purchasing network structure information is network structure information of a heterogeneous information network based on purchasing information of the plurality of sample users on the plurality of sample commodities, nodes of the heterogeneous information network comprise the plurality of sample users and the plurality of sample commodities, edges of the heterogeneous information network are used for connecting the sample users and the sample commodities, and edges of the heterogeneous information network are used for indicating that the sample users have real scoring values on the sample commodities;

the second obtaining module is specifically configured to:

based on the connection relation between the sample user and the sample commodity in the heterogeneous information network, randomly walking by taking the target user as an initial node to determine a plurality of first random walking sequences, randomly walking by taking the target commodity as the initial node to determine a plurality of second random walking sequences; the first random walk sequence comprises a first preset number of nodes; the second random walk sequence comprises a second preset number of nodes;

determining a third type of feature matrix aiming at the target user according to the plurality of first random walk sequences;

and determining a fourth type feature matrix aiming at the target commodity according to the plurality of second random walk sequences.

5. An electronic device is characterized by comprising a processor, a communication interface, a memory and a communication bus, wherein the processor and the communication interface are used for realizing mutual communication by the memory through the communication bus;

a memory for storing a computer program;

a processor for implementing the method steps of any of claims 1 to 3 when executing a program stored in the memory.

6. A computer-readable storage medium, characterized in that a computer program is stored in the computer-readable storage medium, which computer program, when being executed by a processor, carries out the method steps of any one of the claims 1-3.

Images