CN111737427A - A MOOC forum post recommendation method integrating forum interaction behavior and user reading preference - Google Patents

Abstract

The invention discloses a method for recommending a mu lesson forum post by fusing forum interaction behavior and user reading preference, which comprises the following steps: 1) constructing a target scoring matrix, a user interaction scoring matrix and a user interaction frequency matrix of the posts by the user; 2) decomposing the user interaction scoring matrix to obtain a user target function and a user behavior characteristic matrix; 3) calculating the number of times of user interaction behaviors, extracting embedded features, and obtaining a user matrix based on a user behavior feature matrix and a user embedded matrix; 4) extracting a post theme through a noise reduction self-encoder to obtain an object target function, and decomposing a target scoring matrix to obtain a scoring matrix target function; 5) optimizing the objective function of the scoring matrix, the objective function of the user and the objective function of the article, and providing a recommendation list for the user. According to the method, through deep learning and probability-based matrix decomposition, subtle interaction among the users, posts and users is integrated into the model, the cold start problem is relieved, and accurate recommendation is realized by combining reading preference of the users.

Description

A MOOC forum post recommendation method integrating forum interaction behavior and user reading preference

technical field

The invention relates to the technical field of teaching and education data mining and natural language processing, in particular to a MOOC forum post recommendation method integrating forum interaction behavior and user reading preference.

Background technique

In the era of high-speed dissemination of network information and rapid development of computer technology, online learning platforms enable a new generation of netizens to study courses from prestigious universities anytime, anywhere. Massive Open Online Courses (MOOCs) have attracted thousands of scholars. By the beginning of 2018, the number of students studying in my country exceeded 70 million. Although there are a large number of students, due to the openness of the MOOC itself, the level of students is uneven, and the learning objectives are different, which makes the MOOC's dropout rate high and the participation low. As an important module to promote students' knowledge exchange and course participation, MOOC forums play a certain role in reducing the dropout rate.

There are problems of unbalanced information load and "information trek" in MOOC forums. Although most MOOC platforms organize forums with sub-sections at this stage, there is no guarantee that students can choose corresponding sub-sections for their published content, and the problem of information confusion still exists. On the other hand, the single sorting mechanism of the forum makes many new questions flooded by other posts before they are fully answered, and the unanswered rate is high. Responding to forum posts one by one brings a large information load to teachers and teaching assistants. Ideally, the information load of the forum is balanced, and students can have lively discussions, instead of relying solely on teachers to answer questions; students can be a resource to answer each other's questions, assist each other in the forum, and become a learning group sharing knowledge. . MOOC forums need a personalized recommendation technology that combines students' reading preferences and interaction to realize information "routing". The existing MOOC forum post recommendation algorithms use traditional methods such as LDA topic model, word co-occurrence statistics, and associated words to extract user reading preferences. However, traditional methods such as topic model are difficult to combine user behavior information for end-to-end recommendation, and the recommendation is accurate. The degree is not high. On the other hand, the prior art lacks considering the scenarios of low user activity recommended by MOOC forum posts, and does not propose a solution to the cold start problem.

The present invention proposes a recommendation method that integrates forum interaction behavior and user reading preference, and aims to improve the technical defects of the existing MOOC forum post recommendation technology that recommendation accuracy is not high and the user's cold start is not fully considered. Deep learning and probability-based matrix decomposition integrate the information of subtle interactions between users and posts, and users and users into the model to more completely describe user preferences, alleviate the cold start problem, and achieve accurate recommendation.

SUMMARY OF THE INVENTION

The purpose of the present invention is to improve the technical defects of the existing MOOC forum post recommendation technology that the recommendation accuracy is not high and the user's cold start is not fully considered, and proposes a MOOC forum post recommendation method that integrates the forum interaction behavior and the user's reading preference. While considering users' reading preferences, through deep learning and probability-based matrix decomposition, the information of the subtle interaction between users and posts, users and users is integrated into the model, which more completely describes user preferences and alleviates the cold start problem. Course students recommend forum posts of interest to promote forum interaction.

In order to achieve the above purpose, the technical solution provided by the present invention is: a method for recommending MOOC forum posts that integrates forum interactive behavior and user reading preference, including the following steps:

1) Use the user's browsing records in the forum to construct the user's target score matrix, user interaction score matrix and user interaction frequency matrix for posts;

2) Introduce matrices U _a and U _b , use the user interaction frequency matrix as a constraint item, decompose the user interaction rating matrix, obtain the user objective function, and add U _a and U _b as the user behavior characteristic matrix;

3) Count the times of various interactive behaviors of the user, construct a user embedding matrix with the corresponding user ID, splicing the user behavior characteristic matrix and the user embedding matrix, and obtain the user matrix U;

4) Integrate the user's historical posts, introduce the item matrix V, extract the subject of the post through the noise reduction autoencoder, obtain the item objective function, and use the user matrix U and the item matrix V to decompose the target scoring matrix to obtain the scoring matrix objective function;

5) Optimize the scoring matrix objective function, user objective function and item objective function to provide users with a recommendation list.

In step 1), the target rating matrix is a two-dimensional matrix, the horizontal direction of the matrix is the user ID, the column item of the matrix is the post ID, the value of the matrix identifies whether the user is interested in the post or not, and the user's score for the post is set to 1 , indicating that the user has browsed the corresponding post, and the default user may be interested in the subject of the post, and setting the user's score to 0 means that the user has not browsed the corresponding post, and the default user may not be interested in the subject of the post ; The user interaction rating matrix and the user interaction frequency matrix are both two-dimensional matrices, which respectively identify whether there has been interaction between users and the total number of interactions. The interaction includes replies, comments, likes and browsing behaviors between users. When there is interaction between users, the value corresponding to the user interaction score matrix is 1, otherwise it is 0, and the value corresponding to the user interaction frequency matrix is the number of times the user has interacted.

In step 2), obtain the user objective function and the user feature matrix, including the following steps:

2.1) Introduce matrices U _a and U _b , assuming that each row of U _a and U _b obeys the following Gaussian distribution:

为超参数λ_a的倒数，

为超参数λ_b的倒数；where i represents each row of the matrix,

is the reciprocal of the hyperparameter λ _a ,

is the reciprocal of the hyperparameter λ _b ;

2.2) Denote the user interaction rating matrix as Q and the user interaction frequency matrix as C, assuming that each value in Q obeys the Gaussian distribution of the following formula:

为矩阵U_a第i行的向量，和

为矩阵U_b第j列的向量，C^-1为用户互动频次矩阵C的逆矩阵，是上式高斯分布的方差；In the formula, Q _ij is the item value of the i-th row and the j-th column of the user interaction rating matrix,

is the vector of the ith row of matrix U _a , and

is the vector of the jth column of the matrix U _b , C ^-1 is the inverse matrix of the user interaction frequency matrix C, and is the variance of the Gaussian distribution of the above formula;

2.3) According to the probability-based matrix decomposition method, the derivation formula is obtained:

In the formula, P(|) represents the posterior probability, and ∝ represents the proportional relationship;

Substitute into the Gaussian distribution of Q, U _a , U _b , and derive the user objective function;

2.4) Add U _a and U _b to obtain the user behavior characteristic matrix.

In step 3), counting the number of various interactive behaviors of the user includes counting the number of likes, pageviews, replies, replies, postings, and followings of the user, and constructs a user embedding matrix with the corresponding user ID , splicing the user behavior feature matrix and the user embedding matrix to obtain the user matrix U, including the following steps:

3.1) Discretize the times of various interactive behaviors to obtain 6 discrete features;

3.2) Take the 6 discretized features and the user ID as the input of the model embedding layer, and obtain 7 embedding vectors;

3.3) after splicing 7 embedding vectors as the user's embedding layer vector, the embedding layer vectors of all users constitute the user embedding matrix U _c ;

3.4) Splicing the user behavior feature matrix and the user embedding matrix U _c , and constructing the user matrix U according to the following formula:

为超参数λ_u的倒数，为高斯分布的方差。In the formula, the symbol [;] represents the splicing operation, U _a + U _b is the user behavior feature matrix, U _c is the user embedding matrix,

is the reciprocal of the hyperparameter λ _u and is the variance of the Gaussian distribution.

In step 4), the item objective function and the scoring matrix objective function are obtained, including the following steps:

4.1) The title of the forum post, the specific description of the post and a series of historical replies are combined into a content post, and the text is preprocessed;

4.2) Convert the content of each post into bit-sequence encoding and input it to the embedding layer, or initialize the text embedding layer directly through the pre-trained word vector, and output the embedded representation of the text;

降噪自编码器网络的权重服从如下高斯分布：4.3) Input the embedded representation of the text into the denoising autoencoder, restore the input information through the denoising autoencoder, and use the middle layer of the denoising autoencoder to extract the topic vector of the text

The weights of the denoising autoencoder network obey the following Gaussian distribution:

为超参数λ_w的倒数，为高斯分布的方差；where W is the weight of each layer of the denoising autoencoder,

is the inverse of the hyperparameter _λw , and is the variance of the Gaussian distribution;

The user preference ε is introduced, and the following item matrix V is constructed through ε and the topic vector, so that the item matrix V contains topic information and user preference information:

ε obeys the following Gaussian distribution:

为超参数λ_v的倒数，为高斯分布的方差；In the formula,

is the reciprocal of the hyperparameter λ _v , and is the variance of the Gaussian distribution;

4.4) Use the noise reduction autoencoder to restore the input information, and substitute the expressions of W, V, ε to obtain the object objective function;

4.5) Decompose the target rating matrix using the user matrix U and the item matrix V by the following formula:

为超参数λ_r的倒数，为高斯分布的方差；In the formula, R _ij is the item value of the ith row and jth column of the target rating matrix; U _i represents the ith row of the user matrix U, V _j represents the jth row of the item matrix V,

is the reciprocal of the hyperparameter _λr , and is the variance of the Gaussian distribution;

According to the probability-based matrix factorization method, the scoring matrix objective function is obtained.

In step 5), the scoring matrix objective function, the user objective function and the item objective function are optimized, within the preset training threshold, the representation of the user matrix U and the item matrix V is corrected, and the final user matrix and the item matrix are dot-multiplied, from In the result of the dot product, the target user's score for a series of items is retrieved, and by sorting the score list of each user, the corresponding top M posts with high scores are obtained, and a recommendation list is provided for the user.

Compared with the prior art, the present invention has the following advantages and beneficial effects:

1. The present invention uses the noise reduction autoencoder to extract the subject information of the text. Compared with the related methods, this method can utilize the advantages of deep learning in natural language processing to better mine the user's reading preference.

2. The present invention uses a probability-based matrix decomposition method to decompose the user forum interaction rating matrix, and can extract user behavior characteristics through the decomposition of the sparse matrix when there are only a few interactions, so as to alleviate the cold start problem to a certain extent.

3. Aiming at the problem that the text topic model is difficult to perform end-to-end recommendation in combination with user behavior information, the present invention integrates forum interaction information with the topic model through deep learning and a probability-based matrix decomposition method to achieve end-to-end accurate recommendation.

4. The model of the present invention is scalable and supports multiple user information embedding.

Description of drawings

Figure 1 is a logic flow diagram of the method of the present invention.

FIG. 2 is a decomposition analysis diagram of a user interaction matrix of the present invention.

FIG. 3 is a diagram of a forum post recommendation model of the present invention.

FIG. 4 is a schematic diagram of extracting text latent vectors by using noise reduction autoencoder according to the present invention.

Detailed ways

The present invention will be further described below in conjunction with specific embodiments.

As shown in FIG. 1 , the method for recommending MOOC forum posts that integrates forum interaction behaviors and user reading preferences provided by this embodiment includes the following steps:

1) Construct the target score matrix, the target score matrix is a two-dimensional matrix, the horizontal direction of the matrix is the user ID, the column item of the matrix is the post ID, and the value of the matrix identifies whether the user is interested in the post, and the user's score for the post is set as 1. It means that the user has browsed the corresponding post, and the default user may be interested in the subject of the post. Setting the user's score to 0 means that the user has not browsed the corresponding post, and the default user may not be interested in the subject of the post. interest.

Since there are many unanswered posts in the forum, the target score matrix will be very sparse, that is, most of the scores in the matrix are 0. The original data needs to be filtered to remove posts with too few words and many special symbols, and set the user pageview threshold. , remove users whose number of browsing posts is less than the threshold, so as to remove some "zombie" users and "spam" posts. After that, the scoring matrix is constructed, and according to a certain negative sampling ratio, the items with a value of 0 in the matrix are randomly deleted. Specifically, a mask matrix with the same rows and columns as the original matrix can be used, 1 is used to identify the items that need to be retained in the original matrix, and 0 is used to identify the items with a score of 0 that need to be deleted from the original matrix, and the mask matrix is calculated in advance according to the negative sampling ratio. The downsampling matrix is obtained by ANDing the mask matrix with the original score matrix. The downsampling matrix is the actual data that is subsequently fed into the model.

Use the filtered data to construct a user interaction rating matrix and an interaction frequency matrix. They are also two-dimensional matrices, which respectively identify whether there has been interaction between users and the total number of interactions. The interaction includes replies, comments, Like and browse behaviors, if the interaction occurs between users, the value corresponding to the user interaction score matrix is 1, otherwise it is 0, and the value corresponding to the user interaction frequency matrix is the number of times the user has interacted.

2) Obtain the user objective function and the user feature matrix, including the following steps:

2.1) Introduce matrices U _a and U _b , assuming that each row of U _a and U _b obeys the following Gaussian distribution:

为超参数λ_a的倒数，

为超参数λ_b的倒数。where i represents each row of the matrix,

is the reciprocal of the hyperparameter λ _a ,

is the inverse of the hyperparameter _λb .

2.2) Denote the user interaction rating matrix as Q and the user interaction frequency matrix as C, assuming that each value in Q obeys the Gaussian distribution of the following formula:

为矩阵U_a第i行的向量，和

为矩阵U_b第j列的向量，C^-1为用户互动频次矩阵C的逆矩阵，是上式高斯分布的方差；将用户互动频次矩阵作为方差，能够使得用户互动的次数越多，对应项的惩罚越大，即用户在论坛中越活跃，对其特征表示的准确性要求越高。In the formula, Q _ij is the item value of the i-th row and the j-th column of the user interaction rating matrix,

is the vector of the ith row of matrix U _a , and

is the vector of the jth column of the matrix U _b , C ^-1 is the inverse matrix of the user interaction frequency matrix C, which is the variance of the Gaussian distribution of the above formula; taking the user interaction frequency matrix as the variance, the more times the user interacts, the corresponding item The greater the penalty, that is, the more active the user is in the forum, the higher the requirement for the accuracy of its feature representation.

2.3) According to the maximum posterior probability, the derivation formula is obtained:

In the formula, P(|) represents the posterior probability, and ∝ represents the proportional relationship;

Note that the above Q can be obtained from the statistics of user interaction behavior and is a known observation, so the above problem is equivalent to using the known observation Q to infer the model parameters U _a and U _b that are most likely to lead to this observation. Therefore, by introducing the prior distribution of U _a and U _b , that is, the Gaussian distribution, maximum likelihood estimation can be used to calculate the maximum posterior probability of U _a and U _b to obtain the user objective function.

2.4) Since U _a and U _b are the results of the decomposition of the user interaction rating matrix, they contain the information of the matrix horizontal and column items, respectively. As shown in Figure 2, the column items of the user interaction rating matrix represent the user's active response behavior. It reflects the active and active characteristics of users in the forum, and the horizontal item represents the response behavior of other users to the user, and describes the passive behavior characteristics of users in the forum. Add U _a and U _b to get the user behavior characteristic matrix, so that the result contains the information of the user's active and passive behavior.

3) Counting the number of users' various interactive behaviors, including counting the number of likes, pageviews, replies, replies, posts, and followers, and constructing a user embedding matrix with the corresponding user ID, splicing user behaviors The feature matrix and the user embedding matrix are obtained to obtain the user matrix U, including the following steps:

3.1) Discretize the times (numerical values) of various interactive behaviors to obtain 6 discrete features; take the binning discretization as an example, assuming that the sample points of the number of likes of the user = {12, 13, 24, 4, 5, 34, 56, 98, 8}, the binning interval is 5, then the discretized data = {3, 3, 5, 1, 1, 7, 12, 20, 2}. The binning interval can be determined according to the histogram of the data, and is generally set to a value that makes the data distribution more uniform.

3.2) Take the 6 discretized features and user ID as the input of the model embedding layer, and obtain 7 embedding vectors; the output dimension of the embedding layer of the model can be adjusted according to the distribution range of the data. If the value after discretization is not large, the dimension It can be set to be smaller. For specific dimension values, please refer to Figure 3.

3.3) The 7 embedding vectors are spliced as the user's embedding layer vector, and the embedding layer vectors of all users constitute the user embedding matrix U _c .

3.4) Splicing the user behavior feature matrix and the user embedding matrix U _c , and constructing the user matrix U according to the following formula:

为超参数λ_u的倒数，为高斯分布的方差。In the formula, the symbol [;] represents the splicing operation, U _a + U _b is the user behavior feature matrix, U _c is the user embedding matrix,

is the reciprocal of the hyperparameter λ _u and is the variance of the Gaussian distribution.

4) Obtain the item objective function and the scoring matrix objective function, including the following steps:

4.1) Combine the title of the forum post, the specific description of the post and a series of historical replies into a content post, and preprocess the text. The specific steps are: first, combine the title of the forum post, the specific description of the post and a series of historical replies into a content post , and then preprocess the text, including punctuation and text separation, case conversion, removal of meaningless punctuation characters, and links.

4.2) Convert the text to bit-sequence encoding and input it to the embedding layer, or initialize the text embedding layer directly through the pre-trained word vector. Take bit-sequence encoding as an example to illustrate the principle of the embedding layer: Suppose post X={How many layers does the OSI model consist of? }, the order of words in X in the text base is {How:899,many:1456,layer:600,does:3245,the:6723,OSI:28,model:876,consist:547,of:2323 }, then the representation of the post X into a sequence sequence is, X={899, 1456, 600, 3245, 6723, 28, 876, 547, 2323}; assuming that the dimension of the embedding layer is d and the size of the thesaurus is m, the embedding layer will is randomly initialized to a matrix of size m × d, so the words in post X can be indexed by bit order to get the corresponding word vector (embedding vector). When the text pre-training vector is used as input, the word vector representation of each word of the text can be obtained through a large-scale public word embedding, such as Glove or Word2vec, and the embedding layer matrix can be initialized. The embedded representation of the text is then output through the embedding layer. The specific output dimension can be defined by yourself, or set with reference to Figure 3.

包含了文本的主题信息。4.3) Input the embedded representation of the text into the denoising autoencoder, as shown in Figure 4, the denoising autoencoder is composed of a multi-layer feedforward neural network, including an encoding layer, an intermediate layer and a decoding layer. The weight is randomly initialized by the Gaussian distribution of the following formula; the denoising autoencoder restores the original information of the text through the decoder, and captures the implicit description of the text through fewer neural units in the middle layer, making the text more abstract and less dimensional. The information representation of , from the interpretability level, the output vector of the middle layer

Contains the subject information of the text.

为超参数λ_w的倒数，为高斯分布的方差；where W is the weight of each layer of the denoising autoencoder,

is the inverse of the hyperparameter _λw , and is the variance of the Gaussian distribution;

The user preference ε is introduced, and the following item matrix V is constructed by ε and the topic vector, so that the item matrix contains topic information and user preference information:

ε obeys the following Gaussian distribution:

为超参数λ_v的倒数，为高斯分布的方差。In the formula,

is the reciprocal of the hyperparameter _λv , and is the variance of the Gaussian distribution.

4.4) Use the denoising autoencoder to restore the input information, and substitute the expressions of W, V, ε to obtain the item objective function.

4.5) Decompose the target rating matrix using the user matrix U and the item matrix V by the following formula:

为超参数λ_r的倒数，为高斯分布的方差；In the formula, R _ij is the item value of the ith row and jth column of the target rating matrix, U _i represents the ith row of the user matrix U, V _j represents the jth row of the item matrix V,

is the reciprocal of the hyperparameter _λr , and is the variance of the Gaussian distribution;

Note that the above R is obtained through statistics and is a known observation, so the above problem is equivalent to using the known observation R, inversely inferring the model parameters U and V that are most likely to lead to this observation. Therefore, by introducing The prior distribution of U and V, so as to use the maximum likelihood estimation to calculate the maximum posterior probability of U and V, and obtain the scoring matrix objective function.

As can be seen from Figure 2 and the above steps, the decomposition and prediction of the target rating matrix involves user reading preference extraction, user feature extraction and user interaction matrix decomposition, and each component is integrated into the model for end-to-end recommendation.

5) Optimize the scoring matrix objective function, user objective function, and item objective function. In order to minimize the objective function, mature optimizers in the industry, such as Adagrad, RMSprop, and SGD, can be used to solve the gradient of the objective function, so that the parameters of the model are reversed. The propagation is updated, and the representation of the user matrix U and the item matrix V is corrected within a certain training threshold. This training threshold can generally be set by observing the stability and convergence of the predicted results, and there is no fixed setting value.

After reaching the threshold, the end-user matrix and the item matrix are dot-multiplied, and the target user's score for a series of items is retrieved from the result of dot-multiplication. By sorting the score list of each user, the corresponding top M scores are obtained. posts, providing users with a list of recommendations.

The above-mentioned embodiments are only preferred embodiments of the present invention, and are not intended to limit the scope of implementation of the present invention. Therefore, any changes made according to the shape and principle of the present invention should be included within the protection scope of the present invention.

Claims (6)

1. The method for recommending the forum section of the admiration lesson integrating the interactive behavior of the forum and the reading preference of the user is characterized by comprising the following steps of:

1) constructing a target scoring matrix, a user interaction scoring matrix and a user interaction frequency matrix of the posts by the user by using browsing records of the user in the forum;

2) introducing a matrix U_aAnd U_bDecomposing a user interaction scoring matrix by using a user interaction frequency matrix as a constraint item to obtain a user target function, and calculating the U_aAnd U_bAdding the data to obtain a user behavior characteristic matrix;

3) counting the times of various interactive behaviors of a user, constructing a user embedded matrix with a corresponding user ID, and splicing the user behavior characteristic matrix and the user embedded matrix to obtain a user matrix U;

4) integrating the historical posts of the user, introducing an article matrix V, extracting the subject of the post through a noise reduction self-encoder to obtain an article objective function, and decomposing a target scoring matrix by using a user matrix U and the article matrix V to obtain a scoring matrix objective function;

5) optimizing the objective function of the scoring matrix, the objective function of the user and the objective function of the article, and providing a recommendation list for the user.

2. The method as claimed in claim 1, wherein in step 1), the objective scoring matrix is a two-dimensional matrix, the horizontal direction of the matrix is a user ID, the columns of the matrix are post IDs, the value of the matrix identifies whether the user is interested in the post, the score of the post by the user is set to 1, which indicates that the user browses the corresponding post, and the default user may be interested in the subject of the post, and the score of the post by the user is set to 0, which indicates that the user does not browse the corresponding post, and the default user may not be interested in the subject of the post; the user interaction scoring matrix and the user interaction frequency matrix are two-dimensional matrices, whether interaction exists between users and the total number of the interaction are respectively identified, the interaction comprises the actions of reply, comment, praise and browse among the users, if the interaction occurs between the users, the value corresponding to the user interaction scoring matrix is 1, otherwise, the value corresponding to the user interaction frequency matrix is 0, and the value corresponding to the user interaction frequency matrix is the number of the interaction occurring between the users.

3. The mu lesson forum post recommendation method integrating forum interaction behavior and user reading preference as claimed in claim 1, wherein in step 2), obtaining a user objective function and a user characteristic matrix comprises the following steps:

2.1) introducing a matrix U_aAnd U_bSuppose U_aAnd U_bObeys the following gaussian distribution:

where i denotes each row of the matrix,

is a hyperparameter λ_aThe inverse number of (c) is,

is a hyperparameter λ_bThe reciprocal of (a);

2.2) let the user interaction scoring matrix be Q and the user interaction frequency matrix be C, assuming that each value in Q obeys the following Gaussian distribution:

in the formula, Q_ijThe term values in the ith row, jth column of the matrix are scored for user interaction,

is a matrix U_aThe vector of the ith row, and

is a matrix U_bVector of j-th column, C^-1The inverse matrix of the user interaction frequency matrix C is the variance of the Gaussian distribution;

2.3) obtaining a derivation formula according to a matrix decomposition method based on probability:

in the formula, P (|) represents posterior probability, and oc represents proportional relation;

substituted into Q, U_a,U_bThe user target function is obtained through deduction according to the Gaussian distribution;

2.4) mixing U_aAnd U_bAnd adding to obtain a user behavior characteristic matrix.

4. The mu lesson forum post recommendation method fusing forum interactive behaviors and user reading preferences according to claim 1, wherein in step 3), the statistics of the various interactive behaviors of the user includes the statistics of the praise times, the browsed quantity, the replied number, the reply number, the post number and the concerned number of the user, the construction of a user embedding matrix with the corresponding user ID, and the splicing of the user behavior characteristic matrix and the user embedding matrix to obtain a user matrix U, and comprises the following steps:

3.1) discretizing the times of various interactive behaviors to obtain 6 discretization characteristics;

3.2) taking 6 discretization features and the user ID as the input of a model embedding layer to obtain 7 embedding vectors;

3.3) splicing 7 embedded vectors to be used as embedded layer vectors of users, wherein the embedded layer vectors of all the users form a user embedded matrix U_c；

3.4) splicing the user behavior characteristic matrix and the user embedded matrix U_cThe user matrix U is constructed according to the following equation:

in the formula, the symbol [;]indicating a splicing operation, U_a+U_bIn the form of a matrix of user behavior characteristics,

is a hyperparameter λ_uThe inverse of (d) is the variance of the gaussian distribution.

5. The mu lesson forum post recommendation method integrating forum interaction behavior and user reading preference of claim 1, wherein in step 4), an object objective function and a scoring matrix objective function are obtained, comprising the steps of:

4.1) combining the title of the forum post, the detailed description of the post and a series of historical replies into a content post, and preprocessing the text;

4.2) converting the content of each post into a bit sequence code, inputting the bit sequence code into an embedding layer, or directly initializing the embedding layer of the text through a pre-training word vector, and outputting the embedded representation of the text;

4.3) inputting the embedded representation of the text into a noise reduction self-encoder, restoring the input information through the noise reduction self-encoder, and extracting the theme vector of the text by utilizing the middle layer of the noise reduction self-encoder

The weights of the noise-reducing self-encoder network obey the following gaussian distribution:

wherein W is the weight of each layer of the noise reduction self-encoder,

is a hyperparameter λ_wThe inverse of (d), is the variance of the gaussian distribution;

introducing user preferences, and constructing an article matrix V by combining the user preferences and the theme vector, so that the article matrix V comprises theme information and user preference information:

obey the following gaussian distribution:

in the formula (I), the compound is shown in the specification,

is a hyperparameter λ_vThe inverse of (d), is the variance of the gaussian distribution;

4.4) restoring input information by using a noise reduction self-encoder, substituting into an expression of W and V to obtain an object objective function;

4.5) decomposing the objective scoring matrix by using the user matrix U and the item matrix V according to the following formula:

in the formula, R_ijThe item value of the ith row and the jth column of the target scoring matrix is obtained; u shape_iI-th row, V, representing the user matrix U_jThe jth row of the commodity matrix V is represented,

is a hyperparameter λ_rThe inverse of (d), is the variance of the gaussian distribution;

and obtaining a scoring matrix target function according to a probability-based matrix decomposition method.

6. The mu lesson forum post recommendation method fusing forum interaction behavior and user reading preference according to claim 1, wherein in step 5), the scoring matrix objective function, the user objective function and the item objective function are optimized, the representation of the user matrix U and the representation of the item matrix V are modified within a preset training threshold, the final user matrix and the item matrix are dot-multiplied, the scores of the target user on a series of items are retrieved from the dot-multiplied result, the scoring lists of each user are sorted to obtain posts corresponding to the top M scores, and a recommendation list is provided for the user.

Images