Learning resource recommendation method and system based on project domain knowledge and user comments

Abstract

The application discloses a learning resource recommendation method and system based on project field knowledge and user comments, and the method comprises the following steps: collecting learner information, learning resource characteristic information and teacher information, wherein the learner information comprises learner description information and learner interaction information with the learning resource, and the learning resource characteristic information comprises learning resource description information and learning resource characteristic information; finding a teacher with the highest similarity to the learner according to the learner information, and obtaining a matching score of a target learning resource according to the teacher characteristics through a convolutional neural network; establishing a learner short-term preference model and a long-term preference model, and fusing the two models to obtain a learner personal preference model; establishing a learner group preference model, and fusing the learner personal and group models to obtain a learner preference model; and establishing a learning resource characteristic information model and a field knowledge model by using various information characteristics of the learning resource according to the learning resource characteristic information, so that the accuracy of learning resource recommendation is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of recommendation system in computer technology, and particularly relates to a learning resource recommendation method and system based on item domain knowledge and user comments. BACKGROUND

[0002] In the process of item recommendation, the key to improving the recommendation quality is to model the item deeply and effectively. In the existing item recommendation method, the design of the user model is focused, and the feature description and extraction of the item model are ignored. Moreover, because the user has a unique preference for the domain and type of the item in the use of the item. The traditional item recommendation method only shallowly arranges the features of the item and matches the user model, and the precision in the domain and type of the user's attention is not enough. How to obtain the domain features of the item and completely describe the features of the user is the key to improving the existing method. In the matching process, how to connect the constructed models of the item and the user to make the method lightweight is also a problem worth thinking about. SUMMARY

[0003] In order to solve the above problems, the present application provides a learning resource recommendation method and system based on item domain knowledge and user comments, which extracts the domain features of the item, matches the user model, recommends the item more suitable for the user, and finally realizes the purpose of improving the user experience in the recommendation process.

[0004] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows: a learning resource recommendation method based on item domain knowledge and user comments, comprising the following steps:

[0005] Step 1, obtaining item domain feature information, user historical interaction information matrix Y, user comment set information, query embedding matrix Q, and user feature information;

[0006] Step 2, constructing a user-item path graph SG according to the user historical interaction information matrix Y;

[0007] Step 3, constructing an item domain feature model using a graph convolutional neural network based on domain grouping through the item domain feature information;

[0008] Step 4, calculating the attention scores of different entities in the user-item path graph through a multi-channel attention mechanism, then adding the attention scores calculated by different channels to obtain the attention weights of different entities in the graph, and integrating the features of all entities in the user-item path graph to obtain a global item model G l ;

[0009] Step 5, based on the attention mechanism, constructing a user-item path graph SG based on the global item model Gi with the item global model G l The two are fused by assigning different weights to obtain the final item model I:

[0010] Step 6, preprocessing the user's review set, i.e. converting the words in the review set into d-dimensional embedding, finding the words in the user's review set from the embedding matrix by giving a query embedding matrix Q, and converting them into corresponding embedding vectors to obtain the user's corresponding review matrix RC. The item domain feature obtained in step 3 is converted into a matrix form to obtain the item feature matrix IC. After obtaining the user review matrix RC, the Dirichlet probability clustering algorithm is used to calculate the domain type to which the user belongs, and then the domain type to which the user belongs is also converted into a domain type matrix SC.

[0011] Step 7, multiplying the user's review matrix RC and the item feature matrix IC to obtain the user-item domain matrix UIS, multiplying the user review matrix RC and the user type matrix SC to obtain the user-item type matrix UIT, and using a convolutional neural network to extract user domain features U t and user type features U c from the user-item domain matrix UIS and the user-item type matrix UIT respectively.

[0012] Based on the attention mechanism, the weights of the user domain features U t and the user type features U c are calculated, and the two are connected and fused to obtain the user feature model U.

[0013] Step 8, connecting the item model I and the user feature model U, taking the connected model as the input of the multi-layer deep neural network, obtaining the interaction features of the target item and the user, and taking the interaction features as the initial recommendation score ps of the candidate item.

[0014] Step 9, directly adding the domain type to which the user belongs obtained in step 7 to the interaction features of the item model I and the user feature model U as the final recommendation score f's of the candidate item.

[0015] Step 10, ranking the recommendation scores f's according to their high and low, and recommending the top N candidate items with the highest scores to the user in turn.

[0016] In step 1, the item domain feature information refers to detailed description information of knowledge points, categories and release time contained in the item; the user review set information includes the user's review set of all items; and the user historical interaction information matrix Y refers to the implicit interaction matrix of the user and the item. If the user and the item have historically interacted, Y ijis set to 1, otherwise 0; the query embedding matrix Q refers to the matrix including existing word information, any word in the review set of the learner can be found in Q.

[0017] In step 2, the specific process of constructing the user-item path graph is as follows:

[0018] In step 2.1, the paths between the user and the item are sampled through the historical interaction information of the user and the item, after the sampling is completed, K paths are selected, and the length of each path is limited to 6, in the specific sampling process, the target user is the starting point, and the item is the end point, the random walk of 6 paths is carried out, finally, k paths between the user and the target item are obtained;

[0019] In step 2.2, after obtaining the k paths, the paths are spliced to form a user-item path graph with the starting point being the target user and the end point being the candidate item, the maximum path length of the k paths being 6, the user-item path graph contains multiple entities.

[0020] Step 3 is as follows: the construction of the item domain feature model is divided into two parts: group propagation and knowledge group aggregation; the domain knowledge points contained in the item are expanded along the paths in the item domain knowledge graph, and each propagated set is recursively accessed;

[0021] In step 3.1, in the group propagation, the characteristics of the item are reflected in the domain knowledge points it contains, the characteristics of the item are represented by mining the characteristics of the knowledge points and using the domain knowledge points to represent the item, which is defined as follows:

[0022]

[0023] Wherein, c represents the item, and e represents the domain knowledge point;

[0024] The initial representation of the item set is the integration of the domain knowledge point entities, after the propagation in the item domain knowledge graph, all different types of entities in the set are traversed, all entities can effectively expand the potential characteristics of the item, and the multi-layer propagation entity of the item is recursively defined as follows:

[0025]

[0026] Wherein, l represents the hierarchical distance between the entity set and the item domain knowledge point seed;

[0027] In step 3.2, the knowledge group aggregation, the domain feature propagation of the item is divided into two stages: the first stage is to determine the attention weight of the entity on different paths using the intra-group attention network with entity granularity, and the second stage is to represent the domain feature of the item based on the inter-group aggregation network of the relationship link, through the two stages, the domain feature of the item is constructed in more detail.

[0028] Step 3.2.1, in the group attention network stage, the tail entity is represented by different connections and head entities, and different paths to the tail entity have different meanings in the feature propagation process. An attention network is used to determine the proportion of weights on different paths.

[0029] First, select a candidate project entity in the project domain knowledge graph Given the candidate project entity The tail domain knowledge entity e is calculated to obtain the attention weight of the tail domain knowledge entity relative to the project, wherein The representation of the tail entity ti is as follows:

[0030]

[0031] Wherein The representation of the tail entity ti is as follows:

[0032] After obtaining the attention weights of all entities in each group in the same layer, the weighted entities in each group are added to obtain the representation of these relationship links. Finally, the weighted representation of the k different domain groups is obtained:

[0033]

[0034] Wherein, pi represents the attention weight of the tail domain knowledge entity, Represent all tail entities connected to the candidate project;

[0035] Step 3.2.2, in the inter-group aggregation network stage, the representation of different relationship links in each propagation layer is obtained through the head entity and the relationship link, and the relationship information in the project domain knowledge graph is directly used. Each group is considered as an extension of the project domain feature in a certain knowledge direction, and the inter-group aggregation network is used to continue to represent the domain feature of the project, and then the domain feature model is obtained.

[0036] Step 4 is: through the multi-channel attention mechanism, the attention scores of different entities and relationships in the user-item path graph are calculated, the entity information contained in the entire user-item path graph is included in the project, then the attention scores are used as the weights of different entities to distinguish the importance of different entities, and finally all entities are integrated into the features of the target project; In the domain knowledge model, the vector embedding of the entity in the user-item path graph is obtained, and the entity embeddings are combined to form the entity embedding matrix of the user-item path graph.

[0037] After obtaining the overall embedding matrix of the user-item path graph, different weights are assigned to the entities in the user-item path graph using a multi-channel attention mechanism, which consists of three channels, namely the interaction channel, the time channel, and the domain channel.

[0038] In the interaction channel, the importance of domain knowledge points to the target item is captured through vector inner product, and different attention scores are assigned to the users according to their knowledge levels, and the attention score expression is as follows,

[0039]

[0040] where v j represents the embedding of the target candidate item, e n represents the embedding of each knowledge point in the user-item path graph, represents the interaction weight of the knowledge point to the target item;

[0041] Finally, the weighted sum of all direct interaction knowledge points of the target item according to their interaction weight vectors is as follows,

[0042]

[0043] In the time channel, according to the interaction time of the domain knowledge point and the candidate item, a multi-layer perceptron is used to obtain the attention score of the directly interacting knowledge point, and the formula is as follows:

[0044] time i = U T tanh(W l e n + b l )

[0045]

[0046]

[0047] where W i ∈ R d×d , U i ∈ R d represent weight matrices, and b l represents bias.

[0048] In the domain channel, the domain attention score of the knowledge point entity in the user-item path graph is obtained using the item domain feature model obtained in the item domain feature modeling, and the attention score of the domain channel part is as follows:

[0049]

[0050] where si It is the output of the project domain feature model.

[0051] Finally, by combining the attention scores from these three channels, a three-channel attention score is calculated for each entity. The three-channel scores are then concatenated to obtain the global project model G. l ,in, Indicates a connection operation.

[0052]

[0053] Step 6, the clustering process, is as follows:

[0054] User set U = {u1, u2, ..., u...} n}, the domain set T = {t1, t2, ..., t} k Each user u in the user set U can be viewed as a sequence of words. Let wi represent the i-th word. Let u have n words. All the distinct words in U form a large set S, where S = {s1, s2, ..., sn}. j};

[0055] The user set U is used as input to the clustering algorithm, and it is clustered into k types. T contains j words in total.

[0056] ① For each user u in U n The probability corresponding to different groups in Indicate u n The probability of the k-th type in T is calculated as follows:

[0057]

[0058] in Indicate u n The Tth term in T corresponds to the middle. k The number of words of each type, where n is the total number of words in un;

[0059] ② For each population T in T k The probability of generating different words in, T represents k The probability of generating the j-th word in T is as follows:

[0060]

[0061] in Represents group T k The number of words containing the j-th word in T, where N represents the number of words in T. kThe number of all words in T, the core formula of LDA is as follows:

[0062]

[0063] By the current θ u and φ t The probability of the occurrence of the word w in the user u is given, where p(t|u) is calculated using θ u and p(w|t) is calculated using By the current θu and φt, p(t|u) can be calculated for any one type Ti in the user u description, and then according to the results, the topic to which the word should correspond is updated, and at the same time, if the update changes the type Ti to which the word corresponds, it will also affect θu and φt in turn;

[0064] Through the Dirichlet clustering algorithm described above, the user u n contains the user category, as follows:

[0065]

[0066] Where, is the different user categories contained by the user, p i indicates the probability weight of belonging to different categories, i=1, 2, 3..., p n T k represents the weight of the target user's preference for a certain type of resource.

[0067] Step 7 is as follows:

[0068] ① Multi-type CNN: After clustering to obtain the different types of users, each type is expressed through a matrix, and the user's review matrix RC i is multiplied by the user's type matrix SC to obtain the different user review type matrices;

[0069] A user belongs to multiple different types, and the target user's review matrix is regarded as a tensor. In the convolution process, the number of filters is consistent with the number of types to which the user belongs. By sliding the filter in the user's multiple review type matrices, the feature information in different review types is extracted, and finally all the feature information is integrated together to obtain the feature result of review type k;

[0070] ② Max-pooling operation: After the convolution operation, each review type contains multiple feature matrices generated by filters, and a max-pooling operation is used to obtain the significant features.

[0071] Step 9 is as follows:

[0072] First, the learner belongs to a variety of categories obtained by LDA clustering algorithm, and the dot product operation is performed on the learner's category and the probability of the category to which the learner belongs, and the calculation process is as follows:

[0073]

[0074]

[0075] wherein, r refers to the field to which the user belongs, u，t T refers to the user's preference for a certain field, k refers to the field contained by the user, standardized in the range of 0 to 1, the predicted score of the candidate item is as follows:

[0076]

[0077] wherein, ps is the initial recommendation score.

[0078] In addition, the present application also provides a computer device comprising a processor and a memory, wherein the memory stores an executable program, and the processor executes the executable program to implement the learning resource recommendation method based on item field knowledge and user comments.

[0079] Further, a computer readable storage medium is provided, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the learning resource recommendation method based on item field knowledge and user comments.

[0080] Compared with the existing learning resource recommendation method, the present application has at least the following advantages:

[0081] The present application is based on the item field feature model for modeling in the item recommendation, realizes personalized recommendation, integrates the type preference into the recommendation method, improves the fine-grained accuracy of the field feature recommendation, considers the importance of the field feature for the item in the item modeling, and simultaneously considers the global model of the item to realize accurate modeling of the item, in the user modeling process, the field feature model and the clustering type feature of the item are fused through the knowledge graph to realize accurate modeling of the user in the field, and in the fusion process of the user model and the learning model, the user model obtained by clustering is directly added to the predicted score to realize the fine-grained field preference of the user and improve the sensitivity of the method to the user's field preference. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure 1 is a flowchart of the learning resource recommendation method based on item field knowledge and user comments. DETAILED DESCRIPTION

[0083] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0084] In the description of the present application, it should be understood that the terms "include" and "contain" indicate the existence of described features, integers, steps, operations, elements and / or components, but do not exclude the existence or addition of one or more other features, integers, steps, operations, elements, components and / or sets thereof.

[0085] Figure 1 The flowchart is based on the field feature preference recommendation, and the embodiments of the present application will be described in detail.

[0086] Step 1, obtaining project field feature information, user historical interaction information matrix Y, user comment set information and query embedding matrix Q. The project field feature information refers to the detailed description information of the project, such as knowledge points, categories and release time; the user comment set information includes the user's comment set for all projects; the user historical interaction information matrix Y refers to the implicit interaction matrix of the user and the project, if the user and the project have historical interaction, Yij is set to 1, otherwise 0; the query embedding matrix Q refers to the matrix including existing word information, any word in the comment set of the learner can be found in Q.

[0087] Step 2, constructing a user-project path graph SG according to the user's historical interaction information. Specifically as follows:

[0088] Step 2.1, sampling the path between the user and the project through the historical interaction information of the user and the project, after sampling, selecting K paths, limiting the length of each path to 6 to avoid too long path with more noise neighbors. In the specific sampling process, starting from the target user, the project as the end point, the path is 6 random walk, finally, k paths between the user and the target project are obtained.

[0089] Step 2.2, after obtaining k paths, the paths are spliced to form a user-project path graph with the starting point as the target user, the end point as the candidate project and the maximum path length of k as 6. The user-project path graph contains various entities, such as the field knowledge and feature attributes of the project.

[0090] Step 3, constructing a project field feature model S based on the field grouping graph convolutional neural network through the project field feature information i . Specifically as follows:

[0091] Project domain feature model S i The construction is divided into two parts: group propagation and knowledge group aggregation. The domain knowledge points contained in the project are expanded along the path in the project domain knowledge graph as a seed set, and each propagated set is recursively accessed.

[0092] Step 3.1, in the group propagation, the project features are generally reflected in the domain knowledge points it contains, and the project features are accurately and completely represented by mining the knowledge point features and using the domain knowledge points to represent the project, which is defined as follows:

[0093]

[0094] Where c represents the project, and e represents the domain knowledge point.

[0095] The initial representation of the project set is the integration of the domain knowledge point entities, and after propagation in the project domain knowledge graph, all different types of entities in the set are traversed, all of which can effectively expand the potential features of the project. The multi-layer propagation entity of the project will be recursively defined as follows:

[0096]

[0097] Where I represents the number of graph convolution layers, and the entity set is expanded to include the propagation triplets of head entities, relationships and tail entities, and the I-layer propagation of the project is defined as follows:

[0098]

[0099] The rich related domain knowledge contained in the project domain knowledge graph is classified and connected to the original propagation layer to better explore the project's domain knowledge preferences. Since the links in the project domain knowledge graph have classification characteristics, the entities in the traversal are sorted according to the relationship links:

[0100]

[0101] Where r represents the link to which the entity belongs in the project domain knowledge graph, and similarly, the reconnected propagation triplets of the project are defined as follows:

[0102]

[0103] Step 3.2, in the knowledge group aggregation module, the project domain feature propagation is divided into two stages: entity-granularity intra-group attention network and inter-group aggregation network based on relationship links, and the project domain features are constructed in more detail through these two stages.

[0104] In the intra-group attention network stage, the tail entities can be represented by different connections with the head entity; in the feature propagation process, different paths reaching the tail entity have different meanings. An attention network is used to determine the proportion of weights on different paths.

[0105] First, select a candidate project entity in the project domain knowledge graph Given the candidate project entity The tail domain knowledge entity e is calculated to obtain the attention weight of the tail domain knowledge entity relative to the project, wherein The representation of the tail entity t

[0106]

[0107] Wherein The representation of the tail entity t i .

[0108] After obtaining the attention weights of all entities in each group in the same layer, the embeddings of all weighted entities in each group are added to obtain the representation of these relationship links. Finally, the weighted representations of the k different domain groups are obtained:

[0109]

[0110] Wherein, p i represents the attention weight of the tail domain knowledge entity, represents all tail entities connected with the candidate project.

[0111] Step 3.2, in the inter-group aggregation network stage, the representation of different relationship links in each propagation layer is obtained through the head entity and the relationship link, which directly uses the relationship information in the project domain knowledge graph. Each group is considered as an extension of the project domain feature in a certain knowledge direction, and a inter-group aggregation network is used to continue to represent the domain features of the project.

[0112] There are k different group relationships in I layer, therefore, the relationship link aggregation embedding is constructed as follows, wherein, (o k , r k ) controls the attention weight of each group embedding, indicating the characteristics of the project domain a k , as follows:

[0113] a k = ξ (o k , r k ) o k

[0114] Then, a neural network similar to the attention mechanism is used to realize the function of the function, and the formula is as follows: ξ (o k , rk ) = Relu(W2(Relu(W1(o k ||r k )+b1))+b2)

[0115] where ReLU is a nonlinear activation function, W and b are trainable weight matrix and bias, || represents concatenation operation. Then the coefficients of the whole relation link are normalized using the softmax function as follows:

[0116]

[0117] Finally, the representation of the item at the l-th layer is taken as the domain feature model of the item, which is represented as follows:

[0118]

[0119] Step 4, through the multi-channel (time, domain, interaction) attention mechanism, the attention scores of different entities and relations in the user-item path graph are calculated through the multi-channel (time, domain, interaction) attention mechanism, the entity information contained in the whole user-item path graph is included in the item, then the attention scores are taken as the weights of different entities to distinguish the importance of different entities, the three-channel attention scores of each entity are calculated by combining the attention scores of the three channels, and the three-channel scores are connected to obtain the weight of each entity. Finally, all entity features are connected together as the features of the target item to establish the global model G1 of the item. In the domain knowledge model, the vector embedding of the entity in the user-item path graph is obtained, and the entity embeddings are combined to form the entity embedding matrix of the user-item path graph, as follows:

[0120] SG (u，i ) = [e1, e2,...e n ]

[0121] where e n represents an entity in the user-item path graph, n is the number of entities in the user-item path graph, and SG (u，i) represents the overall embedding matrix of the user-item path graph.

[0122] After obtaining the overall embedding of the user-item path graph, a multi-channel attention mechanism is used to assign different weights to the entities in the user-item path graph. The multi-channel attention mechanism consists of three channels, namely the interaction channel, the time channel and the domain channel.

[0123] In the interaction channel, the importance of the domain knowledge points to the target item needs to be captured through vector inner product, and different weights are assigned to the users according to their knowledge level, and the attention weight expression is as follows,

[0124]

[0125] where v j represents the embedding of the target candidate item. n represents the embedding of each knowledge point in the user-item path graph, represents the interaction weight of the knowledge point to the target item.

[0126] Finally, all directly interacting knowledge points of the target item are weighted according to the weighted sum of their interaction weight vectors, as follows,

[0127]

[0128] In the time channel, according to the interaction time of the domain knowledge point and the candidate item, a multi-layer perceptron is used to obtain the intrinsic weight of the directly interacting knowledge point, and the formula is as follows, where W i ∈R d×d , u i ∈R d and b1 are weight matrix and bias.

[0129] time i = U T tanh(W 1 e n + b l )

[0130]

[0131]

[0132] where W i ∈R d×d , U i ∈R d and b l are weight matrix and bias.

[0133] In the domain channel, the domain feature model of the item obtained in the item domain feature modeling is used to obtain the domain attention score of the knowledge point entity in the user-item path graph. Therefore, the attention score of the domain channel part is as follows:

[0134]

[0135] where s i is the output of the item domain feature model.

[0136] Then, the attention scores of the three channels are combined to obtain the item global model, where, represents the connection operation.

[0137]

[0138] Finally, all entity features in the user-item path graph are integrated, and the features of different entities are connected to obtain a global model G of the item l .

[0139] Step 5, based on the attention mechanism, a field feature model S of the item is obtained i and the global model G of the item is assigned different weights, and the two are fused to obtain a final item model I: l

[0140] I = W t tanh(W[S i , M i ])

[0141] Where tanh is an activation function, W t , W is a bias matrix;

[0142] Step 6, first, the review set of the user is preprocessed, that is, the words in the review set are converted into d-dimensional embeddings, by giving a query embedding matrix Q, the words in the user's review set are searched from the embedding matrix, and converted into corresponding embedding vectors, and finally the corresponding review matrix RC of the user is obtained, wherein the kth row represents the word k and its corresponding d-dimensional embedding, at the same time, the item field feature obtained in step 3 is converted into a matrix form to obtain an item feature matrix IC, after obtaining the user review matrix RC, the Dirichlet probability clustering algorithm is used to calculate the field type to which the user belongs, and then the field type to which the user belongs is also converted into a field type matrix SC.

[0143] The user set U = {u1, u2,..., u n}, the field set T = {t1, t2,..., t k}, and each user u in the user set U is regarded as a word sequence wi represents the ith word, and it is assumed that u has n words, and all different words involved in U form a large set S, S = {s1, s2,..., s i};

[0144] The user set U is input to the clustering algorithm (assuming that the clustering is k types, and j words are included in T):

[0145] ① For each user u n in U, the probability of corresponding to different groups is where represents the probability of u n corresponding to the kth type in T, and the calculation process is as follows:

[0146] ​

[0147] where represents u n the number of words of the T-th type in T, n is the total number of words in un; k

[0148] ② For each group T k in T, the probability of different words is generated where, represents T k the probability of the j-th word in T is generated, as follows:

[0149]

[0150] where represents the number of the j-th word in T k group T, N represents the number of all words in T k in T. The core formula of LDA is as follows:

[0151]

[0152] The probability of the occurrence of a word w in a user u is given by the current θu and φt. Where p(t|u) is calculated using θu, and p(w|t) is calculated using φt. By the current θu and φt, p(t|u) can be calculated when a word in the description of a user u corresponds to any one type Ti, and then according to these results to update the type Ti to which the word should correspond. At the same time, if the update changes the type Ti to which the word corresponds, it will also affect θu and φt in turn.

[0153] The user categories contained by a user u n are obtained by the Dirichlet clustering algorithm as follows:

[0154]

[0155] where, is the different user categories contained by the user, p i is the probability weight of belonging to different categories, i = 1, 2, 3,..., p n T k represents the weight of the target user's preference for a certain type of resource.

[0156] ​Step 7, multiply the user review matrix RC obtained in step 7 with the item feature matrix IC to obtain the user-item field matrix UIS, and multiply the user review matrix RC with the user type matrix SC to obtain the user-item type matrix UIT. Then, use the convolutional neural network to extract the user field features U t and user type features U c from the user-item field matrix UIS and the user-item type matrix UIT, respectively. The specific process is as follows:

[0157] ① Multi-type CNN: After clustering to obtain different types of users, each type is expressed by a matrix. Multiply the user review matrix RCi with the user type matrix SC to obtain different user review type matrices. The specific process is as follows:

[0158]

[0159] A user belongs to multiple different types. The review matrix of the target user can be regarded as a tensor, i.e., the review feature tensor is composed of multiple review type matrices. In the convolution process, the number of filters is consistent with the number of types to which the user belongs. By sliding the filters in the user's multiple review type matrices, the feature information in different review types is extracted. Finally, all feature information is integrated together. The feature result of review type k is as follows:

[0160]

[0161] wherein, is the type matrix, is the Relu activation function, [i:i+h-1] is the receptive field, Fk is the filter, and the final convolution result is as follows:

[0162]

[0163] ② Max-pooling operation: After the convolution operation, each review type contains multiple feature matrices generated by filters. At this time, it is not possible to obtain the significant features of the target user from them. Therefore, a max-pooling operation is used to obtain the significant features. The user type feature model is represented as follows:

[0164] U c ={max(c 1 ), max(c 2 ),..., max(c n )}

[0165] The user field feature model modeling process is the same as the user type feature model.

[0166] Based on the attention mechanism, the user field features Ut and user type feature U c The weights are connected and fused to obtain the user feature model U:

[0167] U=W t tanh(W[U t , U c ])

[0168] Wherein, tanh is an activation function, W t , W is a bias matrix;

[0169] Step 8, connect the item model I and the user feature model U, and take the connected model as the input of the multi-layer deep neural network to obtain the target item and the interaction feature with the user. The interaction feature is taken as the initial recommendation score ps of the candidate item, and the calculation process is as follows:

[0170]

[0171] Wherein, is a weight matrix, b l is the bias of the lth layer of the neural network, [U, I] is the connection of the user feature model U and the item model I, and l is the number of layers of the neural network model;

[0172] Step 9, directly add the user type obtained in step 7 to the interaction feature of the item model I and the user feature model U as the final recommendation score f′s of the candidate item. The adding process is as follows:

[0173] First, the multiple categories to which the learner belongs are obtained through the LDA clustering algorithm. The belonging category of the learner and the probability of the belonging category are subjected to dot product operation, and the calculation process is as follows:

[0174]

[0175]

[0176] Wherein, r u，t indicates the user's preference for a certain field, and T k indicates the field contained by the user; meanwhile, the is standardized in the range of 0 to 1. The prediction score of the candidate item can be defined according to the following equation:

[0177]

[0178] Step 10, sort the recommendation scores f′s according to the high and low, and recommend the first N candidate items with the highest scores to the user in turn.

[0179] In addition, the application can also provide a computer device, comprising a processor and a memory, the memory is used to store a computer executable program, the processor reads part or all of the computer executable program from the memory and executes, and the processor can realize the learning resource recommendation method based on the project field knowledge and the user comment features when executing part or all of the computer executable program.

[0180] In another aspect, the application provides a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program can realize the learning resource recommendation method based on the project field knowledge and the user comment features when executed by a processor.

[0181] The computer device can be a notebook computer, a desktop computer or a workstation.

[0182] The processor can be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC) or a ready programmable gate array (FPGA).

[0183] For the memory of the application, it can be an internal storage unit of a notebook computer, a desktop computer or a workstation, such as a memory, a hard disk, or an external storage unit, such as a mobile hard disk or a flash card.

[0184] The computer readable storage medium can include a computer storage medium and a communication medium. The computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information such as computer readable instructions, data structures, program modules or other data. The computer readable storage medium can include a read-only memory (ROM), a random access memory (RAM), a solid state disk (SSD) or an optical disk. Among them, the random access memory can include a resistance random access memory (ReRAM) and a dynamic random access memory (DRAM).

[0185] The above is only a preferred embodiment of the application, but the protection scope of the application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the application within the technical range disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A learning resource recommendation method based on project domain knowledge and user reviews, characterized by, Comprising the following steps: Step 1, obtain the project domain feature information, user historical interaction information matrix Y, user review set information, query embedding matrix Q, and user feature information; Step 2, according to the user historical interaction information matrix Y, construct the user-project path graph SG; Step 3, using a domain grouping-based graph convolutional neural network, a project domain feature model is constructed through project domain feature information ; Specifically as follows: The project domain feature model is divided into two parts: group propagation and knowledge group aggregation; The domain knowledge points contained in the project are expanded along the path in the project domain knowledge graph, and each propagated set is recursively accessed; Step 3.1, in the group propagation, the characteristics of the project are reflected in the domain knowledge points it contains, and the characteristics of the project are represented by mining the characteristics of the knowledge points and using the domain knowledge points to represent the project, which is defined as follows: Wherein, c represents the project, and e represents the domain knowledge point; The initial representation of the project set is the integration of the domain knowledge point entities, after propagation in the project domain knowledge graph, all different types of entities in the set are traversed, and all entities can effectively expand the potential characteristics of the project, and the multi-layer propagation entity of the project is recursively defined as follows: wherein, l represents the hierarchical distance between the entity set and the item domain knowledge point seed; Step 3.2, knowledge group aggregation, the domain feature propagation of the project is divided into two stages: the first stage is to use the intra-group attention network with entity granularity to determine the attention weight of the entity on different paths, and the second stage is to use the inter-group aggregation network based on the relationship link to represent the domain feature of the project, through the two stages, the domain feature of the project is more detailed model construction; Step 3.2.1, in the intra-group attention network stage, the tail entity is represented by different connections with the head entity, and in the feature propagation process, different paths reaching the tail entity have different meanings, and an attention network is used to determine the weight proportion of different paths; First, a candidate project entity in the project domain knowledge graph is selected , and the candidate project entity is given , the tail domain knowledge entity e is calculated, and the attention weight of the tail domain knowledge entity relative to the project is calculated, wherein is the representation of the tail entity , as follows: wherein is a representation of a tail entity is a representation of a tail entity After obtaining the attention weights of all entities in each group in the same layer, the weighted entities in each group are added as the representation of these relationship links, and finally, the weighted representation of the k different domain groups is obtained: wherein, attention weights representing tail domain knowledge entities, representing all tail entities connected with the candidate item; Step 3.2.2, in the inter-group aggregation network stage, the representation of different relationship links in each propagation layer is obtained through the head entity and the relationship link, and the relationship information in the project domain knowledge graph is directly used; Each group is considered as an extension of the project domain feature in a certain knowledge direction, and the inter-group aggregation network is used to continue to represent the domain feature of the project, and then the domain feature model is obtained; Step 4, calculate the attention scores of different entities in the user-item path graph through a multi-channel attention mechanism, and then add the attention scores calculated by different channels to obtain the attention weights of different entities in the graph; integrate the features of all entities in the user-item path graph, and connect the features of different entities to obtain a global model of the item ​ Step 5, based on attention mechanism, project field feature model with the project global model Assign different weights, fuse them, and get the final project model I: Step 6, preprocess the user's review set, i.e. convert the words in the review set into d-dimensional embeddings, find the words in the user's review set from the embedding matrix by giving a query embedding matrix Q, and convert them into corresponding embedding vectors to obtain the user's corresponding review matrix RC, convert the project domain feature obtained in step 3 into a matrix form to obtain the project feature matrix IC; After obtaining the user review matrix RC, the user's domain type is calculated by using the Dirichlet probability clustering algorithm, and then the user's domain type is also converted into a domain type matrix SC; Step 7, multiplying the user's review matrix RC with the item characteristic matrix IC to obtain a user-item field matrix UIS, multiplying the user's review matrix RC with the user's type matrix SC to obtain a user-item type matrix UIT, using a convolutional neural network to extract user field features and user type features from the user-item field matrix UIS and the user-item type matrix UIT respectively and user type features ​ Based on the attention mechanism, calculate the user domain feature and the user type feature The weights are occupied, and the two are connected and fused to obtain a user feature model U; Step 8, connect the item model I with the user feature model U, take the connected model as the input of the multi-layer deep neural network, get the target item and the interaction feature with the user, and take the interaction feature as the initial recommendation score ps of the candidate item; Step 9, directly add the field type to which the user belongs obtained in step 7 to the interaction feature of the item model I and the user feature model U as the final recommendation score f's of the candidate item; Step 10, sort the recommendation scores f's in descending order, and recommend the top N candidate items with the highest scores to the user in turn. 2.The learning resource recommendation method based on project domain knowledge and user reviews according to claim 1, characterized in that, In step 1, the item field feature information refers to detailed description information of knowledge points, categories and release time contained by the item; the user review set information includes a user review set of all items; the user historical interaction information matrix Y refers to an implicit interaction matrix of a user and an item, and if a user and an item have historically interacted, Y ij is set to 1, otherwise 0; the query embedding matrix Q refers to a matrix including existing word information, and any word in the review set of the learner can be found in Q. 3.The learning resource recommendation method based on project domain knowledge and user reviews according to claim 1, characterized in that, In step 2, the specific process of constructing the user-item path graph is as follows: Step 2.1, sample the path between the user and the item through the historical interaction information of the user and the item, after sampling is completed, select K paths, limit the length of each path to 6, in the specific sampling process, start from the target user, end at the item, and perform random walk with path length 6, finally, get k paths between the user and the target item; Step 2.2, after obtaining the k paths, splice the paths to form a user-item path graph with the starting point being the target user and the ending point being the candidate item, and the maximum path length of the k paths being 6, the user-item path graph contains multiple entities. 4.The learning resource recommendation method based on project domain knowledge and user reviews according to claim 1, characterized in that, Step 4 is specifically: Calculate the attention score of different entities and relationships in the user-item path graph through the multi-channel attention mechanism, integrate the entity information contained in the entire user-item path graph into the item, then take the attention score as the weight of different entities to distinguish the importance of different entities, and finally integrate all entities into the feature of the target item; In the domain knowledge model, obtain the vector embedding of the entity in the user-item path graph, combine the entity embeddings to form an entity embedding matrix of the user-item path graph; After obtaining the overall embedding matrix of the user-item path graph, use the multi-channel attention mechanism to assign different weights to the entities in the user-item path graph, the multi-channel attention mechanism consists of three channels, which are the interaction channel, the time channel and the domain channel; In the interaction channel, capture the importance of the domain knowledge points to the target item through vector inner product, assign different attention scores to the knowledge level of the interacting user, and the attention score expression is as follows, wherein, represents an embedding of a target candidate item, represents an embedding of each knowledge point in a user-item path graph, represents an interaction weight of a knowledge point to a target item; Finally, weight the sum of all direct interaction knowledge points of the target item according to the vector of the interaction weight, as follows, In the time channel, according to the interaction time of the domain knowledge point and the candidate item, use a multi-layer perceptron to obtain the attention score of the directly interacting knowledge point, and the formula is as follows: wherein, , represents a weight matrix, represents a bias, In the domain channel, use the item domain feature model obtained in the item domain feature modeling to obtain the domain attention score of the knowledge point entity in the user-item path graph, and the attention score of the domain channel part is as follows: wherein, is the output of the item domain feature model, Finally, the three-channel attention scores are combined to calculate the three-channel attention score of each entity, and the three-channel scores are concatenated to obtain the item global model wherein, denotes a concatenation operation, 。 5.The learning resource recommendation method based on project domain knowledge and user reviews according to claim 1, characterized in that, The clustering process in step 6 is specifically as follows: User set , Domain set , Each user u in user set U is treated as a sequence of words , wi represents the ith word, let u have n words, all different words involved in U form a large set S, ; The user set U is input to the clustering algorithm, and is clustered into k types, and T contains j words in total: , the probability of corresponding to different groups , where represents the probability of corresponding to the kth type in T, the calculation process is shown as follows:​ in express The middle corresponds to the first in T. The number of words of each type, where n is the total number of words in un; ii. For each T, the population , the probability of generating different words where, denotes the probability of generating the jth word in T, as follows: where representing a group contains the number of times the jth word in T, N represents all words in T, the core formula of LDA is as follows: by the current u and t The probability of a word w occurring in a user u is given by u which is computed using φt, and p(w|t) is computed using φt by the current u and t, p(t|u) can be computed for any one of the topics Ti in the user u profile, and then the topic that the word should correspond to is updated based on the results, The user is obtained by the Dirichlet probability clustering algorithm described above The user categories contained are as follows: wherein, is a different user category that the user belongs to, denotes the probability weight of belonging to different categories, i = 1, 2, 3,..., represents the weight of the target user's preference for a certain type of learning resource. 6.The learning resource recommendation method based on project domain knowledge and user reviews according to claim 1, characterized in that, Step 7 is specifically as follows: ①Multi-type of CNN: After clustering the different types of users, each type is expressed by a matrix, and the user's review matrix RC i is multiplied by the user's type matrix SC to obtain the different user review type matrix; A user belongs to multiple different types, and the review matrix of the target user is regarded as a tensor. In the convolution process, the number of filters is consistent with the number of types to which the user belongs. The filter slides in the multiple review type matrices of the user to extract feature information in different review types, and finally all the feature information is integrated together to obtain the feature result of the review type k. ②Max-pooling operation: After the convolution operation, each review type contains multiple feature matrices generated by filters. A max-pooling operation is used to obtain the significant features. 7.The learning resource recommendation method based on project domain knowledge and user reviews according to claim 1, characterized in that, Step 9 is specifically as follows: First, the LDA clustering algorithm is used to obtain multiple categories to which the learner belongs. The dot product operation is performed on the categories to which the learner belongs and the probabilities of the categories to which the learner belongs to calculate the process as follows: where, refers to the user's domain of interest, refers to the user's preference for a certain domain, refers to the domains contained by the user, and standardized in the range of 0 to 1, the predicted score of the candidate item is as follows: wherein, is the initial recommendation score.

8. A computer device, comprising: The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the learning resource recommendation method based on project domain knowledge and user reviews according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the learning resource recommendation method based on project domain knowledge and user reviews according to any one of claims 1-7.

Images