A personalized group recommendation method fusing student behavior information

Abstract

The present application relates to the technical field of group recommendation, in particular to a personalized group recommendation method fusing student behavior information, which comprises the following steps: 1) data preprocessing; 2) student multi-behavior modeling; 3) student preference and item structure acquisition; 4) group consensus modeling; 5) model prediction and optimization. The present application utilizes the behavior information of students and designs a group consensus strategy to provide effective learning resources for group cooperative learning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of group recommendation, in particular, to a personalized group recommendation method fusing student behavior information. BACKGROUND

[0002] The rapid development of online education makes students and classes learn and teach online, which promotes online course platforms to store more and more educational data. How to reasonably use these data resources to promote students or groups to learn more efficiently becomes a challenge that needs to be solved urgently.

[0003] The huge educational information resources are easy to make students lose direction, so how to fully utilize the student's behavior information, accurately capture the student's needs, and provide accurate recommendation services for students is an urgent need and challenge. In addition, in the current user learning process, the user's needs are becoming more and more diversified, and group cooperative learning is essential in education. How to recommend personalized learning resources to groups and improve the effectiveness of learning is an important challenge.

[0004] The group of online cooperative learning not only has very sparse group-project interaction, but also the needs of the group change with the learning state of the members, which poses great challenges to group recommendation methods. The personalized recommendation method can recommend resources for individual students, but it cannot be well recommended for groups, so it is necessary to fuse the behavior characteristics of students, construct group consensus, and design a group recommendation method to make group recommendation more effective. SUMMARY

[0005] The present application provides a personalized group recommendation method fusing student behavior information, which can utilize student behavior information and design group consensus strategies to provide effective learning resources for group cooperative learning.

[0006] According to the personalized group recommendation method fusing student behavior information, the following steps are included:

[0007] 1) data preprocessing;

[0008] 2) student multi-behavior modeling;

[0009] 3) student preference and project structure acquisition;

[0010] 4) group consensus modeling;

[0011] 5) model prediction and optimization.

[0012] As a preferred, in step 1), the method of data preprocessing is:

[0013] The following operations are performed on the student learning data:

[0014] 1-1) The various behaviors K generated by students in the learning process are classified and processed;

[0015] 1-2) The interaction of students and videos, as well as other information related to students and videos, will be constructed as auxiliary information into a knowledge graph G = { (U, V, r) | U E, V E, r R} representing the relationship r connecting entities , r, )| , E, r R} represents the relationship r connecting entities and , where E, R represent the set of entities and the set of relationships, respectively; students U and projects V are matched with an entity in the knowledge graph, i.e. U E, V E; since the interaction of users and projects is divided into multiple behaviors, there are also multiple subgraphs in the knowledge graph , where users are composed of students and projects are composed of videos;

[0016] 1-3) Similar to users, the interaction of groups and videos is constructed as a knowledge graph , which is different from users in that there is only one interaction behavior in the group.

[0017] As a preferred, in step 2), the method for modeling the multiple behaviors of students is:

[0018] The interaction of students with videos under a specific behavior is abstracted as a relationship in the interaction subgraph , behavior relationship matches a relationship in the knowledge graph, i.e. R; the following operations are performed:

[0019] 2-1) Initialize the entities and relationships of each subgraph to obtain the embedding representations e and r of the entities and relationships; then model the specific behavior relationship of the student;

[0020] For users, the specific behavior relationship is aggregated by aggregating the projects it interacts with, i.e.:

[0021]

[0022]

[0023] where is the behavior relationship embedding of the student, represents the number of projects that the student interacts with under behavior k, is the embedding of the project, is a learnable weight coefficient, is a weight matrix, biasing vector;

[0024] 2-2) For the interaction under certain behavior, only the information that is different from other behaviors can provide a useful perspective for the final selection of students, that is, to alleviate the problem of over-smoothing; for this purpose, mutual information is used to minimize the mutual information between the representations of any two different behaviors to quantify their independence and provide different perspectives for the intention of the group; specifically represented as:

[0025]

[0026] where s(·) is a function that measures the correlation between any two behavior representations, which is set as a cosine similarity function here; τ is a temperature hyperparameter in the softmax function; for another behavior relationship embedding of the student, is a loss function that quantifies the independence between different behaviors.

[0027] As a preferred, in step 3), the method for obtaining the structure of student preference and item is:

[0028] 3-1) Since the student's interaction with the item under different behaviors is in the corresponding knowledge graph, in order to obtain the high-order implicit relationship between the user and the item, multi-layer information transmission is performed in the graph, and the first layer output will be used as the input of the next layer to obtain the embedding of the student and the item under each behavior, which is as follows:

[0029]

[0030]

[0031] where, and represent the embedding representation of the message passing to the student and the item under behavior k, the aggregation layer, which retains the unique features of the student or the item under each interaction behavior; for the student neighbor node under behavior k, which is represented as:

[0032] )

[0033] where, represent the adjacent item nodes connected with the student in the knowledge graph of behavior type k, represent the embedding representation of the adjacent item connected with the student ; finally, the node embedding of the student is in the form of aggregation as follows:

[0034] ​ =

[0035] wherein, is a weight matrix, is a bias vector; similar operations are performed for the item nodes as for the student nodes;

[0036] 3-2) Since different behaviors have different effects on user selection purposes, the neural network is trained to adaptively adjust the coefficients to aggregate students and items respectively, obtaining student preference embedding and item embedding representations enhanced by multiple behaviors, for further group consensus modeling:

[0037]

[0038]

[0039] wherein, is a weight matrix, is a bias vector, is an activation function; similar operations are performed for the item nodes as for the student nodes; after obtaining the embedding representations of students and items that aggregate behavior information, the last layer of each is taken as the final embedding representation of the student and the item, respectively, i.e. = , = .

[0040] As a preferred, in step 4), the method for group consensus modeling is:

[0041] 4-1) For the modeling of the consensus of the group, first, the multi-layer propagation function in the knowledge graph is used to aggregate entity information to obtain the general preference of the group, which is used to capture the similarity and cooperation relationship between groups, and obtain the consensus at the group level; Since the interaction between the group and the item is only a single type of interaction, the behavior modeling is represented by the interaction relationship , and similar operations are performed as for the embedding of the student under a single behavior:

[0042]

[0043] wherein, is the embedding representation of group g at layer, and the last layer is taken to model the consensus at the group level, i.e. = ;

[0044] 4-2) To obtain the group consensus at the member level, first project the embedding representation of each member into multiple dimensions, and aggregate the representations of different members using the underlying dependency relationship; the information aggregation layer is built on the basis of the attention neural mechanism; the explicit correlation score between the group member embeddings is represented as , which is calculated as follows:

[0045]

[0046] where, is the projection latent space, and is the transformation matrix corresponding to the s-th projection space, used for embedding projection between different member students and , d is the embedding dimension; then the softmax function is applied to ; then, the behavior embeddings of group members are connected by connecting the representations of different learning subspaces to recalibrate them as follows:

[0047]

[0048] where, represents the concatenation operation of vectors, is the transformation matrix, represents the member embedding representation after calibration, N is the number of members in the group g; then, the preference information of group members is aggregated to obtain the consensus preference of students, guided by the multiple behavior information of members and the interaction items of the group, which is calculated as follows:

[0049]

[0050]

[0051] where, is the importance score between the item and the behavior, is the interaction behavior of the group student to the item , the item is the interaction item of the group, if the student has not interacted with the item , then is the embedding representation of the group-item interaction relationship r, is the consensus at the member level;

[0052] Finally, the group-level consensus and the member-level consensus are added to model the consensus of the group, which is calculated as follows:

[0053]

[0054] As preferred, in step 5), the model prediction and optimization method is as follows:

[0055] In the training process of group-project interaction, since the student-project interaction also contains rich interaction information, the interaction information of the user is considered in the training process, and the group is trained in the same process; the preference score of the student and the group to the project is obtained by using the inner product of the embedding representation between the two; the user and the group are optimized by using a pairwise loss function, and mutual information is used for multi-behavior modeling; specifically, the loss function of the user is:

[0056] +

[0057] Wherein is the predicted preference score of the student to the selected positive item , is the predicted score of the user to the negative item which will not be selected, denotes the training set of the user.

[0058] The loss function of the group is:

[0059] +

[0060] Wherein is the predicted preference score of the group to the selected positive item , is the predicted score of the user to the negative item which will not be selected, denotes the training set of the group.

[0061] The beneficial effects of the present application are as follows:

[0062] (1) The present application structures the data in the education field through a knowledge graph, models the multi-behavior interaction between students and projects, recursively captures the high-order implicit relationship between multi-behaviors in the graph, mines the implicit needs and interest preferences of students, provides personalized learning resources for students, and effectively alleviates the cold start problem of users in the recommendation process.

[0063] (2) The present application obtains the student preference and the project structure, and uses the multi-behavior interaction information of the group members as a guide to explore the implicit interaction and cooperation information between the group members in multiple dimensions, and models the group consensus at the member level. In addition, the interaction between the group and the project is also embedded and propagated in the graph, obtaining the high-order cooperation relationship between the groups, and modeling the consensus at the group level. This will effectively recommend personalized learning resources for groups. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1A flow chart of a personalized group recommendation method fusing student behavior information in an embodiment. DETAILED DESCRIPTION

[0065] For further understanding of the present application, the application will be described in detail with reference to the drawings and embodiments. It should be understood that the embodiments are only used to explain but not to limit the present application.

[0066] EMBODIMENTS

[0067] As shown in the figure, the present embodiment provides a personalized group recommendation method fusing student behavior information, which comprises the following steps: Figure 1

[0068] 1) data preprocessing;

[0069] The method of data preprocessing is as follows:

[0070] The following operations are performed on the learning data of students:

[0071] 1-1) multiple behaviors K generated by students in the learning process are classified and processed; for example, if a student only clicks a video (project) or watches a video for less than 1 minute, the behavior is called a click behavior; if a student watches a video for more than 30 minutes, the interaction with the video is called a target behavior; if a student collects a video, it is called a collection behavior.

[0072] 1-2) the interaction between students and videos and other information related to students and videos will be constructed as auxiliary information into a knowledge graph G={ (U, r, V) | r∈R} representing that the relationship r connects the entity and , , ∈E, r∈R} represents that the relationship r connects the entity and , wherein E and R represent the set of entities and the set of relationships respectively; the student U and the project V are matched with an entity in the knowledge graph, i.e. U⊆E, V⊆E; since the interaction between the user and the video is divided into multiple behaviors, multiple sub-graphs are also correspondingly matched in the knowledge graph; wherein the user is composed of students, and the project is composed of videos;

[0073] 1-3) similar to the user, the interaction between the group and the video is constructed into a knowledge graph , which is different from the user in that there is only one interaction behavior in the group.

[0074] 2) student multi-behavior modeling;

[0075] The method of student multi-behavior modeling is as follows:

[0076] ​The interaction of the student with the video under a specific behavior abstracts to a relation in the interaction subgraph , the behavior relation matches a relation in the knowledge graph, i.e. R; the following operations are performed:

[0077] 2-1) initialize the entities and relations of each subgraph to obtain the embedding representations e and r of the entities and relations; then model the specific behavior relation of the student;

[0078] For the user, the specific behavior relation is aggregated by aggregating the items interacted with, i.e.

[0079]

[0080]

[0081] wherein is the behavior relation embedding of the student, denotes the number of items interacted by the student under behavior k, is the embedding of the item, is a learnable weight coefficient, is a weight matrix, is a bias vector;

[0082] 2-2) for the interaction under a specific behavior, only the information that is different from other behaviors can provide a useful perspective for the final selection of the student, i.e. to alleviate the problem of over-smoothing; for this purpose, mutual information is used to minimize the mutual information between the representations of any two different behaviors to quantify their independence and provide different perspectives for the intention of the group; specifically represented as:

[0083]

[0084] wherein s(·) is a function for measuring the correlation of any two behavior representations, which is set as a cosine similarity function here; τ is a temperature hyperparameter in the softmax function; is another behavior relation embedding of the student, is a loss function for quantifying the independence between different behaviors.

[0085] 3) student preference and item structure acquisition;

[0086] The method for student preference and item structure acquisition is:

[0087] ​3-1) Since the student's interaction with the project under different behaviors is in the corresponding knowledge graph, to obtain the high-order implicit relationship between the user and the project, multi-layer information transmission is performed in the graph, and the embedding of the student and the project under each behavior is obtained, as follows: The output of the layer will be the input of the next layer to obtain the embedding of the student and the project under each behavior, as follows:

[0088]

[0089]

[0090] where, and represent the embedding representation of the message passing to the student and the project under behavior k, the aggregation layer, which retains the unique features of the student or the project under each interaction behavior; the aggregation information representation of the neighbor nodes of the student under behavior k, which is represented as:

[0091] )

[0092] where, represents the adjacent project node connected with the student in the knowledge graph whose behavior type is k; represents the embedding representation of the adjacent project connected with the student ; finally, the node embedding of the student is in the form of aggregation as follows:

[0093] =

[0094] where, is the weight vector, is the bias vector; similar operations are performed for the project node as for the student node;

[0095] 3-2) Since different behaviors have different effects on user selection purposes, adaptive coefficients are trained by neural network to aggregate the student and the project respectively, to obtain the student preference embedding and the project embedding representation enhanced by multiple behaviors, which are used for further group consensus modeling:

[0096]

[0097]

[0098] where, is the weight matrix, is the bias vector, ​is the activation function; for item nodes, similar operations are performed as for student nodes; after obtaining the embedding representations of students and items that aggregate behavior information, the last layer of each is taken as the final embedding representation of students and items, respectively, i.e. = , = .

[0099] 4) Group consensus modeling;

[0100] The method of group consensus modeling is:

[0101] 4-1) For the modeling of group consensus, first, the multi-layer propagation function in the knowledge graph is used to aggregate entity information to obtain the general preference of the group, which is used to capture the similarity and collaborative relationship between groups, and obtain the group-level consensus; since the interaction between the group and the project is only a single type of interaction, the behavior modeling is represented by the interaction relationship , and similar operations are performed to obtain the embedding of the student under a single behavior:

[0102]

[0103] wherein, is the embedding representation of group g at layer , and the last layer is taken to model the group-level consensus, i.e. = ;

[0104] 4-2) To obtain the group consensus at the member level, first, the embedding representation of each member is projected to multiple dimensions, and the representations of different members are aggregated using the underlying dependency relationship; the information aggregation layer is based on the attention neural mechanism; the explicit correlation score between the group member embeddings is represented as , which is calculated as follows:

[0105]

[0106] wherein, is the projection latent space, and are transformation matrices corresponding to the s-th projection space, used for embedding projection between different member students and , and d is the embedding dimension; then, the softmax function is applied to ; then, the behavior embeddings of group members are recalibrated by connecting the representations of different learning subspaces through the following operation connection:

[0107]

[0108] wherein, represents the concatenation operation of vectors, is the transformed matrix, is the member embedding representation after calibration, N is the number of members in group g; then, guided by the information of the behaviors of the members and the interaction projects of the group, the preference information of the group members is aggregated to obtain the consensus preference of the students, which is calculated as follows:

[0109]

[0110]

[0111] wherein, is the importance score between the project and the behavior, is the interaction behavior of the students in the group to the project , the interaction project of the group is if the student has not interacted with the project , then is the embedding representation of the group-project interaction relationship r, is the consensus at the member level;

[0112] Finally, the consensus at the group level is added to the consensus at the member level to model the consensus of the group, which is calculated as follows:

[0113]

[0114] 5) Model prediction and optimization.

[0115] The method of model prediction and optimization is:

[0116] In the training process of group-project interaction, since the student-project interaction also contains rich interaction information, the interaction information of the user is considered in the training process, and the user is trained in the same process as the group; the preference score of the student and the group to the project is obtained by using the inner product of the embedding representation between the two; mutual information is used for multi-behavior modeling; specifically, the loss function of the user is:

[0117] +

[0118] wherein is the predicted preference score of the student to the positive item selected, is the predicted score of the user to the negative item not to be selected, denotes the training set of the user;

[0119] The loss function of the group is:

[0120] +

[0121] wherein is the predicted preference score for a positive item selected by the group is the predicted score for a negative item not selected by the user denotes the training set of groups.

[0122] The above description of the application and its embodiments is illustrative and not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the application, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it, without departing from the purpose of the application, similar structural modes and embodiments can be designed without creativity, which should belong to the protection scope of the application.​​

Claims

1. A personalized group recommendation method integrating student behavior information, characterized in that: Includes the following steps: 1) Data preprocessing; 2) Modeling multiple student behaviors; 3) Student preferences and project structure acquisition; 4) Group consensus modeling; 5) Model prediction and optimization; In step 1), the data preprocessing method is as follows: Perform the following operations on the students' learning data: 1-1) Classify and process the various behaviors K generated by students during the learning process; 1-2) The interaction between students and videos, as well as other information related to students and videos, will be used as auxiliary information to construct a knowledge graph G={( ,r, )| , ∈E, r∈R} indicates that relation r connects entities and Here, E and R represent the set of entities and the set of relations, respectively; student U and project V are both matched with an entity in the knowledge graph, i.e., U⊆ E, V⊆ E; since user-project interactions are divided into multiple behaviors, there are corresponding multiple subgraphs in the knowledge graph. The users are students, and the project consists of videos; 1-3) For the interaction between groups and videos, construct it as a knowledge graph. Unlike users, there is only one type of interaction in a group; In step 2), the method for modeling multiple student behaviors is as follows: The interaction between students and the video under specific behaviors can be abstracted as a relationship in the interaction subgraph. behavioral relationships Matched with a relation in the knowledge graph, namely ⊆ R; Perform the following operations: 2-1) Initialize the entities and relations of each subgraph to obtain the embedded representations e and r of the entities and relations; then model the specific behavioral relations of the students. For users, specific behavioral relationships are derived by aggregating the items they interact with, that is: ； ； in Embedding student behavior relationships Students Projects interacting under behavior k Number, For project embedding, These are learnable weight coefficients. This is the weight matrix. It is the bias vector; 2-2) For interactions under specific behaviors, only by capturing information that differs from other behaviors can a useful perspective be provided for the student's final choice, thus mitigating the problem of oversmoothing. To this end, mutual information is used to minimize the mutual information between the representations of any two different behaviors, quantifying their independence and providing different perspectives on the group's intentions; specifically represented as: ； Where s(·) is a function that measures the association between any two behavioral representations, and here it is set as a cosine similarity function; τ is the temperature hyperparameter in the softmax function; Embedding another behavioral relationship for students. It is a loss function that quantifies the independence between different behaviors; In step 3), the method for obtaining student preferences and project structure is as follows: 3-1) Since students' interactions with projects under different behaviors are all within the corresponding knowledge graphs, in order to obtain the implicit relationships between higher-level users and projects, multi-layered information transfer is performed in the graph. The output of the first layer will be used as the input to the next layer to obtain the embeddings of students and items for each behavior, as follows: ； ； in, and Let these represent the embedded representations of message delivery to students and projects under behavior k, respectively. The aggregation layer preserves the unique characteristics of each student or project under each interaction behavior; For students The aggregated information representation of neighboring nodes under behavior k is as follows: )； in, This indicates the relationship between students in the knowledge graph of behavior type k. Connecting adjacent project nodes, Indicates to students The embedded representation of connected adjacent items; finally, the student's node embedding is obtained through... The aggregation form is as follows: = ； in, Let j be the weight matrix. This is the bias vector; similar operations are performed on project nodes as on student nodes. 3-2) Since different behaviors have different effects on users' choice of purpose, neural networks are used to train adaptive coefficients. We aggregate students and projects separately to obtain multi-behavior-enhanced student preference embeddings and project embeddings, which are then used for further group consensus modeling. ； ； in, This is the weight matrix. For bias vectors, The activation function is used; similar operations are performed on item nodes as on student nodes; after obtaining the embedded representations of students and items with aggregated behavioral information, the last layer of each is taken as the final embedded representation of students and items, respectively. = , = ; In step 4), the method for group consensus modeling is as follows: 4-1) For modeling group consensus, firstly, multi-layer propagation functions are used in the knowledge graph to aggregate entity information and obtain the group's general preferences, which are then used to capture the similarity and collaborative relationships between groups to obtain group-level consensus; since the interaction between groups and items is only of a single type, its behavior is modeled through interaction relationships. To indicate, to perform the operation: ； in, For group g in The layer embedding representation takes the last layer to model group-level consensus, i.e. = ; 4-2) To achieve member-level group consensus, the embedding representation of each member is first projected onto multiple dimensions, and the underlying dependencies are used to aggregate the representations of different members; the information aggregation layer is built on the basis of attentional neural mechanisms; the explicit correlation score between group member embeddings is represented as... Its calculation is as follows: ； in, For projection potential space, and The transformation matrix corresponding to the s-th projection space is used in different member students. and The embedding projection is then applied, where d is the embedding dimension; subsequently, the softmax function is applied to... Then, the behavioral embeddings of group members are recalibrated by connecting the representations of different learning subspaces using the following operation: ； in, This represents the vector concatenation operation. Let be the transformation matrix. Let N represent the calibrated member embedding, where N is the number of members in group g. Then, guided by the members' multi-behavioral information and group interaction items, the group members' preference information is aggregated to obtain the students' consensus preferences, which are calculated as follows: ； ； in, Score the importance of the project and the behavior. For group of students For the project Interactive behavior modeling, project For group interactive projects, if students have not interacted with the project before... ,but An embedded representation of the group-item interaction relationship r. This is a consensus at the member level; Finally, the group consensus is modeled by adding the group-level consensus to the member-level consensus, and the calculation is as follows: ； In step 5), the method for model prediction and optimization is as follows: During the training process of group-project interactions, since student-project interactions also contain rich interaction information, user interaction information is considered during training and trained collaboratively with groups in the same process. The preference scores of students and groups for projects are obtained using the inner product of their embedding representations. Pairwise loss functions are used to optimize for users and groups, and mutual information is used for multi-behavior modeling. Specifically, the user's loss function is: + ； in For students' positive choices Predicted preference score For users to select negative options they do not want to select The predicted score This represents the user's training set; The loss function for the group is: + ； in Positive terms selected for pairs Predicted preference score For users to select negative options they do not want to select The predicted score The training set for the group.

Images