| With the deep integration of big data,cloud computing,data mining,artificial intelligence,and other technologies with education,online education has broken the space-time constraints of traditional classroom education environments and entered a new stage of educational informatization.In particular,with the rise and development of Massive Open Online Courses(MOOCs)in 2012,MOOCs have attracted nearly1000 universities worldwide to offer courses and hundreds of millions of learners to register for learning.Although online education platforms have massive educational resources,they lack functions such as personalized path planning,effective learning organization,and accurate resource recommendation services,which leads them to face problems such as high dropout rates and low course completion rates.To this end,this thesis takes knowledge concepts,learners,and learning communities as research objects.Then focus on three aspects: how to learn the prerequisite relationships among knowledge concepts,construct a reasonable learning community,and recommend appropriate learning resources to the learning community.The specific research contents and contributions of this thesis are as follows:(1)Proposed a method for learning prerequisite relationships between knowledge concepts based on the MHAVGAE model.Aiming at the problems of time-consuming,costly,and unscalable on large-scale datasets when manually labeling prerequisite relationships between knowledge concepts,this thesis proposes a Multi-Head Attentional Variational Graph Auto-Encoders(MHAVGAE)model to learn prerequisite relationships between knowledge concepts.Specifically,this thesis first models learning resources(e.g.,courses)in online education with the knowledge concepts(e.g.,data models)it contains and the relationships between them as a resource concept graph.Then this thesis proposes a node-level attention mechanism to calculate the attention weights between nodes(learning resources or knowledge concepts)and aggregates the neighboring features of nodes to obtain their potential representation;designs a gated fusion mechanism to fuses the potential representations of learning resources in the resource graph to the knowledge concepts in the concept graph;and reconstructs the features of knowledge concepts and the concept graph to learn the unlabeled prerequisite relationships between knowledge concepts.In addition,to reduce the over-reliance on manual labeling in the training process,this thesis proposes a Resource Prerequisite Reference Distance(RPRD)metric to generate inaccurate prerequisite relationships between knowledge concepts as training samples to extend MHAVGAE by weakly supervised settings.Finally,experiments were conducted on the MHAVGAE model and its extensions on three real datasets.The results show that the MHAVGAE model achieves the evaluation metrics of over 0.72 on ACC,0.74 on F1,0.84 on AP,and 0.82 on AUC,and is superior to current advanced models in multiple evaluation metrics;At the same time,the expansion of MHAVGAE also achieved more than 70% of results of the MHAVGAE model.(2)Proposed a prerequisite relationship driven fair Learning community construction method.Aiming at the problems of unfairness when constructing learning communities in existing studies,this thesis proposes a Prerequisite-Driven Fair Clustering algorithm(PDFC)to achieve fair construction of learning communities.Specifically,This thesis first model entities such as learners,courses,videos,and knowledge concepts in online education and their relationships as heterogeneous information networks(HINs).Then designs meta-paths and prerequisite meta-paths associated with learners(target nodes)in HINs and define the structural constraint,introduce the balance constraint of learners on sensitive attributes,and define the fair clustering as a bi-objective minimization problem under two constraints to ensure fairness construction of the learning community.Next,potential embeddings representation of learners in HINs are learned via an attention-based embedding learning model and a graph convolutional network model under structural and balance constraints,and the performing Cholesky decomposition to obtain orthogonal embeddings representation of learners in terms of structure and attributes.Next,fusing the orthogonal embeddings representation of learners under structure and attributes,converting from a bi-objective minimization to a single-objective minimization problem,and computing the top k feature vectors of the fused orthogonal embeddings of learners and input to the k-means algorithm for clustering to achieve fair construction of learning communities.In addition,this thesis proposes a dynamic update strategy for the adjacency matrix to extend PDFC for the dynamically fair construction of learning communities.Finally,experiments were conducted on PDFC and its extensions on three datasets,and the results showed that over half of the fairness evaluation metrics of PDFC under two constraints were better than the comparison method;At the same time,the extension of PDFC can also dynamically capture changes in clustering results.(3)Proposed a prerequisite relations enhanced attentional learning community recommendation method.Aiming at the problem of "information overload and knowledge disorientation" when learners face massive learning resources in online education,this thesis proposes a Prerequisite Relationship-Enhanced Attention Group Recommendation(PREAGR)method for recommending appropriate knowledge concepts to the learning community.This thesis first model learners,courses,videos,and knowledge concepts in online education and their relationships as heterogeneous information networks(HINs),define the interactions between learners with knowledge concepts on meta-paths and prerequisite meta-paths in heterogeneous information networks and analyze the comparison of the interaction between learners,learning communities,and knowledge concepts before and after the introduction of prerequisite relationships between knowledge concepts.Then propose a Path-Aware Attention Embedding(PAAE)method to capture learners’ preference information on different types of paths based on meta-paths and preconditioned meta-paths and design a gated fusion mechanism to fuse learners’ preferences into integrated preferences and define a loss function between learners and knowledge concepts.Next,design preference aggregators to aggregate learners’ preferences for learning community preferences,define loss functions between learning communities and knowledge concepts,and combine loss functions of learners and knowledge concepts for co-training optimization.Finally,experiments were conducted on the PREAGR method on two datasets,and the results showed that the PREAGR method outperformed most comparative methods in evaluation metrics HR@N and NDCG@N(N=5,10,20),which proves its effectiveness. |
