Data Mining Techniques And Applications For Personalized Learning

Personalized learning aims to suggest appropriate learning recourses and activi-ties for learners to make up their knowledge structures,based on their individual dif-ferences.The suggestions should consider the cognitive levels and learning abilities of learners during the learning process.In recent years,many online learning systems show their proliferation,which not only breaks the restrictions of time and space in traditional class-based learning but also provides learners with rich learning resources(e.g.,exercise,course,and lecture).In fact,more and more learners participate in online learning activities,and meanwhile,leave a large amount of learning data.Obviously,such data contain huge scientific and market value,which can provide powerful sup-port for developing data-driven personalized learning services.Therefore,how to make use of data mining technologies to understand and analyze learners' learning data and achieve personalized learning has become a hot research topic in computers and related interdisciplinary subjects.Nowadays,although some methods based on related theories including cognitive psychology have achieved some successes,the research of person-alized learning in online scenarios still faces many difficulties in some issues including characterizing learning resources,capturing the complex and dynamic learning process,and quantifying learning strategies.To this end,according to the three main roles,i.e.,exercise resources,students,and learning strategies,in the online learning systems,this dissertation systematically carried out a series of exploratory research on data mining methods and applications for personalized learning.Specifically,for exercise resources,we propose deep semantic and structural representation methods,where the effects have been verified in the re-lated application tasks of two typical exercises including language questions and logic questions.For student users,we respectively propose two knowledge tracing models incorporating learning factors and question semantics.For learning strategies,we pro-pose a multi-objective personalized recommendation algorithm,which improves stu-dents' learning efficiency.In particular,all work relies on the domestic-leading online platform called "Zhixue" developed by iFLYTEK Co.,Ltd.Both the research issue and data are derived from actual application scenarios,and all solutions are verified on the real platform,showing practical application value.The main contributions of this dissertation can be summarized as follows.Firstly,we study the deep representation methods for exercise resources.On one hand,for language-based exercises,this dissertation proposes a novel exercise embedding method,called Test-aware Attention-based Convolutional Neural Network(TACNN),based on semantic understanding.Actually,semantic comprehension plays a key role in the representation of language-based exercises.However,traditional meth-ods usually rely on rule-based pattern matching,ignoring the semantic richness and de-pendence of exercise textual contents.Specifically,we decompose this type of exercise representation into two parts:sentence understanding and semantic correlation.We first utilize a convolutional neural network to extract the sentence-level features.Then,we adopt an attention mechanism to quantify the semantic dependency of exercise textural content on different questions.Finally,in the task of predicting difficulty attributes,we propose a pairwise learning method to eliminate the incomparability from different tests.Experimental results demonstrate that our method can effectively improve the accuracy and robustness of exercise difficulty prediction for English reading comprehension.On the other hand,for logic-based exercises,this dissertation proposes a novel exercise embedding method,called Neural Mathematical Solver(NMS),based on structural un-derstanding.Different from language-based exercises,logic-based exercises generally contain some specific structural elements like formulas.However,traditional methods usually treat them directly as text sequences for representation while ignoring their in-trinsic structural hierarchy.Therefore,we first develop an assistant tool to construct a formula dependency graph in a certain exercise,preserving formula-enriched structures.Then,we propose two formula graph networks to infer the structural representation con-sidering different strategies including node-level attention and edge-level attention.Fi-nally,we propose a novel nested sequential model to capture both linguistic semantics and structural correlations.We conduct extensive experiments on the task of automatic answering mathematical problems(a kind of typical logic-based exercises).Experi-mental results on a large-scale dataset show that the method can effectively capture the formula structures for the exercise presentations to improve the performance.Secondly,for student users,we study cognitive diagnosis methods in a dynamic learning process.On one hand,considering the impacts of two learning factors in-cluding knowledge connectivity and memory-forgetting on students' learning activities,this dissertation proposes a novel knowledge tracing model,called Exercise-correlated Knowledge Proficiency Tracing(EKPT).Specifically,we first project both students and exercises into a knowledge space with explicit meaning based on the exercise-knowledge association priors,and meanwhile,establishes the connectivity among exercises.Then,we jointly apply two classical educational theories including the learning curve and forgetting curve to capture the change of students' proficiency levels on knowledge con-cepts in the knowledge space.Experimental results show the improvement of its ac-curacy for cognitive diagnosis.On the other hand,considering the impacts of exercise semantics including common knowledge-level semantics and individual content-level semantics on students' learning activities,we propose a novel knowledge tracing frame-work,called Exercise-aware Knowledge Tracing(EKT),incorporating exercise seman-tics.Specifically,we first design a dynamic memory network to store the common information of knowledge concepts and characterize the change of students' knowledge states.Then,we propose a feature extractor to learn the individual information of tex-tual content in each exercise and fuse it into the state tracking process.Finally,we respectively propose two implementations based on two strategies Markov property and attention mechanism,which can predict students' performance in the future.Extensive experiments demonstrate that EKT can guarantee both accuracy and interpretability of the tracking learning process.Finally,in terms of learning mechanism designing,we propose a personalized rec-ommendation algorithm,called Deep Reinforcement Exercise recommender(DRE),considering multiple learning objectives for adaptively recommending exercises to stu-dents.Traditional methods commonly optimize the single objective,i.e.,recommending non-mastered exercises to address the immediate weakness of students,and however,ignore the collaborative effects of multiple learning objectives including Review&Ex-plore,Smoothness of difficulty and Engagement.To address this problem,we design different reward functions to quantify those learning objectives.Then we propose a deep reinforcement learning method to optimize the recommendation strategy by interacting with students during the learning process for finding the optimal recommendation re-sults.We conduct experiments on both offline and online scenarios,where both results show the effectiveness of our proposed algorithm.