MOOC Recommendation Research On Series And Similarity Analysis

Since the first MOOC platform Coursera was built in 2012,MOOC has experienced rapid development.Though the number of registered learners is more than tens of millions,the dropout rate reaches 90%.Based on the drop-out attribution by MOOC situation analysis since 2012,Recommendation finds proper learning content to interfere with the learning process.But MOOC learning is different from the traditional product purchas-ing.It has more obvious characteristics of feature including complexity,data sparsity,complicate interactions among behaviors,etc.Those characteristics have posted a great challenge on MOOC recommendation greatly.According to the learning theories,this thesis recommends on learning series modeling and measuring with consideration on var-ious distances.It supports collaborative learning and adaptive learning which are proved suitable modes for MOOC learning.Recommendation on them can reduce the negative learning experience that could cause drop-out.The research questions and contributions in this thesis can be summarized as follows:1.MOOC recommendation on series modeling and measuring in the learning process.Frustration was found one of the main reasons of drop-out,but existing MOOC recommendation considers mainly on the selection results of learners and ignores the satisfaction factor Also,most MOOC recommendation cannot change weights of different features in real-time.But the requirement of MOOC learners on different features changes with time.We propose a new recommendation method by weighting different features according to the real-time learning situation in the learning process.First,MOOC learning is related to many psychological factors including satisfaction,performance expectation,and ability belief,etc.Based on Expectancy-Value Theory,we propose an achievement and interest combined feature model for MOOC recommendation.In this way,less drop-out will be caused because of frustration.Second,different features of the learning series are mod-eled separately.And the learning situation is measured with learning intensity as the key of feature trade-off for recommendation to support adaptive learning.Thus,the recommendation result can meet the target learner's requirement on different features in real-time.Less drop-out will be caused for frustration.Next,various distances like knowledge distance are combined for enriching the recommendation results and improving the accuracy.Finally,experiments on the real-world dataset find the sparsity and accuracy are both improved.2 types of drop-out measurements are proposed both decreased by this recommendation method.2.locality sensitive learner grouping based recommendation for efficiency improvement.Group learning is one of the main forms of collaborative learning.In this way the lonely feeling of MOOC learning will be decreased.Existing research on learner grouping is mainly on homogeneity or heterogeneity between learners.Traditional recommendation calculates similarity between learners on the numbers of their learning content in common.But the similarity on local dimensions is ignored.Related similarity calculation and refreshment need to compare all the learner pairs.A multi-layer grouping solution is proposed.Similar learn-ers on diverse local dimensions meet in neighbor groups.Neighbor search is only among the groups that the target learner belongs to.The time cost of the recom-mendation is cut down greatly.Experiments on the realworld data find the recall of recommendation is doubled.It means increasing diversity of recommendation results.Great improvement on the time cost is also achieved especially on the refreshment of similarity which is cut down greatly and needs only one multi-layer function calculation?3.Learning content recommendation based on the analysis of learning series.Few MOOC recommendations consider on the sequence between learning content.They are mainly among uncovered items of the target learner.Existing prerequisite research mainly depends on the text match between the context and the knowledge base.It relies on the knowledge base referred.A new prerequisite coefficient calculation is proposed on feature values which can be measured more easily.Because knowledge maintenance decreases with time forgetting,the forgetting modified learning feature helps to improve the accuracy of recommendation results.On modified features,prerequisite and subsequence are recommended according to the failed and achieved prerequisite locating results on learning series.First,MOOC learning behavior series are located to get the points of prerequisite failure and prerequisite achieved.Second,prerequisite coefficient is calculated with forgetting curve combined.Finally,collaborative filtering for prerequisite and subsequence recommendation is trained on similar qualified learners' series.Experiments on the real-world data verify the accuracy improvement of both prerequisite and subse-quence recommendation.Except for the accuracy improvement,learning performance is also increased greatlyLoneliness,frustration and lack of prerequisite are found the main reasons causing drop-out.By supporting collaborative learning and adaptive learning,the recommenda-tions on series and multi-distances help to solve the problems and improve the learning performance:(1)The series recommendation considers on features related to "frustra-tion" and adaptively weights different feature to meet real-time requirement on features including "frustration";(2)The recommendation on similar learners is improved by locality sensitive grouping.Diverse groups facilitate more collaborative learning with less"loneliness";(3)Recommendations on prerequisite and subsequence fulfill a new adaptive learning on series location.That helps to avoid and solve the problem of "lack of prerequisite".In conclusion,this thesis combines learning mode and recommendation research and develops cross research between recommendation technology and education theory.The three recommendation algorithms on series and similarities form a system for MOOC recommendation.First,by modeling the features on learning series,the recommendation in learning process weights the features by real-time learning situation measuring.Furthermore,locality sensitive learner grouping helps to cut down the time cost on collaborative filtering recommendation.Finally,collaborative filtering is improved for prerequisite and subsequence recommendations on qualified neighbors' learning series.Various real-world datasets are employed for the experiments.Coursetalk is one of the biggest MOOC plat-forms.Crawled dataset of this platform is used for learner grouping to test time cost improvement of recommendation with massive learners.Mic-video Platform of East China Normal University tracks detailed learning process data.The dataset is adopted for exper-iments of adaptive learning process recommendation and prerequisite-related recommen-dation.The experiments demonstrate the outperformance of our recommendation methods on accuracy and time cost.Experiments on learning performance are also designed to verify the improvement of our recommendation on it.