DBN-MCPR:Research On Personalized Recommendation Based On Deep Belief Networks For MOOC

With the booming development of internet technology and education big data,online learning platforms such as MOOC have sprung up.The curriculum resources of MOOC platform have been increased because of the network sharing and the digitization of the curriculum resources.In the face of countless,uneven quality of curriculum resources,it is easy for learners to face the dilemma of rich curriculum resources but difficult to choose resources,and then they got lost in the information ocean.How to help learners to quickly and accurately find suitable learning resources in explosive MOOC resources is a problem that needs to be solved urgently in the field of education big data.The personalized recommendation,as an effective way to solve information overload,provides an new idea for the personalized resource recommendation of the MOOC platform.With many excellent features such as simple calculation,good universality,and high recommendation accuracy,the recommendation based on collaborative filtering is currently the most popular and effective personalized resource recommendation technology.However,with the exponential growth of the amount of data for learners and educational resources under the MOOC environment,the traditional collaborative filtering recommendation technology does not perform well when dealing with sparse data and cold start.Furthermore,the truth that the recommended content is duplicative and that the high dimensionality and non-linear data of online learning users cannot be effectively handled both lead to inefficient resource recommendation.With the advent of the third upsurge in the field of artificial intelligence,deep learning technology is increasingly favored by researchers.Compared with the traditional recommendation method,its efficient performance in function approximation,feature extraction and classification prediction can effectively solve some problems,for instance,the problems of similarity calculation method is too single,sparseness of data,inability to characterize high-dimensional attributes,and the inability of potential interest preferences of users can be exploited.In a word,deep learning technology provides a new way for the recommendation of course resources for online learning platforms.In order to enhance the learner's efficiency and enthusiasm,this paper combines deep learning techniques to explore personalized recommendations and then comprehensive analysis and excavation of demographic characteristics,user behavioral interest preferences,and course content attributes can be achieved.The main research is as follows:① Consider that the calculation of similarity applied to the traditional collaborative filtering recommendation method is too single;it is impossible to deeply mine the potential constraint relationship between the learner and the curriculum resources.This paper proposes a recommendation method based on DBN classification instead of the traditional similarity calculation method,and it adopts DBN's efficient feature abstraction and feature extraction capabilities to fully mine learners' interest in curriculum resources.Moreover,for the traditional collaborative filtering recommendation methods,there are widespread problems of cold start and data sparseness.The thesis deeply mines the learner demographic characteristics and curriculum resource content attribute characteristics and then combines learner behavior characteristics together to build a learner interest model based on the DBN,which effectively solves the problems of cold start and data sparseness,and the inaccurate expression of interest preferences of learners on curriculum resources.② In view of the complexity and imperfection of learners' extensive learning behaviors in the MOOC environment,the existing technical specifications can not fully record learners' behavioral information.This paper maintains a criterion to record learner's behavior data based on XAPI.It records the learner's learning behavior in the form of learning activity flow and uses common verbs suitable for the MOOC platform to record learner learning activities,thus achieves a complete record of learning behavior characteristics for MOOC platform and provides a broader range of behavioral features for the construction of learner interest models.Besides,Aiming at the phenomena of DBN-MCPR slow convergence,over-fitting,and long training period caused by too many parameters and excessive model scale in the training process of DBN classification model,this paper presentes a GPU-based parallel acceleration training model based on Theano to accelerate the training cycle and convergence speed of the model.③ The initial values and parameter setting rules of each parameter of DBN-MCPR were optimized,and experimental analysis was conducted by using the public data set MovieLens and the learner elective data of the starC course MOOC platform of Central China Normal University.Firstly,the published data set was used to verify the classification accuracy of DBN-MCPR,and compared with several traditional recommendation methods to highlight its excellent recommendation accuracy;secondly,utilize DBN-MCPR to train the real data set of the MOOC platform in the starC course to excavate some recommended rules between learners and curriculum resources;then come true the learner's prediction score for the course and verify the impact of training data set,learner feature vector,and GPU acceleration on the performance of DBN-MCPR.Finally,experimental results show that DBN-MCPR has better recommendation efficiency and faster convergence speed.