英文摘要 | With the increasing development of information technology in education field, the explosive growth of online teaching resources results in information overload. It makes teaching staffs and students hard to get useful information, and decreases their efficiency of work and study. Recommender systems, an effective method to overcome information overload, take active ways to filter information that users are interested in by analyzing users’ information and historical records. However, in the recommender systems for education filed, users’interests on teaching resources are changing and dynamic over time, which results from many factors, such as the development of courses, their knowledge levels and understanding capacity. Hence, users’ time-dependent interests must be taken into consideration in online teaching resources recommendation. Moreover, in the informationlized education, constructors of teaching resources library attach importance to the resources content, ignorance of the collections of users’ historical records. It results in that data sparsity is very common in their records dataset. Some typical recommendation methods have poor performance. Such cases are even more worse in the launching of website platform. Time-dependent dynamic recommendation methods are hard to get users’ changing patterns from the historical records. And cold start exsits in a larger number of online teaching resources because they lack of users’ behaviors.
Considering users’ dynamic interests, users’ missing behaviors and extremely sparse data in online teaching resources recommendation, we study the time-dependent recommendation methods for online teaching resources, and apply them into the teaching cloud platform of Beijing Honghe technology company. The details of this paper are listed as follows:
(1) Users’ interests on teaching resources, which are affected by the development of knowledge points, are changing over time. Considering this, we propose an hidden semi-Markov model for collaborative filtering. It uses the hidden states to denote users’ latent interests, and introduces state duration to denote the time duration of users’ latent interest. It tracks the transition and duration of users’ interests with the transitive states and state duration in semi-Markov process. Based on the distributions of states transition and state duration, we use the hidden states in last multiple time periods to derive users’ next interests and the probability for items might be preferred. The experiments show that this method can model the heterogeneity of users’ interests duration (different changing patterns), which explains the reason of our algorithm prior to other methods in time-dependent applications.
(2) Considering data sparsity in users’ historical records, we propose a inhibited hidden Markov model for collaborative filtering. It introduces a binary variable (inhibited or active) into the hidden states to denote users’ latent interests. It uses the inhibited hidden states to represent users’ interests with idle states and model the missing historical records in time which results from data sparsity. It uses the active hidden states to represent users’ interests with active states and model users’ behaviors. It tracks users’ changing interests over time with states transition in Markov process. Moreover, based on the distributions of users’ current states and states transition, we predict the distribution of users’ next active latent states, and the probability of each item might be preferred. The experiments show that it can handle the time discontinuity of users’ records, and improve the performance under different data sparsities.
(3) Furthermore, online teaching platform has the problems of extremely sparse historical data and new teaching resources that lack of users’ records (cold starts) when it just launches its system. Hence, we propose a hybrid recommendation method based on the sequential structure of courses. This algorithm represents users’ interests with their studying courses. It combines decision tree and Markov model to model the prior knowledge of the hierarchical structure of teaching books and the sequential structure of courses. We perform content-based analysis on the keywords of teaching resources and courses, and get their association and the association between courses and users to track users’ dynamic interests, which can solve the cold starts of teaching resources. At last, we make full use of historical records and employ weighted method to combine the above results with association rules mining for users’ records. This method combines prior knowledge of courses sequences, users’ records and resources’ content for recommendation. The experiments show that it can track users’ dynamic changes under the high data sparsity and have a better performance.
At the end, we apply the research into the practice, and develop a hybrid recommender system consisting of three layers: data layer for data analysis, off-line layer with a collection of recommendation methods and online layer for making recommendation in real time. It can recommend teaching resources associated with twelve grades, nine versions and ten subjects to users. And we have integrated this hybrid recommender system into the enterprise software. |
修改评论