Tag-Based Collaborative Filtering Recommendation in PLEM
In the context of the upcoming Professional Responsive Mobile Personal Learning Environment (PRiME) project, we investigate learning analytics and recommendation techniques in Personal Learning Environments (PLEs). Technically, recommendation systems are classified into the following classes, based on how recommendations are made: (a) collaborative filtering and (b) content-based recommendations. The Web 2.0/social software movement over the last years has led to an increasing interest in social recommendation system in Technology Enhanced Learning (TEL). The question is how to leverage the learner’s social activities (e.g. tagging, commenting, rating, and sharing) in PLEs to generate personalized recommendations.
The goal of this thesis is to enhance the Personal Learning Environment Manager (PLEM) application with a social recommendation component. The task was to develop, experiment and evaluate different collaborative filtering techniques to recommend personalized learning elements based on the tagging behavior of the learner. Further goal was to study the applicability and effectiveness of such recommendation techniques in PLE settings.
The PLEM system has been developed using Java programming language, with having Spring framework in the back end, and Google Web Toolkit for the front end interface. The recommendation component was developed by leveraging the Weka API for data minig tasks and Apache Lucene’s SnowballAnalyzer for the normalization and stemming of tags.
This master thesis tries to explain and alleviate the current information overload in Personal Learning Environments, like PLEM. The main idea we support throughout this thesis is to provide relevant, good quality, and personal recommendations of learning entities on the PLEM platform by utilizing the tagging behavior of the users. Studies have revealed that tags capture the interests of users, give meaningful reasons why certain content is relevant for an individual which seems a promising step for boosting the performance and quality of recommendation in PLE.
We chose to investigate and experiment if tag-based collaborative filtering recommendation algorithms can be a good solution of our research problem. We adopted and implemented 16 different algorithms, by considering the most popular memory-based and model-based approaches: k nearest neighbor method (kNN), dimensionality reduction, probabilistic classification, clustering and association rule mining. We have tested both, user-based and item-based CF that provide a list of top 10 recommended items.
The evaluation phase was divided in two parts, statistical offline and with users online. The offline evaluation aimed to find the algorithms that achieve high precision and recall, while the user study intended to examine if these best performing approaches meet the user satisfaction.
Based on our findings, we were able to compare different approaches and identify 4 tag-based collaborative filtering algorithms that are most appropriate for providing recommendation in PLE. In terms of accuracy, novelty and usefulness users rated high the user-based association rule mining. The item-based CF on LSA-reduced dataset and the user-based k-Means clustering handled well the sparsity of the dataset and offered good contextual compatibility. By making use of latent semantic relations or topics of interest, both approaches provided recommendations for most users, even for those of them whose tags differ significantly from the others. The item-based k nearest neighbor approach received positive feedback from our test users, and scored high precision and recall in the offline evaluation as well.
We conclude that tag-based CF techniques help the learner in discovering novel high quality items, and at the same time offer good quality of experience. However, the evaluation results reveal also that the quality of user experience does not correlate with high recommendation accuracy.
Created by. Last Modification: Wednesday, 17. October 2012 13:18:52 by .