Abstract:
Internet is an important channel for the distribution of digital contents such as E-books, audio, video, images, etc. Searching for E-books represents a difficult task for users. Very large volume of information and user's personal preferences are the main reasons which make the recommendation systems as essential applications. The proposed system applies semi-supervised learning techniques to improve the classification model. To evaluate the relevance of the proposed method 650 E-books are used as a dataset of constructing and labelling. All data files are tagged with six categories of meta-data such as book title, author, field, publisher, year of publishing, country of the author. Naïve Bayes, J48 in Weka implementation are the two learning algorithms used to verify whether recommender system depend on the learning algorithm. First, a completely supervised learning method was used. Performance measures precision (P), recall(R), F-score (F) showed no impact by the learning algorithm. The proposed system was trained by a semi-supervised learning scheme where user’s feed-back is obtained. There were few books labelled at the beginning and after several iterations, the collection of tagged data was increased. The system with 5 labelled examples achieves a performance of F=0.455, and after provided feedback on 20 examples, the recommendation system manages to obtain an F=0.754. It is noted after the 70 iterations, recommendation model tends to stabilize, reaching F = 0.915. The proposed method improve the performance of E-book recommendation systems.
