A study on novel recommendation methods based on content/collaborative filtering
Saranya Maneeroj · Medical Entomology and Zoology · 2005
Recent advances in information and networking technology have facilitated the creation, distribution and access of online information. The amount of data in this online information space is increasing far more rapidly than the ability of an individual or online customer to process it. Recommender systems are widely used to help users or customers to get the interesting items among the flood of information. Recent recommender systems commonly use content/collaborative recommendation technique which is a combination of the two most widespread techniques to date, which are Content-Based Filtering (CBF) and Collaborative Filtering (CF). The most important step in generating recommendations is to find a group of individuals that have similar tastes, called neighbors. Once a set of high-quality neighbors is formed, the high-quality recommendations are generated. However, there are many occasions where current recommender systems do not provide satisfactory recommendations because of forming improper or poor neighbors. One is opinion representation and the other is improper neighborhood formation. For opinion representation, the current systems only represent a user’s opinion as a rating value in selecting neighbors. They use the rating values on items for evaluating users’ preferences or user’s opinions. The rating value represents the overall preference of the user. A user might express his/her opinion based on some specific features of the item. For more accurate recommendations, the factors that well represent the user’s opinion should be taken into consideration. For neighborhood formation, the current systems employ either pure CF or pure CBF in finding neighbors. To find neighbors based on pure CF, rating values on the same rated items (co-rated items) are compared to form the neighborhood. However, the co-rated items are difficult to discover. Therefore the votes in the intersection of items rated by each pair of users will be small, which may lead to a poor estimate of neighbors. For finding neighbors based on pure CBF method, each pair of user profiles, which contain correlations between the content of items and the user’s preference, is compared to form the neighborhood. However, it is difficult to extract and implement all the features in order to form a user profile, which can cover the user interest. Accordingly, the quality of neighbors is restricted to a small number of item features. The goal of this research is to propose a novel recommender system, which provides higher quality of recommendations in content/collaborative hybrid filtering area. It is based on an advanced neighborhood formation method, which copes with the poor neighbor problem. Instead of using rating value alone, the method uses “user’s opinions on feature” to represent user’s preference in more detailed features. In forming high quality neighbors, we employ a cascade model of combining neighbors resulted from both CBF and CF to apply “user’s opinions on feature” and to eliminate the current neighborhood formation problems. We first employ CBF method to produce coarse candidate neighbors. We then employ CF with “user’s opinion on features” to refine the neighbors form among the candidate set. An experimental recommender system, called Advanced Yawara system was developed to implement and evaluate the proposed neighborhood formation method. In this research, there are 2 kinds of experiment. In the first experiment, we employed five evaluation metrics to compare the Advanced Yawara and two other content/collaborative hybrid systems. The result of first experiment shows that the Advanced Yawara provides higher quality recommendations than current hybrid systems. In the second experiment, we compared the Advanced Yawara with the “Advanced Yawara with different cascade model” and the “Advanced Yawara without various features of opinion”. The result of second experiment shows that our proposed cascade model is suitable to combine neighbors generated from CBF and CF methods for increasing recommendation quality. It also shows that various features of opinion enhance to increase quality of recommendations. Although the experiment focuses on the movie domain, we realize that it is possible to apply the proposed method to other domains such as music.