Improving the performance of recommender systems using a parallel cbir method
Kötet címe (évfolyam száma)
Nowadays, many technologies based on the computer science have rapidly improved the quality of every part of our lives. These technologies are utilized in many different areas, use different techniques, and are based on different concepts; however, they typically have a single goal in common that is to make life easier and improve its quality. Today, biggest companies are offering their services and products based on the computerized technologies. In the current competitive market, giving the users a better experience working with the website, application, or any other software that a company uses to communicate with its users can be highly valuable to that company. Recommender systems are among the most valuable automatic tools that can improve the user experience with a more general system considerably when embedded inside that system. Since the mid-nineties and slightly after that, when the term “recommender system” coined by Resnick and Varian, researchers have conducted huge studies on this topic and the methods and approaches they have found are various and sundry. A recommender system serves the users in many ways, for instance, it shows the items to the users of a retailer’s website in an order which is probably more compatible with their attitudes. One of the major factors that affects one’s decisions and attitude towards an item is the image that describes the item or is highly related to it. Therefore, it seems intuitive that recommender systems consider images in their methods; however, this not the case in general. In this research we have proposed an approach that measures the similarity of the items considering both its common features, which are considered as numerical and categorical, and the image attached to that item. One big issue that treats the feasibility and practicality of such an approach is the difference between the computation time of similarity measurement considering these two categories of features. In order to make this system practical we have used a distributed approach which considerably reduces the computations related to the image part. The results show the effectiveness of this approach despite the simple similarity measure that we have used to compute the similarity of images.