Measuring the Effect of Elementary Descriptive Attributes on News Recommender Systems

Abstract

Information overload and information obscuring are two most recent challenges in finding relevant information from the Internet. Moreover, specifically in the online news industry, using sensationalist and eye-catching headlines for click-baiting has become a common practice. Finding relevant information is becoming harder than ever before. A solution to overcome the challenge is by utilizing a news recommender system. In most of the news recommender systems, a certain number of data attributes is required to produce appropriate decisions. Unfortunately, this is not always the case, especially when users are not required to register to a news portal. In this case, valuable information that can be used to make decisions, such as users’ preferences, visit or reading histories, will not be available. Therefore, in this paper, we take advantage of the elementary descriptive attributes of news articles, namely titles and keywords. We compare how these attributes affect the decision results, namely the recommended related news. We collected news articles from a news portal, generated two sets of related news using different compositions of descriptive attributes, and compared them to the originally defined set. Our findings indicate that the combination of titles and keywords produces highly relevant news which achieved a mean rating value close to 3 (on a scale of 1 to 5). Whereas the original recommended related news only has a mean rating value around 1. The combination produces better results compared to the original recommended related news which is highly affected by news categories. Our findings also revealed that the presence of specific entities (such as a person, location) in the titles has a significant impact on the outcome. This work has a wide spectrum of potential applications in the future, for example for automatic news aggregation, combating spam, reading context understanding, and so on.