On Mining Rating Dependencies in Online Collaborative Rating Networks

The trend of social information processing sees e-commerce and social web applications increasingly relying on user-generated content, such as rating, to determine the quality of objects and to generate recommendations for users. In a rating system, a set of reviewers assign to a set of objects different types of scores based on specific evaluation criteria. In this paper, we seek to determine, for each reviewer and for each object, the dependency between scores on any two given criteria. A reviewer is said to have high dependency between a pair of criteria when his or her rating scores on objects based on the two criteria exhibit strong correlation. On the other hand, an object is said to have high dependency between a pair of criteria when the rating scores it receives on the two criteria exhibit strong correlation. Knowing reviewer dependency and object dependency is useful in various applications including recommendation, customization, and score moderation. We propose a model, called Interrelated Dependency , which determines both types of dependency simultaneously, taking into account the interrelatedness between the two types of dependency. We verify the efficacy of this model through experiments on real-life data.