Information Retrieval As Structured Prediction
Information retrieval can be viewed as a structured prediction problem. Broadly speaking, structured prediction refers to any type of prediction performed jointly over multiple input instances (e.g., a ranking over a list of documents). Rankings are the most common types of structured outputs within information retrieval. Even binary classification can be thought of as a special case where all instances are independent. Indeed, structured prediction is not a new idea. But until fairly recently, the primary impediment has been a lack of efficient and robust methods for training.
In this talk, I will focus on approaches which leverage the structural SVM learning framework. I will first describe the structural SVM formulation along with a general training algorithm that can be applied to a variety of structured prediction problems. I will then describe two applications in information retrieval. The first extends structural SVMs to optimize for mean average precision. The second extends structural SVMs to minimize subtopic loss in diversified retrieval.
This is joint work with Thorsten Joachims, Filip Radlinski and Thomas Finley.