講座內(nèi)容：Prediction models and their efficient evaluation arise as an important and timely direction of fundamental and applied research. It is apparent that there are no ideal models. It is impossible to envision a situation where any model can deliver an ideal fit to experimental data. It is also needless to say that the quality of any model is of paramount importance to any application. To offer a sound and realistice valuation of the quality of constructed models, it is legitimate and intuitively appealing to admit that the prediction results come in a certain non-numeric way.

In this talk, we focus on the concepts of predictive models with prediction results being realized in the form of information granules. We develop a comprehensive algorithmic setting behind the design of information granules of prediction and offer a detailed quantification of quality of information granules. With regard to the nature of prediction information granules, two facets of their quality are identified and characterized, namely a coverage criterion and a specificity criterion. Their roles are presented along with a composite index and its ensuing optimization through an optimal allocation of information granularity.

For the completeness of the presentation and its suitable positioning in a general context, some well-known constructs of statistically-guided prediction in linear regression such as confidence intervals, confidence curves, and prediction intervals are briefly reviewed.The generic idea of prediction intervals is generalized to embrace information granules of prediction including constructs such as fuzzy sets. In the sequel, we introduce a concept of a granular parameter space and a granular output space yielding a granular nature of prediction.