Evaluation of Prediction

• Multiple URL ranking mechanisms:
  • Random order, Original order, and Similarity order
• Need to evaluate predictive accuracy over a real data set.
• Prefetching systems can use more than the top prediction.
  • It may have time/resources to prefetch more than one.
  • It may already have the top prediction in a local cache.
  • We test accuracy using the top one, three, and five predictions.
• We will also evaluate the case in which the prefetching system places objects into an infinite cache.
  • Current prediction failures may be useful later.