Machine Learning Prediction andWeb Access Modeling

History-based machine learning technique is efficient in prediction and improving Web server performance. To generalize the history-only prediction to algorithms that include other sources such as page size and priority levels in determining pre-load pages, we present, in this paper, a new prediction scheme that considers not only multiple attributes for page selection, but also the computational complexity side of the algorithm. The idea is an extension to our earlier matrix application in machine learning Web cache pre-fetching. We use real world data to test the efficiency of the new algorithm. Results show that system performance measured by hit rate is greatly increased by prediction and prefetching, especially for small size caches. In addition, we introduce a user access model that is based on sequence and group user actions to simulate the request pattern. Data generated from the input model are compared with that from the real world.