Caveats

Pre-fetching is pointless if the user can read at a rate faster than data can arrive at his/her browser. We assume that this is not the case.

In general, we have observed that the profiles are learned for the first time over 5 to 7 sessions. Learning changes in user patterns online is dependent on the weights assigned. In our case, we placed heavy emphasis on a user's history. This means it takes longer for the system to adapt to changes. At the same time, the system is less sensitive to random bursts of visits to certain sites that will not be visited again. That is, we set the system to do long term adaptation. We found that this model suited our access patterns.

The cache hit ratio, defined as the number of hits over the number of user requests, for the above tests averaged at 62%. The accuracy of pre-fetches, defined as the number of hits over the number of pre-fetches made, averaged at 50%. However, note that surf does not visit previous pages. In normal browsing, we have found that the ``previous'' button is used rather often, and this increases the hit ratio as well as the pre-fetch accuracy. Note that the pre-fetch accuracy can actually be greater than one if a single pre-fetch can service multiple requests. In our own browsing, we found that the hit ratio hovers at around 75% and the pre-fetch accuracy hovers at around 70%. In this case, the low pre-fetch accuracy is due to visits of new sites in daily Web surfing.

Note that all these numbers are specific to our access patterns and for our learned profiles. We expect that these numbers may change for other usage patterns - in particular, the tuning of the constants in determining user profiles played a key role in improving the efficiency of pre-fetching. However, while we believe that significant work needs to be done in order to automatically tune the system to match user access patterns, we do believe that our system can provide perceptible improvements in the experience of Web browsing.

We also noted with interest that our own access patterns have changed as a result of using the proxy system. However, we felt that a user's access pattern will naturally change when the network condition changes. For example, if the network is slow, the user is usually apprehensive about clicking and then waiting. Since our proxy system basically changes the network conditions as far as the browser is concerned, we felt that a change in our access pattern is tolerable.