The wireless web offers the potential of access to vital information anywhere at any time. However, hand-held devices with small screens connected to the web with slow and expensive network connections offer new challenges. In particular, it is difficult on such devices for a user to scan through lists of e-mail messages, restaurant listings or news stories to find personally relevant information. Services that can personalize information for the user based on the user's location, current task, and a profile of the user's interest can make the wireless web live up to its potential.
Keyword-based systems are an alternative used by some services that are moving into wireless web from the content delivery to pagers. They rely on the user creating a profile that consists of a number of keywords indicating the type of content they would like to receive. Due to the complexity of text input on the cell phones, these services also require a user to create an account or profile on the Internet that determines what content will be delivered. We believe few users will be able to construct profiles with such keywords. Like dialog boxes, keyword-based systems do not help in determining the order in which content is sent to users.
There are two basic approaches that may be used to infer profiles for making recommendations to users: collaborative filtering and content-based filtering. Collaborative filtering monitors the behavior of all users and tries to find users with similar tastes. It then recommends items to an individual if similar individuals like it. Systems by companies such as FireFly and NetPerceptions have been used to recommend CDs, movies, and books. This approach requires many users to rate an item and users to rate many items before it can reliably make recommendations. Collaborative filtering is appropriate for personalizing location-based services such as restaurant recommendation on the web. Applications such as recommending nearby restaurants and retail outlets can be enhanced adaptive personalization technology so that the best nearby options are presented rather than simply the closest options. However, it not appropriate to e-mail and unlikely to work well for changing events such as news delivery.
Content-based recommendation creates a statistical profile of the user's interests in terms of the words and phrases that distinguish items of interest to the user from other items. It is appropriate for news, e-mail and restaurant recommendation other filtering where a text or structured description of content is available. We have developed such a system at the University of California, Irvine and evaluated the system on a group of over 3000 users on news delivery. Our results show an increase in news readership by over 40% when headlines are sent in a personalized order to user.
The adaptive personalization approach with implicit feedback has two important advantages. First, from a user's viewpoint, it is easier to use. The user just uses the system and a profile is automatically constructed. The user need not be told that there is an adaptive server since it requires no user configuration. That is, it is a marketing decision to decide whether to advertise this and not a technical necessity. Second, it allows for additional content-based criteria to be used in determining which stories to send to a user. In particular, in addition to sending relevant news stories or e-mail messages, we also desire to send a diverse group of items. This can be achieved by measuring the similarities among current items and making sure that the items on a screen are relevant and dissimilar to one another. By ensuring some diversity among the items, the most of the bandwidth is devoted to items of known interest to the user, but some bandwidth is reserved for items that are novel and can then determine whether the user is interested in following a new area.