HP Labs India

We are researching techniques to create a highly accurate user profile, including synching across multiple devices, embodying informational and transactional interests of the user. These profiles could be used in a number of applications such as : Personalized search, information filtering for services such as RSS feeds, news sources and personalized navigation interfaces.

Prior efforts in user profile construction have used text data sources like user visited web pages, search keywords and user documents to build word based profiles. Our effort is to build concept based profiles using ontologies like DMOZ and Wikipedia that incorporate additional sources of information such as user generated metadata and location information.

Most forms of information sourcing today do not scale very well. Portals are similar to broadcast channels in that they choose content that is likely to be of broad interest to everyone. Search engines do not account for the users personal interests. It is not feasible to specify and track every RSS feed that one may be interested in. Also, a lot of websites do not carry RSS feeds.

The key research question is how to source information in a pull based manner without explicit effort on the user’s part. In other words, can the users interests and consumption patterns be modeled and used to retrieve relevant information without explicit request or action from the user? This may be topic specific, for example - a user interested in gardening or event specific, for example, keep a user updated on the progress of a typhoon".

Matching the user profile to the suitably represented content involves computing some kind of similarity measure between the profile and the content representation. A commonly used similarity measure for documents and profiles represented using term vectors is the cosine similarity measure. Matching becomes complex when there is insufficient information for computing a similarity measure, for example when videos have very little text description. We are investigating the use of semantic web technologies for matching in these kinds of scenarios.

This project aims at behavioral matching of ads based on a user profile. There could be various metrics for ad selection: click-through, intrusiveness, subsidy provided to user etc. Ad delivery to the client device removes some of the restrictions in server side settings like limited screen real estate,only limited number of bidders succeeding, contextual nature etc. It however introduces news problems: where and when should the ad be shown, how to control the number of exposures etc.

Although the ad matching problem has been extensively studied in the case of contextual ads, few studies exist on ad selection, matching and delivery to client devices.

We are researching issues such as

1. Efficient means of ad delivery (push versus pull), ad caching
2. Bidding techniques, techniques to limit sponsorship/subsidies in a distributed setting
3. Ad matching based on user profile/context/activity under different selection criteria

4. Multidevice ads - Actions on one device, for example reading something on the PC results in related ads on a different device like on TV.

Carrying out tasks on the Web can be a quite complex process. Consider a user who wants to book a ticket. The user has to find the travel website, navigate to the booking page, perhaps even register, fill in address details in a form, select a payment gateway, login to the gateway and approve a transaction and finally print out a receipt of the transaction. If the user wants to repeat this process on a website, he has to start the process all over again.

We are researching methods and interfaces to simplify the transaction experience for users. This includes autonomous agents that poll the web on behalf of the user and make recommendations and easier interfaces to manage transactions across multiple web sites.

Click here for recent Publications listing.