Student Projects

The HP Labs Bristol Semantic Web Group regularly host students. Some of these are undergraduates undertaking an industrial placement as part of their degree, some are carrying out the research portion of their Masters thesis based in labs, others are taking a sabbatical during their PhD. Placements last between 4 and 12 months.

NOTE The 2007 Application Process has finished. This page is for reference only.

Proposed projects include

• Attribute-rule based classifiers for semantic instances
• Adaptive user profiles

It should be emphasised that these are intended to be useful guidance, rather than a definitive list. Not all of these projects will necessarily run, and for those that do the exact details may vary from those presented below. We may also devise new projects not on this list, and welcome original suggestions from prospective students.

If you would like to apply, please contact the semantic web group at semweb@hpl.hp.com with:

1. your resumé;
2. a covering letter indicating your research interests and availability; and
3. your project preference - either taken from the list above, or a description of your own proposal.

Additional details about the process can be found on this page

(top)

Attribute-rule based classifiers for semantic instances

Various data-mining algorithms, such as the decision tree learner C4.5 learn to classify problem instances based on evaluation of rules against instance attributes. Instances within a semantic web have both explicit and implicit attributes at various distances of remove by inference across the network of semantic relationships. So, for example a weather reading W has an explicit attribute for every property r for which there is a triple (W, r, x) (e.g. r = "temperature"); it may also have an implicit attribute for every pair of properties (r1,r2) such that there are triples (W, r1, L), (L, r2, x) (e.g., r1 = "location", L = "London", r2 = "altitude"); and so on. In general the set of all such attributes is extremely large, and may indeed cover the entire database.

The goal of this project will be to design, implement and test algorithms that extend chosen attribute-based decision algorithms to scenarios in which attribute sets are organized into typed hierarchies, as described above.

The challenges lie in deciding which parts of the semantic network to prioritize exploration of, how to prevent over-fitting, and how to decide when to terminate. Query context or user profile might help with these questions, but it is likely that the most promising approach, since it is widely used in standard machine learning, will be the use of information theoretic measures to decide when a choice is potentially valuable. Understanding how to adapt information-theoretic measures to semantic contexts will be a valuable step towards the goal.

(top)

Adaptive user profiles

We are interested in the problem of constructing rich, structured user profiles from implicit user behaviour - that is, profiles constructed from observing users conducting routine tasks.

One such task is tagging of URLs using systems such as del.icio.us. A user's tags can be represented in a number of ways, from a flat list to a tag cloud to something more richly structured.

We have work in progress (an existing student project) which explores the construction, visualisation and application of such profiles. This project (not yet complete) has already yielded promising results which we would like to build on.

Promising directions include:

• investigating alternate methods for profile construction
• Using a wider variety of data sources for tracking implicit user behaviour.
• Using the profile to build personalized navigation interfaces over various corporate information sources
• Looking at the flexibility and adaptiveness of the profile in response to changing user behaviour

(top)