My research is within the field of Artificial Intelligence (AI).
AI is concerned with programs that provide
solutions (or approximations to solutions) to problems which are
'difficult' to solve by traditional methods. The difficulty stems
from the presence in the problem of disorder, uncertainty, lack of
precision or inherent intractability.
I am best known for my work in Case-Based Reasoning (CBR) and
Recommender Systems.
But I have also made contributions to areas such as natural language
processing, machine learning and ant algorithms.
Case-Based Reasoning
Case-Based Reasoning (CBR) is an approach to problem solving which
can be successful in domains where "similar problems have similar
solutions". In such domains, new problems can be solved by
transferring and adapting solutions that were used to solve
similar problems in the past.
Our research addresses questions such as:
How do we measure similarity and how can we efficiently retrieve the most similar cases?
How can these systems learn from experience, and how well do they learn?
How can these system maintain their knowledge, especially if it
is changing over time?
How can these systems explain their actions and conclusions?
How can these systems reason with unstructured information (e.g. text), as well
as structured information?
We have applied our research to problems in several domains:
In using CBR for spam filtering, we faced challenges such as handling
unstructured information,
and using learning to cope with concept drift (i.e. the fact that
spam is a moving target).
In forestry management, we used CBR to estimate timber volumes, and we integrated the estimation software with optimization software to compute high yield cutting plans.
We have used CBR as a form of meta-reasoning, assisting the developers of AI systems to choose methods for explaining their systems' predictions.
Recommender Systems
Recommender systems help us to decide which goods, services or
information to consume based on what they infer about our
preferences.
Our research addresses questions such as:
In a set of recommendations, how do we balance relevance with other properties such as diversity, novelty and serendipity?
How can these systems explain their recommendations?
How can we elicit or infer long-term user preferences and short-term user preferences explicitly (by asking them) and implicitly both from users' actions but also from other sources, such as product reviews?
How can we make recommendations to groups of users, such as a family watching TV together or friends going to the cinema together, in a way that balances the preferences of members of the group? How do we balance the multiple interests of an individual?
How do we strike the balance between obtaining as much preference data from users as we can while minimizing the burden we place on users and invasions of their privacy? How do we give recommender systems a sense of their own competence, and hence their own limitations? How can recommender systems quantify the uncertainty in their data and in the recommendations that they make?
We have applied our research in several settings:
We have fielded experimental recommender systems for small-scale online retailers, where we typically face the challenge that customers rarely make return visits, making it difficult to profile their preferences.
We have made several contributions to research into music playlists, including work on recommending songs to extend a playlist, predicting the listening context of a playlist, and presenting connections between consecutive songs in a playlist.