For 2013, the panel chose the following winning dissertation:
Feature Selection Via Joint Likelihood
Adam C. Pocock, University of Manchester
The field of feature selection has many different competing algorithms, selection criteria and measure functions, with little theoretical justification for the choice of one measure over another. In this thesis we focus specifically on feature selection algorithms which use information theoretic criteria and provide a solid theoretical justification for the use of such criteria. We begin by considering feature selection as a process which minimises a loss function, specifically the model likelihood. From this choice of loss function we show that the previous 20 years of research in information theoretic feature selection can be re-derived by making different factorisation assumptions on that likelihood. We also present experimental results showing how the different factorisation assumptions affect classification performance. Then using our unified perspective we present two further results, a natural way of incorporating prior knowledge via our probabilistic interpretation, and a way to incorporate misclassification costs by choosing a cost-sensitive likelihood.
About the author:
Adam Pocock studied for his PhD in the School of Computer Science at the University of Manchester, working in both the Machine Learning and Optimisation group and the Advanced Processor Technologies group, where he was supervised by Dr Gavin Brown and Dr Mikel Luján. His PhD was the product of 8 years at Manchester where he also gained an MSc and a BSc. He currently works in the Information Retrieval and Machine Learning group at Oracle Labs. His research interests lie in probabilistic explanations of feature selection using information theory, and in applying probabilistic inference to massive datasets.
The runners up for 2013 are:
Budget-Limited Multi-Armed Bandits
Long Tran-Thanh, University of Southampton
The multi-armed bandit (MAB) is a classical problem in decision theory, and presents one of the clearest examples of the trade-off between exploration and exploitation in reinforcement learning. However, the standard MAB gives an incomplete description of many real-world scenarios. In this thesis, we introduce the budget-limited bandit model, a variant of the standard bandits, in which making a decision (and its physical execution) is costly, and is limited by a fixed budget. This model is motivated by a number of real-world applications, such as wireless sensor networks, crowdsourcing systems, or online advertisement. We propose three new classes of algorithms to tackle this problem, for which we provide both rigorous theoretical and extensive experimental analysis. We also demonstrate the practical usefulness of the model by applying it to a real scenario. In particular, we use a budget-limited bandit approach to maximise the total information collection of wireless sensor networks.
About the author:
Long Tran-Thanh obtained his PhD from the Agent, Interaction and Complexity Research Group at the University of Southampton, supervised by Professor Nick Jennings and Professor Alex Rogers. Long is currently a Research Fellow at the University of Southampton working on the ORCHID project that investigates how humans and agents can effectively collaborate. His main research interests lie in investigating different problems of sequential decision making with constraints and their applications in a variety of areas of multi-agent systems, such as coalitional game theory, decentralised coordination, or crowdsourcing systems.
PlayPhysics: An Emotional Student Model for Game-based Learning
Karla Cristina Munõz Esquivel, University of Ulster, Magee
Game-based learning (GBL) environments introduced a new generation of Intelligent Tutor- ing Systems (ITSs) that provide personalised instruction by being constantly aware of stu- dent reactions to the system. Student motivation, attitudes, self-efficacy and affective state have been the key focus of such developments. Current models of student emotion have shown promise in laboratory environments. However, the problem of accurately recognising and inferring student emotions within learning environments persists. The majority of already existing computational models of student emotion employ cognitive theories that are not de- rived from the learning context.
Control-value theory (Pekrun et al. 2007) assumes that control and value appraisals are the most meaningful for determining emotions in educational settings. Our proposed computational emotional student model uses the Control-value theory for reasoning about learners’ emotions in GBL environments settings. The main hypothesis is that this model will recognise student achievement emotions, i.e. emotions relevant to the educational context, with reasonable accuracy (not random). The definition, implementation and evaluation of our computational emotional student model in PlayPhysics, an emotional game-based learning environment for teaching physics, are discussed. Our emotional model is implemented with a dynamic sequence of Bayesian networks for representation of learners’ achievement emotions. Probabilistic Relational Models (PRMs) are employed to facilitate their derivation. The Necessary Path Condition algorithm is employed in combination with Pearson correlations and Binary and Multinomial logistic regression for defining network structure. The Expectation Maximisation (EM) learning algorithm is employed for network parameter learning. Our model employs answers to questions in- game dialogues, contextual variables and physiological variables for recognising student emotion.
Results show a fair accuracy of classification of student achievement emotions for the PlayPhysics’ emotional student model when only contextual and behaviour variables are considered (values of Cohen’s Kappa in a range larger than 0.2 but lower than or equal to 0.4), which then improves when physiological variables, i.e. Galvanic Skin Response (GSR), are incorporated (values of Cohen’s Kappa in a range larger than 0.4 and lower than or equal to 0.6). Our emotional model provides enhanced understanding about the factors involved in reasoning about emotion. PlayPhysics GBL environment is assessed to attain an enhanced understanding of the student experience of achievement emotions.
Future work may focus on creating further game challenges, identifying enhanced predictors for control and value, e.g. using sentiment analysis and analysis of facial expressions. Numerous applications, in areas ranging from biology to e-commerce, are envisioned for the application of our approach to create intelligible and dynamic genetic and emotional consumer data models.
About the author:
Dr. Karla Cristina Munõz Esquivel obtained her Ph.D. in Computer Science at the University of Ulster Magee Campus in July, 2013 having worked in both the Schools of Creative Arts & Technologies and Computing & Intelligent Systems under the supervision of Professor Paul McKevitt and Dr. Tom Lunney. Dr. Munoz, originally from Mexico, came to Magee in 2007 to complete an M.Sc. in Computing and Intelligent Systems with a scholarship from the EU AlBan programme. She was awarded the 8over8 Ltd. prize for best overall performance on the Master's course in 2008 and obtained a Vice Chancellor's Research Scholarship to fund her Ph.D. studies. Her Ph.D. work has been published in international journals and conferences and as a poster at `SET for Britain', House of Commons, London (March, 2011). Dr. Munoz is currently employed as a Software Design Engineer at ASML in The Netherlands, the world's leading provider of lithography systems for the semiconductor industry. Her main research interests include Serious Games, User Modelling, Affective Computing, Intelligent Tutoring Systems, Biofeedback, Human-Computer Interaction and Sentiment Analysis.
The 2014 panel members are:
- Simon Dobson (St Andrews, Chair)
- Russell Beale (Birmingham)
- Joemon Jose (Glasgow)
- Perdita Stevens (Edinburgh)
- Steve Pettifer (Manchester)
- Iain Phillips (Loughborough)