STOR-i Students win Nick Smith Prize
![Photo of Postgraduate Statistics Building, Lancaster University](https://cisweb.lancaster.ac.uk:443/img/cwip/cisweb.lancaster.ac.uk/EventsMedia/PSC%20photo-637062980060941187.jpg?mode=crop&width=874&height=289¢er=0.50%2c0.50)
The Nick Smith Prize is awarded to a Statistics PhD student in their second year on the basis of their excellence in research and involvement with the department. We are very pleased to announce that for the academic year 2018-9, the Postgraduate Research Committee has nominated Alex Fisch (Supervisors: Idris Eckley and Paul Fearnhead) and Henry Moss (Supervisor: David Leslie and Paul Rayson) as the joint winners of the Nick Smith Prize for their excellence in research. The prize commemorates a member of staff who tragically died in a climbing accident.
Henry Moss is a second year student, working to develop statistical and machine learning methods for natural language processing (NLP), and in particular to develop methods for automatic decision-making in response to written and spoken text. As part of this research, Henry has been developing more efficient parameter-tuning methods for machine learning models. This is needed since state of the art NLP models have multiple parameters to tune, and it can take days to fit the model to find out if any particular parameter choice performs well. Henry identified that the ubiquitous cross-validation approach to evaluate model performance and tune parameters is deeply flawed due to high variability which is especially prevalent in the imbalanced datasets of NLP problems; he derived a solution which he published at COLING, one of the premier computational linguistics conferences. This work has now developed into using Bayesian optimisation to carry out efficient and cost-sensitive parameter tuning, which will be published in the near future in the machine learning literature. In parallel, Henry has developed bandit-based techniques to quickly find the highest-performing model for an NLP task which has been published at ACL, another of the top NLP conferences . Henry is currently on secondment at Amazon Research in Cambridge, but is looking forward to returning to somewhere with a few more hills to cycle up.
Alex Fisch is working in the area of computationally and statistically efficient methods for anomaly detection. To date he has completed two projects in this area, which have led to substantial papers submitted to top statistical journals. Each body of work covers statistical modelling, computationally efficient algorithms (available as R packages), and substantial statistical theory.
Both papers look at detecting collective anomalies in time-series data. These are (generally short) regions of the series where the data shows unusual behaviour. The first paper looks at a method for univariate data, and the second paper extends this to multivariate data. Both methods are able to detect multiple collective anomalies in time that is linear in the amount of data, and are robust to point outliers. Alex has been able to show consistency of the method (in that it will detect all collective outliers with no false positives) under relatively weak conditions. The proof of this was particularly challenging in the multivariate case (and has not previously been shown before for earlier methods for this problem). Alex was also able to develop appropriate penalties for his method that mean it can detect both sparse and dense anomalies (i.e. ones that affect a few series a lot, or affect most series a little) — something that is known to be challenging in related changepoint problems.
Whilst these are both excellent pieces of work, the aspect of his research that makes Alex stand out is the level of independence he has shown, including having the key original ideas for these two projects. He is also prolific, with other work including collaborating with Lawrence Bardwell on an online version of this anomaly detection work; writing a J. Stat. Soft. paper for the anomaly package that contains implementations of his algorithms; an extension of these ideas to deal with auto-correlated noise; and substantial progress on a new project looking at developing robust Kalman filter algorithms.
His project is co-funded by BT, and he has also been active in working with them at implementing his algorithms on BT data. We have had very positive feedback from BT about Alex’s contribution in this area, as his work also showing considerable real-world promise.
Congratulations to both Henry and Alex!
Back to News