Ignacio Deza
I develop rigorous statistical methods for problems where standard machine learning fails: sparse data, heavy-tailed noise, and rare critical events. This methodology works across any complex system.
Methodological Approach
When Standard Methods Break
Most machine learning assumes abundant data, Gaussian noise, and common events. Real industrial systems often violate all three assumptions. My work develops alternatives grounded in statistical physics that handle these pathological cases rigorously.
Physics-Based Methods for Non-Physics Problems
Complex systems theory, non-equilibrium statistical mechanics, and information theory provide powerful frameworks for understanding structure in messy data. These methods transfer remarkably well to subscription markets, maintenance prediction, and any system with many interacting components.
Rapid Domain Absorption
A strong theoretical foundation enables quick mastery of domain-specific knowledge. I've moved from climate dynamics to aerospace maintenance to energy forecasting by recognising universal patterns across superficially different problems.
Academic Position
Senior Lecturer in Data Science, Programme Leader
University of the West of England, Bristol
I lead the Applied Statistics and Fundamentals of Machine Learning module and contribute to curriculum design for the BSc in Data Science. My teaching emphasises rigorous foundations combined with practical problem-solving - the same approach I bring to research and applied work.
Problem Classes I Specialise In
Rare Event Prediction
Equipment failures, subscriber defection, extreme weather events - when the outcome you care about is infrequent but critical, standard classification breaks down. Requires careful handling of class imbalance and hybrid statistical-deep learning architectures.
Applications: Aerospace maintenance, subscription churn, climate extremes
High-Dimensional Sparse Data
Many sensors, few measurements. Feature spaces with thousands of dimensions but only dozens of samples. Requires dimensionality reduction methods that preserve physical interpretability and avoid overfitting.
Applications: Sensor networks, industrial IoT, medical diagnostics
Non-Gaussian, Heavy-Tailed Distributions
When your data exhibits power laws, Lévy flights, or other heavy-tailed behavior, Gaussian assumptions produce systematically wrong answers. I develop and implement methods that respect the true statistical structure.
Applications: Financial time series, turbulence, network traffic
Complex Network Dynamics
Systems where interactions between components matter as much as the components themselves. Requires methods from network science and information theory to extract causal structure from observational data.
Applications: Climate teleconnections, supply chains, neural systems
Open Source Software
qNoise
PythonGenerator of non-Gaussian coloured noise implementing Lévy processes and arbitrary correlation structures. Used internationally for stochastic modelling, signal processing, and testing algorithm robustness against realistic noise conditions.
Published: SoftwareX, 2022 • DOI:10.1016/j.softx.2022.101034
Selected Research
Assessing the direction of climate interactions by means of complex networks and information theoretic tools
J.I. Deza, M. Barreiro, C. Masoller
Chaos 25(3), 033105 (2015) • 51 citations • DOI
Maintenance automation using deep learning methods: A case study from the aerospace industry
P.J. Mayhew, H. Ihshaish, I. Deza, A. Del Amo
ICANN 2023 • DOI
Estimating defection in subscription-type markets with down-sampled representation
M. Roberts, I. Deza, H. Ihshaish, Y. Zhu
Working Paper (2023)
The nonequilibrium potential today: A short review
H.S. Wio, J.I. Deza, et al.
Chaos, Solitons & Fractals 165(1), 112778 (2022) • DOI
Wind power ramp characterisation and forecasting using numerical weather prediction and machine learning
H. Ihshaish, I. Deza, R. Sharp
NeurIPS 2021 Workshop: Tackling Climate Change with Machine Learning
Complete list: UWE Research Profile • Google Scholar
Background & Professional Development
Ph.D. in Statistical Physics
Universitat Politècnica de Catalunya, Spain
European Horizon 2020 Fellow • Completed Feb. 2015
Training & Knowledge Transfer
Beyond academic teaching, I work with technical teams on focused professional development: advanced statistical methods, deep learning architectures for non-standard problems, and translating complex methodology into practical implementations.
My approach emphasises storytelling - making sophisticated technical concepts accessible to mixed audiences including engineers, executives, and domain experts without technical backgrounds.
Collaboration & Contact
I'm interested in working on genuinely difficult data problems - situations where conventional approaches have been tried and failed. If your challenge involves sparse measurements, rare but critical events, heavy-tailed distributions, or complex interacting systems, I may be able to help.
Contact: Ignacio.Deza@uwe.ac.uk