Titulo Estágio
Neural Networks for Automatic Rule Tuning
Áreas de especialidade
Sistemas Inteligentes
Sistemas de Informação
Local do Estágio
Feedzai/Remote
Enquadramento
Rule systems are still ubiquitous in fraud detection systems, for example due to compliance reasons. More often than not such rule systems become very complex and are difficult to maintain. This poses problems when the rules thresholds need to be tuned to match the client needs (e.g. achieving a certain alert rate). Because rules are often combinations of various conditions, this tuning can become very cumbersome.
Objetivo
In this project, the objective is to leverage neural networks to guide the rules tuning. For this purpose, we aim to build standard neural ‘blocks’ that encode the same logic operations found in rules. We can then translate any rule system into a neural network, with the advantage of being differentiable. For a specific metric to be adjusted, we can backpropagate the gradients to reach the threshold parameters of each neural block. The gradient information can then guide us on which thresholds to change and in which direction.
This exciting project will therefore be investigating novel approaches to make the rules tuning process easier, and the student will be able to make custom neural networks and gain expertise into deep learning.
Plano de Trabalhos - Semestre 1
- Literature review on transformations between rule systems and neural networks
- Internal survey on the types of rule scenarios used in our domain and the various use-cases (anti money laundering, credit card fraud, etc.)
- Define experimental setup: datasets to be used and rule systems to be tuned (starting from simpler systems to more complex)
- Familiarize with the manual rule tuning solution
- Preprocessing datasets and preparing evaluation tools
- Intermediate report writing.
Plano de Trabalhos - Semestre 2
- Implementation and testing of neural network blocks mimicking rule conditions
- Build a rule_system-to-neural network tool
- Implement strategy to refine rule parameters exploiting gradient information (in collaboration with analysts)
- Test and evaluate strategy on various rule systems (from simpler to more complex). Benchmark vs manual rule tuning system.
- Validate results with analysts, refine solution based on product & analyst feedback
- Document results and writing dissertation
Condições
The data sources are identified and available in the Feedzai Research data repository. The internship agreements ensure access to the Feedzai Research data
Paid internship for the duration of the project (1000€/month) and according to time allocation.
Orientador
Jacopo Bono
jacopo.bono@feedzai.com 📩