AutoGFE-automatic feature engineering with graph neural network

Studiengänge: Informatik, Informationstechnik

Background:

Feature engineering is the task of improving predictive modelling performance on a dataset by transforming its feature space. This task is primarily manual and time-consuming. Automating this task, therefore, has been the trend of the study.

The approaches to automatic machine learning that have been proposed so far can be broadly classified into two categories: 1) extension-selection based approaches, which apply all transforms at once to the base dataset, followed by one step of feature selection; 2) evaluation based approaches, which rely on transformed feature space exploration through evaluation-guided search. However, extension-selection based approaches demand a high computational capacity and computer memory. Most of the guidance-based methods, on the other hand, are too limited in the input information of their surrogate models, using mostly the conversion sequence itself, with a small percentage using some global information, such as the length of the conversion sequence, the remaining time duration, etc.

Describing feature engineering with tree structure, where each edge represents a transformation and each node represents the performance improve cased by the associated edge.We can illustrate historical evaluation process with graph. Through Graph neural network, we can extract information in the graph and used to predict the next available candidate.

Your jobs:

  • Research state of the art feature engineering and graph extraction method
  • Implement a graph neural network to extract information from the graph
  • Implment a reinformcement algorithm to train the surrogate model
  • Implment the feature engineering algorithm
  • Evaluate the feature engineering algorithm with different dataset
  • Write thesis

We can provide:

  • Intensive support
  • A pleasant working atmosphere and constructive cooperation

We expect:

  • Independent thinking and working
  • Knowledge of Python (Pytorch)
  • Knowledge of graph neural networks
  • Knowledge of reinforcement learning

If interested? please contact: Yiran Huangyhuang@teco.edu

Referenz:

[1] An empirical analysis of feature engineering for predictive modeling: https://arxiv.org/pdf/1701.07852.pdf

[2] Cognito Automated Feature Engineering for Supervised Learning: https://www.computer.org/csdl/pds/api/csdl/proceedings/download-article/12OmNxwWoLu/pdf

[3] Feature engineering for predictive modeling using reinforcement learning: https://ojs.aaai.org/index.php/AAAI/article/view/11678/11537