Adaptive machine learning and concept drift

Classical machine learning and data mining methods rely on the assumption that data distribution stays the same during model training and operation. In reality this is is often not the case. As the world is continuously changing, so evolve data that describe it. My research background is in machine learning methods that can diagnose themselves and adapt to changing data distribution over time.

The main outcome of my PhD thesis research was pointing out a distinction between the point in time that change takes place, and optimal training data history for updating predictive models. I have also made contributions in forming optimisation and evaluation criteria for predictive models with evolving streaming data, (inter)active learning for streaming data and adaptive data preprocessing.

Selected publications

Position papers and surveys

  • An overview of concept drift applications by Žliobaitė et al. 2016, a book chapter. DOI PDF
  • A Survey on Concept Drift Adaptation by Gama et al. 2014 in ACM Computing Surveys. DOI PDF
  • Open Challenges for Data Stream Mining Research by Krempl et al. 2014 in SIGKDD Explorations. DOI PDF
  • Next challenges for adaptive learning systems by Žliobaitė et al. 2012 in SIGKDD Explorations. DOI PDF
  • Learning under Concept Drift: an Overview by Žliobaitė 2010 in arXiv.

Evaluation methodologies

  • Evaluation methods and decision theory for classification of streaming data with temporal dependence by Žliobaitė et al. 2015 in Machine Learning. DOI PDF
  • Controlled Permutations for Testing Adaptive Learning Models by Žliobaitė 2014 in Knowledge and Information Systems. DOI PDF
  • How good is the Electricity benchmark for evaluating concept drift adaptation by Žliobaitė 2013 in arXiv.

Active learning

  • Active Learning with Drifting Streaming Data by Žliobaitė et el. 2014 in IEEE Transactions on Neural Networks and Learning Systems. DOI PDF

Cost-sensitive adaptation

  • Towards cost-sensitive adaptation: when is it worth updating your predictive model? by Žliobaitė et al. 2015 in Neurocomputing. DOI PDF

Adaptive preprocessing

  • Adaptive Preprocessing for Streaming Data by Žliobaitė and Gabrys 2014 in IEEE Transactions on Knowledge and Data Engineering. DOI PDF

Handling missing values

  • Optimizing regression models for data streams with missing values by Žliobaitė and Hollmén 2015 in Machine Learning. PDF DOI

Handling concept drift

  • Adaptive Training Set Formation by Žliobaitė, PhD thesis PDF
  • Combining similarity in time and space for training set formation under concept drift by Žliobaitė 2011 in Intelligent Data Analysis. DOI PDF
  • Theoretical Window Size for Classification in the Presence of Sudden Concept Drift by Žliobaitė and Kuncheva 2010, Technical Report PDF
  • On the Window Size for Classification in Changing Environments by Kuncheva and Žliobaitė in 2009 in Intelligent Data Analysis. DOI PDF
  • Determining the Training Window for Small Sample Size Classification with Concept Drift by Žliobaitė and Kuncheva 2009 in IEEE ICDM workshops. DOI PDF
  • Expected Classification Error of the Euclidean Linear Classifier under Sudden Concept Drift by Žliobaitė 2008 in IEEE FSKD. DOI PDF

Change detection

  • BLPA: Bayesian Learn-Predict-Adjust Method for Online Detection of Recurrent Changepoints by Maslov et al. 2017 in IJCNN. PDF
  • Modelling Recurrent Events for Improving Online Change Detection by Maslov et al. 2016 in SIAM SDM. DOI PDF
  • Change with Delayed Labeling: when is it detectable? by Žliobaitė 2010 in IEEE ICDM workshops. DOI PDF