BOTL

This git repo has data, data generators and code for the Bi-directional Online Transfer Learning framework (BOTL).

Running code

To run code either use

• python3 controller2.py ....
• python3 runResults.py .... (this runs controller2.py for multiple repeat iterations)

Use python3 runResults.py --help to display run options

Common Options

• --domain type of dataset used: Following, Heating, Sudden, Gradual
• --type concept drift detection strategy: RePro, ADWIN and AWPro have been implemented
• --window window size of the concept drift detector
• --ReProAcc, --ReProProb and --ADWINDelta parameters required by RePro, ADWIN and AWPro
• --runid used when debugging (change output in controller2.py to somethign other than dev/null)
• --numStreams number of domains in the framework
• --ensemble version of BOTL/how models are combined - see below for options
• --perfCull predictive performance culling threhold parameter used by P-Thresh (BOTL-C.I), MI-Thresh (BOTL-C.II) and CS-Thresh
• --miCull Mutual Information culling threshold parameter used by MI-Thresh (BOTL-C.II)
• --paCull Pricipal Angle/conceptual similarity culling threshold parameter used by CS-Thresh
• --variance total variance captured by the PCs used to represent base models, used by CS-Thresh and CS-Clust
• --learner list of types of models to be used, so far SVRs and RRs can be used

<!--- `--replacement` is model replacement it to be used when determining whether to transfer-->

BOTL variants

Different variants of BOTL have been implemented and are specified by the --ensemble parameter

• BOTL:
  • --ensemble OLS
  • BOTL with no base model selection
  • introduced in [1], used in [2] and [3]
• P-Thresh:
  • --ensemble OLSFE
  • use --perfCull to set predictive performance culling threshold parameter
  • BOTL with predictive performance thresholding to select base models
  • introduced in [1] as BOTL-C.I, used in [2] and [3]
• MI-Thresh:
  • --ensemble OLSFEMI
  • use --miCull to set mutual information culling threshold parameter
  • BOTL with mutual information and predictive performance thresholding to select base models
  • introduced in [1] as BOTL-C.II, used in [2] and [3]
• CS-Thresh:
  • --ensemble OLSFEPA
  • use --paCull to set conceptual similarity culling threshold parameter
  • BOTL with conceptual similarity and predictive performance thresholding to select base models
  • introduced in [3]
• CS-Clust:
  • --ensemble OLSKPAC2
  • use --variance to determine how much variance is captured within the PCs that represent the uderlying concept
  • BOTL with parameterless conceptual clustering to select base models
  • introduced in [3], uses STSC [4] to create clusters of similar models

Available data and data generators and BOTL implementations:

• Following distance data for 6 journeys (2 drivers).
• Drifting hyperplane data generator
• Smart home heating simulation (with real world weather data)

Note the underlying framework is the same for all three implementations. For ease of reproducibility all three versions have been added.

AWPro

AWPro is a concept drift detection algorithm that combines aspects of RePro [5] and ADWIN [6] that better suit the BOTL framework. AWPro was first introduced in [2].

Parameter Analysis

Parameter analysis has been done to consider (see parameterAnalysis.pdf) to impact of the parameter values of underlying concept drift detection strategies, and how they impact the BOTL framework.

Source Code

The BOTL framework has been created using various code from other sources. ADWIN and AWPro implementations (which uses ADWIN as a basis for drift detection) are based upon the implementation available: https://github.com/rsdevigo/pyAdwin. This code is included in datasetBOTL/BiDirTansfer/pyadwin/

Other work relating to future variations of BOTL use Self-Tuning Spectral Clustering has been created based on the implementation available: https://github.com/wOOL/STSC. This code is used in datasetBOTL/BiDirTransfer/Models/stsc*.py

References

<a id="1">[1]</a> McKay, H., Griffiths, N., Taylor, P., Damoulas, T. and Xu, Z., 2019. Online Transfer Learning for Concept Drifting Data Streams. In BigMine@ KDD.

<a id="2">[2]</a> McKay, H., Griffiths, N., Taylor, P., Damoulas, T. and Xu, Z., 2020. Bi-directional online transfer learning: a framework. Annals of Telecommunications, 75(9), pp.523-547.

<a id="3">[3]</a> McKay, H., Griffiths, N. and Taylor, P., 2021. Conceptually Diverse Base Model Selection for Meta-Learners in Concept Drifting Data Streams. arXiv preprint arXiv:2111.14520.

<a id="4">[4]</a> Zelnik, M.L. and Perona, P., 2015. Self-tuning spectral clustering. Advances in Neural Information Processing Systems, pp.1601-1608.

<a id="5">[5]</a> Yang, Y., Wu, X. and Zhu, X., 2005, August. Combining proactive and reactive predictions for data streams. In Proceedings of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining (pp. 710-715).

<a id="6">[6]</a> Bifet, A. and Gavalda, R., 2007, April. Learning from time-changing data with adaptive windowing. In Proceedings of the 2007 SIAM international conference on data mining (pp. 443-448). Society for Industrial and Applied Mathematics.