transfertools

transfertools is a small Python package containing recent transfer learning algorithms. Transfer learning strives to transfer information from one dataset, the source domain, to a related dataset, the target domain. Several constraints and assumptions can be placed on the domains, inspiring different algorithms to do the information transfer. The package contains four transfer learning algorithms.

Installation

Install the package directly from PyPi with the following command:

pip install transfertools

OR install the package using the setup.py file:

python setup.py install

OR install it directly from GitHub itself:

pip install git+https://github.com/Vincent-Vercruyssen/transfertools.git@master

Contents and usage

Transfer learning aims to transfer information from a source domain Ds to a related target domain Dt. A domain consists of a dataset with attributes X and labels Y. Thus, the source domain is Ds = {Xs, Ys} and the target domain is Dt = {Xt, Yt}. The fundamental assumption is that the source and target domain live in the same feature space. Different flavors of transfer learning methodologies exist. Unsupervised transfer learning, for instance, disregards label information and only uses Xs and Xt to determine what information to transfer. Supervised transfer learning uses the full domains Ds and Dt to do the transfer. Semi-supervised transfer learning uses the full source domain Ds and the target attributes Xt to do the transfer.

The actual information that is transferred also differs between methods. Domain adaptation techniques transform the source (and target) domains such that they match more closely (according to different criteria) and then combine all the data points to construct Dcombo. Instance selection techniques select a subset of the source data that should be transferred to the target data to construct Dcombo. After transfer, a classifier can be constructed using Dcombo.

Instance selection techniques

The transfertools package contains two instance selection transfer techniques tailored to anomaly detection:

1. The LocIT (localized instance transfer) algorithm works in a completely unsupervised manner. It transfers the instances in Ds that have matching localized distributions in both domains [1]. This algorithm can also be used in other applications than anomaly detection.
2. The CBIT (cluster-based instance transfer) algorithm works in a semi-supervised manner. It transfer the instances in Ds that fall inside a cluster defined on the target data [2].

Given a source domain {Xs, Ys} and a target domain {Xt, Yt}, the algorithms are applied as follows:

from transfertools.models import LocIT, CBIT

# train
transfor = LocIT()
transfor.fit(Xs, Xt)

# predict
Xs_trans = transfor.transfer(Xs)

# ... or immediately
Xs_trans = transfor.fit_transfer(Xs, Xt)

Domain adaptation techniques

The transfertools package contains two instance domain adaptation techniques:

1. The CORAL (correlation alignment) algorithm is an unsupervised transfer learning technique that aligns the first and second order statistics of the source and target data [3].
2. The TCA (transfer component analysis) algorithm is an unsupervised transfer learning technique that projects the source and target data onto a lower-dimensional subspace [4].

Given a source domain {Xs, Ys} and a target domain {Xt, Yt}, the algorithms are applied as follows:

from transfertools.models import TCA, CORAL

# train
transfor = CORAL()
transfor.fit(Xs, Xt)

# predict
Xs_trans = transfor.transfer(Xs)

# ... or immediately
Xs_trans = transfor.fit_transfer(Xs, Xt)

Package structure

The transfer learning algorithms are located in: transfertools/models/

For further examples of how to use the algorithms see the notebooks: transfertools/notebooks/

Dependencies

The transfertools package requires the following python packages to be installed:

• Python 3
• Numpy
• Scipy
• Scikit-learn

Contact

Contact the author of the package: [email protected]

References

[1] Vercruyssen, V., Meert, W., and J. Davis. (2020) Transfer Learning for Anomaly Detection through Localized and Unsupervised Instance Selection. In 34th AAAI Conference on Artificial Intelligence, New York. To be published

[2] Vercruyssen, V., Meert, W., and Davis, J. (2017) Transfer learning for time series anomaly detection. In CEUR Workshop Proceedings, vol. 1924, pp. 27-37.