OASIS: a tool for efficient evaluation of classifiers¶
OASIS is a tool for evaluating binary classifiers when ground truth class labels are not immediately available, but can be obtained at some cost (e.g. by asking humans). The tool takes an unlabelled test set as input and intelligently selects items to label so as to provide a precise estimate of the classifier’s performance, whilst minimising the amount of labelling required. The underlying strategy for selecting the items to label is based on a technique called adaptive importance sampling, which is optimised for the classifier performance measure of interest. Currently, OASIS supports estimation of the weighted F-measure, which includes the F1-score, precision and recall.
When should I use OASIS?¶
OASIS is particularly useful when:
- you have a test set, but you don’t yet have ground truth labels
- ground truth labels can be obtained sequentially (this constraint may be relaxed in a future update)
- F1-score, precision, or recall is a sufficient measure for the classifier’s performance
- the classification problem demonstrates a high degree of class imbalance (examples include entity resolution, text classification and many problems in the medical domain)
Note that the final point to do with class imbalance does not need to be satisifed in order for OASIS to provide accurate estimates of classifier performance. It merely describes when OASIS is expected to excel over simpler methods, such as uniform sampling (see Passive sampler).
Where can I find out more?¶
Details about the interface for OASIS, including the required inputs are given in the OASIS section of this documentation. For more information about the algorithm itself and a proof of its consistency, please refer to the following paper:
N. G. Marchant and B. I. P. Rubinstein, In Search of an Entity Resolution OASIS: Optimal Asymptotic Sequential Importance Sampling, arXiv:1703.00617 [cs.LG], Mar 2017. link
License and disclaimer¶
OASIS is released under the MIT license. Please see the LICENSE file included with the source.
The software is provided “as is”, without warranty of any kind, express or implied, including but not limited to the warranties of merchantability, fitness for a particular purpose and noninfringement. In no event shall the authors or copyright holders be liable for any claim, damages or other liability, whether in an action of contract, tort or otherwise, arising from, out of or in connection with the software or the use or other dealings in the software.