March 2017 DSSG hack for the Deutsche Krebsgesellschaft (DKG).
We aim at building a multi-label model that is able to predict the labels for a given RIS article, that includes features like:
- abstract
- title
- authors
- ...
We've built a django webservice that allows the DKG to interact with our model via RIS file uploads.
The service has the following features at the moment:
- upload training RIS file - triggers model selection on given data.
- upload test RIS file - produces keyword predictions for the given articles.
On each of the predictions, a user can give either positive or negative feedback, e.g. add another label or remove a predicted label respectively.
In order to improve the mult-label model, the service is able to receive the feedback of a user.
We've implemented different strategies of prioritization for this active learning setting. See this article for a survey of active learning.
- truncate classifications, used for removing noisy or underrepresented labels from the dataset
- top_level_labels to extract top level label lists from the CSV strings
- multilabel_cancer_classification pipeline for multi label classifications
- multiabel_test interactive pipeline for multi label classifications, including truncated classifications
The module cleaning_classification_labels implements a cleaning pipeline for classifications which should be applied to rectify the labels a bit.
Also there is a notebook which we added in the beginning of the hack, features.ipynb, looking at different attributes of the features and also doing initial classification on useful label.