Please enable JavaScript.

Coggle requires JavaScript to display documents.

ML Solution (Software Options (Python (Is it possible to get the already…

- - - - There are several topics that appear in the topicDictionary.txt file which are not associated with any articles in training data. For example, tunisiaattack2015, turkeycoupattempt, sanbernardinoshooting, parisattacks, orlandoterrorattack.
        Not surprising when these relate to events that occurred after the period covered by training data!
        Can you confirm if unseen tags like these will occur in the test data and form part of accuracy score calculations?
        If so, any thoughts to share on applying a supervised learning approach in this setting?
    - - In a developing crisis situation any historical training data is inevitably ‘stale’ to some extent. Current new topics of interest may not have been seen in the past, or seen in a context whereby the articles referring to them were tagged with different topics. Handling this real-world complexity of document classification is an essential part of the challenge and therefore we will be calculating the F1 score across all of the topics and test articles whether topics are associated with training articles or not
      - Clearly, the absence of training examples for some topics will impact a fully supervised learning approach. However, you may wish to consider the relationship between the topics that have no training examples and the topics that do have training examples. You may also wish to consider augmenting a supervised learning approach with unsupervised methods such as clustering of documents, term matching and/or keyword extraction.
- - - - How should training data be used?
        
        Proportion for use and a proportion for testing?