User permissions required: 'View Sources' AND 'Review and label
What is 'Balance' and why is it important?
The Balance rating presented in the Model Rating in Validation is a reflection of how balanced the reviewed data (i.e. the training data) in a dataset is, when compared to the dataset as a whole.
It takes into account a number of contributing factors (as shown below), including:
- The similarity of the reviewed data to the unreviewed data (shown as a percentage score)
- The proportion of reviewed data that has been reviewed through random sampling (i.e. 'Shuffle' mode)
- The proportion of data that has been reviewed using 'Rebalance' (see below for detail)
- The proportion of data that has been reviewed whilst using 'Text search'
An example 'Balance' component of the Model Rating
It's important that the proportion of data reviewed through random sampling is high (ideally 20%+) and the proportion of reviewed data labelled using search is low.
The balance rating is most heavily influenced, however, by the similarity score that measures the similarity of the unreviewed data to the reviewed data.
This similarity score is calculated by a proprietary labelling bias model that compares the reviewed and unreviewed data to ensure that the labelled data is representative of the whole dataset. If the data is not representative and has been labelled in a biased manner, model performance measures can be misleading and potentially unreliable.
Labelling bias in the platform is typically the result of an imbalance of the training modes used to assign labels, particularly if too much 'text search' is used and not enough 'Shuffle' mode. It can still occur, however, even if a high proportion of 'Shuffle' mode is used. Training specific labels in modes like 'Teach label' can naturally lead to a slight imbalance in the reviewed data. The platform helps you identify when this happens and helps you address it in a quick and effective way.
What is 'Rebalance' and how do you use it?
'Rebalance' is a training mode that helps to reduce the potential imbalances in how a model has been labelled, i.e. labelling bias, which mean that the reviewed data is a not as representative of the whole dataset as it could be.
The 'Rebalance' training mode shows verbatims that are underrepresented in the reviewed set.
Labelling the verbatims (as you would in any other training mode) presented in this mode will help address imbalances in the training data and improve the model's balance score.
Top Tip: Rebalance is typically most effective when used little and often. Labelling a small number of verbatims (between 10 and 20) in this mode and allowing the model to retrain before refreshing and labelling more examples is the best way to maximise the impact it will have on the model's balance score.
The 'Rebalance' training mode on a demo dataset
If you find that you have a high similarity score but the Balance rating is still low, this is likely because you have not labelled enough of the training data in 'Shuffle' mode. If this is the case, the platform will suggest labelling a random selection of verbatims as the priority recommended action. Training in this mode gives the platform additional confidence that the dataset has not been labelled in a biased manner and that the training data is a representative sample.
How much 'Rebalance' should I use?
You should continue to use 'Rebalance' iteratively to improve the similarity score for your model, which will in turn increase your 'Balance' rating.
Once this reaches a 'Good' rating in Validation, it is up to you how much more you would like to increase the similarity score before stopping training in 'Rebalance'.
You can aim to optimise this rating as much as possible, but continued training will always be a case of diminishing returns. A 'Good' rating should typically be considered an acceptable level of performance for a good model.
Previous: Understanding and increasing Coverage | Next: When to stop training your model