liveBook · Manning

Label shift

As explained in Build AI Drug Discovery Pipelines

Label shift is a distribution shift in which the distribution of target values differs between training data and evaluation or application data.

Label shift occurs when target-value distributions differ between training data and unseen application data. In solubility modeling, training only on highly soluble molecules can leave a model unprepared for molecules spanning the full solubility range. section 4.1.2

Inspect target distributions before modeling. Similar solubility distributions in training and validation data suggest no visible shift there, while a broader application range may require models that extrapolate better. section 4.1.2

Figure 4.1 Comparison of solubility distributions for train and validation data sets.

Compare solubility distributions across training and validation data to identify potential target-distribution discrepancies. section 4.1.2

How Build AI Drug Discovery Pipelines covers it

How deep it goes
Defines label shift with a solubility example and connects it to extrapolation, but stops short of mitigation algorithms.
Examples
Uses highly soluble training molecules versus broadly soluble application molecules as the central example.
What it assumes
Assumes basic training, validation, test, and target-value terminology while defining distribution shift through solubility.
What you can do after
Evaluate whether target distributions align before selecting models for solubility prediction.
How it connects
Connects label shift with distribution shift, solubility prediction, validation data, and model extrapolation.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Build AI Drug Discovery Pipelines. The section links point to the parts of the book it draws on.