Principle:Snorkel team Snorkel Slicing Function Definition

Knowledge Sources	Slice-based Learning: A Programming Model for Residual Learning in Critical Data Slices Snorkel Slicing Tutorial
Domains	Data_Slicing, Robustness, Model_Evaluation
Last Updated	2026-02-14 20:00 GMT

Overview

A mechanism for programmatically defining critical data subsets (slices) that require special attention during model training and evaluation.

Description

Slicing Function Definition extends the labeling function paradigm to identify critical data subsets rather than assign labels. A slicing function (SF) takes a data point and returns a binary indicator: 1 if the data point belongs to the slice, 0 if not.

Slices represent subsets of data where model performance is particularly important, such as:

• Short text messages in a spam classifier
• Rare categories in a product classifier
• Edge cases identified by domain experts

The SlicingFunction class inherits directly from LabelingFunction, reusing its preprocessing and resource infrastructure but with binary output semantics.

Usage

Use this principle when you need to identify and monitor model performance on critical data subsets. Define slicing functions to capture subpopulations where failures are costly, where data is underrepresented, or where domain-specific patterns require special handling.

Theoretical Basis

A slicing function $s_j$ defines a binary partition:

$s_j:{𝒳}\to \{0,1\}$

where $s_j(x)=1$ indicates membership in the slice. The collection of slicing functions creates a slice matrix $S\in \{0,1{\}}^{n\times p}$ where $p$ is the number of slices.

Unlike labeling functions, slicing functions:

• Do not abstain (always return 0 or 1)
• Define overlapping subsets (a data point can be in multiple slices)
• Are used for model conditioning, not label generation

Related Pages

Implemented By

• Implementation:Snorkel_team_Snorkel_SlicingFunction_Init