Principle:Snorkel team Snorkel Slicing Function Application

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

Overview

A process for applying slicing functions to a dataset to produce a slice matrix encoding membership of each data point in each defined slice.

Description

Slicing Function Application executes all defined slicing functions against a dataset, producing a slice matrix $S\in \{0,1{\}}^{n\times p}$ where each entry indicates whether a data point belongs to a particular slice. The output is a NumPy record array with named columns corresponding to each slicing function.

This step reuses the LF applier infrastructure (PandasLFApplier) but produces a record array instead of a plain array, preserving slice names as column metadata for downstream use.

Usage

Use this principle after defining slicing functions and before initializing a slice-aware classifier. The slice matrix is required for both data preparation (adding slice labels to dataloaders) and evaluation (scoring per-slice performance).

Theoretical Basis

Given $p$ slicing functions $\{s_1,\dots ,s_p\}$:

$S_{i,j}=s_j(x_i)\in \{0,1\}$

The slice matrix $S$ defines overlapping subsets. Unlike the label matrix, $S$ has no abstentions; every data point is classified as in or out of each slice.

Related Pages

Implemented By

• Implementation:Snorkel_team_Snorkel_PandasSFApplier_Apply