ssepy: A Library for Efficient Model Evaluation through Stratification, Sampling, and Estimation in Python
Given an unlabeled dataset and model predictions, how can we select which instances to annotate in one go to maximize the precision of our estimates of model performance on the entire dataset?
The ssepy package helps you do that! The implementation of the ssepy package revolves around the following sequential framework:
- Predict: Predict the expected model performance for each example.
- Stratify: Divide the dataset into strata using the base predictions.
- Sample: Sample a data subset using the chosen sampling method.
- Annotate: Acquire annotations for the sampled subset.
- Estimate: Estimate model performance.
See our paper here for a technical overview of the framework.
In order to intall the package, run
pip install ssepyAlternatively, clone the repo, cd into it, and run
pip install .You may want to initialize a conda environment before running this operation.
Test your setup using this example, which demonstrates data stratification, budget allocation for annotation via proportional allocation, sampling via stratified simple random sampling, and estimation using the Horvitz-Thompson estimator:
import numpy as np
from sklearn.cluster import KMeans
from ssepy import ModelPerformanceEvaluator
np.random.seed(0)
# Generate data
total_samples = 100000
true_performance = np.random.normal(0, 1, total_samples) # Ground truth
# Unobserved target
print(np.mean(true_performance))
annotation_budget = 100 # Annotation budget
# 1. Proxy for ground truth
proxy_performance = true_performance + np.random.normal(0, 0.1, total_samples)
evaluator = ModelPerformanceEvaluator(proxy_performance=proxy_performance, budget=annotation_budget) # Initialize evaluator
# 2. Stratify on proxy_performance
evaluator.stratify_data(clustering_algorithm=KMeans(n_clusters=5, random_state=0, n_init="auto"), features=proxy_performance) # 5 strata
# 3. Allocate budget with proportional allocation and sample
evaluator.allocate_budget(allocation_type="proportional")
sample_indices = evaluator.sample_data(sampling_method="ssrs")
# 4. Annotate
sampled_performance = true_performance[sample_indices]
# 5. Estimate target and variance of estimate
estimate, variance_estimate = evaluator.compute_estimate(sampled_performance, estimator="ht")
print(estimate, variance_estimate)For the difference estimator under simple random sampling, run
evaluator = ModelPerformanceEvaluator(proxy_performance=proxy_performance, budget=annotation_budget) # initialize sampler
sample_indices = evaluator.sample_data(sampling_method="srs") # 2. sample
sampled_performance = true_performance[sample_indices] # 4. annotate
estimate, variance_estimate = evaluator.compute_estimate(sampled_performance, estimator="df") # 5. estimate
print(estimate, variance_estimate)The difference estimator is also implemented in the ppi_py package. They implement the prediction-powered estimator that corresponds to the difference estimator in case of mean estimation. Their package has more functionalities than what we offer, so check it out if you're interested in using this estimator.
The repo comes with a series of examples contained in the examples folder:
sampling-and-estimation.ipynbfor an example on how to stratify, sample, and estimateoracle-estimation.ipynbon the computation of the (oracle) efficiency of the estimators under various survey designs, assuming we had access to all ground truth variables. This file contains the core part of the code underlying the results in the paper
The supported sample designs are: (SRS) simple random sampling without replacement, (SSRS) stratified simple random sampling without replacement with proportional and optimal/Neyman allocation, (Poisson) sampling. All sampling methods have associated (HT) Horvitz-Thompson and (DF) difference estimators.
Feel free to reach out if you find any bugs or you would like other features to be implemented in the package.