Code for VLDB 2019 paper Snuba: Automating Weak Supervision to Label Training Data
Reef is an automated system for labeling training data based on a small labeled dataset. Reef utilizes ideas from program synthesis to automatically generate a set of interpretable heuristics that are then used to label unlabeled training data efficiently.
Reef uses Python 2. The Python package requirements are in the file requirements.txt. If you have Snorkel, can set a flag here as True but there is a simple version of learning heuristic accuracies in this repo as well.
The inputs to Reef are the following:
- A labeled dataset, which contains a numerical feature matrix and a vector of ground truth labels (currently only supports binary classification)
- An unlabeled dataset, which contains a numerical feature matrix
The following is the overall workflow Reef follows to label training data automatically. The overall process is encoded in [1] generate_reef_labels.ipynb and the main file program_synthesis/heuristic_generator.py
- Using the labeled dataset, Reef generates heuristics like decision trees, or small logistic regression models. The synthesis code is in
program_synthesis/synthesizer.py.- A heuristic is generated for each possible combination of
cfeatures, wherecis the cardinality. For example, withc=1and 10 features, 10 heuristics will be generated. - For each generated heuristic, a beta parameter is calculated. This represents the minimum confidence level at which the heuristics will assign a label. This is done by maximizing the F1 score on the labeled dataset.
- A heuristic is generated for each possible combination of
- These heuristics are passed to a pruner that selects the best heuristic by maximizing a combination of the F1 score on the labeled dataset and diversity in terms of how many points it labels that previously selected heuristics don’t.
- The selected heuristic and previously chosen heuristics are then passed to the verifier which learns accuracies for the heuristics based on the labels the heuristics assign to the unlabeled dataset.
- Finally, Reef calculates the probabilistic labels the heuristics assign to the labeled dataset and pass datapoint with low confidence labels to the synthesizer. We repeat this procedure in an iterative manner.
The tutorial notebooks are based on a text-based plot classification dataset. We go through generating heuristics with Reef and then train a simple LSTM model to see how an end model trained with Reef labels compares to an end model trained with ground truth training labels.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent