Vowpal Wabbit Documentation

Vowpal Wabbit is a high-performance machine learning system designed for online learning. Key capabilities include:

• Flexible Input Format: Supports free-form text interpreted via bag-of-words and multiple namespaces.
• Speed: Optimized for fast online learning, with sparse gradient descent (GD) as the baseline.
• Scalability: Features a bounded memory footprint independent of data size and uses the hashing trick to bound the feature set size.
• Feature Interaction: Supports internal pairing of feature subsets to enable linear complexity on cross-products, which is particularly useful for ranking problems.
• Reinforcement Learning: Includes implementations for contextual bandit algorithms.

The Wikipara dataset is a generation scheme designed to test model performance under varying class counts and training examples per class. It uses Wikipedia pages as classes and paragraphs from those pages as examples.

Structure:

• Class: Defined by the Wikipedia page ID.
• Training Examples: A subset of paragraphs taken from the page.
• Test Examples: A separate subset of paragraphs taken from the same page.

This dataset is particularly useful for testing scenarios with high cardinality (e.g., 10,000 classes) and extremely low data availability per class (e.g., 3 training examples and 1 test example).

To run the advertising demo using the Criteo Conversion Logs dataset, follow these steps:

1. Download the dataset from http://labs.criteo.com/downloads/2014-conversion-logs-dataset/ (you must agree to the terms and conditions).
2. Untar the downloaded file.
3. Move the data.txt file into the demo/advertising/ directory.
4. Execute the naive_baseline.py script to run the baseline model.

Demos covering neural networks and feature engineering include:

• dna/: DNA sequence classification with feature engineering.
• normalized/: Demonstrations of feature normalization effects.
• random-noise/: Robustness experiments using random noise features.

The MovieLens demo demonstrates different modeling approaches for the movielens-1M dataset, comparing linear models against various low-rank interaction configurations.

To run the full comparison suite, use the make shootout command. This will produce results for:

• linear: A baseline model with only user and item bias (no interactions).
• lrq: A linear model augmented with rank-7 interactions (e.g., --l2 1.25e-7 --lrq um7).
• lrqdropout: A linear model with rank-14 interactions and dropout (e.g., --lrq um14 --lrqdropout).
• lrqdropouthogwild: The lrqdropout configuration trained in parallel on multiple cores without locking (lock-free parallel learning).

Note: The first run of make shootout produces extensive output; subsequent runs will show cached results.

To build VW-WASM locally, you must install and activate the Emscripten SDK (emsdk). Follow these steps to clone, install, and activate the latest version:

1. Clone the emsdk repository.
2. Install and activate the latest version.
3. Source the environment variables.

git clone https://github.com/emscripten-core/emsdk.git
cd emsdk
./emsdk install latest
./emsdk activate latest
source ./emsdk_env.sh

Vowpal Wabbit provides several classification examples:

• dbpedia/: Text classification using the DBpedia ontology dataset.
• mnist/: Handwritten digit recognition using the MNIST dataset.
• ocr/: Optical character recognition demo.

To demonstrate Searn's performance on an entity-relation recognition task, use the provided make command. This command automates the following steps:

1. Downloads the preprocessed entity-relation dataset.
2. Trains a joint model for predicting entity and relation types.
3. Computes statistics on the test set.

For more information regarding the dataset used in this demo, refer to the CogComp resource page: http://cogcomp.cs.illinois.edu/page/resource_view/43

make er.perf

When modifying code or adding new tests, you may need to update the expected output files used for comparison.

• Update/save new expected outputs: make <testName>.expected
• Generate correct outputs for a test without actually running it: make <testName>.pretend

If you have a pre-trained candidate policy (even one trained outside of VW) and want to evaluate it against production bandit data, use the legacy --eval method.

1. Create an evaluation file (e.g., eval.dat).
2. For each production data instance, prepend the line with the action your candidate policy would have chosen.
   • Format: <candidate_action> <production_data_line>
   • Example: If production data is 1:2:0.5 | feat_a and candidate chooses action 2, write 2 1:2:0.5 | feat_a.
3. Run VW with the --eval flag.

Warning: VW will prevent you from using biased estimators like dm (direct method) with the --eval flag to avoid incorrect results.