AlphaD3M is an AutoML system that automatically searches for models and derives end-to-end pipelines that read, pre-process the data, and train the model. AlphaD3M leverages recent advances in deep reinforcement learning and is able to adapt to different application domains and problems through incremental learning.

AlphaD3M provides data scientists and data engineers the flexibility to address complex problems by leveraging the Python ecosystem, including open-source libraries and tools, support for collaboration, and infrastructure that enables transparency and reproducibility.

This repository is part of New York University's implementation of the Data Driven Discovery project (D3M).

Support for Many ML Problems

AlphaD3M uses a comprehensive collection of primitives developed under the D3M program as well as primitives provided in open-source libraries, such as scikit-learn, to derive pipelines for a wide range of machine learning tasks. These pipelines can be applied to different data types and derive standard performance metrics.

Learning Tasks: classification (semi-supervised, binary, multiclass, and multi-label), regression (univariate, and multivariate), time series (forecasting, hierarchical forecasting, and classification),  image-based problems (object detection, remote sensing, and image recognition), graph-based problems (collaborative filtering, community detection, graph matching, link prediction, and vertex classification),  multi-instance learning and clustering.

Data Types: tabular, time series, hierarchical (grouped, multi-index) time series, geospatial, images, multi-spectral imagery, relational, text, graph, audio, video.

Data Formats: D3M, raw CSV, raw text files, OpenML, and scikit-learn datasets.

Metrics: accuracy, F1, macro F1, micro F1, mean squared error, mean absolute error, root mean squared error, object detection AP, hamming loss, ROC-AUC, ROC-AUC macro, ROC-AUC micro, jaccard similarity score, normalized mutual information, hit at K, R2, recall, mean reciprocal rank, precision, and precision at top K.

Installation

You can use AlphaD3M through d3m-interface.  d3m-interface is a Python library to use D3M AutoML systems.

Docker/Singularity containers (full version)

AlphaD3M and the D3M Core will be deployed as a container. This version works with Python 3.6 through 3.8, and supports all the ML tasks and data types mentioned above. You need to have Docker or Singularity installed on your operating system.

For this version, you just need to install d3m-interface:

$ pip install d3m-interface

The first time d3m-interface is used, it automatically downloads a Docker image containing AlphaD3M and the D3M Core.

PyPI (lightweight version)

Currently, this version has support for classification and regression tabular problems. This package works with Python 3.8.

To install, run these commands:

$ pip install d3m-interface
$ pip install alphad3m
$ pip install d3m-common-primitives d3m-sklearn-wrap dsbox-corex dsbox-primitives sri-d3m

The last command installs the primitives available on PyPI.

⚠ If you have issues with the installation of 'opencv-python-headless' (dependency of d3m-common-primitives). We would recommend you to install it directly from its Python wheels. ⚠

Documentation

The documentation of our system can be found here. To help users get started with AlphaD3M, we provide Jupyter Notebooks in our public repository that show examples of how the library can be used. We also have documentation for the API.

Usability, Model Exploration and Explanation

AlphaD3M greatly simplifies the process to create predictive models. Users can interact with the system from a Jupyter notebook, and derive models using a few lines of Python code.

Users can leverage Python-based libraries and tools to clean, transform and visualize data, as well as standard methods to explain machine learning models.  They can also be combined to  build customized solutions for specific problems that can be deployed to end users.

The AlphaD3M environment includes tools that we developed to enable users to explore the pipelines and their predictions:

PipelineProfiler, an interactive visual analytics tool that empowers data scientists to explore the pipelines derived by AlphaD3M within a Jupyter notebook, and gain insights to improve them as well as make an informed decision while selecting models for a given application.

Visual Text Explorer, a tool that helps users to understand models for text classification, by allowing to explore model predictions and their association with words and entities present in the classified documents.

How AlphaD3M works?

Inspired by  AlphaZero, AlphaD3M frames the problem of pipeline synthesis for model discovery as a single-player game where the player iteratively builds a pipeline by selecting actions (insertion, deletion and replacement of pipeline components). We solve the meta-learning problem using a deep neural network and a Monte Carlo tree search (MCTS). The neural network receives as input an entire pipeline, data meta-features, and the problem, and outputs action probabilities and estimates for the pipeline performance. The MCTS uses the network probabilities to run simulations which terminate at actual pipeline evaluations. To reduce the search space, we define a pipeline grammar where the rules of the grammar constitute the actions.  The grammar rules grow linearly with the number of primitives and hence address the issue of scalability. Finally, AlphaD3M performs hyperparameter optimization of the best pipelines using SMAC.

For more information about how AlphaD3M works, see our papers:

AlphaD3M: Machine Learning Pipeline Synthesis

VIDA-NYU / d3m / AlphaD3M
GitLab.com