plugml

easy-to-use and highly modular machine learning framework based on scikit-learn with postgresql data bindings

why plugml

• focus on the main problems, i.e. the machine learning, not the data munging
• supports many use cases out of the box, e.g. document retrieval, document similarity
• simply define the structure of your SQL table and automatically get your data
• no boilerplate code anymore for handling your numerical, categorical and text features
• high modularity allows to easily plug your custom algorithms; no fighting against the framework
• plugml is used for prototyping models for our production app, so it is part of the lifecycle of an actually used product

features

• directly load data from a SQL table, no more boilerplate code
• easy data preprocessing and filtering with user-defined functions
• out-of-the-box support for numerical, categorical and text features
• high performance as it utilizes scikit-learn, sparse scipy matrices, ...
• plug your custom feature extractors with minimal effort
• automatic feature scaling and imputation; custom weights
• dimensionality reduction supporting very large features
• kNN queries

roadmap

v0.3

• theanets autoencoders as compressors (feed-forward)

v0.4

• theanet recurrent autoencoder
• use recurrent autoencoder for nlp instead of bag of wards (or rather provide option)

v0.5

• introduce labels to data
• infrastructure for supervised ml

v0.6

• theanets rnns for regression/classification

long-term ideas

• automatic data extraction based on sql schema
• intelligent application of ml techniques based on data structure
• auto-identify semantic entities and relationships from data structures
• (match natural language queries to ml-specific actions, i.e. non-ml specialist interface to data analysis)

get started (to be shown in ipynb)

1. load data with dao
2. define data
3. define features
4. combine in extractor
5. compress features
6. load in KNN

notes

subject to signifant changes until at least v1.0