Skip to content
isaacmg edited this page Aug 31, 2020 · 27 revisions

Project Utility

  • Provide users an easy to use platform that contains recent research models applicable to any time series task.
  • Allow users to easily track experiments and explore best hyper-parameters.
  • Enable users to easily combine multiple modalities of data to improve performance.
  • Facilitate explainability with easy to use dashboard so both ML users and non-technical stakeholders can gain insights into model predictions.

Current Objectives

Objective Overview Priority Related Issues
Meta-data incorporation We want to enable users to incorporate static meta-data into forecasts. To do this we need to add modules for auto-encoders, semi-supervised embeddings techniques as well as model synthesis Medium
  • Meta-Data Embedder Training Module
  • Meta-Data embedder fusion params
  • Deployment/extended evaluation interface We want to allow users to deploy models easily to production environments without having to refactor outputs. We also want to enable continuous evaluation as new temporal data appears. High
  • Inference mode
  • Continue Evaluation
  • Platform Integration We need to enable integration with other commonly used platforms such as Amazon Web Services, Microsoft Azure and GCP High
  • Dataverse and G-Cloud integration
  • Multi-task learning support Multi-task learning often enables more robust models that can address multiple business needs at once. Medium
  • Add multitask support in decoder
  • Adding additional base models There are tons of time series forecasting models available. Currently, our repository only has a fraction. Medium
  • Add XC Attention Model
  • Refactor DA-RNN
  • Adding more loss functions There are tons of useful loss functions that can increase performance at forecasting/classification Medium
    Documentation Documentation is essential for any widely used repo. We want to add easy to use docs to enable user of flow to easily use and extend our repo to suit their business and research needs. Medium
    Interpretability We want to make predictions interpretable to various stake holders (both technical and non-technical) High
    Bug Fixes Obviously existing bugs that lead to incorrect results or unexpected behavior should be fixed. Very High
  • Fix DA-RNN GPU Bug
  • Long term outlook