Please enable JavaScript.
Coggle requires JavaScript to display documents.
Automating Machine Learning (Criteria for AutoML excellence (resource…
Automating Machine Learning
ML & its Life Cycle
not always linear - may need to adjust previous conclusions
ML can outperform even teams of data scientists
necessity to understand ML conceptually
AutoML is not automatic - requires subject area expert to identify target, problem, establish parameters
saves time on process design, improves efficiency
AirBnb Example
4 areas where AutoML had positive impact on productivity
feature engineering
algorithm selection , hyperparameter tuning
exploratory data analysis
model diagnostics
Tools and Platforms
context-specific tools
implemented within another system for specific purpose
can be embedded within existing sales platform
general platforms
designed for general purpose ML
open-source
developed by and for computer scientists/data engineers
usually requires Python/R experience
commercial
usually provided for a price
datarobot is example of this
Criteria for AutoML excellence
understanding and learning
improve analysts understanding of concept
ML system should explain process
system should support thinking around business decisions
resource availability
ML needs ability to access existing data within organization
allow easy use of the resulting model generated by ML
support for ML system should be available
ease of use
easily understandable, fast integration to existing platform
analyst needs all prerequisite knowledge
visualization intuitive
process transparency
needs to be possible to drill into ML system, analyze decisions it made
without transparency, understanding/learning criteria becomes difficult
productivity
reaching accurate conclusions, explaining them more quickly
generalizable across contexts
AutoML should work for all target types, data
sizes, and different time perspectives
system should be capable of
handling small, medium, and big data
handle both cross-sectional and longitudinal data
accuracy
most important criteria
selecting best features/models to use for situation
recommend actions
ML needs to "know" context of situation for model