1. Show suggestions and borderlines Hierarchical Clustering
Interactive Machine Learning at Scale with CHISSL
Dustin Arendt1, Emily Grace2, and Svitlana Volkova2
1Visual Analytics Group, 2Data Science and Analytics Group
Structured Input
Boston Dataset1
Task: housing price (regression)
14 hand engineered features
Visualized using radar glyphs
The distribution is the median housing price
CHISSL
Computer-Human Interaction
Semi-Supervised Learning
What?
CHISSL is an O(n) incremental transductive learning algorithm and user interface to:
Rapidly organize unlabeled data into groups defined by example
The model learns user intent and makes recommendations
The model can be corrected in a feedback loop.
The user can export the recommendations as labels to train the model
How?
Train inductive model
Export transduction
Show suggestions and borderlines
1-NN classifier
User Provides Label
Parent pointer array
Hierarchical Clustering
Representation
Client Loop
Server
Text Input
VAST Challenge 2014 text dataset2
Task: text exploration by selecting a keywords and coloring by topic
The distribution is the date of publication
Tip of the Iceberg
Only a few carefully chosen representative instances are shown per group to avoid overwhelming the user.
Drag and Drop
Users train the model by dragging an instance to the group they feel it is most similar to.
Responsive
The model learns from the users example after each user interaction and re-predicts within milliseconds.
User friendly
Users can choose what to group, how many groups there are, and what groups mean.
ABOUT
Pacific Northwest
National Laboratory
The Pacific Northwest National Laboratory,
located in southeastern Washington State,
is a U.S. Department of Energy Office of
Science laboratory that solves complex
problems in energy, national security, and
the environment, and advances scientific
frontiers in the chemical, biological, materials,
environmental, and computational sciences.
The Laboratory employs nearly 5,000 staff
members, has an annual budget in excess
of $1 billion, and has been managed by
Ohio-based Battelle since 1965.
For more information on the science
you see here, please contact:
Dustin Arendt
Pacific Northwest National Laboratory
Richland, WA 99352
(509)
Sequence Input
VAST Challenge 2014 GPS dataset2
Task: patterns of life, an icon encoding the predominant activity, e.g., home, work, food, each hour for 1 day
The distribution is the date the sequence occurred
Dynamic Graph Input
VAST Challenge Dataset2
Task: sub-network clustering
Instances are day x ego-network pairs
The features are the degree distribution of the sub-network
Why?
You can’t crowdsource every problem:
Requires domain expertise
Experts are rare and their time is valuable
Classes are well-defined or known ahead of time
Better than active learning because the user:
Chooses what to label
Can refine the task on the fly
Can understand model performance in real-time
Next?
Evaluation against Active Learning:
Computational
User study
Applications:
Trajectories & time series
Cybersecurity and insider threat
Engineering:
Enrichment of streaming data
File Name // File Date // PNNL-SA-#####
Datasets