Visual Analytics Researcher

Helping users extract Insights from data

Do they look like this?

What you really
need to know

Insights!

What are insights?

• Deep understanding
• Meaningful
• Non obvious
• Actionable
• Based on data

Ingredients

• Data
• Tasks
• Users 🫵🏼

How to extract insights

My insights toolbelt?

What do I use?

• Visual Analytics
• Trees
• Graphs
• Photos
• Accesibility
• Big Data

Navio

You get a new dataset
what do you do?

Moma Dataset

You have these ideas (Tasks)

• What's in the data, what are the main characteristics of the attributes and the rows? (Summarize)
• How many art pieces come from the US? How many of those are photos? (Navigate)
• What are some interesting and unexpected distributions? (Explore)

State of the art: SPLOMS

State of the art: Parallel Coordinates

Navio vs Parallel Coordinates

Navio vs SPLOM

Navio and Machine Learning

UMAP Playground

MLExplore.js

• Interpret and interact with TSNE+Kmeans

Trees

Work with Michael L. Pack, Audra Buck-Coleman, Catherine Plaisant, Ben Shneiderman.

[TRC2015] [Vis2013][Thesis2013] [TRB2013] [DRS2013] [Vis2011] [CHI2011] [TRB2011]

Treeversity Architecture

• DB Postgresql. Millions of rows
• Backend Python-Django
• Rest API
• Frontend HTML5 + Js + D3.js + Crossfilter. Thousands of rows

Responsive interactions allow for visual analytics

TreeVersity2 Demo

Trees

Networks

Work with Aaron Wilson, Juan Liu, Dan Davies, Peter Jarvis and Eric Bier

[IEEEVIS_VIP2016] [AIMagazine2016] [AVI2016] [IAAI2015]

Plenty of work has been done on the field

Our objectives

• Apply state of the art of netViz
• Report what worked (and didn't) in our industry setting
• We open sourced some of our components for the community to reuse

Some netviz tools

• NetworkX
• Gephi
• Cytoscape
• Pajek
• NetworkX
• Many others

Why develop another tool?

• Integration
• Web-based
• Scalability
• Industry

Best practices

Customer's needs

• Get an overview of the network
• Explore nodes of interest's ego-networks on demand

Best practice #1: offer overviews and ego-network modes

Overview mode

Too much information

Rank-by-relevance framework

Best practice #2 filter and display the most important nodes

Rank-by-relevance framework 2

Best practice #3 include neighborhoods in the selection

NetClustering.js

Force-in-a-box

Best practice #4 Cluster and separate groups

Jump into clusters

Jump into clusters

Jump into clusters

Jump into clusters

Best practice #5 Allow users to filter by cluster

Overview

Ego-network view

Ego-distance mode

Best practice #7 Allow node expansion on demand

Node pining

Best practice #8 Allow custom layouts

Validation

Fraud, waste and abuse

Police chat networks

Drug latent communities

Network Explorer Architecture

• Java Backend+Frontend (already in place)
• Javascript plugin tightly connected
• Application in production
• Millions of rows, dozens of users

Contributions

• Best practices
• Node Navigator
• Open Sourced force-in-a-box layout
• Open Sourced in-browser clustering
• Rank-by-relevance framework

Current work

US Senate Voting Patterns

How do Colombian Senators vote?

Photos

Work with Cati Boulanger, Sanjay Kairam, Frank Bentley, Joseph ‘Jofish’ Kaye, David A. Shamma

[AVI2016_2] [CHI2016] [CHI2016_2]

The problem

How to display groups of photos?

What if the photos have also statistics?

What are the best practices for representing photos with associated statistics with treemaps?

Experiment 1:

Photo count

Ex1. Demo

We measured

1. Time to find photo
2. User preference

Ex1. Demographics

154 Mechanical Turk Responses (67 women)

Ex1. Results

Avg time to find depending on photo position

Photo count sweet spot

For a 390x440px phototreemap

Experiment 2:

Layout

Demo

We measured

1. Time to find photos
2. Number of errors
3. Accuracy
4. User preference

Ex2. Demographics

63 responses (27 women)

Time to find photo (ms)

Errors

Accuracy

User preference

81.3% of the participants preferred the scrollable layout

Experiment 3:

User Study

Design

• 7 participants
• Users viewed their own photos
• 5 point likert scale on usefulness and pleasingness for 3x2 configurations
• Labels vs no Labels
• Size and sorting by views, comments or favorites

Configuration preference

Conclusions

• Small number of photos is better but not too small
• Cropping and aspect ration has a big impact
• Scrolling can be preferable to cropping
• Labels are important

Future/Current Work

• Explore cropping compensation techniques
• Usage of treemap deeper levels for categorization
• Zoomable techniques

Future/Current work

Zoomable Photo Treemap

Photo widget to show many photos at once with numeric values and categorical groupings

Flickr's Magic View

PhotoRing

How to better allow exploration and navigation of millions on photos

Other photo experiments

Photo Ranking

BTactile

How to better enable people to support blind students in learning graphics

Other topics

• Time-based Visualization
• Network Visualization
• Social Networks
• Real time/incremental visualization
• Big Data visualization

Tactile Graphics

• https://BTactile.com
• Machine Learning for Tactile Graphics
• DUTO.org - IRIS

PhD

Silicon Valley