Tessera: Discretizing Data Analysis Workflows on a Task Level
Jing Nathan Yan Ziwei Gu Jeffrey M. Rzeszotarski
Cornell University
jy858@cornell.edu
zg48@cornell.edu
jeffrz@cornell.edu
Exploratory Visual Analysis(EVA) is NOT EASY
Massive Data
Data analysts
…and tracking the process can be even harder!
SESSION
Goal1
Goal2
Subgoal
Subgoal
…
Subgoal
…
e1
e2
e3
e4
…….
ek
e5
How do we map log events to (sub)-goals?
Filter Year>= “2000”, Count number of movies, plot histograms of Counts setting X-axis=distributor.
e1
Task-level analyses of log data reveal useful insights, but logs are complex, heterogenous, and large in scale
Our new approach, Tessera, employs data, cognitive and temporal features to discretize analyst log data
Filter Year>= “2000”, Count number of movies, plot histograms of Counts setting X-axis=distributor.
e1
e2
e3
e4
e1
en
…….
ek
….
e5
SIMILARITY
Session Event Log
Filter Year>= “2000”, Count number of movies, plot histograms of Counts setting X-axis=distributor.
e1
e2
e3
e4
e1
en
…….
ek
….
e5
SIMILARITY
Our new approach, Tessera, employs data, cognitive and temporal features to discretize analyst log data
Session Event Log
Tessera outperformed two competing approaches in two different datasets gathered from think-aloud sessions
In cases where users were analyzing both small and large datasets, Tessera was able to identify meaningful segments of user-directed activity.
Contributions
Tessera Framework
Abstract EVA event logs into goal-directed segments
Combine data transformation, cognitive and temporal features
Robust to different datasets and parameters settings.
A New Benchmark dataset
A new dataset combining think-aloud reports from participants
Contains more EVA operations and recruited more participants
Larger and collected through a more commonly used EVA platform.