3. Our recent work with RS interfaces:
System Type API
SmallWorlds Music / Movies Facebook
TasteWeights Musical Artists / Jobs Facebook, Twitter,
DBPedia, LinkedIn
TopicLens Twitter users and topics / Static / Twitter API
Movies
WigiPedia Semantic Labels DBPedia / MediaWiki
TopicNets People, Documents, PDF Documents /
Topics Structured RDF
documents
4. Inspectability and Control Elements:
System Inspectability Control
SmallWorlds Column Graph, Circular Graph, Node-repositioning
List View Drop-down menus
TasteWeights List Views, Slider positions, Item/user sliders, Locks,
Background Opacity, On-hover domain sliders.
edges, Provenance view for re-
ranking
TopicLens Graph and River View, 3D view, Side panel controls (buttons
Many on-hover actions. Zoom and sliders).
Graph “spinning”, node clicks,
Sorting. ( UI only No data-
level controls )
WigiPedia Wiki Page, Node-link graph, Node selection (click). Button
Pop-up list views, edge panel.
highlighting, tabular view.
Node dragging (interpolation
technique)
TopicNets Graph view: Zoom, Click, Drag, Huge amount of control. 10+
List views, Table views, Charts panels of functions. Full
graph interaction, Layout
algorithms etc.
5. Inspectability Elements:
Inspectability Mechanism Advantage Disadvantage
Node-Link Graph Good provenance. Easy to Scales badly, gets cluttered
inspect paths, neighbor links quickly (abstraction /
etc clustering can help)
List Views Simple, can be reranked with Hard to display connectivity
provenance annotations.
Interactive (hover, click, zoom Can handle lots of information. Hidden functionality. Usually
etc) Create a “game-like” feel. needs some training / learning
Keep user interested. curve, or good
annotation/help tools
Tabular Views Easier to understand than a Hard to display complex
graph. connectivity / provenance
Text-based Simple, Lots of detail Boring? Does not scale well.
6. Control Elements:
Control Mechanism Advantage Disadvantage
Node-Link Graph (rating using Communicates impact of user Not initially intuitive, difficult
node-drags) input very well to rerank vertically (crossed
edges)
Node-Link Graph (for data Very useful for selecting a Edges cause clutter quickly
selection) subset from a general esp. for large graphs.
overview
Slider List Views Clean look, Users are familiar Difficult to resize, less freedom
with slider input, can be than node-link views.
reranked easily with
provenance data shown
Right-click Useful for node-specific Hidden functionality. Usually
functionality needs some training / learning
curve, or good
annotation/help tools
Control panels (buttons, Easier to understand than a Can get cluttered quickly
sliders etc) graph, can be labeled more depending on system
easily. complexity.
7.
8. TopicLens: Exploring Content and Network Structure in
Parallel
(Devendorf, O’Donovan, Hollerer)
Hybrid Network Views
River and Graph
representations displayed
in parallel.
River shows statistical
(topic modeled)
information about
network selections.
Graph view shows
information about the
underlying network.
9. Visual Analysis of Dynamic Topics and
Social Network in Parallel
Network Data Sources (APIs
etc)
LDA over Topology
Content Analysis
LDA Topics Sub-Networks
Sub-Networks
Graph View
River View
10. TopicNets: Exploring Topic Relations in
Social Networks
LDA “Topic Models” useful
for understanding relations in
large volumes of text.
Visualization and Interaction
can help a user gain insights
into topic modeled data.
LDA can be iteratively
applied to tailor the
information space to a users
requirement.
Gretarsson, O’Donovan et al.
2011 (ACM Trans. On the Web)
11. TopicLens is a General solution: New
York Times Article Example
Currently integrating this tool with many different credibility models- crawled 7 topic-specific data sets from twitter.Social credibility model, with many different components (eg: Retweet analysis)Content-based credibility mode (using NLP)Bayesian model built on ground truth collected data.
Map different network facets onto different components of the visualization
TopicNets: Previous WorkTopic modeled view of 10,000 research papers from UCSD / CALIT2Clustering of colored nodes reflects the clustering of topics from within the published research. Interestingly, it mirrors the departmental structure of the institution quite well.
We have tested this system on many data sets. Examples include:News Articles (New York Times), A collection of awarded NSF grant proposals. The US Health Bill, A PHD Thesis, A collection of three visualization conferences proceedings over three years. Twitter data (Libya, LondonRiots, and Obama data sets). Facebook preference data for 15,000 users.
Concept of this interface can be generalized. Any “credibility” metric can be applied.This example is from topic modeled data from TwitterNOTE: User 21 appears “credible” wrt the selected topic “wikileaks”. However, this “credibility” is not backed up by his friends and follower groups on TwitterUser 9 however, appears credible and has this credibility backed up by the underlying network.
User-centric or topic-centric views:Military relevant example: By switching to a user-centric view, we can perform interesting searches over the microblog network. Lets say a user shows interest in a “remote garage door opener”. This appears completely normal, until the system highlights that this user appears prevelant in the context of the topic “home made explosives”. At this point, an alarm is raised and an analyst uses the interface to examine other topics from the target user’s underlying network. This form of visual-interactive topic-based analysis is unique to TopicLens.
River view shows average preference for item I across the selected user’s friend group
Provides a quick and easy analysis of the distribution of topic popularity