2009 13th International Conference Information Visualisation
The NIH Visual Browser: An Interactive Visualization of Biomedical Research
Bruce W. Herr II, Edmund M. Talley, Gully A.P.C. Burns, David Newman, Gavin LaRowe
ChalkLabs, National Institute of Neurological Disorders and Stroke, Information Sciences
Institute, University of California Irvine/NICTA Australia, ChalkLabs
{, , , , }
Abstract analysts, program officers, and researchers an intuitive,
This paper presents a technical description of the easy-to-use tool for navigating and analyzing large-scale
methods used to generate an interactive, two- funding data.
dimensional visualization of 60,568 grants funded by the Although still in development, this tool currently
National Institutes of Health in 2007. The visualization is provides:
made intelligible by providing interactive features for • an orthogonal view of the entire collection
assessing the data in a web-based visual browser, see • a zoom-able component for viewing
http://www.nihmaps.org. The key features include deep iteratively smaller sub-sets of the collection
zooming, selection, full-text querying, overlays, color- • an interactive interface for selecting groups
coding schemes, and multi-level labeling. Major insights, of documents for analysis
broader applicability, and future directions are • a system of layers, overlays, color-coding
discussed. schemes, and labels
• selection components for various fields
Keywords--- information visualization, scientific such as institute, topic, or keyword
funding, topic modeling, information navigation.
2. Data
The data used for textual analysis were titles and
1. Introduction abstracts from NIH grants funded in 2007. In addition,
we extracted information on Principal Investigator and
Within successful biomedical research, there are two funding Institute or Center (from among 27 distinct NIH
elements that accompany every project: proposals and Institutes and Centers) for each grant. The data is
publications. Finding ways to analyze these elements in a publically available on the NIH website [2] (specific
framework that is intuitive and convenient for biologists project data available on request). After removing
is a very difficult undertaking. The scope and range of duplicate abstracts (~1% of the corpus) and grants with
on-going projects and related data sets are not easily no abstracts (~10%), we analyzed a total of 60,568
identifiable, let alone computable within large-scale abstracts associated with 25 separate Institutes.
analyses.
By providing a gestalt-like frame of reference and an 3. Technique
expeditious visual entry point for large information
spaces that is not available via textual reports, 3.1. Topic Modeling
spreadsheets, or database results, visualization is now
well established as both an academic discipline and a We used topic modeling [3] to analyze the NIH
tool for improving scientific research. It is already well abstracts, and to provide a robust means for computing
entrenched in the biomedical community where it is used similarities between abstracts. The topic model is a
to facilitate, enhance, and drive analytical and decision- state-of-the-art Bayesian statistical machine learning
making processes [1]. technique for modeling collections of text documents.
In this case study, we present a technical description The intuition behind the topic model is that individual
of an interactive mapping tool based on a comprehensive documents exhibit multiple concepts or topics. The topic
collection of abstracts from proposals funded by the model simultaneously learns a set of topics that together
National Institutes of Health (NIH). The data provides a describe a collection, and the mixture of topics within
significant view of federally funded biomedical research each document in the collection. These topics – which
extant within the United States. The primary purpose of are learned automatically from the data with no prior
this tool is to provide NIH administrators, portfolio
978-0-7695-3733-7/09 $25.00 © 2009 IEEE 511
505
DOI 10.1109/IV.2009.105
, knowledge – are usually semantically meaningful, and DrL is ideally suited for large graphs; meaningful
highly interpretable. results have been obtained on graphs with up to 2 million
While the topics themselves produce a useful nodes [23]. The input for DrL is a similarity network
taxonomy or ontology to categorize individual abstracts, where the weight of each edge in the network represents
they also provide a reliable basis for computing how similar the two connected nodes are. When DrL is
similarity between any two abstracts. We ran a total of run, it produces a 2d layout where nodes that are similar
six independent topic models, initially learning T=200, are placed close to one another. This typically results in
T=400 and T=600 topics (each topic setting was run local clustering of related nodes and a global structure of
twice from different random initial conditions). The related clusters.
topic models were learned using Gibbs sampling, a For this project, quite a bit of time was spent on
standard technique for this model [5]. For each of the six getting a good layout. This is due to the fact that there
runs, the Gibbs sampler was run for 500 iterations. At are so many variables in topic generation, grant-grant
the end of 500 iterations, very similar topics in each run distance computation (0 is identical, infinity is very
were manually merged using hierarchical agglomerative dissimilar), distance to similarity conversion (0 is
clustering, resulting in models having T = 183, 186, 343, dissimilar, 1 is very similar), and layout parameters. A
349, 529 and 535 topics. These six topic models were change in parameters at any of these stages could
run for a further 500 iterations of the Gibbs sampler to drastically alter the layout. Further, we wanted the layout
reach stationarity. to come out where major research areas are strongly
clustered with strands of grants between clusters
3.2. Document Similarity representing ‘bridges’ which consist of projects laid out
in a gradient between two topically defined research
The layout and visualization of the 60,568 grants is areas. These project bridges have the potential to add
completely based on pair-wise similarities computed for extra dimensionality to the layout, since their topical
each pair of grants. While we did compute all 2 billion content provides concrete information on the
similarities, we only kept similarities that were above a relationships between clusters. The layout we eventually
threshold. Since there is no gold standard for measuring chose for this project is shown in Figure 1.
similarity between two text documents, we took a multi-
resolution approach of computing a similarity score
based on general and specific similarity. This was
achieved by computing the similarity based on all six
topic model runs, with topic resolutions ranging from
183 to 535 topics. Averaging over six different topic
model runs also improves the similarity estimate.
The input to the graph layout algorithm is a list of
pairs of nodes with an associated weight, given by the
similarity. We computed the similarity between grant
abstracts di and dj using the average symmetric Kullback-
Leibler divergence function [6], computed using the
respective mixtures of topics Pr(t| di), and Pr(t| di):
Pr(t | d i )
KL* (di , d j ) = ∑ (Pr(t | di ) − Pr(t | d j ) )log
t Pr(t | d j )
This symmetric version (denoted by *) of KL-divergence
was converted to a [0,1] similarity for input into the
graph layout algorithm, which is described in the next
section.
Figure 1. Final layout of the NIH grants
3.3. Graph Layout - DrL
3.4. Visual Encoding
Layout is a critical factor in all visualization
applications, especially for large-scale data sets. For Making sense of the network layout without any
laying out the data, the Distributed Recursive Graph labeling, coloring, or sizing is nearly impossible. So, to
Layout (DrL) algorithm developed by Shawn Martin was help in understanding we chose to color the nodes
used for the data [7]. At its core, DrL is a force-directed according to the institute which funded the represented
layout algorithm based on VxOrd [8] and optimized for grant. Unfortunately, there are 25 different institutes in
large graphs. It offers both a parallel and sequential the dataset, so there are many colors which are too
version of the above force-directed algorithm, a recursive similar to one another to tell apart. To deal with this we
multi-level version for very large graphs, and the ability had to do several things. First, we chose colors such that
to add new nodes to a previously drawn graph. similar colors were as far apart as possible and near
colors used complimentary colors. While not perfect, this
506
512
The NIH Visual Browser: An Interactive Visualization of Biomedical Research
Bruce W. Herr II, Edmund M. Talley, Gully A.P.C. Burns, David Newman, Gavin LaRowe
ChalkLabs, National Institute of Neurological Disorders and Stroke, Information Sciences
Institute, University of California Irvine/NICTA Australia, ChalkLabs
{, , , , }
Abstract analysts, program officers, and researchers an intuitive,
This paper presents a technical description of the easy-to-use tool for navigating and analyzing large-scale
methods used to generate an interactive, two- funding data.
dimensional visualization of 60,568 grants funded by the Although still in development, this tool currently
National Institutes of Health in 2007. The visualization is provides:
made intelligible by providing interactive features for • an orthogonal view of the entire collection
assessing the data in a web-based visual browser, see • a zoom-able component for viewing
http://www.nihmaps.org. The key features include deep iteratively smaller sub-sets of the collection
zooming, selection, full-text querying, overlays, color- • an interactive interface for selecting groups
coding schemes, and multi-level labeling. Major insights, of documents for analysis
broader applicability, and future directions are • a system of layers, overlays, color-coding
discussed. schemes, and labels
• selection components for various fields
Keywords--- information visualization, scientific such as institute, topic, or keyword
funding, topic modeling, information navigation.
2. Data
The data used for textual analysis were titles and
1. Introduction abstracts from NIH grants funded in 2007. In addition,
we extracted information on Principal Investigator and
Within successful biomedical research, there are two funding Institute or Center (from among 27 distinct NIH
elements that accompany every project: proposals and Institutes and Centers) for each grant. The data is
publications. Finding ways to analyze these elements in a publically available on the NIH website [2] (specific
framework that is intuitive and convenient for biologists project data available on request). After removing
is a very difficult undertaking. The scope and range of duplicate abstracts (~1% of the corpus) and grants with
on-going projects and related data sets are not easily no abstracts (~10%), we analyzed a total of 60,568
identifiable, let alone computable within large-scale abstracts associated with 25 separate Institutes.
analyses.
By providing a gestalt-like frame of reference and an 3. Technique
expeditious visual entry point for large information
spaces that is not available via textual reports, 3.1. Topic Modeling
spreadsheets, or database results, visualization is now
well established as both an academic discipline and a We used topic modeling [3] to analyze the NIH
tool for improving scientific research. It is already well abstracts, and to provide a robust means for computing
entrenched in the biomedical community where it is used similarities between abstracts. The topic model is a
to facilitate, enhance, and drive analytical and decision- state-of-the-art Bayesian statistical machine learning
making processes [1]. technique for modeling collections of text documents.
In this case study, we present a technical description The intuition behind the topic model is that individual
of an interactive mapping tool based on a comprehensive documents exhibit multiple concepts or topics. The topic
collection of abstracts from proposals funded by the model simultaneously learns a set of topics that together
National Institutes of Health (NIH). The data provides a describe a collection, and the mixture of topics within
significant view of federally funded biomedical research each document in the collection. These topics – which
extant within the United States. The primary purpose of are learned automatically from the data with no prior
this tool is to provide NIH administrators, portfolio
978-0-7695-3733-7/09 $25.00 © 2009 IEEE 511
505
DOI 10.1109/IV.2009.105
, knowledge – are usually semantically meaningful, and DrL is ideally suited for large graphs; meaningful
highly interpretable. results have been obtained on graphs with up to 2 million
While the topics themselves produce a useful nodes [23]. The input for DrL is a similarity network
taxonomy or ontology to categorize individual abstracts, where the weight of each edge in the network represents
they also provide a reliable basis for computing how similar the two connected nodes are. When DrL is
similarity between any two abstracts. We ran a total of run, it produces a 2d layout where nodes that are similar
six independent topic models, initially learning T=200, are placed close to one another. This typically results in
T=400 and T=600 topics (each topic setting was run local clustering of related nodes and a global structure of
twice from different random initial conditions). The related clusters.
topic models were learned using Gibbs sampling, a For this project, quite a bit of time was spent on
standard technique for this model [5]. For each of the six getting a good layout. This is due to the fact that there
runs, the Gibbs sampler was run for 500 iterations. At are so many variables in topic generation, grant-grant
the end of 500 iterations, very similar topics in each run distance computation (0 is identical, infinity is very
were manually merged using hierarchical agglomerative dissimilar), distance to similarity conversion (0 is
clustering, resulting in models having T = 183, 186, 343, dissimilar, 1 is very similar), and layout parameters. A
349, 529 and 535 topics. These six topic models were change in parameters at any of these stages could
run for a further 500 iterations of the Gibbs sampler to drastically alter the layout. Further, we wanted the layout
reach stationarity. to come out where major research areas are strongly
clustered with strands of grants between clusters
3.2. Document Similarity representing ‘bridges’ which consist of projects laid out
in a gradient between two topically defined research
The layout and visualization of the 60,568 grants is areas. These project bridges have the potential to add
completely based on pair-wise similarities computed for extra dimensionality to the layout, since their topical
each pair of grants. While we did compute all 2 billion content provides concrete information on the
similarities, we only kept similarities that were above a relationships between clusters. The layout we eventually
threshold. Since there is no gold standard for measuring chose for this project is shown in Figure 1.
similarity between two text documents, we took a multi-
resolution approach of computing a similarity score
based on general and specific similarity. This was
achieved by computing the similarity based on all six
topic model runs, with topic resolutions ranging from
183 to 535 topics. Averaging over six different topic
model runs also improves the similarity estimate.
The input to the graph layout algorithm is a list of
pairs of nodes with an associated weight, given by the
similarity. We computed the similarity between grant
abstracts di and dj using the average symmetric Kullback-
Leibler divergence function [6], computed using the
respective mixtures of topics Pr(t| di), and Pr(t| di):
Pr(t | d i )
KL* (di , d j ) = ∑ (Pr(t | di ) − Pr(t | d j ) )log
t Pr(t | d j )
This symmetric version (denoted by *) of KL-divergence
was converted to a [0,1] similarity for input into the
graph layout algorithm, which is described in the next
section.
Figure 1. Final layout of the NIH grants
3.3. Graph Layout - DrL
3.4. Visual Encoding
Layout is a critical factor in all visualization
applications, especially for large-scale data sets. For Making sense of the network layout without any
laying out the data, the Distributed Recursive Graph labeling, coloring, or sizing is nearly impossible. So, to
Layout (DrL) algorithm developed by Shawn Martin was help in understanding we chose to color the nodes
used for the data [7]. At its core, DrL is a force-directed according to the institute which funded the represented
layout algorithm based on VxOrd [8] and optimized for grant. Unfortunately, there are 25 different institutes in
large graphs. It offers both a parallel and sequential the dataset, so there are many colors which are too
version of the above force-directed algorithm, a recursive similar to one another to tell apart. To deal with this we
multi-level version for very large graphs, and the ability had to do several things. First, we chose colors such that
to add new nodes to a previously drawn graph. similar colors were as far apart as possible and near
colors used complimentary colors. While not perfect, this
506
512
