NASIG 2012 - Discovering the World's Research (ITHAKA portion)
1. Discovering the World’s Research
Ron Snyder
Director of Advanced Technology, ITHAKA/JSTOR
NASIG Annual Conference - 2012
June 9, 2012
2. Who we are
ITHAKA is a not-for-profit organization that helps the academic
community use digital technologies to preserve the scholarly
record and to advance research and teaching in sustainable ways.
We pursue this mission by providing innovative services that aid in
the adoption of these technologies and that create lasting impact.
JSTOR is a research platform that enables
discovery, access, and preservation of scholarly
content.
3. JSTOR Factoids
• Started in 1997
• Journals online: 1,604
• Articles online: 7.5 million
• Disciplines covered: 60
• Participating institutions: 7,800
• Countries with participating institutions: 167
4. JSTOR site activity
User Sessions (visits)
» New Sessions (per hour):
70k peak, 38k average
» Simultaneous Sessions:
44k peak, 21k average
Page Views
» 3.5M per day, 6.7M peak
Content Accesses
» 430k per day, 850K peak
Searches
» 456k per day, 1.13M peak
5. ITHAKA/JSTOR Discovery Initiatives
• Overhaul of JSTOR Search Infrastructure
• Coming Soon (Summer 2012), watch for it…
• Analytics and data warehouse
• Ingesting, organizing, and analyzing billions of usage
events since JSTOR inception
• Improved external discoverability
• Various SEO, Google/GS, MS-Academic projects
• Local Discovery Integration (LDI) Pilot
• Machine-based document classification
7. Problem Statement:
» Research has shown time and again that both students and faculty are
beginning their research at places other than the library OPAC, most
notably Google/Google Scholar and discipline-specific electronic
databases, and that the trend is continuing
Starting point for research, identified by faculty in 2003, 2006, and 2009 (2009 Faculty Study, ITHAKA)
100%
90%
2003
80%
2006
70%
2009
60%
50%
40%
30%
20%
10%
0%
The library building online librarygeneral-purpose specific engine
Your A catalog A search electronic research resource
8. Where is discovery happening?
Where JSTOR ‘sessions’ originated | Jan 2011 – Dec 2011
9. Problem Statement:
» As web-scale discovery services are being purchased and
implemented by institutions, the value of those implementations
are somewhat limited because they are (for the most part) only
addressing that limited population of researchers who begin at
a library-designated starting point (e.g. OPAC)
JSTOR usage | Australia | 2010 Nov.
JSTOR Google/Google Scholar Known Linking Partner Library
16%
6%
9%
76%
2%
10. Research Behavior: Students
What is the easiest place to start research
according to students?
Library Databases
Google
0 10 20 30 40 50 60 70
Source: ProQuest survey of student research habits, 2007
11. Research Behavior: Faculty
Starting Point for Research, identified by faculty in 2003, 2006, and 2009
100%
90%
2003
80%
2006
70%
2009
60%
50%
40%
30%
20%
10%
0%
The library building online librarygeneral-purpose specificengine
Your A catalog A search electronic research resource
Source: ITHAKA 2009 Faculty Survey, 2010
12. Concept:
» If we can more effectively reach the users at the place(s) where
they normally begin their research, then we can begin to more
effectively build their awareness of the resources that the
institution has licensed/purchased for their purposes
» The local discovery integration (LDI) pilot study will attempt to
measure changes in the student/faculty research experience by
„embedding‟ the institution‟s selected web-scale discovery
service in strategically-selected places in the JSTOR interface
where – we believe – the user would naturally want to „cast a
wider net‟ for discovery
2010 JSTOR Usage Highlights
Total Significant Accesses 594,888,001
Articles Downloaded 74,901,344
Articles Viewed 112,751,906
Searches Performed 168,720,887
Inbound Links from Licensed Partners 13,013,904
Inbound Links from Google/Scholar 157,903,053
13. How it works
Links Out
• Search Results
Advanced Search Page
Search Results View
• 3rd Page “Lightbox” pop-up
• Article View - Incoming from Google
• Article View - All other non-Google
• Zero Results Page
We placed links at various places along the research workflow in
JSTOR to allow students and researchers to “Cast a wider net”
14. Search results page
» JSTOR may not be the most appropriate starting place in every
instance, but it is a trusted and familiar interface. This will allow
the user to „flowback‟ to another starting place (e.g. the library)
• Uses the familiar
university logo to grab
attention
• Inserts search terms into
link text to notify user of
customized behavior
• Positioned proximate to
search results; relevant
during the search result
evaluation phase
15. Empty results page
» In this instance, the user has found nothing and the most typical web
response is to hit the „Back‟ button. If we allow the user – at this point –
to execute a search in the local discovery interface, we might improve
the user experience
• One of the key places
where a user is likely to
want to try a different,
broader search
• Larger placement takes
advantage of available real
estate and cognitive space
• Users typically do not
spend time on this page so
it is important to increase
notice-ability and self-
explanation
16. Article page after Google search
» In 2010, over 32M Google/Google Scholar searches brought users
directly to an article page. They may or may not have found what they
really wanted, so we‟d like to give them an alternative discovery choice
• Visible when coming
from a Google or Google
Scholar search
• Captures basic search
terms from the search
• Provides an opportunity
to convert a user from a
Google/Google Scholar
user to a Summon user
17. Article page after JSTOR search
» In 2010, almost 113M articles were viewed in JSTOR. Again, they may
not have found what they really wanted, so we‟d like to give them an
alternative discovery choice
• Visible when coming
from a JSTOR search
• Raises visibility of the
feature by exposing it to a
large number of users
• Inserts search terms
into link text to notify user
of customized behavior
18. Results View: All Pages
Link out from the
bottom of all pages of
the search results view.
This will allow more
opportunities to link out
for students/ researchers
combing through large
sets of results.
19. Results View: 3rd Page
Pop-up on the third page of search results
Prompts the student/ researcher to indicate whether they wish to link out through the LDI. This
will enable us to measure whether students wish to “cast a wider net” or not. In the other link
scenarios we don’t have a baseline of how many students do not notice the link vs. choose not
to use it
21. Results Overview
» Highest usage occurred in Zero Results scenario
Data shown is for all institutions participating in Summon LDI
Date range: July 2011 – February 2012
23. The Problem
JSTOR Corpus
• 60 disciplines
• 1,600 journals
• Nearly 8 million articles
• Disciplines are associated at the Journal level
• All articles in a Journal inherit the Journal assigned
disciplines
• Using this approach many articles have incomplete
and/or incorrect discipline tagging hindering discovery
• How to assign disciplines to articles?
24. Topic Models
• Human classification and tagging is not feasible
• A machine-based classification process is desired
• Topic models are a way of finding structure in a set of
documents
• They allow is to find “latent” themes
• A topic model is not a topic map
• Some topic modeling approaches include
• Latent Semantic Analysis (LSI/LSA) (Deerwester 1990)
• Probabilistic LSA (Hoffmann 1999)
• Latent Dirichlet Allocation (LDA) (Blei 2003)
25. Topic Modeling – our approach
LDA – Latent Dirichlet Allocation
• A generative probabilistic model for analyzing
collections of documents
• A Bayesian model where each document is modeled
as a mixture of topics (disciplines)
• Models semantic relationships between documents
based on word co-occurrences
26. The Process
• We select the most representative documents from
each JSTOR discipline to build a topic model (from
the vocabulary of the document sample)
• This sampling and vocabulary modeling is the most important part
of the process!
• We’re still experimenting with this, but find the citation network
provides a good means for identifying core documents in a
discipline
• Also considering whether usage data might be leveraged here
• Each document in the corpus is then analyzed and
compared to the topic model to determine how well it
matches each topic
• A probability distribution is generated providing discipline weights
• The top weighted discipline(s) are associated with each article
27. Application
• On-site discovery
• Will be a key element of our overhauled search
infrastructure, tentatively scheduled for beta release mid-summer
• Use in article-level discipline/subject/topic mappings
for better integration with aggregated indexes
• Will support a richer data feed for Summon, for instance
Editor's Notes
So, how do we take a good idea (web-scale discovery) and make it better?How do we take the basic principle – which is good and valuable – and use it in such a way so that it achieves a broader impact?
Replace w data for all JSTOR (requested)
For several years, the anecdotal evidence that librarians had been witnessing first-hand was beginning to be verified by user studies published by OCLC and others, as well as student surveys reported on by ProQuest and others.The evidence was overwhelming: the gateway function that libraries had played for so long – and valued so much – being THE gateway to academic research – was quickly being overtaken by web-based search engines like Google.It was one thing when undergraduates started to migrate away from the library …
A number of organizations had been following this trend closely - including my own (ITHAKA … which is the organizational umbrella under which JSTOR and Portico reside). We were taking a longitudinal look at faculty views about the library – and other pertinent scholarly communications issues – and comparing those view with similar survey data from librarians.One noticeable disconnect in these surveys – as you might imagine – as the perception of the “library as gateway”. Librarians believe it to be hugely important and faculty less so (science faculty much less so than humanities faculty). And students? Even less than that.Yet, the dollars being spent on access services in libraries – both software and people – were (and continue to be) tremendous. Are those expenditures aligned properly with the expectations of the users, and if they are, then how do we more effectively leverage those investments to reach a broader audience?
Search Result Page: Design Notes-- Link is proximate to the first search result so that it is part of the evaluation workflow (e.g. user looks at first result, decides it is no good, sees the link)-- Uses branding element that the user is familiar with … should be something that all students / users will recognize-- Customized text in the link indicates to the user that this is session-specific and workflow-specific behavior-- Uses a link which is unobtrusive and not confusing, yet noticeable (vis-à-vis a search box, which we already have too many of) Requirements Used-- Uses 16x45px max logo
No results page: design notes-- Mimics the JSTOR search box above, clearly indicating behavior to the user-- Button size is near the maximum size it could be and still look like a clickable button-- Key place where users are likely to “cast a wider net” (exhausted all JSTOR search results)-- Takes advantage of larger real estate and simpler page design to drive users toward this feature Requirements Used-- Uses canonical name for button text (25 characters max)-- Uses larger logo, 250x50px max
Article Page – JSTOR search: Design notes-- Customized text in the link indicates to the user that this is session-specific and workflow-specific behavior-- Article page is already very stuffed, esp. with CSP/Publisher stuff, so we were forced to go with something more minimalist here-- Not necessarily a core workflow for users, searching from the article page, but gives us the opportunity to expose the feature to a wide audience-- Only difference from JSTOR article page is the missing “Back To Search Results” linkRequirements Used-- Uses canonical name for link text (25 characters max)
Article Page – JSTOR search: Design notes-- Customized text in the link indicates to the user that this is session-specific and workflow-specific behavior-- Article page is already very stuffed, esp. with CSP/Publisher stuff, so we were forced to go with something more minimalist here-- Not necessarily a core workflow for users, searching from the article page, but gives us the opportunity to expose the feature to a wide audienceRequirements Used-- Uses canonical name for link text (25 characters max)
Article Page – JSTOR search: Design notes-- Customized text in the link indicates to the user that this is session-specific and workflow-specific behavior-- Article page is already very stuffed, esp. with CSP/Publisher stuff, so we were forced to go with something more minimalist here-- Not necessarily a core workflow for users, searching from the article page, but gives us the opportunity to expose the feature to a wide audienceRequirements Used-- Uses canonical name for link text (25 characters max)