Ar Xiv org 250 000 documents 47 000

  • Slides: 16
Download presentation
Ar. Xiv. org 250, 000 documents 47, 000 registered users 1 million+ downloads per

Ar. Xiv. org 250, 000 documents 47, 000 registered users 1 million+ downloads per year Cost Per Paper $10000 Commercial Journal $1000 Non-Profit Journal $10 ar. Xiv

Goal: Process increasing number of submissions at constant or declining cost

Goal: Process increasing number of submissions at constant or declining cost

ar. Xiv has an active core of users: 10% of users are responsible for

ar. Xiv has an active core of users: 10% of users are responsible for about 1/3 of all submissions, 50% of all users have logged in (to submit or update a paper) in the past 1. 5 years

Authentication and Access Control Recently moved from an http authentication/Berkeley database system to a

Authentication and Access Control Recently moved from an http authentication/Berkeley database system to a system based on cookies and a relational database. Currently, all registered users (who haven’t been suspended) can submit to all subjects classes in all archives – the original submitter or somebody with the paper password can update the paper. People are allowed to register depending on their E-mail address: abc@university. edu can register, but xyz@company. com can’t unless company=ibm, lucent, …; this list is hard to maintain (we have to block popular ISPs in every country), exceptions are dealt with manually at great cost (each case takes detective work), and there are many people in. edu (alumni, non-research staff) who shouldn’t be able to submit. Because registration and submission are linked, user database can’t be used to offer other services: e-mail notification, personalization.

Endorsements and Trust Management Administrators Grandfathered Users In new system, everyone will be able

Endorsements and Trust Management Administrators Grandfathered Users In new system, everyone will be able to register. Users who registered under the old system will still be able to upload to any archive or subject class, but new users will need to be endorsed by an author with a publication history in that category. Burden shifts from one senior staff person to 47, 000 registered users. User database can be used

Endorsee t n me d e s r o En Endorser e d o

Endorsee t n me d e s r o En Endorser e d o c

Web-based interface for administrators: • View user history and publications • Monitor endorsement process

Web-based interface for administrators: • View user history and publications • Monitor endorsement process • Manage authority records • Disable ability to submit or endorse • Keep “institutional memory”

Future Directions • Flexible Submission Queue (Currently submissions are published the following evening –

Future Directions • Flexible Submission Queue (Currently submissions are published the following evening – we can’t easily delay a submission) • Validating Metadata Form (Force users to clean up entry errors, so administrators don’t have to) • Automatic Protection (Suspicious submissions and endorsements will be automatically delayed) • New Search Engine based on Lucene • Retrofit e-mail notification (current awareness) to use new user database.

Classifying Articles with the Support Vector Machine Paul Ginsparg Paul Houle Thorsten Joachims Jae-Hoon

Classifying Articles with the Support Vector Machine Paul Ginsparg Paul Houle Thorsten Joachims Jae-Hoon Sul Goal: identify papers in existing archives that are relevant to a new subject archive, q-bio (Quantitative Biology)

Active Training of SVM Training: q-bio Training: not q-bio Other far from margin Other

Active Training of SVM Training: q-bio Training: not q-bio Other far from margin Other close to margin SVM finds maximum-margin hyperplane. We do first training run on one year of data, then identify other papers that lie close to the dividing line. We iteratively classify these by hand to refine the classification

Classifer performance improves as the size of a category increases.

Classifer performance improves as the size of a category increases.

Time Series Analysis of Content and Usage Information Paul Ginsparg Jon Kleinberg

Time Series Analysis of Content and Usage Information Paul Ginsparg Jon Kleinberg

Kleinberg’s algorithm uses a hidden Markov model to detect bursts of word usage in

Kleinberg’s algorithm uses a hidden Markov model to detect bursts of word usage in ar. Xiv titles, reveals intellectual trends in the last decade of high-energy physics theory.

Announcement Cited by other papers Web Link Added Review papers have a distinctive pattern

Announcement Cited by other papers Web Link Added Review papers have a distinctive pattern of use: an initial spike after announcement, followed by a long nearly-constant tail.