Learning Center
Plans & pricing Sign in
Sign Out
Get this document free



									                                                           Vijay Krishnan
                                                       Founder, Infoaxe. Inc
                                                                         Mobile No: (650) 796-6388
                                   Mountain View, CA-94040               Email:

1     Work Experience
    • Founder, Infoaxe Inc. Jan 2008-Present.
      Infoaxe enables internet users to conveniently search and re-visit content they have viewed in the past. Infoaxe currently has more
      than 4 million registered users added in under two years. Infoaxe has raised total funding of $3.9 Million from world renowned
      Silicon Valley Venture Capital firms and Angel investors such as Draper Fisher Jurvetson, Labrador Ventures, Band of Angels,
      Amidzad Partners and Stephen Oskoui.
    • Sr. Scientist Associate at the Data Mining and Research Group, Yahoo! Sunnyvale, CA. Aug 2007-Jan 2008.
    • Summer Intern at the Data Mining and Research Group, Yahoo! Sunnyvale, CA. June-Sept 2006.

2     Education
    • Stanford University(2005-2007): Masters(M.S) in Computer Science. GPA = 3.92/4.00
    • IIT Bombay(2000-2005): Bachelors(B.Tech) and Masters(M.Tech) degrees in Computer Science and Engineering from the Indian
      Institute of Technology Bombay, under the Dual-Degree1 program. GPA = 9.18/10.00.
Research Interests: Data Management, Data/Text/Web/Sequence Mining, Graphical Models, Large Scale Machine Learning, Web
Search, Information Retrieval, Mathematical Modeling and Natural Language Processing.

3     Patents and Publications
    1. Method for efficiently building compact models for large multi-class text classification. United States Patent # 20090274376.
    2. Vijay Krishnan and Rashmi Raj. Web Spam Detection with Anti-Trust Rank. In AIRWeb 2006, the ACM SIGIR 2006
       Workshop on Adversarial Information Retrieval on the Web.∼vijayk/krishnan-raj-airweb06.pdf
    3. Vijay Krishnan and Christopher D. Manning. An Effective Two-Stage Model for Exploiting Non-Local Dependencies in Named
       Entity Recognition. In COLING-ACL 2006, Sydney, Australia, July 2006.∼vijayk/krishnan-manning-colacl06sub.pdf
    4. Vijay Krishnan, Sujatha Das and Soumen Chakrabarti. Enhanced Answer Type Inference from Questions using Sequential
       Models In HLT/EMNLP 2005, Vancouver, B.C, Canada, October 2005.∼soumen/doc/emnlp2005/382.pdf
    5. Vijay Krishnan. Shortcomings of Latent Models in Supervised Settings. In ACM SIGIR 2005, Salvador, Brazil, August 2005.∼vijayk/p106-krishnan.pdf
    6. Shipra Agrawal, Vijay Krishnan and Jayant Haritsa. On Addressing Efficiency Concerns in Privacy-Preserving Mining. In
       the 9th Intl. Conf. on Database Systems for Advanced Applications (DASFAA), Jeju Island, Korea, March 2004.
    7. David J. Ruau, Michael Mbagwu, Joel T. Dudley, Vijay Krishnan, Atul J. Butte. Comparison of Automated and Human
       Assignment of MeSH Terms on Publicly-Available Molecular Datasets. In AMIA 2011 Summit on Translational Bioinformat-
       ics, San Francisco, CA.

4     Honors/distinctions
    • I am one of the 12 silicon valley entrepreneurs featured in Cees Quirijns’s new book ”Startup Best Practices”. The book’s
      Kindle edition can be purchased on Amazon at
      sr 1 1?ie=UTF8&m=AG56TWVU5XWC2&s=books&qid=1292047500&sr=8-1. The book can also be downloaded for free from http://
    • Served as research paper reviewer for the Decision Support Systems Journal.
    • Ranked 119th in Google India Code Jam 2005 out of over 14,000 participants from six countries in South Asia. This was a
      programming contest organized by Google India.
    • Awarded the Google India conference travel award and also the ACM student travel award, for attending SIGIR 2005 and
      presenting my paper titled Shortcomings of Latent Models in Supervised Settings.
  1 It is a Five year integrated (B.Tech + M.Tech) program. The first four years are almost identical to the B.Tech program except for extra research

emphasis. The fifth year comprises minimal course work and is devoted entirely to research.

    • Served as research paper reviewer in the World Wide Web(WWW) 2005 conference committee.
    • Presented my research papers titled Enhanced Answer Type Inference from Questions using Sequential Models at HLT-EMNLP 2005
      in Vancouver, Canada and An Effective Two-Stage Model for Exploiting Non-Local Dependencies in Named Entity Recognition at
      COLING-ACL 2006 in Sydney, Australia.
    • Obtained the A+ grade in the Text Retrieval and Web Search course taught by Prof. Chris Manning and Prof. Prabhakar Raghavan,
      at Stanford.
    • Obtained 98 percentile in the Computer Science subject GRE (taken in Nov 2004).
    • Obtained the highest grade AA in all Data/Text Mining courses I have taken at IIT Bombay, namely Hypertext Mining and Retrieval
      (Prof. Soumen Chakrabarti), Advanced Topics in Data Mining and Data Warehousing and Mining (Prof. Sunita Sarawagi).
    • Obtained the highest grade AA in all my research projects so far, namely the Seminar (6th semester), the Mini-Project (7th semester)
      and Dual Degree Project Stages I, II and III.

5     Industry and Research Projects
5.1    Founder of Infoaxe Inc. Web History Search Engine (Jan-08 to Present)
Infoaxe has over 4 million registered users added in under two years. Infoaxe’s Web History Search Engine allows users to easily keep track
of Web Pages browsed that they would like to search and get back to later without going through the hassle of bookmarking or tagging.
Social features on the site make it extremely easy and fun to share webpages with friends as real-time feeds.
    I handle multiple aspects of product design, technology for fast incremental indexing, search ranking, metrics driven optimization of
virality and user engagement, media relations, partnerships, finances, day to day operations and the different business aspects of running
the company.

5.2    Search Query Categorization (Data Mining and Research Group at Yahoo! - Summer
       2006 and Aug 2007 to Jan 2008)
Categorized search queries to a huge taxonomy of nearly 800 classes with over 700,000 manually labeled train queries. Used an SVM based
classifier and improved on Yahoo!’s existing Naive Bayes based categorizer by about 200% and lifted classification accuracy from about
20% to about 60%. Faced many additional challenges primarily those of scalability due to lack of published literature using huge training
data for query classification. Multiple groups at Yahoo! including the data insights team and the search ranking teams, put my classifier
to use and obtained improvements on their respective tasks. Also categorized search queries, ad snippets and webpages to a much larger
taxonomy with over 6500 leaf level classes. My methods lifted the query classification accuracy from 12% to over 90% with similar gains on
classifying ad snippets and webpages. I also invented a method to learn significantly sparser SVM models than the one-vs-rest SVM and
my methods could learn SVM models that could be encoded in 20MB rather than the 10GB that would be required my one-vs-rest SVM
models, with no discernable change in accuracy from the one-vs-rest SVM. My method also ensured that as a direct result of learning of
sparse models, both SVM training and classification could be made faster by at least a factor of 10.
     This in turn led to across the board benefits in multiple Yahoo products involving search relevance, ad matching for queries, ad matching
for content, user profiling and behavior targeting based on a user’s search history logs etc. This work led to the publication of a US patent
titled Method for Efficiently Building Compact Models for Large Multi-class Text Classification.

5.3    Web Spam Detection using HyperLink Analysis (Fall 2005)
We exploited the intuition that the pages that point to spam pages are very likely to be spam pages themselves. We start with a seed set
of human labeled spam pages and propagate Anti-Trust scores backward along the incoming links. This is done by running topic specific
pagerank on the transpose of the web graph, with the seed set chosen as the teleport set. Our code was in Java. This work has been
accepted in AIRWeb 2006. Project Advisors: Prof. Jeff Ullman and Dr. Anand Rajaraman, Stanford University.

5.4    Named Entity Recognition (Oct 2005- March 2006)
Named entity recognition (NER) is a subtask of information extraction that seeks to locate and classify atomic elements in text into
predefined categories such as the names of persons, organizations, locations etc. We improved NER performance by encoding non-local
dependencies. For instance, if the token Washington is identified as a person in one part of a document, it is likely to be a person in
another occurrence in the document, as well. We used a Conditional Random Field (CRF) based NER system using local features to
make predictions and then train another CRF which uses both local information and features extracted from the output of the first CRF.
We obtained 13.3% relative error reduction when compared to the 9.3% relative error reduction offered by the best systems that exploit
non-local information. Additionally, our running time for inference is just the inference time of two sequential CRFs, which is much less
than that of other more complicated approaches that directly model the dependencies and do approximate inference. Our code was in
Java. This work has been accepted as a full research paper in COLING-ACL 2006. This was a solo project. Project Advisor: Prof.
Chris Manning, Computer Science Department, Stanford University

5.5    Question Classification for Question Answering(QA)(Autumn 2004-05)
We addressed the problem of classifying a question to its Answer type(or atype). We would like to classify the question, Who was the first
person to run the mile in under four minutes? to the class Person which can be used to shortlist candidate answers in a QA system. We

got good improvements over the best existing approach of using word bigram features in an SVM classifier. We first trained a Conditional
Random Field (CRF) to tag a segment of the question as relevant to the classification task. Additional features derived from the CRF
output was fed to the SVM which enabled us to lift accuracy from 79.2% to 86.2% on a benchmark dataset. We also had good results in
classifying questions to the wordnet noun hierarchy. This work was accepted as a full research paper in HLT/EMNLP 2005. Our code
was in Java. Project Advisor: Prof. Soumen Chakrabarti, IIT Bombay.

5.6     Parameter estimation and supervision in Latent Models(2004-05)
Our experiments with supervised document classification using the Aspect Model, and obtained rather unsatisfactory results. Deeper
analysis showed that the aspect model is unable to reward certain natural features and partitions on a small synthetic corpus. I also
analyzed the scenario of using tempered EM and gave strong evidence suggesting that it would not plug the above shortcomings. In all, I
showed that the Aspect Model has some inherent weaknesses in its formulation. A possible improvement to the Latent Dirichlet Allocation
model was also suggested. This work was accepted as a poster paper in SIGIR 2005. Project Advisor: Prof. Soumen Chakrabarti,
IIT Bombay

5.7     Privacy Preserving Association Rule Mining from Large Databases (Summer 2002, 2003)
Data mining services require accurate input data for their results to be meaningful, but privacy concerns may influence users to provide
spurious information. We improved on the method of Rizvi and Haritsa(VLDB 2002) which is data distortion scheme for association rule
mining that simultaneously provides both privacy to the user and accuracy in the mining results was proposed. However, mining the
distorted database can be orders of magnitude more time-consuming as compared to mining the original database. We addressed this
issue and demonstrated that by (a) generalizing the distortion process to perform symbol-specific distortion, (b) appropriately choosing the
distortion parameters, and (c) applying a variety of optimizations in the reconstruction process, runtime efficiencies that are well within
an order of magnitude of undistorted mining can be achieved. Thus our algorithm makes privacy preserving mining a practically viable
process. Our code was in C++ and Perl. This work was published in the proceedings of DASFAA 2004 as a full research paper. Project
Advisor: Prof. Jayant R. Haritsa, Indian Institute of Science, Bangalore.

5.8     Combining Heterogeneous Features in Discriminative Classification(2003-04)
Discriminative classification algorithms, especially Support Vector Machines(SVM) are known for their high document classification ac-
curacy. They typically do better than classification algorithms based on generative models. In real-life data, there is usually additional
information available over the raw term counts which we call heterogeneous features. Heterogeneous features naturally arise in hypertext,
hierarchies, relational databases. Methods to improve classification accuracy further by exploiting the additional heterogeneous features
that are often available with real-life data, were explored. Our code was primarily in Perl. Project Advisor: Prof. Soumen Chakrabarti,
IIT Bombay

6     Course Projects
    1. Optimizing Jakarta Lucene’s Document Scoring: Used the IR Engine, Jakarta Lucene to efficiently index and query on a
       document corpus. Tuned document scoring attributes such as boost factor for the title, applying stopwording and porter stemming,
       document normalization and term frequency factor. Source code: Java. Project Advisors: Dr. Prabhakar Raghavan and Prof.
       Chris Manning, Stanford University.
    2. Implementing the Naive Bayes Classifier and certain variants Implemented the smoothed Multinomial Naive Bayes Classifier
       and classified the 20-newsgroups dataset with Chi-square feature selection and n-fold cross validation. Also implemented the different
       variants of Naive Bayes proposed by Rennie et. al (ICML 2003), and verified their results. Source code: Java. Project Advisors:
       Dr. Prabhakar Raghavan and Prof. Chris Manning, Stanford University.
    3. Overload Control of E-Commerce Web Servers:(Autumn 2004-05) Improved on the LIFO queuing policy during overload
       conditions for e-commerce web servers. Implemented heuristics to make decisions that maximize expected revenue per unit CPU
       time. Httpperf load generator used for experiments. Source code: PERL. Project Advisor: Prof. Varsha Apte, IIT Bombay.
    4. Natural Language Processing Experiments(Spring 2003-04) Tokenized text into logical units, compound words, dates in
       various formats, telephone numbers etc. Analyzed token-type ratios for different text categories. Implemented WordNet based
       heuristics for Word Sense Disambiguation. Source code: PERL. Project Advisor: Prof. Pushpak Bhattacharyya, IIT Bombay.
    5. Automated Music Generator(Spring 2002-03) Music Generator trained on available pieces of Mozart’s music. Smoothed
       Markov-chain state transition probabilities estimated. Pleasant sounding and original music obtained. Source code: Java. Project
       Advisor: Prof. Pushpak Bhattacharyya, IIT Bombay.
    6. Pascal-like Compiler(Spring 2002-03) Compiler written by doing Lexical Analysis, Syntactic Analysis, Semantic Analysis and
       Code Generation. Source code: Lex. Yacc Project Advisor: Prof. Amitabha Sanyal, IIT Bombay.
    7. Medical Information System(Autumn 2002-03) Secure system for doctors, lab assistant, nurses and patients to share, manage,
       supervise and access hospital records and medical data. Features for online appointments and managing drugs inventory. Source
       code: Java (JDBC and Servlets). Project Advisor: Prof. S. Sudarshan, IIT Bombay.
    8. Peer to Peer Networking software(Spring 2001-02) Improving an existing software for security, user interface. Also added
       peer to peer chat, conferencing and chess playing facility. Source code: Java. Project Advisor: Prof. G. Sivakumar, IIT Bombay.
    9. Sensitivity Analysis on the VRPTW(Autumn 2001-02) Sensitivity Analysis applied on the Vehicle Routing Problem with
       Time Windows(VRPTW). Algorithm applied to the postal delivery system in Mumbai to get a good routing. Source code: Java.
       Project Advisor: Prof. S. Arunkumar, IIT Bombay.

    10. Chess Puzzle Solver(Spring 2000-01) Constraint logic program to identify possibility of forced ”Check-mate” within three
        moves with initial chess board setting. Source code: MIT Scheme(LISP). Project Advisor: Prof. Abhiram G. Ranade, IIT
    11. Rook and King Chess Endgame(Autumn 2000-01) Finding an efficient checkmate for a King and Rook against a lone King
        starting at any position. Source code: Fortran. Project Advisor: Prof. G. Sivakumar, IIT Bombay.

7      Systems Experience
•   Programming Languages : C, C++, Java, Perl, Fortran, Pascal, Scheme, VHDL, SQL, Lex and Yacc.
•   Platforms : Linux, HP Unix, Solaris, DOS, Windows 95/98/NT/XP.
•   I have also worked with Intel 8085/x86 architecture Assembly Languages.
•   I have used the Matlab and SAS packages and the Eclipse IDE, for Java Programming.
•   I have used Jakarta Lucene for efficient indexing and querying of documents.

8      Courses
     1. Data/Text/Web Mining, Machine Learning and Statistics:
        • Stanford University: Machine Learning, Text Retrieval and Web Search, Data Mining, Probabilistic Models in AI(Bayesian and
        Markov Networks), Convex Optimization, Modern Applied Statistics and Natural Language Processing.
        • IIT Bombay: Data Warehousing and Mining, Hypertext retrieval and mining, Advanced topics in Data Mining(Sequence and
        Graphical Models), Probability, Statistical Theory and Stochastic processes, Artificial Intelligence, Categorical Data Analysis and
        Regression, Theory of Estimation, Natural Language Processing, Information Theory and Coding.
     2. Finance/Entrepreneurship courses at Stanford University: Investment Science Honors, Technology Venture Formation.
     3. Other relevant courses at IIT Bombay: Data Structures and Algorithms, Databases and Information Systems, Design and
        Analysis of Algorithms, Linear Optimization, Software Systems Lab(Perl, Java, C and shell scripts), Operating Systems, Computer
        Networks, Applied (Approximation) Algorithms, Geometric(Randomized) Algorithms


To top