# Machine Learning Meetup Notes: 2010-04-28

From Noisebridge

(Difference between revisions)

Mschachter (Talk | contribs) |
Mschachter (Talk | contribs) |
||

Line 6: | Line 6: | ||

***Dual problem is a quadratic programming problem that is easier to solve than the primal problem | ***Dual problem is a quadratic programming problem that is easier to solve than the primal problem | ||

***After dual problem is optimized, only the support vectors (the data points whose langrangian multipliers are > 0) are needed to make predictions for new data (and their associated multipliers) | ***After dual problem is optimized, only the support vectors (the data points whose langrangian multipliers are > 0) are needed to make predictions for new data (and their associated multipliers) | ||

− | *Thomas talked about the KDD conference and their data competition [https://pslcdatashop.web.cmu.edu/KDDCup/rules_data_format.jsp | + | *Thomas talked about the KDD conference and their data competition [https://pslcdatashop.web.cmu.edu/KDDCup/rules_data_format.jsp] |

− | ] | + | |

*Sai skyped in and talked a bit about his use of libSVM for classification of user history on his website cssfingerprint.com | *Sai skyped in and talked a bit about his use of libSVM for classification of user history on his website cssfingerprint.com | ||

*We talked a little bit about libSVM [http://www.csie.ntu.edu.tw/~cjlin/libsvm/] | *We talked a little bit about libSVM [http://www.csie.ntu.edu.tw/~cjlin/libsvm/] |

## Revision as of 09:13, 29 April 2010

- Mike S presented a mathematical overview of SVMs
- Started with introduction to linear classification [1]
- Discussed the kernel trick [2]
- Loosely derived the loss function and dual loss function for support vector machines [3]
- Emphasized two important aspects of SVMs:
- Dual problem is a quadratic programming problem that is easier to solve than the primal problem
- After dual problem is optimized, only the support vectors (the data points whose langrangian multipliers are > 0) are needed to make predictions for new data (and their associated multipliers)

- Thomas talked about the KDD conference and their data competition [4]
- Sai skyped in and talked a bit about his use of libSVM for classification of user history on his website cssfingerprint.com
- We talked a little bit about libSVM [5]