Attempt on a Kaggle competition, Personalized Web Search Challenge (hosted by Yandex)

URL: http://www.kaggle.com/c/yandex-personalized-web-search-challenge

Deadline: Friday, January 10, 2014

Team Members

• Yosuke Sugishita
• David Kim
• Possibly Brendan and David Hsiao
• Idea: Should we make this open to people in Data Science Club and local data meetups on Meetup.com? If we have too many people (I think 4-5 people in one team is a limit), we can make multiple teams and still work together.

Ideas on our team name

• Asian Revolution
• West Coasters
• Canadian Kimchi Roll

File structure

• script
  • file_manupulation
  • analysis
• lib
  • Functions / classes to use in other scripts.
• test
  • Scripts to test functions.
• data
  • Contains all the data, like test and train. Not committed due to the large size of the files. Download them directly from Kaggle.

About branches / pull requests

All the code must be reviewed by at least by one other person before being pulled into the master. Make a branch, write code, test, and send a pull request. Use short, descriptive names for branches.

Never directly work on the master.

Tools

• Version Control
  • Git with scm_breeze (https://github.com/ndbroadbent/scm_breeze)
  • Note, scm_breeze is a must. It's a huge productivity booster.
• Language(s)
  • Python
• Editors
  • Vim
  • PyCharm might be good. The same company's Ruby IDE is awesome.
  • Any other editors you like?
• Database?
  • Looks like some people on Kaggle tried to use databases, but it didn't work out very well:
  • http://www.kaggle.com/c/yandex-personalized-web-search-challenge/forums/t/6183/handling-703-000-000-urls/
  • http://www.kaggle.com/c/yandex-personalized-web-search-challenge/forums/t/6353/someone-else-using-r-mysql-as-database-need-some-feedback

Notes on possible strategies (more on the wiki)

Two ways to look at this problem:

1. Collaborative filtering (recommender) problem
   • Netflix Prize winners' solution: http://www2.research.att.com/~volinsky/papers/ieeecomputer.pdf
2. We can also look at the past clicks a certain user has performed.
   • The user is probably more (or less) likely to click the pages they already clicked and liked. => Need to test this.

Our first strategy is based on 2. (Low-hanging fruits! Yay!)

• https://github.com/yosukesugishita/personalized_search_challenge/wiki/Initial-Model:-Take-advantage-of-multiple-visits
• Here is the paper that inspired this strategy: http://people.csail.mit.edu/teevan/work/publications/papers/wsdm11-pnav.pdf

Here is the paper I got inspiration from for this strategy: http://people.csail.mit.edu/teevan/work/publications/papers/wsdm11-pnav.pdf

Some notes on the data

The train file is big (16GB when uncompressed)

We need to think about how to handle this. Perhaps use a database, like sqlite or MySQL? I (Yosuke) suspect we can try our first strategies with a randomly-sampled subset of the data. How would we go about it?

Train and test

In the competition, the first 27 days are used as train data, and the last 3 days as test data. (http://www.kaggle.com/c/yandex-personalized-web-search-challenge/data)

Perhaps we can locally test our model using the first 24 days train and the next 3 days as test.