You are viewing...

HN News Recommender v0.3

Updated on March 04, 2012 at the 19th hour
Posted under:

DISCLAIMER: Expressed views on this blog are my own.

Google Chrome Extension: Get it here. Updated Git source here as well:

What is it? Visit here.

What's new?
  • In the options page, there are two new options. Color and reading frequency.
  • Increased accuracy of recommendations.
  • For more details look below!
What's next?
  • Moving from local storage to IndexedDB to overcome the space limitation. (v 0.4)
  • Possible expansion of recommendations to other sites such as reddit.
  • Possible renaming of the extension if the expansion happens.
  • Firefox Extension as well.
  • Reading frequency refers to how much you tend to read per day. Selecting average means means we give some weight towards your not interested articles and some weight toward your interests. What does this mean exactly? The notion is that some people may not look at as many pages as others, so weights of not interested words should not count as much, but I found that you get far more recommendations
  • Colors - you can choose the text and highlight colors of your recommendations, so if you don't like the default. You can set it there. In a newer version, I will add a color picker and preview text.
  • Feature elimination using the assumption of normally distributed features. This is an interesting one, I figured it would be better to get rid of particular words without defining a set of words to exclude from the calculations. Words like 'a' ,'this', and etc. do not add much meaning to the article, so I would consider them outliers. There are also words that you don't see very often and would be considered another outlier. These outliers would have extraordinary high or low counts, but let's say we assumed that the average word count in a class is the mean of every feature (you might see technology, sql, php, and such more often together) , then we can find z-scores to see how far away the feature is from the mean and get rid of words that surpass the z-score that has been set. I hope this will increase the accuracy of the recommendations.

Read, if you'd like: Improving the HackerNews News Recommender
You just read "HN News Recommender v0.3". Please share if you liked it!
You can read more recent posts here.