Saturday, September 3, 2011

Content recommendations relevant to the user interest in version 1.0.0

After a week since the first release v1.0.0 is available.

From now the plugin tracks user behavior and tries to suggest relevant posts. Prediction accuracy depends on amount of content on the blog and how well the content is described by tags, categories and the title.

The recommendations engine got smarter and rewards posts which haven't been read yet. It also queries database in a more effective way to improve performance.

Sunday, August 28, 2011

Critical bug fixed in 0.9.2

First iteration came sooner then I expected. A critical (and embarrassing) bug was breaking links to the recommended content.

Saturday, August 27, 2011

Post-Miner 0.9.1 released!

After few days of work and learning WordPress API the plugin is officially released. It has been posted to the wordpress.org and can be found at http://wordpress.org/extend/plugins/post-miner/.

While writing release notes I realized recommendations engine will work well only with English language. The reason for that is PorterStemmer library used by the plugin. PorterStemmer is very popular algorithm for stemming words and it works only with English.

I hope the community will like it and come back with some feedback (I'm sure many things could be done better).

Friday, August 26, 2011

Welcome to the Post-Miner dev blog

Hello World! I created this blog to write about the Post-Miner, ongoing development and related subjects.

Post-Miner is a content recommendations engine for WordPress which will help your audience discover relevant posts on your blog without searching. Idea behind the plugin is to generate more page views and in the future stream traffic between blogs.

While I'm waiting for the WordPress.org team to grant me an access to the extensions SVN let me tell you few things about the plugin.

On activation Post-Miner transforms all posts into "phrase vectors". If you are not familiar with the term it means all words from the title, categories and tags receive a weight (float number from 0..1). This generalization is helpful for data mining and quick calculations.
Post content won't be index for a while. It adds more complications and the main goal right now is to make the software stable.

Current version recommends only similar posts. Recommendations are calculated with Euclidean distance which is a simple but effective method.

In close future I'm planning to add:
* user interest vector
* some collective intelligence to help with ranking and add some variety to the recommendations