Collaborative Filtering and Drupal

By kyvinh@hcilib.sourceforge.net on 18 Dec 2002 at 16:28 UTC

The new Recommendation Module for Drupal is a "collaborative filtering" engine. This means that the module predicts interesting nodes, according to your personal tastes.

The innovative principle is based on "automated user-based collaborative filtering". Check, for example, Amazon.com's "Customers who bought this book also..." or the MovieLens university research project.

Think of it as a "word-of-mouth" system, where friends would recommend you interesting nodes.
---
As an example, let's say that you rate a couple of nodes (as "Not Recommended", "Recommended", or "Highly Recommended"). After rating a few nodes, the module will match your profile with the other users on the website. Now knowing who your "friends" are, the module can list the nodes that you might be interested in.

The module can be downloaded at http://hcilib.sourceforge.net/kyvinh/recommendation.zip and is compatible with Drupal 4.1RC. For developers, the package includes a DEVELOP file that summarizes the code.

Additional pointers:
Research papers and links to working systems like the great MovieLens or the pathetic TV Recommender.

Comments

Context-sensitive?

Dries commented 19 December 2002 at 10:27

Are recommended posts, related posts or not necessarily? That is, can the module show a block in the sidebar with context-sensitive recommendations? Say, will it attempt to recommend financial stories when reading a finance related story, or will it recommend a story about say, the fact they found fossil evidence that man and bonobos share a common ancestor?

No, difference between user- and item-based filtering

kyvinh@hcilib.sourceforge.net commented 19 December 2002 at 10:58

These recommenders have been extensively researched academically. And in conclusion, they derived 2 primary models of recommendations.
The recommendation module is a "user-based" recommender, whereas your description fits the "item-based" recommenders.

This research by a very authoritative source on Information Filtering (Pr. Herlocker) outlines the differences between the 2 systems. User-based filtering (or Collaborative Filtering) permits (1)the filtering of hard-to-analyze data, (2) filtering based on quality and taste and not 'objective' meta-data, and (3) recommendations of serendipitous articles--articles that you did not expected but that you hopefully like.

Oops, I should have answered your question directly: No, the recommended posts are not necessarily related posts. However, the use of such module is usually within a domain of knowledge. For example, on drupal.org, it will be unlikely for you to be recommended posts on bonobos
So, the "related posts" block will be the job for another module

Follow-up on the use of technology

kyvinh@hcilib.sourceforge.net commented 19 December 2002 at 10:36

A lot of questions arise from this recommender system.
How would you use such recommendation engine? Should such technology replace good old-fashioned word-of-mouth? Or how would Drupal interact with it? Doesn't Drupal already give many ways to interact and promote interesting pages?