Aside from some performance improvements, there are two new functionality updates in this release.
Firstly, a user tag based affinity modelling and item recommendation algorithm. This provides a simple use case where clients have keyword tagged items. The offline modeling stage associates users to tags they have interacted with more than the average user based on activity data. The online recommender then utilizes Seldon’s existing continually updated activity based stats to recommend currently popular items that are being interacted with that match the user’s tags utilizing the associated weights to those tags. This recommender is therefore ideal for high churn media websites.
A second change is to allow multiple dimensions in recommendation calls. Seldon already provides several ways items can be included and filtered in item recommendations calls. One such way is that Seldon allows items to be placed in so-called “dimensions” to allow clients to split items into 1 of N classes. For example on a news site an article might be in the sports, news or celebrity section. Recommendation calls can now specify a list of dimensions to return recommendations in, so for example only recommend articles which are in the sports or celebrity section.
Version 0.94 is now available on Github and the AWS AMI and Vagrant virtual machine builds have been updated.