We have just released the new version of Reco4J. We work a lot on it in order to add new interesting algorithms, like SVD++ and Binary Matrix Factorization, and to improve the neo4j/reco4j adding new powerful features like pre and post filtering and named recommendation.

The Reco4J project will be presented by Luigi Giuri on Thursday, 26 September 2013, with a conference session at JavaOne 2013 in San Francisco.

After less than one year from the foundation, Reco4J has caught the interest of the biggest conference in the world about modern software development. With a so special and skilled attendance, the session will give both an overview of the concepts behind the project design and a hands-on demo of how to start using Reco4J. Join us.

The main goal of this version is to improve the general quality of the source code and simplifing the use of the API, adding new interesting features. We add the concept of Model, Predictor and UserItemDataset to add versatility to our infrastructure. In particular we add context-aware functionalities implementing prefilter and postfilter (implemented by means of cypher query) mechanism to filter out users and items before and after the recommendation process. We improve the update mechanism to intercept also the change in the recommendation value or the delete of a rank relationship. Moreover we implement in the reco4j-neo4j project a lot of examples that show how to use our software in several case.

Today we release a neo4j java client that allows accessing neo4j functionalities by means of a really simple library. You don't have to know how to parse/decode the response from json, how to submit request to get nodes, add nodes, add properties and so on. It was implemented to access Reco4J REST interface (Reco4J plugin for Neo4J) but now it is a general purpose Neo4J client. It is in an early release, any bug reporting or contribution is really appreciated.

In the new version we improve the performance of the Matrix Factorization recommender. We run several tests on the movie lens 100K database with both the implementation, the old and the new one.

These are the details about the source database:

Items:        1682

Users:         943

Ratings:     80000

We calculate the time required to build the recommender (with 20 features) on such database, with the following result:

Previous Version:

TotalTime: 451985 milliseconds

Current implementation:

Total Time: 205279 milliseconds

The main goal of this version is improve the performance of the previous version and to provide a new recommender based on Mahout functionalities. In particular we improve the performance of Matrix Factorization  Recommender that allows to prepare the recommender very quckly and provide recommendetion even faster.