Reco4j

This is the reco4j tumblr blog. Reco4j is an open source graph-based recommendation engine. The main site is: http:// www.reco4j.org

[reco4j]

Version 0.5.0 available today!

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.

Click here to get the list of the new features.

[reco4j]

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.

Find session details here: http://goo.gl/mJN1lO

[reco4j]

Version 0.4.0 available today!

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.

More details can be find here: http://www.reco4j.org/release-notes.jsp

[reco4j]

Neo4J Java Client

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.

Details for download can be found here: http://www.reco4j.org/download.jsp

[reco4j]

First tests with the new release

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 improvement is of more than 2 times. In both cases the MAE is 0.736437751264315.

[reco4j]

Version 0.3.0 available

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.

More details can be find here: http://www.reco4j.org/release-notes.jsp

[reco4j]

Reco4j Terracotta Cache Test

Starting from #Neo4j 1.9M4 it is really simple and straightforward to implement a custom cache for nodes and relationships. We deploy a cache based on #Terracotta #BigMemory. We start testing it, but our test shows that the best performance are reached using internal cache. BigMemory adds little overhead and we can't use off-heap storage (since at the moment NodeImpl and RelationshipImpl can't be serialized) to use all BigMemory functionalities. In any case we will use it as Reco4J cache, since it is a really powerful tool for second level cache.

[reco4j]

Reco4J + Neo4J + Mahout deployment started and first test

In these days, after the release of the version 0.2.0, we are working to integrate Apache Mahout as recommendation engine for reco4j. We have just successfully completed the first test session. Now we will work to optimize this integration and to provide all the mahout recommendation's functionalities through Reco4j/Neo4j.

[reco4j]

Version 0.2.0 finally available

The new version of Reco4J is now available. The main goal of this version is to provide reco4j functionalities by means of a Neo4j plugin. This version can be installed as plugin of neoj so that the recommendation can be obtained with REST requests to the neo4j http server.

Moreover a Event Handler has been implemented in order to update recommendation data.It listens if there is a new node or a new relation and recalculate only the part of the recommendation involved into the change.

More details can be find here: http://www.reco4j.org/release-notes.jsp

[reco4j]

Version 0.2.0 is coming ...

We are working on the new version of Reco4j. This new version, among other things, uses a different apporach to graph traversal; now it is not based on Gremlin but on native call to Neo4J Kernel, this improves radically the performance. Moreover this reduces the number of dependecy of the reco4j-neo4j project simplifing the deploy.

[reco4j]

Someone is interested ...

In the first 10 days from the beginning of the project we received 800 different contact. This is very important for us, now we know that this project can be useful and interesting. We will work hard to accelerate the developement providing new versions with new functionalities in the next days. We will try to deploy one version per month.

[reco4j]

The first version available

Today we have released the first version fo Reco4J. This is a very early version that contains only some first funcionalities. It allows to make recommendation using its API, leverging on collaborative filtering and matrix factorization algorithms. Our principal goal is to provide a very powerful and complete set of recommendation funcionalities on a preexisting graph database. The current version uses the Neo4J as graph database management system.