The Analytics1305 Blog

September 2010 Release Now Available

aToF — Mon, 30 Aug 2010 18:24:18 +0000

The September 2010 Release of the Analytics1305 Machine Learning Library is now available for download. The following new features are available:

QuicSVD: A new addition to our library; it is a fast singular value decomposition method for dimensionality reduction.
Support Vector Machines: Added polynomial kernel to the list of non-linear kernels.

As usual there have been bug fixes and updates to the online documentation.

Analytics1305 Development Team.

Some issues with duplicate points

aToF — Tue, 13 Jul 2010 21:46:18 +0000

We are continuously testing our library with several datasets. A recent test uncovered a problem of kmeans and other algorithms that use trees. The problem appears when the dataset contains duplicate points. Indexing a tree might fail if the number of duplicates is larger than the leaf_size. As a temporary fix you can increase your leaf_size or remove the duplicates. We have fixed the problem and we will release the new binaries beginning of August.

The problem is due to a compiler (gcc) optimization that turns out to be unsafe. Windows version does not have this problem.

Problem with Linux binaries fixed

aToF — Fri, 09 Jul 2010 15:03:56 +0000

Some of you had problems running the linux binaries due to dynamic linking with lapack and boost. We fixed that problem with static linking. The library doesn’t have any dynamic linking dependencies anymore, so it should run fine on any system. The new binaries are available on the download page. If you keep facing problems email us at support@analytics1305.com

July 2010 Release Now Available

aToF — Wed, 07 Jul 2010 14:06:40 +0000

The July 2010 Release of the Analytics1305 Machine Learning Library is now available for download. As promised the following new features are available:

K-Means: online preprocessing, better seeding with kmeans++, automatic determination of K using BIC and/or XMeans
New Algorithms: Bregman Divergence Neighbors which is ideal for documents, count data and histograms.
Linear Regression: Online mode.

Additionally we have included a Support Vector Machine algorithm which uses the Sequential Minimal Optimization algorithm. The binaries support the linear and non-linear Gaussian kernel.

Analytics1305 Development Team.

July 2010 Release Note

aToF — Sun, 20 Jun 2010 23:49:02 +0000

Dear Friends,

We are pleased to announce the July 2010 release of the Analytics1305 Machine Learning Library. The release will be available for download on Friday the second of July. This release will include the following new features:

K-Means: online preprocessing, better seeding with kmeans++, automatic determination of K using BIC and/or XMeans
New Algorithms: Bregman Divergence Neighbors which is ideal for documents, count data and histograms.
Linear Regression: Online mode.

As always our releases include bug fixes as we isolate them.

Nearest Neighbor Classifier Bug Fixed

Varchas — Fri, 04 Jun 2010 18:28:27 +0000

There was a bug in the windows version of the nearest neighbor classifier that has been fixed. Specifically, the program would die quietly after performing nearest neighbors whilst reading the input labels. This has been fixed.

Analytics1305 Releases Binaries for Download

aToF — Thu, 27 May 2010 16:29:31 +0000

Dear Friends,

We have received numerous requests from our users to make our algorithms available for downloading and we are pleased to announce our first (beta) release. Users can now download binaries for our machine learning algorithms natively compiled for the following operating systems:

Windows
Fedora
Ubuntu
Debian

The current release only supports 32 bit platforms with 64 bit planned for a future release. As you well know our algorithms are orders of magnitude faster than those available on R, Weka and SPSS. For the current release the following binaries are available:

K-Nearest Neighbors (With Classifier)
K-Means Clustering
Kernel Density Estimation (With Non-Parametric Bayes Classifier built-in and Automatic Metric Learning)
Linear Regression (With Stepwise and VIF Selection on top of basic linear regression)
Non-Negative Matrix Factorization with Sparsity Constraints for Missing Data.

Our algorithms support both dense and sparse data along with progressive modes. We have also revised our documentation with a renewed focus on usage examples. As always we support the release with our forums.

Here is a link to the download page.

Linear Regression, Quic-SVD, Non-negative Matrix Factorization Released on Amazon Ec2

Dong — Wed, 21 Apr 2010 18:11:42 +0000

We are pleased to announce fast implementations of linear regression, singular value decomposition, and non-negative matrix factorization. They are now available on Amazon Ec2.

Our implementation of linear regression supports the variable elimination algorithm based on variance inflation factor and stepwise regression based on the AIC criterion.

Our implementation of singular value decomposition uses the Quic-SVD algorithm that uses a tree structure called the cosine tree.

Our implementation of non-negative matrix factorization supports the factorization based on the specified sparsity level.

For more info on the beta release please refer to our documentation here.

Announcing Hadoop Bootstrapped SVM

Varchas — Wed, 17 Feb 2010 19:54:13 +0000

We are pleased to announce a bootstrapped version of the Analytics1305 Support Vector Machine, available on Amazon Ec2, which utilizes the open source implementation of Google’s map-reduce technology embodied as the Apache Hadoop project.

Bootstrapping is a method employed in machine learning to improve model accuracy and stability (confidence in results). Support Vector Machines are the best off-the-shelf classifier technology available today. The combination of the two makes a powerful and versatile classifier for most types of data and we have made every effort to make the whole analysis process from raw data to classification results as simple as possible. For more info on the beta release please refer to our documentation here.

SVM Classifier Released On Amazon Ec2

Varchas — Fri, 25 Dec 2009 23:18:04 +0000

We are very excited to announce a Support Vector Machine (SVM) classification algorithm available on our free public cloud machine as well as free public AMI. It is a 2-class classifier which implements the SMO optimization algorithm. The salient features are as follows:

Implements 2 class Sequential Minimal Optimization (SMO)
Uses 1305 file formant and thus can handle categorical and sparse data formats
Generated models can be saved and used for classification at any time
Offers 2nd order working set selection and caching techniques for speed

Our benchmarks beat the popular LibSVM and SVM-Light tools for standard problems from the UCI machine learning datasets repository. Please find documentation and samples on how to run the SVM here.