SD - Solutions

Learning About Data Relationships (LADR, pronounced "ladder") is a high performance and scalable software solution which addresses the challenge of finding related entities within a dataset. This challenge is present in the financial industry (market monitoring and analysis), the scientific industries (drug discovery and gene analysis) and more. It is also present for any business that aims to optimize its processes through the use of business analytics. LADR was designed to handle large datasets which are common in real world complex systems. LADR is scalable and can run in parallel on multiple processing units. It can be adapted to any computing platform and customized to meet any domain-specific needs.

How Does it Work?

LADR performs analysis based upon a statistical structure called a Bayesian network. Such networks decribe relationships between variables in a system. Finding the Bayesian network that best matches the input data is the goal of LADR. This process is called "structure learning". LADR's input is independent sample data, no relationship information is required. All relationship information is computed by LADR. The input data can have a static (time-independent) interpretation or a dynamic (time-dependent or time-series) interpretation. The following example illustrates these concepts.

A Static Example:

Input

A	B	C	D
1.5	0.004	30198	-1
2.0	0.050	90017	0
3.6	0.420	98175	-2
-1.8	0.008	51982	-1
5.6	0.030	87192	-1
2.4	0.670	34710	-1
-2.3	0.012	21740	-2
-9.1	0.045	19740	0
3.5	0.001	9179	0
6.7	0.230	19277	-2

Output

A Dynamic Example:

Input

time	A	B	C	D
0	1.5	0.004	30198	-1
1	2.0	0.050	90017	0
2	3.6	0.420	98175	-2
3	-1.8	0.008	51982	-1
4	5.6	0.030	87192	-1
5	2.4	0.670	34710	-1
6	-2.3	0.012	21740	-2
7	-9.1	0.045	19740	0
8	3.5	0.001	9179	0
9	6.7	0.230	19277	-2

Output

Screenshot of what a run of LADR looks like:

Performance

LADR is very efficient in both time and space. For the performance numbers quoted here, a standard PC (2.2 GHz CPU, 4 GB RAM) was used. A static dataset of 37 variables and 200 data points can be computed at a rate of over 37 Million candidate Bayesian networks evaluated per minute. A dynamic dataset of 20000 variables and 100 data points can be computed at a rate of over 9 Million candidate Bayesian networks evaluated per minute. Datasets of 1 Million variables have been successfully computed on this standard PC, taking up only 350 MB of system memory.

Scalability

The algorithms used by LADR are scalable to multiple compute elements. On a standard PC with 4 CPU cores, a near perfect 4X scaling has been observed when using a multi-threaded version of LADR. This scaling also applies to arbitrary compute clusters due to the data-independence of the underlying search algorithms.

Want to Know More?

A great way to get more familiar with LADR is to download the trial version and try it with your data.

We are available to answer any questions you may have about LADR:

info@structureddata01.com for any kind of additional information

sales@structureddata01.com for sales information