A Java Toolbox for the Analysis ofMassIve Data STreams using

Probabilistic Graphical Models

DescriptionWe present a Java toolbox for scalable probabilistic machine learning.

Probabilistic machine learning

Model your problem using a flexi-ble probabilistic language based ongraphical models. Then, fit it withdata using a Bayesian approach tohandle modelling uncertainty.

H x

Multi-core and distributed processing

AMIDST provides tailored parallel anddistributed implementations of Bayesianparameter learning for batch and stream-ing data. This processing is based onflexible and scalable message passing al-gorithms.

Main Features

Probabilistic Graphical Scalable inference Data StreamsModels

Specify your model usingprobabilistic graphical mod-els with latent variables andtemporal dependencies

Perform inference on yourprobabilistic models withpowerful approximate andscalable algorithms.

Update your models whennew data is available. Ap-propriate for learning fromdata streams.

WEKA

Large-scale Data Researchers Interoperability

Use your defined models toprocess massive data sets ina distributed computer clus-ter using Flink or Spark.

Flexible toolbox for re-searchers performing theirexperimentation in machinelearning.

Leverage existing function-alities and algorithms by in-terfacing to other softwaretools.

Real world and highly complex use-cases

Risk prediction in creditoperations

Recognition of trafficmaneuvers

Risk prediction in credit opera-tions. Financial data is collectedcontinuously and reported on amonthly basis. We explore it asan evolving classification prob-lem and co. This work has beenperformed in collaboration withone of our partners, the Spanishbank BCC.

Prototype models for early recog-nition of traffic maneuver inten-tions. Data is continuously col-lected by car on-board sensorsgiving rise to a large and quicklyevolving data stream. This workhas been performed in collabo-ration with one of our partners,DAIMLER.

Identifying global trends in the financial sector

Individual trends forfinancial indicatorsWe can identify seasonal

drifts and gradual changes.

{E[H1,t]}−credit{E[H2,t]}−income{E[H3,t]}−expenses{E[H4,t]}−balance{E[H5,t]}−mortgages{E[H6,t]}−loans

Global trend forfinancial indicatorsHighly correlated with

unemployment rate (UR).Pearson’s correlationcoefficient is 0.96.

Apr

200

7

Aug

200

7

Dec

200

7

Mar

200

8

Jul 2

008

Oct

200

8

Feb

200

9

Jun

2009

Sep

200

9

Jan

2010

Apr

201

0

Aug

201

0

Dec

201

0

Apr

201

1

Sep

201

1

Jan

2012

Jun

2012

Oct

201

2

Mar

201

3

Jul 2

013

Dec

201

3

{UR}{E[Ht]}

Team

Andres Masegosa Ana M. Martınez Darıo Ramos-Lopez Rafael Cabanas Thomas D. Nielsen Helge Langseth Antonio Salmeron Anders L. MadsenDeveloper Developer Developer Developer Scientific Member Scientific Member Scientific Member Scientific Member

NTNU Aalborg University University of Almeria Aalborg University Aalborg University NTNU University of Almeria HUGIN EXPERT S/A

Academic and Industrial partners Contact

• Visit http://amidst.github.io/toolbox/ tosign up to join the growing AMIDST community ordownload the software.

The AMIDST project has received funding from the European Union’s Seventh FrameworkProgramme for research, technological development and demonstration under grant agreement no

619209. More information on amidst.eu.