Kickoff TRACK. Plantilla

1
Sommaire projet TRACK

• Présentation le consortium et les objectifs
• Formalisme
• Données temporelles
• Fusion multi-classifieur semi supervisé
• Applications
• CEE prévision des taux de change
• Nationwide offre de carte de crédit
• Banco Santander utilisation de carte de crédit
• Caisse dÉpargne optimisation ATM
• Conclusion intégration pas à pas des
  connaissances a priori

2
Présentation consortium Track Project
Utilisateurs Concepteurs R D
Nation Wide Anite system
Instituto de Tecnologia del Conocimiento
Caisse dEpargne ISOFT
Banco Santander Ibermatica
3
Plate-forme Décisionnelle Combinaison dagents
/ experts
Instituto Universitario de Tecnología del
Conocimiento Universidad Complutense.
4
FUSION
Concept de la fusion pour le data mining
Cumuler des pièces dinformation (EAReL group)
Dans quel but ? Synthétiser de multiple
information Faire la synergie objective entre
différents opinions Limiter les erreurs de
décision (True Knowledge)
Système Optimisé par fusion dinformation
Fusionner des information issues de résultat de
Software Détecter des gènes de comportements
dans le temps
5
Formalisme management des données temporelles
...
t0
t1
t2
T1
Time
Ft
Ft
C1
?t,t1
?t-1,t
C2
?t-1,t1
C3
Prévision Filtre de Kalmann
6

Formalisme les données
R experts données symboliques
Espace discret
L Õr1,,R Lr
Matrice carrée L x L
N mesures sur T périodes
Projection
7
Modèle de fusion
P(c) M! / (n1! n2! ... nL!) . ?l1,...,L Plnl
M états sur L niveaux dénergie
Algorithm boosting Stochastic Mechanic
Iteration (Boosting Bartlett)
1 0 0 0 1 0 0 0 1
1 1 0 1 1 0 0 0 1
Step T0
Step T-gt 0
Algorithm
Energy Cost U
Minimum Energy U
8
Résultats
"Generating overlapping clusters",
Cole-Wishart-71 "An Improved Algorithm for the
Jardine-Sibson Method of Generating Overlapping
Clusters".
9
The TRACK Project
Active Decision Support Toolkit for the Financial
Sector
Application Time Series Data Analysis
10
The TRACK Project
Active Decision Support Toolkit for the Financial
Sector
Time Series Data View (6)
Dollar Against Yen
11
The TRACK Project
Active Decision Support Toolkit for the Financial
Sector
Time Series Application
Exchange Rate between Money
Against
Models estimating Exchange Rate during 218 days
Model Characteristic Lower 5 standard Volatility
Models
Model Characteristic Upper 5 standard Volatility
Models A Priori Knowledge
Model Characteristic Sub 1 being the average of
all models
Final CEC Review
12
The TRACK Project
Active Decision Support Toolkit for the Financial
Sector
Track Input Source Data Content the zi.csv file
Scoring each volatility Exchange Rate money-model
Statistic Variables
Adding Knowledge Temporal Handling Variables
13
The TRACK Project
Active Decision Support Toolkit for the Financial
Sector
Time Series Normalisation Error Trend (TTDM)
14
The TRACK Project
Active Decision Support Toolkit for the Financial
Sector
10 analyses
Experts add a priori knowledge about what do they
want
Experts with specific questions Expert-Model Is
there a set of model better than other
? Expert-Money Is there a set of money having
special behaviour ? Expert-Info Could we
separate models according to their
characteristics ?
Experts looking for variables explaining
behaviour Expert-Global Scoring the Statistic
variables (Good or Bad) Expert-Cluster 3
classifications supervised (2, 3 and 4
classes) Expert-Alice 3 classifications with
label meaning (5, 6 and 9 classes)
15
The TRACK Project
Active Decision Support Toolkit for the Financial
Sector
Track Output A Priori Data Content the of.csv
file
16
The TRACK Project
Active Decision Support Toolkit for the Financial
Sector
Evaluation des analyses
17
Comparaison SOFI Cluster 4C
18
Modèles retenus
19
The TRACK Project
Active Decision Support Toolkit for the Financial
Sector
Description fusion so.csv file Class number 3
20
The TRACK Project
Active Decision Support Toolkit for the Financial
Sector
S.O.F.I. Results
Dollar Against Yen
Deutsche Mark Against Yen
21
The TRACK Project
Active Decision Support Toolkit for the Financial
Sector
Conclusion de lexpérience
Quels sont les modèles les mieux adaptés à
lévaluation ? - Caractéristiques des taux de
change - Dans le cas où plusieurs modèles sont
considérés bons, le meilleur est la moyenne des
modèles
22
Nationwide End User Application
Nationwides Questions Which of our customers
should be offered credit cards? Which of our
customers will be profitable as credit card
customers?
23
Banco Santander

• Which are the characteristics of the customers
  who cancel their credit/debit cards?
• And which are the characteristics of the canceled
  cards?
• How can we analyze credit/debit card customer
  behavior in order to minimize the number of
  cancellations?

• Regarding Client Data the universe to try should
  include
• Particular people (no companies)
• All titular of Credit/Debit cards
• Historical information (15 Months)

24
Caisse dÉpargne ATM

• Essential complements to
• open interchange fee strategy
• bank trade evolution
• indirect profitability
• direct profitability
• services
• appropriated by offices personnel

25
Activity analysis
Act
20
.

• Customer activity on other banks
• 18 per year
• Non customers activity
• from 11 to 7

Clients
18
non
Act
16
.
clients
14
12
10
8
6
4
2
0
95
96
26
The TRACK Project
Active Decision Support Toolkit for the Financial
Sector
Conclusion
27
The TRACK Project
Active Decision Support Toolkit for the Financial
Sector
Conclusion
Pour faire du datamining - le recueil des
données laborieux est la clé de la réussite -
les outils pour manipuler les données - des
experts métiers savoir quoi sur quoi, qui et
quand.