DB

MG Data mining fundamentals

DataBase and Data Mining Group of Politecnico di Torino

Data analysis

Most companies own huge databases
containing

operational data
Data mining fundamentals
textual documents

experiment results

DB

These databases are a potential

MG
source of useful information
Data Base and Data Mining Group of Politecnico di Torino

Elena Baralis
Politecnico di Torino
DB
MG
2

Data analysis Data mining

Information is hidden in huge datasets
Non trivial extraction of

not immediately evident
implicit

human analysts need a large amount of time for the
previously unknown
analysis
potentially useful

most data is never analyzed at all information from available data
4,000,000

3,500,000 The Data Gap

Extraction is automatic
3,000,000
performed by appropriate algorithms
2,500,000

2,000,000

Extracted information is represented by means of
1,500,000
Disk space (TB) abstract models
since 1995
1,000,000
denoted as pattern
500,000
Analyst
number
0

DB DB
1995 1996 1997 1998 1999

MG MG
3 4
From R. Grossman, C. Kamath, V. Kumar, Data
Mining for Scientific and Engineering Applications

Example: profiling Example: profiling

User/service profiling

Consumer behavior in e-commerce sites

Recommendation systems

Selected products, requested information,

Advertisements

Search engines and portals

Market basket analysis

Query keywords, searched topics and objects

Correlated objects for cross selling

Social network data
User registration, fidelity cards

Facebook, google+ profiles
Context-aware data analysis

Dynamic data: posts on blogs, FB, tweets
Integration of different dimensions

Maps and georeferenced data
E.g., location, time of the day, user interest

Localization, interesting locations for users
Text mining

Brand reputation, sentiment analysis, topic trends

DB
MG
5 DB
MG
6

Elena Baralis
Politecnico di Torino
1
 DB
MG Data mining fundamentals
DataBase and Data Mining Group of Politecnico di Torino

Example: biological data Biological analysis objectives

Microarray
Clinical analysis

expression level of genes in a cellular tissue
detecting the causes of a pathology

various types (mRNA, DNA)
monitoring the effect of a therapy
diagnosis improvement and definition of new specific

Patient clinical records CLID
PATIENT shx013: shv060: shq077: shx009: shx014: shq082: shq083: shx008:
therapies
ID 49A34 45A9 52A28 4A34 61A31 99A6 46A15 41A31

personal and demographic data IMAGE:740604
ISG20 || interferon-1.02
IMAGE:767176
TNFSF13 || tumor-0.52
stimulated-2.34
gene 20kDa
necrosis -4.06
1.44
factor (ligand)
0.57 -0.13
-0.29superfamily,
0.71member1.03
13 -0.67
0.12 0.34
0.22
-0.51
-0.09

Bio-discovery

exam results IMAGE:366315
LOC93343 || **hypothetical
-0.25 -4.08 protein BC011840
0.06 0.13 0.08 0.06 -0.08 -0.05
IMAGE:235135
ITGA4 || integrin,-1.375
alpha 4 (antigen
-1.605 CD49D, 0.155alpha -0.015
4 subunit of0.035
VLA-4 receptor)
-0.035 0.505 -0.865
gene network discovery

Textual data in public collections
analysis of multifactorial genetic pathologies

heterogeneous formats, different objectives

Pharmacogenesis

scientific literature (PUBMed)

lab design of new drugs for genic therapies

ontologies (Gene Ontology)

DB
MG
7 DB
MG
8

Knowledge Discovery Process Preprocessing

data cleaning
reduces the effect of noise
selection
identifies or removes outliers
preprocessing solves inconsistencies
preprocessing
data integration
transformation reconciles data extracted
data selected from different sources
selected data mining data integrates metadata
data preprocessed identifies and solves data
preprocessed interpretation
data value conflicts
data manages redundancy
transformed
data Real world data is dirty
pattern
knowledge Without good quality data, no good quality
KDD = Knowledge Discovery from Data pattern
DB
MG
10 DB
MG
11

Data mining origins Analysis techniques

Draws from

statistics, artificial intelligence (AI)

Descriptive methods

pattern recognition, machine
Extract interpretable models describing data
learning Machine Learning,
Example: client segmentation
Statistics,
database systems AI Pattern


Recognition
Predictive methods

Traditional techniques are not

Exploit some known variables to predict
appropriate because of Data Mining
unknown or future values of (other) variables

significant data volume
large data dimensionality

Example: spam email detection

Database
systems

heterogeneous and distributed
nature of data
From: P. Tan, M. Steinbach, V. Kumar,
Introduction to Data Mining

DB
MG
12 DB
MG
13

Elena Baralis
Politecnico di Torino
2
 DB
MG Data mining fundamentals
DataBase and Data Mining Group of Politecnico di Torino

Classification Classification

Objectives Approaches
decision trees

prediction of a class label
bayesian classification

definition of an interpretable model of a given classification rules
phenomenon neural networks
k-nearest neighbours
training data training data SVM

model model

unclassified data classified data unclassified data classified data

DB
MG
14 DB
MG
15

Classification Classification
Requirements
Applications
accuracy
detection of customer propension to leave a company
(churn or attrition)
interpretability
fraud detection
scalability
classification of different pathology types
noise and outlier

management
training data dati di training

model modello

unclassified data classified data dati non classificati dati classificati

DB
MG
16 DB
MG
17

Clustering Clustering
Approaches

Objectives
partitional (K-means)

detecting groups of similar data objects
hierarchical

identifying exceptions and outliers
density-based (DBSCAN)
SOM

Requirements
scalability
management of
noise and outliers
large dimensionality
interpretability

DB
MG
18 DB
MG
19

Elena Baralis
Politecnico di Torino
3
 DB
MG Data mining fundamentals
DataBase and Data Mining Group of Politecnico di Torino

Clustering Association rules

Applications
Objective

customer segmentation
extraction of frequent correlations or pattern from a

clustering of documents containing similar information transactional database

grouping genes with similar expression pattern

Tickets at a supermarket
counter
Association rule
TID Items diapers beer
1 Bread, Coke, Milk

2% of transactions contains
2 Beer, Bread both items
3 Beer, Coke, Diapers, Milk
30% of transactions
4 Beer, Bread, Diapers, Milk containing diapers also
5 Coke, Diapers, Milk contain beer

DB
MG
20 DB
MG
21

Association rules Other data mining techniques

Applications
Sequence mining

market basket analysis
ordering criteria on analyzed data are taken into

cross-selling account

shop layout or catalogue design
example: motif detection in proteins

Time series and geospatial data
Tickets at a supermarket

temporal and spatial information are considered
counter
Association rule

example: sensor network data
TID Items diapers beer
Regression
1 Bread, Coca Cola, Milk Sensor network

2% of transactions contains
prediction of a continuous value
2 Beer, Bread both items

example: prediction of stock quotes
3 Beer, Coca Cola, Diapers, Milk
30% of transactions

Outlier detection
4 Beer, Bread, Diapers, Milk containing diapers also

example: intrusion detection in network traffic
5 Coca Cola, Diapers, Milk contain beer analysis

DB
MG
22 DB
MG
23

Open issues

Scalability to huge data volumes

Data dimensionality

Complex data structures, heterogeneous data
formats

Data quality

Privacy preservation

Streaming data

DB
MG
24

Elena Baralis
Politecnico di Torino
4