Data Quality: Why Preprocess the Data?

 Measures for data quality: A multidimensional view

 Accuracy: correct or wrong, accurate or not

 Completeness: not recorded, unavailable, …

 Consistency: some modified but some not, dangling, …

 Timeliness: timely update?

 Believability: how trustable the data are correct?

 Interpretability: how easily the data can be understood?

2
 Major Tasks in Data Preprocessing
 Data cleaning
 Fill in missing values, smooth noisy data, identify or remove outliers, and
resolve inconsistencies
 Data integration
 Integration of multiple databases, data cubes, or files
 Data reduction
 Dimensionality reduction
 Numerosity reduction
 Data compression
 Data transformation and data discretization
 Normalization
 Concept hierarchy generation

3
 Data Cleaning
 Data in the Real World Is Dirty: Lots of potentially incorrect data,
e.g., instrument faulty, human or computer error, transmission error
 incomplete: lacking attribute values, lacking certain attributes of
interest, or containing only aggregate data
 e.g., Occupation=“ ” (missing data)
 noisy: containing noise, errors, or outliers
 e.g., Salary=“−10” (an error)
 inconsistent: containing discrepancies in codes or names, e.g.,
 Age=“42”, Birthday=“03/07/2010”
 Was rating “1, 2, 3”, now rating “A, B, C”
 discrepancy between duplicate records
 Intentional (e.g., disguised missing data)
 Jan. 1 as everyone’s birthday?

4
 Incomplete (Missing) Data
 Data is not always available
 E.g., many tuples have no recorded value for several
attributes, such as customer income in sales data
 Missing data may be due to
 equipment malfunction
 inconsistent with other recorded data and thus deleted
 data not entered due to misunderstanding
 certain data may not be considered important at the time of
entry
 not register history or changes of the data
 Missing data may need to be inferred

5
 How to Handle Missing Data?
 Ignore the tuple: usually done when class label is missing (when
doing classification)
 not always effective: values from remaining attributes wasted

 Fill in the missing value manually: tedious + infeasible?

 Fill in it automatically with
 a global constant : e.g., “unknown”, a new class?!
 the attribute mean
 the attribute mean for all samples belonging to the same
class: smarter
 the most probable value: inference-based such as Bayesian
formula or decision tree
6
 Noisy Data
 Noise: random error or variance in a measured variable
 Incorrect attribute values may be due to
 faulty data collection instruments
 data entry problems
 data transmission problems
 technology limitation
 inconsistency in naming convention

7
 How to Handle Noisy Data?
 Binning
 first sort data and partition into (equal-frequency) bins
 then one can smooth by bin means, smooth by bin median,
smooth by bin boundaries, etc.
 Regression
 smooth by fitting the data into regression functions
 Clustering
 detect and remove outliers
 Combined computer and human inspection
 detect suspicious values and check by human (e.g., deal with
possible outliers)

8
Smoothing data by binning: examples

9
 Smoothing data by regression: example

Object temperature melting

4
id time
1 10.3 3.2
melting time

3 2 23.6 2.5
3 54.2 0.9
2
4 11.7 3.0
1 … … …

10 20 30 40 50 60
temperature values in the line

10
Remove outliers with clustering

11
 Data Reduction Strategies
 Data reduction: Obtain a reduced representation of the data set that is
much smaller in volume but yet produces the same (or almost the same)
analytical results
 Why data reduction?
 Curse of dimensionality
 Efficiency
 Data reduction strategies
 Dimensionality reduction, e.g., remove unimportant attributes
Wavelet transforms
 Principal Components Analysis (PCA)
 Feature subset selection, feature creation
 Numerosity reduction (some simply call it: Data Reduction)
 Regression and Log-Linear Models
 Histograms, clustering, sampling
 Data cube aggregation
 Data compression

12
13

Dimensionality Reduction
 Curse of dimensionality
 When dimensionality increases, data becomes increasingly
sparse
 Obj = (x1, … , xn)
 Likelihood of many zeros increases
 Density and distance between points, which is critical to
clustering, outlier analysis, becomes less meaningful
 d(Oi, Oj) = 𝑥1 − 𝑦1 2 + ⋯ + 𝑥𝑛 − 𝑦𝑛 2

 The possible combinations of subspaces will grow exponentially

 There are 2n subspaces in a n-dimensional space
14

Dimensionality Reduction
 Dimensionality reduction
 Avoid the curse of dimensionality
 Help eliminate irrelevant features and reduce noise
 Reduce time and space required in data mining
 Allow easier visualization
15

Dimensionality Reduction
 Dimensionality reduction techniques
 Wavelet transforms (omitted)
 Principal Component Analysis
 Supervised and nonlinear techniques (e.g., feature
selection)
 Principal Component Analysis (PCA)
 Find a projection that captures the largest amount of variation in data
 The original data are projected onto a much smaller space, resulting in
dimensionality reduction. We find the eigenvectors of the covariance matrix,
and these eigenvectors define the new space

x2

x1
16
Principal Component Analysis (Steps)
 Given N data vectors from n-dimensions, find k ≤ n orthogonal vectors
(principal components) that can best represent data
 Normalize input data: Each attribute falls within the same range
 Compute k orthonormal (unit) vectors, i.e., principal components
 Each input data (vector) is represented by a linear combination of the k
principal component vectors
 The principal components are sorted in order of decreasing “significance”
or strength
 Since the components are sorted, the size of the data can be reduced by
eliminating the weak components, i.e., those with low variance (i.e., using
the strongest principal components, it is possible to reconstruct a good
approximation of the original data)
 Works for numeric data only
17
 Attribute Subset Selection
 Another way to reduce dimensionality of data
 Redundant attributes
 Duplicate much or all of the information contained in one or
more other attributes
 E.g., purchase price of a product and the amount of sales tax
paid
 Irrelevant attributes
 Contain no useful information for the data mining task at hand
 E.g., students’ address is often irrelevant to the task of predicting
students' GPA

18
 Heuristic Search in Attribute Selection
 There are 2d possible attribute combinations of d attributes
 Typical heuristic attribute selection methods:
 Best single attribute under the attribute independence
assumption: choose by significance tests
 Best step-wise feature selection:
 The best single-attribute is picked first
 Then next best attribute condition to the first, ...
 Step-wise attribute elimination:
 Repeatedly eliminate the worst attribute
 Best combined attribute selection and elimination
 Optimal branch and bound:
 Use attribute elimination and backtracking
19
 Numerosity Reduction
 Reduce data volume by choosing alternative, smaller forms of
data representation
 Parametric methods (e.g., regression)
 Assume the data fits some model, estimate model
parameters, store only the parameters, and discard the
data (except possible outliers)
 Non-parametric methods
 Do not assume models
 Major families: histograms, clustering, sampling, …

20
 Parametric Data Reduction: Regression
 Linear regression
 Data modeled to fit a straight line
 Often uses the least-square method to fit the line
 Multiple regression
 Allows a response variable Y to be modeled as a linear
function of multidimensional feature vector

21
 y
Regression Analysis
Y1

 Regression analysis:
Y1’
 modeling and analysis of numerical data y=x+1
 A parametric approach

 parameters are estimated so as to give X1 x

a "best fit" of the data
 Most commonly the best fit is  Used for prediction (including
forecasting of time-series data),
evaluated by using the least squares
inference, hypothesis testing,
method, but other criteria have also and modeling of causal
been used relationships

22
 Regression Analysis
 Linear regression: Y = w X + b
 Two regression coefficients, w and b, specify the line
and are to be estimated by using the data at hand
 Using the least squares criterion to the known values of
Y1, Y2, …, X1, X2, ….
 Multiple regression: Y = b0 + b1 X1 + b2 X2
 Many nonlinear functions can be transformed into the
above

23
 Histogram Analysis
 Divide data into buckets and 40
store frequency for each bucket 35
 Partitioning rules: 30
 Equal-width: equal bucket 25
range 20
 Equal-frequency (or equal- 15
depth) 10
5
0

100000
10000

20000

30000

40000

50000

60000

70000

80000

90000
24
 Histogram Analysis: examples
Data set:1,1,5,5,5,5,5,8,8,10,10,10,10,12,14,14,14,15,15,15,15,15,15,18,18,
18,18,18,18,20,20,20,20,20,20,20,21,21,21,21,25,25,25,25,25,28,28,30,30,30

Single value-frequency equal-width

25
 Clustering
 Partition data set into clusters based on similarity, and store
cluster representation (e.g., centroid and diameter) only
 Can be very effective if data is clustered but not if data is
“smeared”
 Can have hierarchical clustering and be stored in multi-
dimensional index tree structures
 There are many choices of clustering definitions and clustering
algorithms
 Cluster analysis will be studied in depth later

26
 Sampling
 Sampling: obtaining a small sample s to represent the whole
data set N
 Allow a mining algorithm to run in complexity that is potentially
sub-linear to the size of the data
 Key principle: Choose a representative subset of the data
 Simple random sampling may have very poor performance in
the presence of skew
 Develop adaptive sampling methods, e.g., stratified sampling

27
 Types of Sampling
 Simple random sampling
 There is an equal probability of selecting any particular item
 Sampling without replacement
 Once an object is selected, it is removed from the population
 Sampling with replacement
 A selected object is not removed from the population
 Stratified sampling:
 Partition the data set, and draw samples from each partition
(proportionally, i.e., approximately the same percentage of
the data)
 Used in conjunction with skewed data

28
29

Sampling: With or without Replacement

Raw Data
Sampling: Cluster or Stratified Sampling

Raw Data Cluster/Stratified Sample

30
31

Data Compression
 String compression
 There are extensive theories and well-tuned algorithms
 Typically lossless, but only limited manipulation is possible
without expansion
 Audio/video compression
 Typically lossy compression, with progressive refinement
 Sometimes small fragments of signal can be reconstructed
without reconstructing the whole
 Dimensionality and numerosity reduction may also be considered
as forms of data compression
32

Data Compression

Original Data Compressed

Data
lossless

Original Data
Approximated
33

Data Transformation
 A function that maps the entire set of values of a given
attribute to a new set of replacement values s.t. each old
value can be identified with one of the new values
 Methods
 Smoothing: Remove noise from data

 Normalization: Scaled to fall within a smaller, specified range

 min-max normalization
 z-score normalization
 normalization by decimal scaling
 Discretization: Concept hierarchy climbing
 Normalization
 Min-max normalization: to [new_minA, new_maxA]
v − minA
v' = (new _ maxA − new _ minA) + new _ minA
maxA − minA

 Ex. Let income range $12,000 to $98,000 normalized to

[0.0, 1.0]. Then $73,600 is mapped to
73,600 − 12,000
(1.0 − 0) + 0 = 0.716
98,000 − 12,000

34
 Normalization
 Z-score normalization (μ: mean, σ: standard deviation):
v − A
v' =
 A

 Ex. Let μ = 54,000, σ = 16,000. Then, 73,600 is mapped to

73,600 − 54,000
= 1.225
16,000
 Normalization by decimal scaling for attr. A
v where j is the smallest integer such that max 𝑣 ′ <1
v'= j 𝑣𝜖𝐴
10
 Ex. Let the recorded values for attribute A range from -943 to 917
 So max(|A|) = 943
 The desired value is j = 3, since 943/1000 = 0.943 < 1

35
 Discretization
 Three types of attributes
 Nominal—values from an unordered set, e.g., color, profession
 Ordinal—values from an ordered set, e.g., military or academic rank
 Numeric—real numbers, e.g., integer or real numbers
 Discretization: Divide the range of a continuous attribute into intervals
 Interval labels can then be used to replace actual data values
 Reduce data size by discretization
 Supervised vs. unsupervised
 Split (top-down) vs. merge (bottom-up)
 Discretization can be performed recursively on an attribute
 Prepare for further analysis, e.g., classification

36
 Data Discretization Methods
 Typical methods: All the methods can be applied recursively
 Binning
 Top-down split, unsupervised
 Histogram analysis
 Top-down split, unsupervised
 Clustering analysis (unsupervised, top-down split or bottom-up
merge)
 Decision-tree analysis (supervised, top-down split)
 Correlation (e.g., 2) analysis (unsupervised, bottom-up
merge)

37
 Simple Discretization: Binning

 Equal-width (distance) partitioning

 Divides the range into N intervals of equal size: uniform grid

 if A and B are the lowest and highest values of the attribute, the width of

intervals will be: W = (B –A)/N.

 The most straightforward, but outliers may dominate presentation

 Skewed data is not handled well

 Equal-depth (frequency) partitioning

 Divides the range into N intervals, each containing approximately same

number of samples
 Highly effective for both sparse and dense data

 Managing categorical attributes can be tricky

38
Binning and histogram for discretization
❑ Sorted data for price (in dollars): 4, 8, 9, 15, 21, 21, 24, 25, 26, 28, 29, 34
❑ binning ❑ histogram
* Partition into equal-frequency * Equal width (10)
(equal-depth) bins:
- Bin 1 (1-10): 4, 8, 9
- Bin 1: 4, 8, 9, 15
- Bin 2 (11-20): 15
- Bin 2: 21, 21, 24, 25
- Bin 3: 26, 28, 29, 34 - Bin 3 (21-30): 21, 21, 24, 25,
26, 28, 29
* Smoothing by bin means:
- Bin 1: 9, 9, 9, 9 – Bin 4 (31-40): 34
- Bin 2: 23, 23, 23, 23  Equal frequency (3)
- Bin 3: 29, 29, 29, 29 - Bin 1: 4, 8, 9

* Smoothing by bin boundaries: - Bin 2: 15, 21, 21

- Bin 1: 4, 4, 4, 15 - Bin 3: 24, 25, 26

- Bin 2: 21, 21, 25, 25 - Bin 4: 28, 29, 34

- Bin 3: 26, 26, 26, 34

39
Discretization by clustering or classification

 Clustering:
 group similar values into a single group

 Each group viewed as a single value

 Classification
 Partition the values into groups so that they are most co-

related to the class labels

 A typical way to evaluate the correlation is via entropy (later)

40
Concept Hierarchy Generation
 Concept hierarchy organizes concepts (i.e., attribute values) hierarchically and
is usually associated with each dimension in a data warehouse
 Concept hierarchies facilitate drilling and rolling in data warehouses to view
data in multiple granularity
 Concept hierarchy formation: Recursively reduce the data by collecting and
replacing low level concepts (such as numeric values for age) by higher level
concepts (such as youth, adult, or senior)
 Concept hierarchies can be explicitly specified by domain experts and/or data
warehouse designers
 Concept hierarchy can be automatically formed for both numeric and nominal
data. For numeric data, use discretization methods shown.

41
 Concept Hierarchy Generation for Nominal Data

 Specification of a partial/total ordering of attributes explicitly at

the schema level by users or experts
 street < city < province < country
 Specification of a hierarchy for a set of values by explicit data
grouping
 {St. John’s, Mount Pearl, Gander} < Newfoundland
 Specification of only a partial set of attributes
 E.g., only street < city, not others
 Automatic generation of hierarchies (or attribute levels) by the
analysis of the number of distinct values
 E.g., for a set of attributes: {street, city, province, country}

42
 Automatic Concept Hierarchy Generation
 Some hierarchies can be automatically generated based on
the analysis of the number of distinct values per attribute in
the data set
 The attribute with the most distinct values is placed at the
lowest level of the hierarchy
 Exceptions, e.g., weekday, month, quarter, year

country 15 distinct values

province_or_ state 365 distinct values

city 3567 distinct values

street 674,339 distinct values

43