Business Intelligence Midterm

CRISP-DM Phases

1. Business Understanding

2. Data Understanding

3. Data Preparation

4. Modeling

5. Evaluation

6. Deployment

Business Understanding Phase

• define business requirements/objectives

• translate objectives into DM problem definition

• prepare initial strategy to meet objectives

Data Understanding Phase

• collect data

• assess data quality

• perform EDA
  

Data Preparation Phase

• clean, prepare, and transform dataset

• prepare for modeling in following phases

• choose cases/variables appropriate for analysis

Modeling Phase

• select and apply 1+ modeling techniques

• calibrate modeling settings to optimize results

• additional data prep if required

Evaluation Phase

• evaluate 1+ models for effectiveness

• determine if objectives achieved

• make decision regarding DM results before deploying

Deployment Phase

• make use of models created

• simple deployment: generate report

• complex deployment: implement additional DM effort in another department

• in business, customer often carries out deployment based on model

Description (DM Task)

• describes general patterns and trends

• easy to interpret/explain

• transparent models

• pictures and #’s

• like scatterplots, descriptive stats, range

Estimation (DM Task)

• numerical predictor/categorical (IV) values to estimate changes in numerical target variables (DV)

• like regression models - predicting something based off another variable

Classification (DM Task)

• like estimation, but target variables (DV) are categorical

• like classification of simple vs. complex tasks, fraudulent card transactions, income brackets

• can accommodate a categorical variable as your target when partitioned into 2+ categories

Prediction (DM Task)

• similar to estimation and classification, but with a time component

• like what is the probability of Hogs winning a game with a certain combination of player profiles, or future stock behavior

Clustering (DM Task)

• similar to classification, but no target variables

• clustering tasks don’t aim to estimate, predict, or classify a target variable

• only segmentation of data

• like focused marketing campaigns

Association (DM Task)

• no target variable

• finding attributes of data that go together

• profiling relationships between 2+ attributes

• understand consequent behaviors based on previous behaviors

• like supermarkets seeing what items are purchased together (affinity analysis)

Learning Types (DM Task)

Supervised

• have a target variable

• know categories (like estimation)

Unsupervised

• exploratory, no target variable

• searching for patterns across variables

• don’t know target variables or categories

• like clustering

Supervised DM Tasks

estimation, classification, prediction

Unsupervised DM Tasks

Clustering, association

ways to handle missing data

1. user-defined constant (0.0 or “Missing”)

2. mode / mean / median

3. random values

4. imputation - most likely value based on other attributes for the record

Min-Max Normalization

• determines how much greater the field value is than the minimum value for the field, scales the difference by the field’s range

• values range from 0 to 1

• good for KNN

• X* = (X - min(X))/(max(X)-min(X))

Z-score Standardization

• takes difference between field value and field value mean, scales difference by field’s stdev

• typically range from -4 to 4

• z = (X - mean(X))/stdev(X)

• can be used to identify outliers if they’re beyond standard range

Skewness

• right skewed → mean > median, tail to the right, positive skewness

• left skewed → mean < median, tail to the left, negative skewness

• cutoff: -2 to 2

• can use natural log transformation to make data more normal

Kurtosis

• skewness in height terms

• cutoff: -10 to 10

Normality vs. Normalization

Normality - to transform variable so its distribution is closer to normal without changing its basic information

Normalization - standardizes the mean/variance or range of every variable and the effect that each variable has on results

Confidence interval

• used when the population has a normal distribution or when n is large

• indicates how precise an estimate is (narrow = precise)

• point estimate ± margin of error

Margin of error

• range of values above and below a sample statistic

• MoE = z-score * (population stdev / sqrt(n))

• as sample size increases, margin of error will decrease, CI will shrink

Multiple regression

• uses multiple independent variables to explain the dependent variable

• y = B₀ + B₁X₁ + B₂X₂ + … + B_nX_n

• Ex. Rating = 59.9 - 2.46*Sugars + 1.33*Carbs

  • Every additional unit of sugar will decrease rating by 2.46, and every additional unit of carbs with increase rating by 1.33

• assumptions:

  • multicollinearity - make a note if metrics above 0.8

  • linearity - if histogram of residuals is normally distributed and P-P plot mostly linear

overfitting

model is trained to be overconfident, become unwilling to learn new patterns if they come up and denying the validation’s results

Backward regression

• starts with the most complex model (all variables) and step by step removes them until it finds the model with the most explanatory power

• consumes the most resources

Forward regression

• starts with the least complex model (one IV) and iterates through models that increase in complexity until it finds the one with the most explanatory power

• consumes the least resources

Parametric vs. non-parametric models

Parametric

• have assumptions on the distributions and characteristics of the data (like skewness/kurtosis), more structured and efficient

Non-parametric

• makes no assumptions, can adapt structure based on the data

k-Nearest Neighbors

• aka instance-based learning or memory-based reasoning

• training set records stored, then classification is performed for new unclassified records based on records they’re most similar to

• can weight by distance with euclidean distance formula (numeric variables)

  • min-max normalization can be used to scale variables so weight is standard

• for categorical variables, assign 1 if value is the same for unclassified records and 0 if different

Decision Tree

• collection of nodes connected by branches, extending downward from a root node to terminate lead nodes

• supervised learning

• target variables must be categorical

CART (classification and regression trees)

• computes “goodness” of candidate splits/optimality measures

• can only do two splits at a time, so the tree will naturally be large

C4.5 Tree

• similar to CART, but uses information gain (highest value) / entropy reduction (lowest value) to select optimal splits

• not limited to two splits, so it is more efficient and smaller