What is BA?

• Skills, Tech & Practices for Continuous Iterative Exploration & Investigation to gain insights and drive biz planning

  Data → Insights → Biz Planning

• Types

  • Descriptive
    • Insights about past
    • Summarise raw data to make interpretable
  • Predictive
    • Understand the future
    • Provide actionable insights based on past data
  • Prescriptive
    • Advise possible outcomes
    • Attempt to quantify effect of future decisions on outcomes

Descriptive Biz Analytics

• Randomness, population and samples
• Data - Categorical or Numerical
• Sample - Random, Systematic, Stratified or Clustered
• Stats
  • Central Tendency: Mean, Mode, Median, Midrange, Quartiles
  • Dispersion: Range, Var, Std Dev, Interquart Range, Coeff of Var
  • Shape: Skew or Dispersion
  • Exploratory: Five number summary, box-and-whiskers
• Data Visualisation
  • Presentation of data visually to amplify cognition
  • Tool: Tableau
  • Designing Data Viz Products: Design → Paper&Pencil → Execution
  • Extra
    • Trend Lines, Forecasting, Clustering, What-If Analysis

Predictive Biz Analytics / Machine Learning

• ML is subset of AI
• What Is:
  • Transforming data into knowledge to produce actionable insights
  • Gives computers ability to learn without being explicitly programmed
• ML Lifecycle
  1. Define Goals
     1. Specify Biz Prob
     2. Define unit of analysis, prediction target
     3. Prioritise model criteria
  2. Data Prep
     1. Find appropriate data
     2. Merge data in a single table
     3. Explore data
     4. Clean data
     5. Feature engineering
  3. Create Model
  4. Interpret Model
  5. Implement Model
• Model Types (step 3-5)
  • Supervised
    • Numeric Prediction → Regression
      • Regression Types
        1. Linear Regression
           1. Univariate Linear Reg
           2. Multivariate Linear Reg
           3. Non-Linear Reg
        2. Decision Tree
           • Built top down from root (contains all instances) to leaf (prediction with smaller SD)
           • Entropy = homogeneity of data, Std Dev (0 = homo)
           • Pros
             • Both Reg and Classification
             • Easy to Interpret
             • Gens Biz Knowledge
           • Cons
             • Prone to overfitting
             • Too sensitive to instances, does not generalise well
        3. Random Forest
           • Set of Decision trees
           • Split training dataset to train diff models, combine at end
        4. Deep Learning
           • Based on large neural networks
           • Learn by example (training set shows examples, connections are made automatically)
        5. Ensembles
           • Superset of random forest
           • Set of any number and type of models
      • Evaluation
        • Mean Absolute Error
        • Mean Squared Error
        • R-Squared
    • Categorical Prediction → Classification
      • Models
        1. Logistic Regression
           • Instead of fit categorical var, we predict probabilities of each category
        2. Decision Tree
           • See above (Regression>DecisionTree)
        3. Random Forest
           • See above (Regression>RandomForest)
        4. Deep Learning
           • See above (Regression>DeepLearning)
        5. Ensembles
           • See above (Regression>Ensembles)
      • Evaluation
        • Confusion Matrix
          • True Positive (TP) - p classified p
          • False Positive (FP) - n classified p
          • True Negative (TN) - n classified n
          • False Negative (FN) - p classified n
          • Calculate measures based on confusion matrix - Eg. FPCost(FP)(Cost(FN)
        • Accuracy = (TP+TN)/Total
        • Precision = TP / (TP+FP)
        • Recall = TP/(TP+FN)
        • F-Measure = (2PrecisionRecall)/(Precision+Recall)
        • Phi Coefficient
  • Unsupervised
    • Training data has examples, not outcome
    • Goal - discovery and pattern finding
    • Not able to be evaluated
    • Algorithms
      1. Clustering
         • Finds "self-similar" groups of instances
         • Centroid: Geometric center of group
         • How: Specify or determine centroids
         • Beware scaling → Normalise
      2. Anomaly Detection
      3. Association Rules
      4. Topic Modeling