Please enable JavaScript.

Coggle requires JavaScript to display documents.

Machine learning - Coggle Diagram

- - - - equal E(x) = mu
        
        unbiased indicator
- - - - model has overfitted
  - - - if my Training data contains more zeros than ones ( 9500,500) then biased model
      - Under sampling
        
        from(9500,500) draw (500,500)
        
        But , we are losing a lot of information
      - Over-sampling
        
        from(9500,500 with repetition) draw (9500,9500)
        
        Some informationget higher weightage
      - SMOTE
  - - - and be useful in the real world.
    - - some orgs test the variance also
  - - - Impute
        
        K-NN
        
        use all variables
      - mean/mode
      - If correlation high, regress
        
        better result than knn
  - - - Z-transformation
        
        X-mu/sigma
        
        for normal??
      - Min max
    - - but sometimes that wont work - as it will eliminate more than 10%
        
        go for observation
        
        business Knowledge
- - - - deviance
        
        null
        
        residual
        
        Deviance difference
        
        ND-RD
        
        apprxo. follows CHi sq
        
        ND>>RD
        
        H1 is true
        
        1 more item...
      - sigmoid transformation
      - Log likelihood
      - maximum likelihood
      - Can we drop a variable? compare models
        
        AIC
        
        Based on number of variables
        
        When we compare two models the lower the better
        
        BIC
        
        Based on number of variables and no of observation
    - - Generate rules after K-means and then use KNN
  - - - if accuracy very high, leakages in your data, check
      - ROC Curve
        
        higher the area under the curve better, max 1, absolutely good/bad decision possible
  - - - otherwise overfitting
      - lower entropy means higher information gain
      - Gini
    - - Maximum Voting Technique
        
        average decisions
        
        average probabilities
    - - bagging + boostrapping
        
        Random Forest
        
        OOB Error
        
        bagging should be applied only to unstable models, otherwise inefficient
- - - - K-means
        
        Once clustering is complete
        
        elbow method --> number of clusters
    - - distance - linkage
        
        single/complete/average/variance
      - Cluster dendrogram
    - - content based
      - filtering based
        
        collaborative filtering
        
        User based
        
        Item based
        
        Amazon
        
        Similarity measure
        
        Jac(X,Y) similarity
        
        X intersection Y/ X union Y
        
        K-NN
        
        Cosine
        
        dort prod(a,b)/|a||b|
      - Model based
        
        Association Rule Mining
        
        Support
        
        P(A intersection B)
        
        min defined
        
        X--> or Y--> X remain same
        
        confidence
        
        P(A cap B)/P(A)
        
        min defined
        
        generally support < confidence
        
        lift
        
        rule useful when lift>1
        
        P(A cap B)/P(A).P(B)
        
        X--> or Y--> X remain same
        
        rare but possible
        
        Neural Networks