L2 Flashcards
Formulation of the problem is a :
Model : an abstract mathematical description, and not the real thing.
Abstraction
The process of removing irrelevant detail from a representation.
The most common form of ML:
- Input: (Features) , x
- Output: (Label), y
- Goal: Learn Mapping, f: x to y.
Supervised (inductive) learning
Given: Training data + desired outputs.
Unsupervised Learning
Given: Training data (without desired output).
Why bother transforming to vector?
Bias: Ordinality Assumption
Label Encoding Pros
Simplicity and Efficiency.
Memory Efficient
Useful for ordered Categories.
Label Encoding Cons
Implies Ordinality
Not suitable for non-ordinal features.
Model Dependency.
Linear Model Pros:
Many real-world processes are approximated with linear mode.
Can be solve analytically
The simplest yet the fundamental model of the other machine learning models.
Non-Linear problem can be transformed to linear one.
Linear Model Con:
May have low effectiveness and capacity.
Model too simple,
Noise in the features
Simple features doesn’t account for all variation,
Error in labels.
Data Representation Normalization - Common Techniques:
Min-Max Scaling: Transforms features by scaling each feature to a given range, usually [0,1] or [-1,1]
Max Abs Scaling: Scales each features by its max absolute value, resulting in values in the range [-1, 1].
Data Representation Normalization - Benefits:
Faster Convergence
Numerical Stability
Equal Importance.
Precision Function
True Positive / (True Positive + False Positive)
- Positive within all the positive predictions (model-wise)
Recall Function
True Positive / (True Positive + False negative)
- Positive within all positive instances (data-wise)
Accurancy Function
(True Positive + True Negative) / Total
- correct prediction over all instance.
