Study_Datascience

1. Linear Regression

• SST(total) = SSR(explained) + SSE(unexplained)

• OLS (ordinary least squares) model - minimum SSE [lowest error]

• R-squared = SSR/SST (measure how much of the total variability of the dataset, ranging 0 to 1 variability)

2. Multiple Linear Regression

• adjusted R-squared (<R-squared) measures how well your model fits the data, but it penalizes the use of variables that are meaningless for the regression

• increases R-squared but decreases adjusted R-squared ⇒ omit the variable

3. Logistic Regression

-> The logistic regression predicts the probability of an event occurring

• Maximum likelihood estimation (MLE) : estimates how likely it is that the model at hand describes the real underlying relationship of the variables

• LL-null ( log likelihood-null) : the log-likelihood of a model which has no independent variables

• LLR(log likelihood ratio) : measures if our model is statistically different from LL-null

4. k-means clustering

• multivariate statistical technique that groups observations on the basis some of their features or variables they are described by

→ maximize the similarity of observations within a cluster and maximize the dissimilarity between clusters

1. choose the num of clusters [The elbow method]
  * minimize the distance between points in a cluster (low WCSS within-cluster sum of squares)
2. specify the cluster seeds
3. assign each point to a centroid
4. adjust the centroids

5. Hierarchical clustering

pros and cons of dendrogram

• Pros

1. Hierarchical clustering shows all the possible linkages between clusters

2. No need to preset the number of clusters like K-means

3. Many methods to perform hierarchical clustering

• Cons

1. Scaleability

2. Computational expensive

6. Basic Neural Network

'Data - Model - Objective function - Optimization algorithm'

• The objective function

: a measure of how well our model’s outputs match the targets.

two types

1. loss (supervised learning)

   • The lower the loss function, the higher the level of accuracy of the model

2. reward (reinforcement learning)

   • The higher the reward function, the higher the level of accuracy of the model