datascience-terms.md

Data Science Terminologies

Feature Extraction

Term Frequency (TF)

Term Frequency-Inverse Document Frequency (TF-IDF)

Mutual Information Weight of a Term in a Document $$w_{ij} = \log_2 \left(\frac{TF_{ij}/S}{\sum_{i'=1}^N TF_{i'j}/S \sum_{j'=1}^NTF_{ij'}/S} + 1\right)$$

Similarity and Dissimilarity Measures

Euclidean distance

Cosine similarity

KL-divergence

0-1 Loss

l_{0-1}=\left\{ \begin{array}{ll} 0 & if h(X)=y\\ 1 & if h(x)\ne y\end{array} \right.

Square Loss

Low-pass filter dealing with noise in time-series data (signals)

Geometric Mean

Trimean (TM)

Interquartile Mean (IQM)

Fisher transformation

Platt Scaling

Jacknifing

Fisher Transformation

Point Biserial

**ANOVA (Analysis of Variance)

Weighted Least-Square Regression

RANSAC (Random Sample Consensus)

Conditional Random Field

Probability Theory

Joint probability $p(x,y)$

Conditional probability $p(y|x)$

Muttual Informaiton (MI) $MI(x,y) = \frac{p(x,y)}{p(x)p(y)}$

Conditional Mutual Information $MI(A;B|C) = MI(A;B,C) - MI(A;C)$

Sampling

Jackknife resampling is a technique that predates bootstrap sampling technique for estimateing the precision of sample statistics (mean, median, ..). Given a dataset of size $N$, Jackknife resampling uses the leave-one-out technique to calculate the desired parameter using the remaining $N-1$.This process is repated for each saple point, and finally the aggregate of all the estimates is the Jackknife estimator of that parameter. (The main difference with bootstrap is that in bootstrap technique, subsets are drawn randomly with replacement.)

Evaluation Metrics

Cohens kappa coefficient

Clustering

Purity

Mutual Information

Rand Index