Linear Regression
Attempting to fit a function (a straight line) to a dataset to observe a linear relationship. Then used to predict further values given the line that we’ve fit.
Validation
Machine Learning engineers will use train-test splits, which means that they will split up the data into a train and a test set. Statisticians will use metrics, prediction intervals and correlations.kjA
Unsupervised learning
Regression is supervised (requires examples to calculate the line), but unsupervised learning is all about anomaly detection and clustering.
Linear Correlation
As one variable increases or decreases, another variable does as well. This is the backbone of the technique is trying to find a relationship like this.
Benefits and Drawbacks of Linear Regression
Benefits
- Can make predictions on unseen data. The line gives us that ability.
Drawbacks
- Real-world data is not always linear.
- There will be error about the line.
- We cannot predict reliably outside of our range of values.
Considering sampling bias
The data may be skewed, given some bias of the sample. The data are usually a sample from a population and never indicative of the entire population itself, so ensure that you take that into account.
Performing a basic linear regression in Julia
First, we require some basic packages to support our work, change the the package mode with ] in the REPL and add some packages:
- DataFrames
- CSV
module chapter5
using DataFrames
using CSV
using Downloads
using Plots
using Flux
function example51()
# First download the file.
dataset = Downloads.download("https://bit.ly/3goOAnt")
# Then read the file into a DataFrame
df = DataFrame(CSV.File(dataset))
# Extract our values with dataframe indexing.
X = df[:, 1:end-1]
Y = df[:, end]
# Fit a linear regression - https://fluxml.ai/Flux.jl/stable/models/overview/
model = Dense(1 => 1) # One input variable, one output variable
# TO BE CONTINUED
end
end