Up until this point in the semester, we’ve focused on analyzing datasets. We’ve looked into the early stages of the data science lifecycle, focusing on the programming tools, visualization techniques, and data cleaning methods needed for data analysis.
This lecture marks a shift in focus. We will move away from examining datasets to actually using our data to better understand the world. Specifically, the next sequence of lectures will explore predictive modeling: generating models to make some prediction about the world around us. In this lecture, we’ll introduce the conceptual framework for setting up a modeling task. In the next few lectures, we’ll put this framework into practice by implementing various kinds of models.
A model is an idealized representation of a system. A system is a set of principles or procedures according to which something functions. We live in a world full of systems: the procedure of turning on a light happens according to a specific set of rules dictating the flow of electricity. The truth behind how any event occurs are usually complex, and many times the specifics are unknown. The workings of the world can be viewed is its own giant procedure. Models seek to simplify the world and distill them it into workable pieces.
Example: We model the fall of an object on Earth as subject to a constant acceleration of \(9.81 m/s^2\) due to gravity.
Why do we want to build models? As far as data scientists and statisticians are concerned, there are three reasons, and each implies a different focus on modeling.
To explain complex phenomena occurring in the world we live in. Examples of this might be:
In these cases, we care about creating models that are simple and interpretable, allowing us to understand what the relationships between our variables are.
To make accurate predictions about unseen data. Some examples include:
When making predictions, we care more about making extremely accurate predictions, at the cost of having an uninterpretable model. These are sometimes called black-box models, and are common in fields like deep learning.
To measure the causal effects of one event on some other event. For example,
This is a much harder question because most statistical tools are designed to infer association not causation. We will not focus on this task in Data 100, but you might take other advanced classes on Causal Inference if you are intrigued!
Most of the times, we aim to strike a balance between building interpretable models and building accurate models.
In general, models can be split into two categories:
Note: These specific models are not in the scope of Data 100 and exist to serve as motivation.
Deterministic physical (mechanistic) models: Laws that govern how the world works.
Probabilistic models: models that attempt to understand how random processes evolve. These are more general and can be used describe many phenomena in the real world. These models commonly make simplifying assumption about the nature of the world.
The regression line is the unique straight line that minimizes the mean squared error of estimation among all straight lines. As with any straight line, it can be defined by a slope and a y-intercept:
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
# Set random seed for consistency
np.random.seed(43)
plt.style.use('default')
#Generate random noise for plotting
x = np.linspace(-3, 3, 100)
y = x * 0.5 - 1 + np.random.randn(100) * 0.3
#plot regression line
sns.regplot(x=x,y=y);
For a pair of variables \(x\) and \(y\) representing our data \(\mathcal{D} = \{(x_1, y_1), (x_2, y_2), \dots, (x_n, y_n)\}\), we denote their means/averages as \(\bar x\) and \(\bar y\), standard deviations as \(\sigma_x\) and \(\sigma_y\).
A variable is represented in standard units if the following are true:
To convert a variable \(x_i\) into standard units, we subtract from it its mean and divide by its standard deviation. For example, \(x_i\) in standard units is \(\frac{x_i - \bar x}{\sigma_x}\).
The correlation (\(r\)) is the average of the product of \(x\) and \(y\), both measured in standard units.
\[r = \frac{1}{n} \sum_{i=1}^n (\frac{x_i - \bar{x}}{\sigma_x})(\frac{y_i - \bar{y}}{\sigma_y})\]
def plot_and_get_corr(ax, x, y, title):
ax.set_xlim(-3, 3)
ax.set_ylim(-3, 3)
ax.set_xticks([])
ax.set_yticks([])
ax.scatter(x, y, alpha = 0.73)
r = np.corrcoef(x, y)[0, 1]
ax.set_title(title + " (corr: {})".format(r.round(2)))
return r
fig, axs = plt.subplots(2, 2, figsize = (10, 10))
# Just noise
x1, y1 = np.random.randn(2, 100)
corr1 = plot_and_get_corr(axs[0, 0], x1, y1, title = "noise")
# Strong linear
x2 = np.linspace(-3, 3, 100)
y2 = x2 * 0.5 - 1 + np.random.randn(100) * 0.3
corr2 = plot_and_get_corr(axs[0, 1], x2, y2, title = "strong linear")
# Unequal spread
x3 = np.linspace(-3, 3, 100)
y3 = - x3/3 + np.random.randn(100)*(x3)/2.5
corr3 = plot_and_get_corr(axs[1, 0], x3, y3, title = "strong linear")
extent = axs[1, 0].get_window_extent().transformed(fig.dpi_scale_trans.inverted())
# Strong non-linear
x4 = np.linspace(-3, 3, 100)
y4 = 2*np.sin(x3 - 1.5) + np.random.randn(100) * 0.3
corr4 = plot_and_get_corr(axs[1, 1], x4, y4, title = "strong non-linear")
plt.show()
When the variables \(y\) and \(x\) are measured in standard units, the regression line for predicting \(y\) based on \(x\) has slope \(r\) and passes through the origin.
\[\hat{y}_{su} = r \cdot x_{su}\]
\[\frac{\hat{y} - \bar{y}}{\sigma_y} = r \cdot \frac{x - \bar{x}}{\sigma_x}\]
Starting from the top, we have our claimed form of the regression line and we want to show that its equivalent to the optimal linear regression line: \(\hat{y} = \hat{a} + \hat{b}x\)
Recall:
At a high level, a model is some way of representing a system. In Data 100, we’ll treat a model as some mathematical rule we use to describe the relationship between variables.
What variables are we modeling? Typically, we use a subset of the variables in our sample of collected data to model another variable in this data. To put this more formally, say we have the following dataset \(\mathcal{D}\):
\[\mathcal{D} = \{(x_1, y_1), (x_2, y_2), ..., (x_n, y_n)\}\]
Each pair of values \((x_i, y_i)\) represents a datapoint. In a modeling setting, we call these observations. \(y_i\) is the dependent variable we are trying to model, also called an output or response. \(x_i\) is the independent variable inputted into the model to make predictions, also known as a feature.
Our goal in modeling is to use the observed data \(\mathcal{D}\) to predict the output variable \(y_i\). We denote each prediction as \(\hat{y}_i\) (read: “y hat sub i”).
How do we generate these predictions? Some examples of models we’ll encounter in the next few lectures are given below:
\[\hat{y}_i = \theta\] \[\hat{y}_i = \theta_0 + \theta_1 x_i\]
The examples above are known as parametric models. They relate the collected data, \(x_i\), to the prediction we make, \(\hat{y}_i\). A few parameters (\(\theta\), \(\theta_0\), \(\theta_1\)) are used to describe the relationship between \(x_i\) and \(\hat{y}_i\).
Notice that we don’t immediately know the values of these parameters. While the features, \(x_i\), are taken from our observed data, we need to decide what values to give \(\theta\), \(\theta_0\), and \(\theta_1\) ourselves. This is the heart of parametric modeling: what parameter values should we choose so our model makes the best possible predictions?
To choose our model parameters, we’ll work through the modeling process.
Our first step is choosing a model: defining the mathematical rule that describes the relationship between the features, \(x_i\), and predictions \(\hat{y}_i\).
In Data 8, you learned about the Simple Linear Regression (SLR) model. You learned that the model takes the form: \[\hat{y}_i = a + bx_i\]
In Data 100, we’ll use slightly different notation: we will replace \(a\) with \(\theta_0\) and \(b\) with \(\theta_1\). This will allow us to use the same notation when we explore more complex models later on in the course.
\[\hat{y}_i = \theta_0 + \theta_1 x_i\]
The parameters of the SLR model are \(\theta_0\), also called the intercept term, and \(\theta_1\), also called the slope term. To create an effective model, we want to choose values for \(\theta_0\) and \(\theta_1\) that most accurately predict the output variable. The “best” fitting model parameters are given the special names \(\hat{\theta}_0\) and \(\hat{\theta}_1\) – they are the specific parameter values that allow our model to generate the best possible predictions.
In Data 8, you learned that the best SLR model parameters are: \[\hat{\theta}_0 = \bar{y} - \hat{\theta}_1\bar{x} \qquad \qquad \hat{\theta}_1 = r \frac{\sigma_y}{\sigma_x}\]
A quick reminder on notation:
In Data 100, we want to understand how to derive these best model coefficients. To do so, we’ll introduce the concept of a loss function.
We’ve talked about the idea of creating the “best” possible predictions. This begs the question: how do we decide how “good” or “bad” our model’s predictions are?
A loss function characterizes the cost, error, or fit resulting from a particular choice of model or model parameters. This function, \(L(y, \hat{y})\), quantifies how “far off” a single prediction by our model is from a true, observed value in our collected data.
The choice of loss function for a particular model depends on the modeling task at hand. Regardless of the specific function used, a loss function should follow two basic principles:
Two common choices of loss function are squared loss and absolute loss.
Squared loss, also known as L2 loss, computes loss as the square of the difference between the observed \(y_i\) and predicted \(\hat{y}_i\): \[L(y_i, \hat{y}_i) = (y_i - \hat{y}_i)^2\]
Absolute loss, also known as L1 loss, computes loss as the absolute difference between the observed \(y_i\) and predicted \(\hat{y}_i\): \[L(y_i, \hat{y}_i) = |y_i - \hat{y}_i|\]
L1 and L2 loss give us a tool for quantifying our model’s performance on a single datapoint. This is a good start, but ideally we want to understand how our model performs across our entire dataset. A natural way to do this is to compute the average loss across all datapoints in the dataset. This is known as the cost function, \(\hat{R}(\theta)\): \[\hat{R}(\theta) = \frac{1}{n} \sum^n_{i=1} L(y_i, \hat{y}_i)\]
The cost function has many names in statistics literature. You may also encounter the terms:
We can substitute our L1 and L2 loss into the cost function definition. The Mean Squared Error (MSE) is the average squared loss across a dataset: \[\text{MSE} = \frac{1}{n} \sum_{i=1}^n (y_i - \hat{y}_i)^2\]
The Mean Absolute Error (MAE) is the average absolute loss across a dataset: \[\text{MAE}= \frac{1}{n} \sum_{i=1}^n |y_i - \hat{y}_i|\]
Now that we’ve established the concept of a loss function, we can return to our original goal of choosing model parameters. Specifically, we want to choose the best set of model parameters that will minimize the model’s cost on our dataset. This process is called fitting the model.
We know from calculus that a function is minimized when (1) its first derivative is equal to zero and (2) its second derivative is positive. We often call the function being minimized the objective function (our objective is to find its minimum).
To find the optimal model parameter, we:
We repeat this process for each parameter present in the model. For now, we’ll disregard the second derivative condition.
To help us make sense of this process, let’s put it into action by deriving the optimal model parameters for simple linear regression using the mean squared error as our cost function. Remember: although the notation may look tricky, all we are doing is following the three steps above!
Step 1: take the derivative of the cost function with respect to each model parameter. We substitute the SLR model, \(\hat{y}_i = \theta_0+\theta_1 x_i\), into the definition of MSE above and differentiate with respect to \(\theta_0\) and \(\theta_1\). \[\text{MSE} = \frac{1}{n} \sum_{i=1}^{n} (y_i - \hat{y}_i)^2 = \frac{1}{n} \sum_{i=1}^{n} (y_i - \theta_0 - \theta_1 x_i)^2\]
\[\frac{\partial}{\partial \theta_0} \text{MSE} = \frac{-2}{n} \sum_{i=1}^{n} y_i - \theta_0 - \theta_1 x_i\]
\[\frac{\partial}{\partial \theta_1} \text{MSE} = \frac{-2}{n} \sum_{i=1}^{n} (y_i - \theta_0 - \theta_1 x_i)x_i\]
Step 2: set the derivatives equal to 0. After simplifying terms, this produces two estimating equations. The best set of model parameters \((\hat{\theta}_0, \hat{\theta}_1)\) must satisfy these two optimality conditions. \[0 = \frac{-2}{n} \sum_{i=1}^{n} y_i - \hat{\theta}_0 - \hat{\theta}_1 x_i \Longleftrightarrow \frac{1}{n}\sum_{i=1}^{n} y_i - \hat{y}_i = 0\] \[0 = \frac{-2}{n} \sum_{i=1}^{n} (y_i - \hat{\theta}_0 - \hat{\theta}_1 x_i)x_i \Longleftrightarrow \frac{1}{n}\sum_{i=1}^{n} (y_i - \hat{y}_i)x_i = 0\]
Step 3: solve the estimating equations to compute estimates for \(\hat{\theta}_0\) and \(\hat{\theta}_1\).
Taking the first equation gives the estimate of \(\hat{\theta}_0\): \[ \frac{1}{n} \sum_{i=1}^n y_i - \hat{\theta}_0 - \hat{\theta}_1 x_i = 0 \\ \left(\frac{1}{n} \sum_{i=1}^n y_i \right) - \hat{\theta}_0 - \hat{\theta}_1\left(\frac{1}{n} \sum_{i=1}^n x_i \right) = 0 \\ \hat{\theta}_0 = \bar{y} - \hat{\theta}_1 \bar{x} \]
With a bit more maneuvering, the second equation gives the estimate of \(\hat{\theta}_1\). Start by multiplying the first estimating equation by \(\bar{x}\), then subtracting the result from the second estimating equation. \[ \frac{1}{n} \sum_{i=1}^n (y_i - \hat{y}_i)x_i - \frac{1}{n} \sum_{i=1}^n (y_i - \hat{y}_i)\bar{x} = 0 \\ \frac{1}{n} \sum_{i=1}^n (y_i - \hat{y}_i)(x_i - \bar{x}) = 0 \]
Next, plug in \(\hat{y}_i = \hat{\theta}_0 + \hat{\theta}_1 x_i = \bar{y} + \hat{\theta}_1(x_i - \bar{x})\): \[ \frac{1}{n} \sum_{i=1}^n (y_i - \bar{y} - \hat{\theta}_1(x - \bar{x}))(x_i - \bar{x}) = 0 \\ \frac{1}{n} \sum_{i=1}^n (y_i - \bar{y})(x_i - \bar{x}) = \hat{\theta}_1 \times \frac{1}{n} \sum_{i=1}^n (x_i - \bar{x})^2 \]
By using the definition of correlation \(\left(r = \frac{1}{n} \sum_{i=1}^n (\frac{x_i-\bar{x}}{\sigma_x})(\frac{y_i-\bar{y}}{\sigma_y}) \right)\) and standard deviation \(\left(\sigma_x = \sqrt{\frac{1}{n} \sum_{i=1}^n (x_i - \bar{x})^2} \right)\), we can conclude: \[r \sigma_x \sigma_y = \hat{\theta}_1 \times \sigma_x^2\] \[\hat{\theta}_1 = r \frac{\sigma_y}{\sigma_x}\]
Just as was given in Data 8!
Remember, this derivation found the optimal model parameters for SLR when using the MSE cost function. If we had used a different model or different loss function, we likely would have found different values for the best model parameters. However, regardless of the model and loss used, we can always follow these three steps to fit the model.
At this point, we’ve:
Now, what are some ways to determine if our model was a good fit to our data? We will delve into this more in the next chapter, but there are three main ways for evaluating a model.
Statistics:
Performance metrics:
Visualization: