Logistic regression, despite its name, is a linear model for classification rather than regression. Logistic regression is a supervised classification algorithm where the dependent variable (label) is categorical, i.e. yes or no. It takes a linear combination of features with weights (parameters) and feeds it though a nonlinear squeezing function (sigmoid) that returns a probability between 0-1 of which class it belongs to. We learn the model, finding the proper weights, by using gradient descent. Those newfound weights will then be able to classify new data.
Once Upon A Time
You’re a dying prophet. You start a cult to have them continue your teachings when you pass away and you place a secret shrine at the top of this special hill.
You only want specific people to reach this shrine. People you deem as the Chosen Ones.
The special hill is near impossible to climb from the start through brute force. So, you implement a special machine with a special rock. The special machine has a scale you step on and acts like an escalator that will move you up the hill. The scale looks for a certain weight, and the closer you are to that weight, the further the machine will take you up the hill.
The special rock is similar to Thor’s hammer, in that only people ‘worthy’ enough can operate it. Over decades the special rock has learned who is ‘worthy’ and who is not through seeing millions of people try to get to the top. The special rock gets you closer to the weight the more ‘worthy’ you are. In order for someone to use the machine, a person must step onto the scale on the machine with the special rock.
If the special machine takes someone at least 50% (threshold) up the hill, then you predict that they can make it to the top (positive class), and they’re classified as worthy. The higher the machine takes them, the more confident you are that they can make it. If the machine doesn’t reach 50%, you assume that they’re not going to make it to the top, and are not worthy (negative class). There are some gritty people that could’ve made it to the top even if the machine goes less than 50%, but we assume they can’t if the machine doesn’t meet the threshold (misclassification).
Logistic regression is the go-to method for binary classification problems. It’s a discriminate model because it estimates the probability directly from the training data by minimizing error.
We can’t use linear regression because it is extremely sensitive to anomalies and will perform poorly, as some of the predictions will be outside [0,1] and change our threshold. For example, in our plot below, we can see that the threshold for whether we predict if someone is alert or not is near the middle.
But what happens when we have another piece in our training set that received more sleep? It ends up severely lowering our threshold and ruining our model.
So, logistic regression maps the linear combination of weights and features to a s-shaped function that returns a probability between 0 and 1.
Given where and is discrete and binary, like or .
Your task is to learn a classification function: . While we want a binary output, we still know we have limitations. For example, we can’t predict Credit Card Default with any certainty. But suppose we want to predict how likely is a customer to default, where we output a probability between 0 and 1 that a customer will default. That makes sense and would be suitable and practical. We can then set the threshold to make it binary.
In this case, the output is real (regression) but is bounded (classification) between 0 and 1.
How do we guarantee that where ? We use the sigmoid function:
Recall that our function can be written as if we append 1 to all .
For example, appending 1 to x we get:
Taking the dot product of and we get:
In other words, cast the output to bring the linear function quantity between 0 and 1. (Note, one can use other S-shaped functions). If it is above 0.5, we predict 1. Else, we predict 0.
How do we find ? Just like linear regression, we want to minimize the risk/cost function:
Our risk function makes sense because we are trying to minimize the difference between our prediction and the actual answer (labels).
Remember, is now the logistic function so the is not the quadradic function we had when was linear. Instead, the cost is a complicated non-linear function. Consequently, there are many local optima, hence gradient descent may not find the global optimum.
So, we need a different function that is convex. The new convex function is:
Putting these functions into one compact function (because or ):
Notice that we can derive the same cost function with a probabilistic interpretation:
which can be compactly written as:
Recall that are independently generated. So, the probability of getting in distribution from the corresponding is:
This gives us the likelihood of the parameters for training examples:
It is easier to work with logs so we take the log of the function and minimize the negative log likelihood:
which gives us the same cost function as before.
Our goal is to find the right parameter/weights ’s to minimize the cost function . To do so, we use gradient descent.
Before talking more, make sure you understand partial derivatives/gradients.
Gradient Descent is an iterative, optimization method that repeats until converge (it has found the minimum). For instance, let’s say we have a convex cost function and our goal is to find the that minimizes it. If we plot it, it looks like so:
Recall that in gradient descent, we start off by initializing our weights randomly, which puts us at the black dot at the top right on the diagram above. Taking the derivative, we see the slope at this point is a pretty big positive number. We want to move closer to the center, because we want to find the minimum cost, so naturally, we should take a pretty big step in the opposite direction of the gradient of the function at that point. If we repeat the process enough, we soon find ourselves nearly at the bottom of our curve and much closer to the optimal weight. Convergence means our gradient/slope is at or near zero because this means we found the minimum.
Notice that this only easily happens if our cost function is convex (bowl-shaped), like the cost function in linear regression, because then our descent will always find the global minima. This is why we had to change our cost function earlier to a convex one, because the original one was a complicated non-linear function like so:
This becomes a problem because we always start by initializing the ’s in a random spot, making it possible that gradient descent won’t find the global minima and end up in a local minima which represents suboptimal solutions.
So running gradient descent, we repeat until convergence:
- Simultaneously update for all ’s:
After some calculus, we get:
Remember that choosing the right step size is very important. If the step size is too small, it will take an extremely long time to converge. If it’s too large, you’ll overshoot the global minimum.
With the right , gradient descent will converge, finding the optimal weights.
Given a training set with features, # of iterations , and learning rate the algorithm for logistic regression works like this:
- Initialize all weights to for all .
- For iterations:
- Run gradient descent to find the optimal ’s with the formula:
- Run gradient descent to find the optimal ’s with the formula:
You’re Vegas and you want to predict whether an NBA team will win a ring or not. You’re given data on 3 teams, their features, and their label (whether they have won a championship). Their features are number of Hall of Famers, and number of All-Stars, i.e. means the 4th team has 7 Hall of Famers and 6 All-Stars.
Our learning rate is set to , iterations set to 100, and our training data is given as follows:
Remember that we append a 1 to all our examples so we can also solve for . Our updated data is:
First we initialize all ’s to zero:
Using gradient descent, we look for our betas with the algorithm:
Now looking for and :
When we calculate we see that it also comes to .25. So your new weights are . This is after one iteration.
After 100 iterations, we get:
These weights correctly find a separating line between our positive and negative examples.
For our new example , with 3.5 Hall of Famers and 4 All-Stars, we want to predict whether this is a championship team. We simply place our features through the sigmoid with our newfound ’s:
, therefore we predict that this team will indeed win a championship!
Logistic regression is a model where the dependent variable is categorical, i.e. 1 or 0, dead or alive, etc. The inputs along with weights are taken and squeezed through a non-linear function such as a sigmoid. This outputs a value between 0 and 1 which corresponds to a probability of how likely the class belongs to it. In order to find the weights you use gradient descent, similar to linear regression.
- Logistic Regression is a discriminative classifier. It assumes some functional form and estimates parameters of directly from training data. It cannot obtain a sample of the data because is not available.
- If your training set is extremely large, you could use a variation of gradient descent called stochastic gradient descent.
- Perceptron uses step function, while logistic uses a nonlinear function. They both also update their weights differently.
- Doesn’t perform well when feature space is too large.
- Relies on transformations for non-linear features.
- Convenient probability scores for observations.
- Wide spread industry comfort for logistic regression solutions.
- Make sure features are on a similar scale, which can be done using variance or standard deviation of the variable.