Technically, linear algebra is the space of mathematics that deals with vector spaces. Colloquially, linear algebra involves the manipulations of variables with many, many dimensions. It's a tricky business but absolutely essential to any data scientist. By the end of this lecture, you should be able to
It bears mentioning: if your chosen field is data science in some form, you will use linear algebra on a daily basis. Period. There is no escaping it.
There's enough to linear algebra for it to consume entire courses, so I can't possibly do it justice in one lecture. So I strongly recommend taking a linear algebra course as early as possible.
If you need any clarifications at all, or want to be pointed in a direction to learn more, please ask me! I am more than happy to help.
There's a lot of jargon associated with linear algebra. We'll touch on some of the basic terms. Head over to the Wikipedia Linear Algebra page for similar information.
For starters: use two dimensions ($x$ and $y$ values) and three dimensions ($x$, $y$, and $z$ values) as references to anchor your understanding. When we refer to "lines" and "planes", think about what they look like in these contexts.
To a certain extent, linear algebra is a generalization of the 2D (and 3D?) Cartesian plotting you've already done before now. Now we're considering lines that can have arbitrary numbers of variables, instead of just two ($x$ and $y$).
A scalar is a number. You've seen many of them and will see many more. Just a number. Hooray!
Scalars can come with or without units. So long as it's just one number...
In [1]:
x = 5 # scalar
y = 4924.2352353 # scalar
z = (1, 2, 3) # NOT a scalar!
...it's a scalar.
A vector is a collection of scalars that describe some object. Think of a vector as a generalized scalar: whereas one's velocity can be measured with a single number (55mph), one's physical position on a baseball field needs at least two numbers ($x$ and $y$).
This is a picture of a two-dimensional vector and its opposite:
Vectors in linear algebra are pretty much your bread-and-butter. These are your data points, your observations, your model descriptors, and the cornerstone of any algorithm you may implement.
It can be helpful to think of vectors and arrays as interchangeable, since both have some finite number of elements. For vectors, these elements are always scalars.
The number of elements in a vector is the number of dimensions. Vectors with three elements
are three-dimensional.
Vectors have two very important properties that ultimately distinguish them from scalars:
These vectors are three-dimensional, have equal magnitude (they're the same length), but different directions.
Computing the direction of the vector is easy: it's just the vector itself. In 2D or 3D, you can map a vector by starting at the origin and drawing a line toward the point indicated by the elements of the vector. The 2D vector (5, 5) points up and to the right at a 45 degree angle. The vector (5, 0) points to the right along the $x$-axis.
Computing the magnitude is a little trickier, but not much. After all, if you drew the point (5, 5) and the line connecting it to the origin (0, 0), how would you compute that distance between those two points?
In [2]:
import numpy as np
vector = np.array([5, 5])
magnitude = np.sqrt( (vector[0] ** 2) + (vector[1] ** 2) )
print("Magnitude: {:.2f}".format(magnitude))
More generally: for each element in a vector, compute its square. Sum up all the squares, and take the square root of that sum.
2D: magnitude = $\sqrt{x^2 + y^2}$
3D: magnitude = $\sqrt{x^2 + y^2 + z^2}$
Pythagorus was on a roll!
A space (usually referred to as a vector space) can, abstractly, be thought of as a sort of "enclosure" of possibilities, which occasionally takes physical form as vectors.
Spaces are (usually) infinite in size, but finite in dimension. Take a two-dimensional space: only two dimensions ($x$ and $y$), but you can imagine these two dimensions ranging from $-\infty$ to $+\infty$, thereby making the two-dimensional space infinite.
Spaces are the culmination of many vectors; vectors live in a space, and vectors can define a space. One common activity for data scientists and researchers is to compute vectors that define, or at least approximate, some space of interest.
Some other possibly-unhelpful analogies might include:
In essence, a "space" is a potentially-infinite realm of possibility in which we typically observe a finite (and often very small, relatively speaking) number of vectors that happen to originate and live within that space. But even though the space itself is inifinite, we can often do a decently good job of describing its shape, size, contours, and general structure with a relatively small number of well-chosen vectors from that space.
Like with the visiting extraterrestrials, even without the context of having been born and raised on this planet and familiar with the culture and human condition, you could probably give a few very basic but nonetheless informative pieces of key data that will help them find the Capitol.
Same with quantum multiverses, and trying to distinguish one from another: there are a theoretically-infinite number of alternate universes, but you probably don't need to store the entire universe on a hard drive in order to differentiate it from another. You could probably devise a much simpler representation that still adequately captures the full universe to a certain degree.
So it is with spaces: you will (almost) always begin every data science problem with the assumption that your data live in some space. Even though the space is likely infinite, your data (you assume) provide a fairly reasonable approximation of the characteristics of that space--at least, reasonable enough for you to conduct some analytics.
I hope I've beaten this dead horse to your satisfaction. If you're still confused, I wouldn't blame you, so please ask me questions on #csci1360e-discussions.
There are a lot of applications of linear algebra, but the one most likely to resonate with the majority of you is using linear algebra to solve systems of equations.
I'm sure you've seen the following ad nauseum since middle school:
Find x.
$y = 2x + 5$
(no, don't draw an arrow to the $x$ with the caption "there it is!")
Systems of equations are sets of multiple such equations that involve the same variables.
As you may recall, you need at least as many equations as you have unique variables in order to solve* a system of equations. In this case: two variables, $x$ and $y$, and two equations.
* (even then, there may not be a solution, but at least there's the possibility of a solution)
How does one solve a system of equations? There are a handful of strategies.
1: Substitution. Solve for one of the variables in one equation, and plug that into the second equation.
2: Elimination. Rewrite the two equations in such a way that when you add them together, a variable is eliminated entirely. Then substitute the remaining variable.
You may have noticed a few things about the previous equations:
This is where the linear aspect of linear algebra comes from. This Math StackExchange answer does a great job of summarizing linear algebra from a high level, the key bit being:
"Linear algebra is so named because it studies linear functions. [...] Roughly, this means that inputs are proportional to outputs and that the function is additive. We get the name 'linear' from the prototypical example of a linear function in one dimension: a straight line through the origin."An important caveat to keep in mind: while you will represent the vast majority of your data as systems of linear equations, do not fool yourself into thinking this is a 100% accurate representation.
Linear algebra is a way of modeling, or representing, things. But it has limitations, the biggest one being that everything is a line. You can't explicitly model complex dependencies between variables.
Of course the benefit of using linear models is that they're (relatively) simple. Need to take a derivative of an equation? Oh, it's just a line? Poof! Need to integrate an equation? Again, just a line? Poof ($+C$)!
By essentially reducing everything down to Taylor polynomials, it becomes very easy to manipulate the equations on a large scale, at the cost of some precision. Exactly how much precision depends on the system, and as data scientists it will be up to you to figure out whether or not the trade-off is worthwhile!
Substitution and elimination work well, but can you imagine doing this sort of thing for systems that involve hundreds of variables and thousands of equations?
And if you do go into data science...that's an optimistic scenario. Quite often, you deal with hundreds of thousands of variables, and millions to billions of equations.
Enter vectors and matrices!
Think of a vector as an array of the coefficients for an equation.
$x + y = 35$
The corresponding vector for this equation would be: [1, 1]
$0.45x + 0.65y = 18.55$
The corresponding vector for this equation would be: [0.45, 0.65]
A matrix, then, is just a bunch of vectors stacked up on top of each other. Stack our two vectors we made from the two equations, and we get
Then, of course, we have our vector of variables: $\vec{x} = [x, y]$
And our solution vector: $\vec{y} = [35, 18.55]$
Which, at last, gives us our linear equation!
This form of the system is ubiquitous in linear algebra, and you're almost always interested in solving for the variables $\vec{x}$.
Most of your job as data scientists will consist of actually setting up this system; that's far and away the hardest part. Once you have the system in this form, there are quite a few cool ways of solving it (some of which we'll touch on in the last week).
But figuring out how to set up the system is non-trivial. What variables are important enough to include? As an example:
I want to predict tomorrow's weather: rainy, sunny, cloudy, or windy. What variables would be good predictors? Today's temperature would probably be a good place to start. The current season as well. Maybe also whether today was rainy, sunny, cloudy, or windy. Probably the last weeks' worth of all that data, too. Maybe historical data from previous years as well?
...and we've already hit hundreds of variables. The equations themselves are our observations: previous data points we have where we can fill in values for all these variables and for which we already know the weather (this is called training data).
There's no hard-and-fast system for choosing what variables to consider, and how many equations to include. In the era of "big data", the general strategy is to include as many equations (observations, data points) as possible.
Variables are a little trickier. Some may help more than others; some may be correlated with each other, which can potentially result in perniciously misleading statistical properties. Some may send you in the wrong direction--which can be helpful if you realize you're heading in the opposite direction!
Some questions to discuss and consider:
1: Recall A4, the BONUS in particular. This discussed an alternative strategy from Q1 for data normalization. It involved making each data point a unit vector, or a vector with magnitude 1. Give an explanation for why this could be a useful normalization step.
2: Starting from the linear equation $A\vec{x} = \vec{y}$, can you perform the $A\vec{x}$ multiplication (matrix-vector) by hand? If you need a refresher, please ask in #csci1360e-discussions.
3: A linear equation, such as $A\vec{x} = \vec{y}$, can also be viewed as a mapping between vector spaces. Explain what the vector spaces of the equation are, and how the mapping is performed.