In this introductory lecture, the instructor outlines a three‑video series on Principal Component Analysis (PCA), describing its purpose as a set of tools for studying and visualising large quantitative data tables. The talk defines the data structure (individuals in rows, quantitative variables in columns), introduces notation for means and standard deviations, and surveys a wide range of disciplinary applications (sensory analysis, ecology, economics, genetics, marketing, sociology). A running example of 10 wines evaluated on 27 sensory attributes (plus two preference scores and a qualitative label) is presented throughout. The speaker explains the two complementary perspectives of PCA – analysing individuals (rows) to find similar or extreme cases, and analysing variables (columns) to uncover linear relationships via the correlation matrix. Because correlation matrices become unwieldy with many variables, PCA creates synthetic (principal component) variables that summarise the main directions of variability, enabling easier visualisation and interpretation of both individuals and variables. The lecture concludes by emphasizing that simultaneous study of rows and columns enriches interpretation and sets the stage for practical implementation in the next videos.
Chapters
This week we have for you 3 videos which together present the main details of principal component analysis. Principal component analysis is a set of tools which allow us to study and visualize large datasets. We will present the methods from a theoretical as well as a practical point of view. The outline of this week's work is as follows. We will first define the types of data we can use principal component analysis on. Then we'll look at examples of situations in which principal component analysis can be performed. Then we'll define some useful notation. Next, we'll focus on the individuals, then on the variables. At the end, we will spend some time looking at how to interpret the results and provide interpretation aids. So what kind of data are we looking at? Principal component analysis, also known as PCA, applies to data tables where rows can be considered like individuals and columns like quantitative variables. Let x_ik be the value taken by individual i for variable k. i varies from 1 to capital I, the number of individuals, and k from 1 to capital K, the number of variables. Overline x_k is the mean of variable k calculated over all individuals, and s_k the standard deviation of the sample for variable k. Here we use 1 over y, which means that the standard deviation is calculated on the data and we don't try to estimate the standard deviation of the population. Data tables, with individuals in rows and variables in columns, can be found in many different areas, which means that we can perform PCA on quite a diverse range of datasets. Here I have listed a few applications with examples. Here is a first example involving sensory analysis where products are described by a set of variables, often called sensory attributes, like sourness, bitterness, sweetness, and so on. Thus, we have a table with various products for example different wines, and each is going to have a score for sourness, bitterness, sweetness, etc. So the aim of PCA is to study this data table. Here is an example from ecology where individuals are rivers and variables are different pollutants. So we end up with a data table with rivers in rows and pollutants in the columns. In economics, we might have years in the rows and economic indicators in the columns. We can then track the evolution of economic indicators through the years. Instead of years for rows, it could be countries if we want to compare the economic situation of several countries. Often in genetics, patients are represented in terms of their genes. This kind of dataset can be large since there are lots of genes. In marketing, we could have a set of brands and several measures of satisfaction. And in sociology, we could have different social classes and different activities, with the table entries being the average time spent by individuals of each social class at a given activity. Clearly, tables like these, with individuals in rows and variables in columns, are found across many fields. We are going to work with a running example throughout this course involving wine. In the rows, we have 10 wines and in the columns, 27 quantitative variables. These are sensory attributes like sweetness, bitterness, fruity odor, and so on. The values in the data table correspond to the average score given by several judges for the same wine and descriptive variable. So for example, the wine S. Michaud has a mean score of 4.3 for the fruity odor. That is to say, this is the average over all judges. And so on for all wines and all variables. In our dataset, we have also 2 quantitative variables that correspond to preference: preference in terms of odor and overall preference. We have also a qualitative variable representing different wine labels. There are 2 labels here. The first is Sauvignon and the second is Vouvray. We will see later how we can consider this information in the analysis. The aim of doing PCA here is to characterize the wines according to their sensory characteristics. So first, we will focus on the 27 sensory characteristics for characterizing the wines. How we can study this data table? This data table can be studied in different ways. We can see it as a set of rows and try to look for differences between rows. We can also look at the data table as a set of columns and investigate the similarities or links between columns. To study individuals, that is the rows, we need to know when 2 individuals are close and when they are different from the point of view of all the variables. If there are many individuals, which are the most similar and the most dissimilar? Are there any groups of individuals which are homogeneous in terms of their similarity? In addition, we may want to look for axes of variability which can separate extreme individuals from more normal ones. In our example, 2 wines are considered similar if they are evaluated similarly for all sensory characteristics. In practice, this means that the 2 wines will consistently vary in the same direction with respect to the mean for many characteristics and can thus be said to have the same sensory profile. More generally, we may want to know whether or not there are groups of wines with similar profiles That is, similar sensory profiles which might separate extreme wines from more averaged ones. Following the approach taken to study individuals, might it also be possible to interpret the data in terms of the variables? For instance, which variables provide similar information to each other? Between variables, rather than similarity, we in fact talk about relationships. And the most well-known studied relationships between variables are linear. Indeed, PCA focuses on linear relationships between variables. More complex connections also exist such as quadratic, logarithmic, and exponential ones, but these are not studied in PCA. This may seem restrictive, but in practice, many relationships can be considered linear, at least as a first approximation. In exactly the same way as for individuals, creating groups of variables may provide useful information. When we have only a very small number of variables, it is possible to draw conclusions from the correlation matrix. This matrix holds all of the linear correlation coefficients between pairs of variables. However, when working with a large number of variables, the correlation matrix is huge And it's therefore essential to have a tool capable of summarizing the most important relationships between variables in a visual way. The aim of PCA is to draw conclusions from the linear relationships between variables by detecting the main directions of variability. As we will see, these conclusions can be enriched using the synthetic variables generated by PCA. It then becomes easier to characterize the data using a small number of synthetic variables rather than all the original ones. In our example, the correlation matrix brings together the 351 correlation coefficients, so trying to group variables using the correlation matrix would be rather tedious. The average itself is a synthetic variable, but This is defined a priori. Here, we want to define indices from the data, so a posteriori. Obviously, since we are working with the same data table, looked at from either the point of view of rows or columns, there exists a link between the 2. Therefore, simultaneously studying individuals and variables will only improve their respective interpretations. When studying individuals, we can build groups of individuals, but then we want to characterize these groups, and to do this, we would like to use the variables. For instance, we would like to be able to say that certain products are similar because they are acidic and bitter, whereas others are similar because they are sweet. So for this, we need an automated method especially if we have a lot of variables. Similarly, when there are groups of variables, it may not be easy to interpret the relationships between large number of variables in a group, so we could think about making use of specific individuals, that is individuals who are extreme with respect to these relationships. For example, the connection between height and weight could be illustrated by the contrast between 2 extreme individuals. A dwarf and a giant. In summary, PCA aims to delve into and decrypt data. In PCA, we can visualize the data with simple plots, which give nice summaries of the data. Essentially, we take the information contained in a data table and view it graphically. Let's now start to get into the details. We said earlier that we can look at our data table as a set of rows or as a set of columns. If we see it as a set of rows, we aim studying individuals. If we see it as a set of columns, we aim studying variables. Studying individuals can be seen as considering a cloud of points, where each point corresponds to an individual. This point cloud lives in a space with many dimensions. If there are K variables, it lives in a K-dimensional space. When studying variables, each is a point in an I-dimensional space, that is to say there are I coordinates for each variable. Therefore, I will see a cloud of variables in R^I. We have seen what data tables look like in PCA applications and what kinds of questions we can ask with PCA. In the next videos, we will see how to put PCA into practice.
