
Courses
Courses
Choosing a course is one of the most important decisions you'll ever make! View our courses and see what our students and lecturers have to say about the courses you are interested in at the links below.

University Life
University Life
Each year more than 4,000 choose NUI Galway as their University of choice. Find out what life at NUI Galway is all about here.

About NUI Galway
About NUI Galway
Since 1845, NUI Galway has been sharing the highest quality teaching and research with Ireland and the world. Find out what makes our University so special – from our distinguished history to the latest news and campus developments.

Colleges & Schools
Colleges & Schools
NUI Galway has earned international recognition as a researchled university with a commitment to top quality teaching across a range of key areas of expertise.

Research
Research
NUI Galway’s vibrant research community take on some of the most pressing challenges of our times.

Business & Industry
Guiding Breakthrough Research at NUI Galway
We explore and facilitate commercial opportunities for the research community at NUI Galway, as well as facilitating industry partnership.

Alumni, Friends & Supporters
Alumni, Friends & Supporters
There are over 90,000 NUI Galway graduates Worldwide, connect with us and tap into the online community.

Community Engagement
Community Engagement
At NUI Galway, we believe that the best learning takes place when you apply what you learn in a real world context. That's why many of our courses include work placements or community projects.
MA500=CS4102+CS4103
Geometric Foundations of Data Analysis
Covid19 update
From 6pm March 12 the lectures, continuous assessment, and summer examination for this module will be delivered online.

Graham's remaining lectures
Slides for Graham's lectures 19, 20, 21, 22, 23, 24 will be uploaded, as usual, shortly after the lecture slot. A recording of Graham delivering the lecture, to an empty class, will also be uploaded. 
Emil's remaining lectures
Details to be announced. Remaining homeworks
The fourth homework should be submitted by 23 March. The fifth homework should be submitted by 3 AprilSemester exams
The plan is to deliver our Maths exams more or less as usual. They'll be made available on Blackboard at the time scheduled in the official exam timetable. Students will answer them at home. They'll be given two hours to write out the solutions. Then they'll be given a short period of time (30 mins say) to scan or photograph the answers and upload them to Blackboard. The onus will be on the student to make sure that the uploaded file is easily readable by the grader.
Module Content
Data analysis is a process of inspecting, cleansing, transforming, and modeling data with the goal of discovering and communicating useful information, informing conclusions, and supporting decisionmaking. Data analysis involves many overlapping areas of study, such as descriptive statistics (concerned with summarizing information in a sample), statistical inference (concerned with inferences about a population based on properties of a sample), machine learning (concerned with performing tasks using algorithms that rely on patterns and inference), data mining (sometimes considered a subfield of machine learning and concerned with discovering patterns in large data sets), information visualization (concerned with the visual representations of abstract data to reinforce human cognition), and several other areas. See this interesting review for comments about the related term data science.Geometry is concerned with questions of shape, size, relative position of figures, isometry, and the properties of space. It underlies much of data analysis, as can be seen from textbooks such as those by [Kendall], [Le Roux], [Kirby], [Tossdorff], [Hartmann], [Outot], [Tierny], [Edelsbrunner], [Patrangenaru], [Biau], [Wichura], [Dryden]. The recent online textbook Mathematical Foundations for Data Analysis by Jeff M. Phillips again emphasizes the importance of a geometric understanding of techniques when applying them to data analysis.
This module focuses on some geometric methods used in data analysis. It covers the geometric and algorithmic aspects of these methods, as well as their implementation as Python code on Linux computers, and their application to a range of different types of data. The first half of the course emphasizes geometric aspects of classical techniques such as least squares fitting, principal component analysis, hierarchical clustering, nearest neighbour searching, and the JohnsonLindenstrauss Theorem for dimensionality reduction. The second half of the course covers more recent techniques that have been developed over the last two or three decades, and emphasizes topological aspects as well as geometric aspects. The second half of the course makes use of R interfaces to Python Mapper and to efficient C++ routines for persistent homology.
Part I = CS4102: Classical Techniques (5 ECTS, first 24 lectures)
 Least Squares Fitting
 Principal Component Analysis
 Hierarchical Clustering and Persistence
 Nearest Neighbours and the Johnson–Lindenstrauss Theorem
 Topological Preliminaries
 Mapper Clustering
 Persistent Homology
Module Coordinates
 Lecturer: Graham Ellis & Emil Sköldberg
 Lectures:
Mon 10.00am, GE, ADB1020
Tue 12.00m, GE, ADB1020
Wed 10.00am, ES, ADB1019 (or ADB1020)
Fri 14.00pm, ES, ADB1019 (or IT206)  Tutorials: Wednesday and Friday lectures will often take the format of a tutorial and so no formal tutorials are scheduled.
 Recomended text: Part I is based on chapters from the textbook Multivariate Analysis by Sir Maurice Kendall and chapters from the online textbook Mathematical Foundations for Data Analysis by Jeff M. Phillips . Part II is based on the survey An introduction to Topological Data Analysis: fundamental and practical aspects for data scientists by Frédéric Chazal and Bertrand Michel, and on the recent book Topological Data Analysis for Genomics and Evolution Topology in Biology by Rabadán and Blumberg.
 Problem sheet: available here. (A list of examtype problems for selfstudy is available here.)
 Module Website: Information and module documents will be posted to this site, which is linked from the Blackboard MA500=CS4102+CS4103
Geometric Foundations of Data Analysis pages. Blackboard will also be used for announcements and for posting grades.
Module Assessment
Part I will be assessed by a 2hour written exam (52%) and three continuous assessment assignments (16% each).
Part II will be assessed by a 2hour written exam (50%) and two continuous assessment assignments (25% each).
Each exam will consist of four questions, with full marks for four correct answers.
Each assignment will consist of a data analysis problem that needs to be tackled using the Python programming language or the R programming environment, and submitted (by email to both lecturers) as a PDF document.
Supplementary Material and News
Emil's Lecture Notes
These will be posted here.Exam Details
Exam hints/help for MA500/CS4102/CS4103 exams can be found here and here.Lecture Notes
Lecture Notes (Click number to download notes for the lecture.) 
Lecture
Summaries 
1 
Began by explaining the terms "geometry", "data analysis", "statistics", "probability". Then explained how to find the line y=b_{0} + b_{1} x that "best fits" a collection of data points (x_{i},y_{i}) for i=1,2,...,n. We took "best fit" to mean the line that minimizes the sum of the squares of the residuals e_{i} = y_{i}  b_{0}  b_{1}x_{i . } 
2 
Explained how to determine the best (in the least squares sense) hyperplane that fits data points (y_{i},x_{i,1} + ... + x_{i,p1}) in R^{p} for i=1,2, ..., n.
The normal equations were explained using linear algebra and geometry. 
3 
Explained how to determine the best (in the least squares sense) polynomial of degree at most d that fits data points in R^{2}.
Introduced the coefficient of determination Proved that the coefficient of determination satisfies 0<= r^{2} <= 1 for p=2. Then gave the matrix notation for computing R^{2} for p>=2 and stated, without proof, that for p>=2 again 0<= R^{2} <=1. One often says that "R^{2} percent of the variation is explained by considering the p1 independent variables". 
4 
Gave a theorem about the F^{*} = MSR/MSE statistic which is needed for the first homework.
These notes from last year start with a worked example involving this theorem about F^{*} . The notes also contain a second statistical theorem and example that is needed for the first homework. Ask me if you need any explanation about these notes.
Then described the linear algebra setting needed for the description of Principal Component Analysis that will be given in the next lecture. 
5 
Explained that the aim of Principal Component Analysis is to find an orthogonal matrix A for which ACA^{t} is diagonal, where C is the covariance matrix of your data points v_{1}, ..., v_{n} in R^{p}. 
6 
Gave an example of PCA in gait analysis.
Then started towards a proof of the Spectral Theorem. 
7 
Proved the spectral theorem: any real symmetric nxn matrix has n linearly independent eigenvectors.
This is the basis of Principal Component Analysis 
8 
Gave an introduction to hierarchical cluster analysis, dendrograms, and barcodes. Illustrated these notions through two examples: 1) clustered five objects for which pairwise distances were given; 2) counted the number of objects in the following digital photo, and counted the number of these objects that have holes in them. During the lecture I illustrated dendrroagrams and barcodes using the GAP software system for computational algebra. The code for reproducing the examples is here. 
9 
Described the SmithWaterman algorithm for measuring the similarity s(V,W) between two sequences V, W of letters. It finds a maximal scoring local alignment and returns the score.
By scaling, we can assume that 0<= s(V,W) <= 1 on any finite set of data. We can then define the dissimilarity measure In general we do not suppose that d(V,W) satisfies all three axioms of a metric. But when it does we should expect it to be more easily interpreted. Showed the Clustal Omega online resource for determining the (dis)similarity of genetic sequences, and for returning the output in the form of a dendrogram (or phylogenetic tree). 
10 
Began by explaining how cluster analysis and barcodes can be used to investigate the geometric shape of a data set S in R^{n}. A computer demonstration was given for a subset of points S in R^{2} corresponding to the following digital image of a starfish. The corresponding barcode is Then described a singlelikage hierarchical clustering algoithm for producing a barcode from the matrix of distances/dissimilarities between n objects. 
11 
Began by illustrating the single linkage hierarchical clustering algorithm described last lecture.
Then explained the knearest neighbour problem (kNN problem) and gave an example where it needs to be solved. Explained that the naive direct solution to the kNN problem can be improved in many situations. One solution uses Voronoi tessellations of Euclidean space. I'll say more about this next lecture. 
12 
Began with a review of how to construct Voronoi cells as intersetions of halfspaces. Here is an example of a 3dimensional Voronoi cell.
Described an algorithm, involving Voronoi cells, for finding the closest point in a finite set S to a vector v in R^{n}, the finite set S being a collection (or database) of points in R^{n}. Stated the JohnsonLindenstrauss Theorem and the Norm Preservation Proposition. End of CS4102 
13 
Start of CS4103Started to talk about the paper "Topology based data analysis identifies a subgroup of breast cancers with a unique mutational profile and excellent survival" by Nicolau, Levine and Carlsson.Then gave the definition and examples of a simplicial complex. 
14 
Explained what is meant by the geometric realization K of a simplicial complex K.
Then explained how, from any n×n matrix of pairwise distances between n objects, and any ε>0, we can construct a geometric realization K_{ε} of a simplicial complex. We considered a sample S of 250 readings x_i=(h_{i,0}, h_{i,2}, h_{i,4}) of water levels at Galway harbour, where h_{i,0} is the initial level when person i reaches the harbour, h_{i,2} is the level 2 hours later, and h_{i,4} is the level 4 hours after the initial reading. We constructed a 250x250 matrix of distances d_{i,j}=x_{i}x_{j} and used this to construct a simplicial complex K_{ε} for various thresholds ε>0. The graph of one of these simplicial complexes was as follows. 
15 
Began with 750 points randomly selected from two disjoint quarters of a torus in R^{3}. Any linear transformation R^{3} > R^{2} (obtained say from PCA or the JohnsonLindenstrauss theorem) would lose geometric information. However, we saw that the geometric information seems to be retained when the points are mapped to the vertices of the clique complex K_{ε} for various values of ε.
Introduced the notion of homotopy between two maps f,g:X>Y. Introduced the notion of homotopy equivalence between two topological spaces. Proved that the circle is homotopy equivalence to the projective plane minus the origin. 
16 
Stated Leray's nerve theorem: the geometric realization NU of the nerve of an open cover of a compact space X is homotopy equivalent to X, provided that any subcollection of the open sets in the cover has empty or contractible intersection. 
17 
Explained how Leray's nerve theorem motivates the use of the clique complex K_{ε} in data analysis.
Next lecture we see how these ideas lead to the Mapper clustering procedure for representing a matrix of distances between data points as a simplicial complex. The procedure was illustrated by taking 1000 points in the plane, sampled at random from the image of the starfish in Lecture 10. The following code produced the following 1dimensiona simplicial complex as a representation of the data.
gap> HapExample("1.3.5");

18 
Described the Mapper clustering procedure. Next lecture we'll see some computer examples of this. 
19 
An audio version of the lecture slides is available here.
Recapped what we've done so far in topological data analysis, and then defined the nth Betti number of a simplicial complex. Ended up with an example in which I calculated the 0th and 1st Betti numbers of a simplicial complex. 
20 
An audio version of the slides is available here.
Introduced, in an informal way, the notion of persistence of 1dimensional holes in a data set, and explained how this persistence can be represented using barcodes. Ended up with the definition of the degree n homology vector space H_{n}(K) of a simplicial complex K. For finite K this vector space has finite dimension equal to the Betti number β_{n}(K). 
21 
An audio version of the slides is available here.
Started with a computer example of persistent homology analysis. Then defined persistent homology and persistent Betti numbers. Ended with an explanation of an algorithm, involving only matrix column operations, for computing persistent Betti numbers and bar codes. 
22 
An audio version of the slides is available here.
Went through an example of the use of persistent homology to analyze the space of natural images. Finished off with a statement of a stability theorem. A rough, nonmathematical, statement is: if data sets are changed just a small amount then the resulting barcodes only change a small amount. 
23 

24 