Tda presentation

TOPOLOGICAL DATA ANALYSIS

HJ van Veen· Data Science· Nubank Brasil

TOPOLOGY I

• "When a truth is necessary, the reason for it can be found by analysis, that is, by resolving it into simpler ideas and truths until the primary ones are reached." - Leibniz

TOPOLOGY II

• Topology is the mathematical study of topological spaces.

• Topology is interested in shapes,

• More specifically: the concept of 'connectedness'

TOPOLOGY III• A topologist is someone who does not see the

difference between a coffee mug and a donut.

HISTORY I

• “Nothing at all takes place in the universe in which some rule of maximum or minimum does not appear.” - Euler

• Seven Bridges of Koningsbrucke: devise a walk through the city that would cross each bridge once and only once.

HISTORY II

HISTORY III• Euler's big insights:

• Doesn’t matter where you start walking, only matters which bridges you cross.

• A similar solution should be found, regardless where you start your walk.

• only the connectedness of bridges matter,

• a solution should also apply to all other bridges that are connected in a similar fashion, no matter the distances between them.

HISTORY IV

• We now call these graph walks ‘Eulerian walks’ in Euler’s honor.

• Euler's first proven graph theory theorem:

• 'Euler walks' are possible if exactly zero or two nodes have an odd number of edges.

TDA I• TDA marries 300-year old maths with

modern data analysis.

• Captures the shape of data

• Is invariant

• Compresses large datasets

• Functions well in the presence of noise / missing variables

TDA II• Capturing the shape of data

•Traditional techniques like clustering or dimensionality reduction have trouble capturing this shape.

TDA III• Invariance.

• Euler showed that only connectedness matters. The size, position, or pose of an object doesn't change that object.

TDA IV• Compression.

• Compressed representations use the order in data.

• Only order can be compressed.

• Random noise or slight variations are ignored.

• Lossy compression retains the mostimportant features.

• "Now where there are no parts, there neither extension, nor shape, nor divisibility is possible. And these monads are the true atoms of nature and, in a word, the elements of things." - Leibniz

MAPPER I

• Mapper was created by Ayasdi Co-founder Gurjeet Singh during his PhD under Gunnar Carlsson.

• Based on the idea of partial clustering of the data guided by a set of functions defined on the data.

MAPPER II• Mapper was inspired by the Reeb Graph.

MAPPER III• Map the data with overlapping intervals.

• Cluster the points inside the intervals

• When clusters share data points draw an edge

• Color nodes by function

MAPPER IV

MAPPER VDistance_to_median(row) x y z

1.5 1.5 1.5 1.5

1.5 -0.5 -0.5 -0.5

0 1 1 1

0 1 0.9 1.1

3 2 2 2

3 2.1 1.9 2

Y

MAPPER VI• In conclusion:

FUNCTIONS• Raw features or point-cloud axis / coordinates

• Statistics: Mean, Max, Skewness, etc.

• Mathematics: L2-norm, Fourier Transform, etc.

• Machine Learning: t-SNE, PCA, out-of-fold preds

• Deep Learning: Layer activations, embeddings

CLUSTER ALGO’S• DBSCAN / HDBSCAN:

• Handles noise well.

• No need to set number of clusters.

• K-Means:

• Creates visually nice simplicial complexes/graphs

SOME GENERAL USE CASES

• Computer Vision

• Model and feature inspection

• Computational Biology / Healthcare

• Persistent Homology

COMPUTER VISION• Demo

MODEL AND FEATURE INSPECTION

• Demo

COMPUTATIONAL BIOLOGY• Example

PERSISTENT HOMOLOGY• Example

SOME FINANCE USE CASES

• Customer Segmentation

• Transactional Fraud

• Accurate Interpretable Models

• Exploration / Analysis

CUSTOMER SEGMENTATION• Demo

TRANSACTIONAL FRAUD• Example of spousal fraud

ACCURATE INTERPRETABLE MODELS

• Create: global linear model

• Function: L2-norm

• Color: Heatmap by ground truth and animate to out-of-fold model predictions

• Identify: Low accuracy sub graphs

• Select: Features that are most important for sub graphs

• Create: Local linear models on sub graphs

• Stack: Decision Tree

• Compare: Divide-and-Conquer and LIME

• DEMO

EXPLORATION / ANALYSIS• Demo

QUESTIONS?

FURTHER READING• Google terms:

• Ayasdi, Topological Data Analysis, Robert Ghrist, Gurjeet Singh, Gunnar Carlsson, Anthony Bak, Allison Gilmore, Simplicial Complex, Python Mapper.

• Videos:

• https://www.youtube.com/watch?v=4RNpuZydlKY

• https://www.youtube.com/watch?v=x3Hl85OBuc0

• https://www.youtube.com/watch?v=cJ8W0ASsnp0

• https://www.youtube.com/watch?v=kctyag2Xi8o