Marketing Science 2017 - IFTTT TDA - presentation version.pptx

Visualizing Emergent Identity of Assemblages in the

Internet of Things:

A Topological Data Analysis Approach

Paper&presented&at&2017&INFORMS&Marketing&Science&Conference,&Los&Angeles,&CA,&June&10,&2017

Tom$Novak$and$Donna$Hoffman,$The$George$Washington$University

Agenda

Use Topological Data

Analysis (TDA) to

operationalize

DeLanda’s (2002,

2006, 2011, 2016)

concept of an

assemblage’s

possibility space

(a.k.a. the market

structure of

underlying consumer

needs).

› The Internet of Things (IOT) and

IFTTT

› Topological Data Analysis (TDA)

› Analysis of IFTTT Recipes with TDA

› Future Directions

2

The IoT and IFTTT

3

IFTTT Codes IoT and Web

Interactions

4

Internet&of&Things&(IoT)

The$wide$range$of$everyday$objects$and$products$in$the$real$world$

that$are$enhanced$with$programmable$sensors$and$actuators$that$

communicate$with$other$devices$and$consumers$through$the$

Internet.$$$--Hoffman$and$Novak$2016

IFTTT.com

Anatomy of an IFTTT Recipe

5

If any new post on Blogger then create a link post on Facebook

Trigger

Event

Trigger

Channel

Action

Event

Action

Channel

IFTTT (If-This-Then-That) Recipe

5 Years of IFTTT

6

Each day, 20 million IFTTT recipes are run by IFTTT users (Lunden 2015).

Users can choose to make their recipes public by publishing them. From 2011

to 2016, a total of 331,391 IFTTT recipes have been published.

Of these published recipes, 20,675 are unique. Variants of these unique

recipes have been published anywhere from 1 to 9273 times.

The 20,675 unique published IFTTT recipes use:

• 347 different trigger channels using 1110 different trigger events

• 297 different action channels using 591 different action events

Research Question: What is the topological structure of the 20,675 published

IFTTT recipes?

An Assemblage Emerges from the Interaction of its

Components

7

Consumer

Alexa

Hue

IFTTT

Assemblage

Interacting&Components

The Possibility

Space and

Assemblages

TDA provides an

empirical way of

visualizing the

mechanism-

independent

possibility space,

given a population

of individual

assemblages.

Two requirements for emergence of an

assemblage (DeLanda 2011):

› Mechanism-dependent: ongoing

recurrent processes involve

interaction among components of an

assemblage, through their exercised

paired capacities.

› Mechanism-independent: a

mathematical topological structure, the

possibility space, contains points of

attraction that guide the recurrent

processes of assemblage. This leads to

the emergence of populations of

assemblages that wind up in the space

place, although “different trajectories may

be attracted to the same final state”

(DeLanda 2002, p 7).

8

Topological Data

Analysis (TDA)

9

Topological Data Analysis (TDA)

TDA (Carlsson 2009; Lum, et.al. 2012; Singh, Memoli and Carlsson

2007) uses computational topology techniques on complex high-

dimensional data to produce a three-dimensional topology of simplicial

complexes (discrete, combinatorial objects) in which groups of data are

represented as nodes that contain rows that are similar to each other

in the high-dimensional topological space and the edges connect nodes

that share rows.

10

How TDA Creates Topological Models

11

X

y

Bin 1

Bin 2

Bin 3

Bin 4

Bin 1

Bin 2

Bin 3

Bin 4

Slide images adapted courtesy of Ayasdi, Inc. http://ayasdi.com/

Step 1

Rectangular data.

Many rows, 2

columns x and y.

When plotted,

these data define

a circle.

Step 2

Map data

onto a single

number using

a function (or

lens). Here,

the lens is

the y

coordinate.

Step 3

Sort data into

overlapping bins,

based upon the

value of the lens.

Then, look at

values of the

original variables

in these bins.

Step 4

Cluster the original

data (i.e. x and y

values) within each

bin. Bins 1 and 4

have one cluster.

Bins 2 and 3 have

two clusters. This

results in 6 nodes.

Step 5

Connect nodes by

an edge if they

have rows in

common. The

shape of the

topological model

has meaning and

represents the

shape of the data.

Analysis of IFTTT

Recipes with TDA

12

TDA Implementation

Open source

› Python Mapper (Müllner and Babu 2013; Singh, Mémoli, and Carlsson 2007)

› Kepler Mapper in Python (Triskelion 2015, proof of concept for Ayasdi flavor of TDA)

› Dionysus (in C++ with Python bindings) based on Zomorodian and Carlsson (2005)

and Edelsbrunner,

› Letscher and Zomorodian (2000) on computing persistent homology

› Package TDA - R interface for GUDHI, Dionysus, PHAT

› TDA Mapper – R package using Mapper

› JavaPlex library implements persistent homology for MATLAB and java-based systems

› CTL = C++ library for computational topology

› And others, see GitHub

Commercial

› Ayasdi (Workbench Web platform and Python SDK)

13

IFTTT Recipes – Preparation for TDA

20,675 unique IFTTT published recipe text strings from 2011-2016 of the form:

IF say a specific phrase using_channel Amazon Alexa

THEN boost your hot water using_channel Hive Active Heating

Created binary variables for ngrams that had a frequency > 50, ignoring standard

stop words:

› 1 grams (triggers n=364 and actions n=264): “Alexa”

› 2 grams (triggers n=451 and actions n=353): “Amazon Alexa”

› 3 grams (triggers n=407 and actions n=324): “Amazon Alexa add”

PCA on 2163 ngrams (of these, 880 were non-redundant)

› 362 eigenvalues > 1

› First 100 eigenvalues explained 51.04% of variance (first 2 explained 1.91%)

› Obtained scores on first 100 components (variance = eigenvalue)

14

Assumptions for TDA of Our IFTTT Data

Euclidean distance was used as the metric defining distance

between the rows of the data matrix.

› Why used? Each row contains scores on the first 100 principal

components of ngrams, so Euclidean distance has a natural

interpretation.

Neighborhood Lenses 1 & 2 (coordinates of a k-nearest neighbors

graph of the data embedded in two-dimensions)

› Why used? Magnifies differentiation among groups by locally

adapting distance/closeness.

Note: other metrics and lenses were also tried, but produced less clearly

interpretable topological modes.

15

Topological Model of 5 years of

Published IFTTT Recipes from 2011-2016

32

71%$of$IFTTT$recipes$are$in$

nodes$of$one$large$

connected$component

10%$of$IFTTT$recipes$

are$in$nodes$that$

are$singletons

19%$of$IFTTT$recipes$are$in$nodes$of$13$small$connected$components

Ayasdi 6.7$Workbench$and$Python$SDK$used$to$generate$topological$models

16

367 nodes contain 14,692 IFTTT recipes (71% of all recipes).

Recipes can be in more than one node.

Nodes are connected by an edge if they have a recipe in common.

Topological Model of 5 years of

Published IFTTT Recipes from 2011-2016

17

Topological Models – Varying Resolution

(number of bins used by TDA to generate nodes)

Low$Resolution$Model$(resolution$=$25)

High$Resolution$Model$(resolution$=$60)

18

Topological Models – Varying Gain

(degree of overlap of IFTTT recipes within nodes)

Low$Gain$Model$(gain$=$1.4)

High$Gain$Model$(gain$=$2.8)

19

T=DateTime

A=RescueTime,$

LIFX,Wemo

T=Android$Device,$

Reddit,$Android$

Battery

A=Android$Device,$

Android$Wear

T=Facebook$Pages,$

Facebook,$Tumblr,$

Twitter

A=Facebook,$

Dropbox,$OneDrive,$

Box,$Tumblr,$

WordPress,$Blogger

z

T=Gmail

A=IF,Hue,$

Boxcar,$Wemo

T=Alexa,$Wemo,$

SmartThings,$Ring,$

ScoutAlarm,$Dlink

Motion,$Arlo

A=SmartThings,$$

Harmony,$

Manything

T=Location

T=Google$Calendar,$

Stocks,$Square

A=Skype,$

Pushbullet

T=Fitbit,$

Jawbone$UP,$

Misfit,$

Withings,$Nike+

A=Slack,$Hue,$

IF,$Gail,$SMS

T=Pocket,$WordPress,$

Tumblr,$Blogger,$Ebay,$

Instapaper,$Craigslist

A=Evernote,$Delicious,$

Diigo,$Pinboard,$

Tumblr,$WordPress,$

Blogger$

A=OneNote

T=Android$

SMS

A=Android$

Device

T=Youtube,$

Foursquare,$Vimeo,$

Flickr,$DailyMotion

A=Pocket,$Instapaper,$

Bitly,$Tumblr,$FB$

Pages,$Blogger

T=Inoreader,$

Tumblr,$Flickr,$

Twitter

A=Buffer

A=Google$Drive

T=Twitter,$Life$

360

A=Ecobee,$

Tumblr

T=Automatic,$

Spotify,$

Withings,$Dash,$

Toodledoo,$

FitBit

A=Google$Cal,$

Google$Glass,$

Jawbone$Up,$

SMS

A=Todooist,$

Beminder,$

Toodledo

T=Email

A=Hue,$

Wemo,

ManyThing

T=Weather,$

Space

A=Hue,$Nest,$

LittleBits,$LIFX,$

WeMo,$IF

T=IFTTT,$GitHub,$

Wemo,$Yo,$

Particle,$Stripe,$

Smappee

A=email,$Gmail,$

Pusbullet,$IF$

notifications,$

SMS

T=Alexa$

(say$specific$

phrase)

T=500px

A=Feedly

Interpretation via

IFTTT Channels

Groups$of$nodes$identified$with$network$clustering$using$Ayasdi’s Community$Algorithm$(Louvain$modularity$optimization).

20

IMAGE&AND&

VIDEO&SOCIAL&

CONTENT

SOCIAL&

MEDIA

HOME&IOT

INFORMATION

TASKS

WEARABLE&IOT

TDA Identifies 6 Broad Groups of

IFTTT Recipes

21

Year 1

Year 3

Year 5

22

487 Amazon Alexa IFTTT Recipes

”Social Alexa”

Consumer talks

to Alexa

”Alexa Automation”

Alexa triggers other actions

Red$shading$indicates$that$>$50%$of$recipes$in$a$node$use$Amazon$Alexa.$

TDA$identifies$two$categories$of$Alexa$IFTTT$recipes$occupying$different$locations$in$the$

topological$model.$

SOCIAL&

MEDIA

HOME&IOT

33

TDA of Subset of 487 Alexa IFTTT Recipes

z

Social$Alexa

247$Recipes

Alexa$Automation

168$Recipes

Other$Alexa$Recipes

72$Recipes

Future Directions

23

24

Removing Noise in Data Provides a Clearer Topological Model

TDA$of$2163$IFTTT$ngrams

(Hamming$Metric)$

TDA$of$100$principal$components

of$2163$IFTTT$ngrams

(Euclidean$Distance$Metric)$

25

Preliminary Analysis of Manually Coded IFTTT Rules

TDA$of$455$IFTTT$codes

Trigger$noun$and$verb$phrase$codes

Action$noun$and$verb$phrase$codes

(Hamming$Metric)$

28

Analysis of Manually Coded IFTTT Rules

THING TRIGGERS

THING ACTIONS

WEB TRIGGERS

WEB ACTIONS

WEB TRIGGERS

THING ACTIONS

THING TRIGGERS

WEB ACTIONS

Future

Research

Directions

Methodological

› Comparison of alternative approaches for

processing structured text data (IFTTT

recipes): N-grams, word2vec, doc2vec,

topic models, latent semantic analysis,

human coding.

› Comparison of alternative visualization

approaches: TDA, network analysis, PCA,

MDS, hierarchical and k-means clustering

Substantive

› Why are certain IFTTT recipes more

frequently published, favorited and

added?

› What are the dynamics of how new IFTTT

recipes emerge over time?

29

30

Acknowledgments

The Ayasdi 6.7 software platform for

topological data analysis (ayasdi.com)

was used to construct all topological

models of the IFTTT data. The authors

acknowledge the support of Devi

Ramanan, Global Head – Product

Collaborations, Ayasdi Inc., Menlo Park,

CA

IFTTT public recipe data from 2011-

2016 were provided by, and used with

permission of, IFTTT.com, San

Francisco, CA.

31

postsocial.gwu.edu