StarZero All shows Try the podcast clipper →

Dec 5, 2013 · 26m · mad

Vaclav Petricek, eHarmony // Data Driven NYC 19 // October 2013

Vaclav Petricek · 20m spoken Andy Beveridge · 50s spoken Misha Sobolev · 23s spoken Matt Turck · 21s spoken Margaret Oest · 13s spoken
0:00 / 0:00
▶ Watch on YouTube →

gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions

Vaclav Petricek, Director of Machine Learning at eHarmony, presents how data science, psychological profiling, empirical behavioral modeling, and distributed graph optimization are combined to predict long-term marital compatibility and optimize online matchmaking.

How this conversation actually went

Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 1.7% of the talking time here. How this is scored →

Matt as informed peer 0.4 Guest teaching 5.4 Guest disagreement 0.4 Matt pushing back 0.4
05100:0010:0020:000:43–4:12 · Matt as informed peer 0/10 Predicting Long-Term Compatibility Vaclav presents eHarmony's 150-question survey model and psychological dimensions like obstreperousness used to predict long-term marital satisfaction. As a pure monologue presentation segment, host participation scores are zero.4:12–8:09 · Matt as informed peer 0/10 Affinity Matching and User Communication Signals Vaclav details signal analysis on user profile features including distance decay, height differentials, and food preferences. The host does not speak during this presentation segment.8:09–15:02 · Matt as informed peer 0/10 Computer Vision and Machine Learning Infrastructure Vaclav gives a technical overview of eHarmony's machine learning stack, including Vowpal Wabbit, Hadoop, RStudio, and genetic algorithm packages. The host is inactive during this monologue segment.15:02–18:14 · Matt as informed peer 0/10 Match Distribution Optimization and PNAS Study Findings Vaclav explains match distribution as a network flow problem and presents findings from a published PNAS study on eHarmony marriage and divorce rates. Host scores remain zero for this presentation monologue.18:14–26:07 · Matt as informed peer 2/10 Interactive Audience Question and Answer Session Host Matt Turck moderates an audience Q&A session where attendees question Vaclav on subscription business models, international models, and self-selection bias in academic studies. Vaclav responds collaboratively while clarifying technical and methodology points.0:43–4:12 · Guest teaching 5/10 Predicting Long-Term Compatibility Vaclav presents eHarmony's 150-question survey model and psychological dimensions like obstreperousness used to predict long-term marital satisfaction. As a pure monologue presentation segment, host participation scores are zero.4:12–8:09 · Guest teaching 5/10 Affinity Matching and User Communication Signals Vaclav details signal analysis on user profile features including distance decay, height differentials, and food preferences. The host does not speak during this presentation segment.8:09–15:02 · Guest teaching 6/10 Computer Vision and Machine Learning Infrastructure Vaclav gives a technical overview of eHarmony's machine learning stack, including Vowpal Wabbit, Hadoop, RStudio, and genetic algorithm packages. The host is inactive during this monologue segment.15:02–18:14 · Guest teaching 6/10 Match Distribution Optimization and PNAS Study Findings Vaclav explains match distribution as a network flow problem and presents findings from a published PNAS study on eHarmony marriage and divorce rates. Host scores remain zero for this presentation monologue.18:14–26:07 · Guest teaching 5/10 Interactive Audience Question and Answer Session Host Matt Turck moderates an audience Q&A session where attendees question Vaclav on subscription business models, international models, and self-selection bias in academic studies. Vaclav responds collaboratively while clarifying technical and methodology points.0:43–4:12 · Guest disagreement 0/10 Predicting Long-Term Compatibility Vaclav presents eHarmony's 150-question survey model and psychological dimensions like obstreperousness used to predict long-term marital satisfaction. As a pure monologue presentation segment, host participation scores are zero.4:12–8:09 · Guest disagreement 0/10 Affinity Matching and User Communication Signals Vaclav details signal analysis on user profile features including distance decay, height differentials, and food preferences. The host does not speak during this presentation segment.8:09–15:02 · Guest disagreement 0/10 Computer Vision and Machine Learning Infrastructure Vaclav gives a technical overview of eHarmony's machine learning stack, including Vowpal Wabbit, Hadoop, RStudio, and genetic algorithm packages. The host is inactive during this monologue segment.15:02–18:14 · Guest disagreement 0/10 Match Distribution Optimization and PNAS Study Findings Vaclav explains match distribution as a network flow problem and presents findings from a published PNAS study on eHarmony marriage and divorce rates. Host scores remain zero for this presentation monologue.18:14–26:07 · Guest disagreement 2/10 Interactive Audience Question and Answer Session Host Matt Turck moderates an audience Q&A session where attendees question Vaclav on subscription business models, international models, and self-selection bias in academic studies. Vaclav responds collaboratively while clarifying technical and methodology points.0:43–4:12 · Matt pushing back 0/10 Predicting Long-Term Compatibility Vaclav presents eHarmony's 150-question survey model and psychological dimensions like obstreperousness used to predict long-term marital satisfaction. As a pure monologue presentation segment, host participation scores are zero.4:12–8:09 · Matt pushing back 0/10 Affinity Matching and User Communication Signals Vaclav details signal analysis on user profile features including distance decay, height differentials, and food preferences. The host does not speak during this presentation segment.8:09–15:02 · Matt pushing back 0/10 Computer Vision and Machine Learning Infrastructure Vaclav gives a technical overview of eHarmony's machine learning stack, including Vowpal Wabbit, Hadoop, RStudio, and genetic algorithm packages. The host is inactive during this monologue segment.15:02–18:14 · Matt pushing back 0/10 Match Distribution Optimization and PNAS Study Findings Vaclav explains match distribution as a network flow problem and presents findings from a published PNAS study on eHarmony marriage and divorce rates. Host scores remain zero for this presentation monologue.18:14–26:07 · Matt pushing back 2/10 Interactive Audience Question and Answer Session Host Matt Turck moderates an audience Q&A session where attendees question Vaclav on subscription business models, international models, and self-selection bias in academic studies. Vaclav responds collaboratively while clarifying technical and methodology points.

speaking balance: gold is Matt, purple is the guest (3 minute bins)

0:00 · Matt 5% · guest 95%0:00 · Matt 5% · guest 95%3:00 · Matt 0% · guest 100%3:00 · Matt 0% · guest 100%6:00 · Matt 0% · guest 100%6:00 · Matt 0% · guest 100%9:00 · Matt 0% · guest 100%9:00 · Matt 0% · guest 100%12:00 · Matt 0% · guest 100%12:00 · Matt 0% · guest 100%15:00 · Matt 0% · guest 100%15:00 · Matt 0% · guest 100%18:00 · Matt 5% · guest 95%18:00 · Matt 5% · guest 95%21:00 · Matt 0% · guest 100%21:00 · Matt 0% · guest 100%24:00 · Matt 8.6% · guest 91.4%24:00 · Matt 8.6% · guest 91.4%
Sharpest disagreement ▶ 23:38 Andy Beveridge corrects Vaclav on selection bias studies

An audience member directly counters Vaclav's assumption regarding university selection bias studies, pointing out that published research on the topic already exists.

 Hardest push from Matt ▶ 21:58 Andy Beveridge challenges eHarmony PNAS claims on selection bias

An audience member questions whether eHarmony's reported lower divorce rates stem from user self-selection bias rather than algorithmic superiority.

 Biggest teaching moment ▶ 5:25 Empirical height and distance signals in user matching

Vaclav educates the audience with precise empirical data showing how distance decay plateaus around 60 miles and optimal user response occurs with a 4 to 8 inch height differential.

 Matt holds his own ▶ 19:58 Matt Turck probes data science team size

Host Matt Turck steps in during Q&A to ask Vaclav specifically about the size and composition of eHarmony's engineering and machine learning teams.
the scores for every segment, with the reasoning behind each
ChapterTopicMatt as informed peerGuest teachingGuest disagreementMatt pushing backWhy
Predicting Long-Term Compatibility 0500 Vaclav presents eHarmony's 150-question survey model and psychological dimensions like obstreperousness used to predict long-term marital satisfaction. As a pure monologue presentation segment, host participation scores are zero.
Affinity Matching and User Communication Signals 0500 Vaclav details signal analysis on user profile features including distance decay, height differentials, and food preferences. The host does not speak during this presentation segment.
Computer Vision and Machine Learning Infrastructure 0600 Vaclav gives a technical overview of eHarmony's machine learning stack, including Vowpal Wabbit, Hadoop, RStudio, and genetic algorithm packages. The host is inactive during this monologue segment.
Match Distribution Optimization and PNAS Study Findings 0600 Vaclav explains match distribution as a network flow problem and presents findings from a published PNAS study on eHarmony marriage and divorce rates. Host scores remain zero for this presentation monologue.
Interactive Audience Question and Answer Session 2522 Host Matt Turck moderates an audience Q&A session where attendees question Vaclav on subscription business models, international models, and self-selection bias in academic studies. Vaclav responds collaboratively while clarifying technical and methodology points.

Statements from this episode (12)

Disclosure
Petricek: eHarmony asks users about 150 questions to model personality
“We ask actually about 150 questions these days, which give us a very good snapshot of the personalities of people.”
Vaclav Petricek Dec 5, 2013 ▶ 2:12
Assertion Not checkable as stated
Petricek: Obstreperousness is predictive of long-term marital satisfaction
“One of these dimensions that's quite predictive is obstreperousness.”
Vaclav Petricek Dec 5, 2013 ▶ 2:32
Disclosure
eHarmony only introduces couples predicted to be in upper quartile of satisfaction
“And we only introduce them if they fall into the upper quartile. If there is a very high probability that they will be in the upper quartile in terms of satisfaction.”
Vaclav Petricek Dec 5, 2013 ▶ 4:02
Assertion Not checkable as stated
eHarmony data shows communication peaks when men are 4-8 inches taller
“The highest probability of communication between two people is when the height difference is between four to eight inches, when the male is four to eight inches taller.”
Vaclav Petricek Dec 5, 2013 ▶ 6:39
Assertion Not checkable as stated
eHarmony finds matching vegetarians yields the highest positive success lift
“For vegetarians, this is the biggest positive lift versus, like, a normal average match that we make.”
Vaclav Petricek Dec 5, 2013 ▶ 7:38
Assertion Not checkable as stated
eHarmony has data on about 40 million registered users
“We have this data on about forty million users who have registered on eHarmony you know, since the beginning, and we know all these attributes about them”
Vaclav Petricek Dec 5, 2013 ▶ 10:17
Disclosure
eHarmony relies on Vowpal Wabbit as its main machine learning tool
“Our workhorse is Wopal Wabbit, up there on the right, which came out of Yahoo Research.”
Vaclav Petricek Dec 5, 2013 ▶ 12:21
Assertion Supported
eHarmony accounted for a quarter of online dating marriages from 2005-2012
“So about a quarter of all the marriages from online dating had met on eHarmony.”
Vaclav Petricek Dec 5, 2013 ▶ 17:00
Assertion Supported
One in three U.S. marriages between 2005 and 2012 began online
“One in three marriages in this period, so that was from 2005 to 2012, met online.”
Vaclav Petricek Dec 5, 2013 ▶ 17:05
Assertion Contradicted
eHarmony couples had nearly half the divorce rate of offline couples
“The couples that met on eHarmony also had a low divorce rate, so nearly half of what the marriages that met actually offline.”
Vaclav Petricek Dec 5, 2013 ▶ 17:59
Disclosure
Petricek: eHarmony has fewer than 10 people on machine learning
“Ah, so eHarmony has about 200 people in Santa Monica. We have a few offices around the world, and in terms of technology, we have about 50% tech, but still, like, a lot of people are building the site, infrastructure, and so on, so matching is about twenty-ish…”
Vaclav Petricek Dec 5, 2013 ▶ 20:03
Disclosure
Petricek: eHarmony builds localized compatibility models for each country
“What we do is whenever we were rolling out in a new country, we first redo the compatibility model. So we take a new sample, a representative sample of marriages, and we have specific models for that locale in terms of compatibility. And then on the large-scal…”
Vaclav Petricek Dec 5, 2013 ▶ 21:24
Made with StarZero

Turn any episode into a week of clips.

This entire site, over 400 conversations transcribed, diarized, checked and made playable, runs on the StarZero media pipeline. Drop in your own episode and the podcast clipper finds the moments worth sharing, cuts them, captions them, and reframes them for every feed.

Try the podcast clipper Learn more