Top Banner
msg.ai Learning customer service through topic modeling of tweets Peter Frick
12

Rank my tweets

Feb 18, 2017

Download

Data & Analytics

Peter Frick
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: Rank my tweets

msg.aiLearning customer service through

topic modeling of tweetsPeter Frick

Page 2: Rank my tweets

Can machine learning improve social customer interactions?

Page 3: Rank my tweets

Leveraging LDA to learn tweet topicsAirline

tweets

Page 4: Rank my tweets

Leveraging LDA to learn tweet topics

Terrible!! The worst flight ever! The pretzels and peanuts were bad

Airline tweets

Page 5: Rank my tweets

Leveraging LDA to learn tweet topics

Terrible!! The worst flight ever! The pretzels and peanuts were bad

topic3

topic1

topic2

Latent Dirichlet

Allocation (LDA)

Airline tweets

WorstFailTerrible

Page 6: Rank my tweets

Leveraging LDA to learn tweet topics

Terrible!! The worst flight ever! The pretzels and peanuts were bad

Airline tweets

Latent Dirichlet

Allocation (LDA)

topic3

topic1

topic2

WorstFailTerrible

Page 8: Rank my tweets

Using topics to estimate airline customer service

Topic 2

Tweets passing filter

Ranking airlines by Topic 2“bad customer service”

Page 9: Rank my tweets

Using topics to estimate airline customer service

Ranking airlines by relativebad customer service

Page 10: Rank my tweets

Predicted customer service matches recently published results

Source: 2015 J.D. Power North America AirlineSatisfaction Traditional Airline Rankings

J.D. Power rankings Ranking airlines by relativebad customer service

Page 11: Rank my tweets

PredictionModel input

020

4060

8010

0

Varying µ−2

02

46

Popu

latio

n do

ublin

gs

DIP rate(doublings/h)

Time in erl (d)−0.04 0.00 0.04

020

4060

8010

0

0 5 10 15 20 25 30

−20

24

6Dens

ity 2µ

µ

0.5µ

0.5σ

σ

2σVarying σ

PredictionModel input

020

4060

8010

0

Varying µ

−20

24

6Po

pula

tion

doub

lings

DIP rate(doublings/h)

Time in erl (d)−0.04 0.00 0.04

020

4060

8010

0

0 5 10 15 20 25 30

−20

24

6Dens

ity 2µ

µ

0.5µ

0.5σ

σ

2σVarying σ

Peter Frick

Page 12: Rank my tweets

Data pipeline

Msg.ai tweets

LDA  ModelCleaned  data

Pandas

NLTK

Topic  visualization

Kaggle tweets

pyLDAvis

Topic  Prediction