6 - Supervised learning, classification

Question 1

Customer Relationship Analytics

Sentiment analysis

Answer 1

Sentiment analysis can offer direct insights on the firm’s relationship to (potential) customers

Also called: opinion analysis

What is new:

• automated analysis
• of large amounts of text data
• collected over the web
• with objective predictive power (risk of bias)

Question 2

Where can I find data on sentiments?

Answer 2

Traditional: surveys, interviews, written complaints

• Advantages: sentiment expressions are somewhat structured or at least targeted
• Disadvantages: limited responses, possibly sugar-coated, possibly too negative

New: “Harvesting” social networks

• Advantage: customers freely express their sentiments in large quantities
• Disadvantages: expressions are not structured, difficult to control, probably still biased

Question 3

Excursus: Collecting Data from a population

Population

Answer 3

The complete set of elements we study

• also called: sampling frame
• e.g. all people in the UK, all BMW cars, all patients going through a medical treatment, …

Question 4

Excursus: Collecting Data from a population

Census

Answer 4

Collecting data from every population member

• expensive and lengthy
• in the US: every 10 years, every household is contacted

Question 5

Excursus: Collecting Data from a population

Referendum

Answer 5

Collection of data form every population member on a voluntary basis

• expensive
• e.g.: 2016: “Should the UK remain a member of the EU or leave?”

Question 6

Collecting data from sample

Sample

Answer 6

Subset of members selected from a population

• Exhibits characteristics typical of those possessed by the population of interest
• data is collected from the sample with the objective to analyse and make inferences about the population
• sample must be “well-selected” to “well represent” the population (we are always interested in the population)

Question 7

Bias

Whats a bias?

Answer 7

Also: tendency, prejudice

can be positive or negative

Question 8

Bias

Biased sample

Answer 8

• measurements, observations or responses are likely to be unrepresentative of the population as a whole, because of the way the sample is chosen
• e.g. test Whiskas only on cats from Whiskas employees

Question 9

Bias

Analysis of a biased sample

Answer 9

• results may be very misleading

- a severely biased sample is usually worthless

Question 10

Bias

Biased sample

What does that mean for sentiment analysis?

Answer 10

• different channels can offer different messages

- people are more likely to say bad things

Question 11

Web scraping: extracting sentiments from online sources

Answer 11

• still includes manual efforts to varying extends
• can be automated via a variety of tools or code in Python or Scrapy
• efficiency depends on the fit of the search pattern
• ethically and legal “grey” area: customers do not direct their expressions toward the company, but toward an audience of peers

Question 12

Approaches to sentiment analysis

Which perspective should be taken?

Answer 12

needs to be defined beforehand

Focus

• domain
• document or sentence
• context
• aspect or holistic

Question

• quality
• reason
• intensity
• subjectivity
• seriosity (e.g. irony)
• originality (customer vs. competitor)

Question 13

What is an opinion?

Answer 13

opinion = (entity, aspect, sentiment, owner, date)

• aspects can be dresses directly or indirectly
• sentiments can be stated in a subjective or objective way
• the statement’s author is not always the sentiment’s owner
• date information can be used to trace trends