Abstract
Most traditional market research relies on prompted and test-based methods such as surveys and focus groups. Whilst these methods have some value they are not perfect ‘one-stop’ techniques as they are only one part of the research paradigm.
The plethora of unprompted and observable online consumer data is something that can be integrated alongside these traditional methods, complimenting them and helping provide a more complete view.
However there are significant challenges in how to best extract, explore and extrapolate the insights from data of this type. Elegant, yet sophisticated solutions are needed - which is what I did with this project using TrustPilot as an example.
Case Study - ASOS
Using ASOS TrustPilot reviews as a case study..
ASOS is one of the UK’s biggest e-commerce retailers. They also have a lot of feedback on TrustPilot so they lend themselves perfectly as a good case study.
Context & Project Aims
Why does this matter to ASOS and what do we want to do?
A short introduction outlining the context of this project, what we want to achieve and why this should matter to ASOS.
Data EXTRACTION
How do we get the data?
We will extract ASOS TrustPilot reviews at scale. We will format the data to provide structure.
To do this we will build a custom web scraping class in Python that implements multi-threaded html parsing.
Information EXPLORATION
How can we distil the scraped ASOS data and convert it into useable information that will allow us to interpret and generate insights from?
We will use Natural Language Processing techniques that will allow us to make sense of the data. We can look at what is being said (words) and how it is being said (meaning).
To do this we will implement spaCy to allow us to explore part of speech analysis and we will use word2vec to explore linguistic context in review text. We will visualise the information using Plotly.
EXTRAPOLATION of Insight
What can ASOS implement that will enable us to act on this insight in a smart and timely manner?
We will build AI that can predict if a review is likely to be rated positive or negative off the basis of what people say in their review.
To do this we will build a deep learning classifier using Tensor Flow. Review text will be represented using Google’s Universal Sentence Encoder.