Identifying the signals that make a customer review helpful

An applied research project on 50,000 Amazon reviews and the signals that predict their helpfulness.

PythonXGBoostNLPCustomer insights

50,000

Reviews analysed

Variables

XGBoost

Model

The work

Challenge

Understand which parts of a product review genuinely help other customers make a decision, beyond the star rating alone.

Approach

Build a feature set combining transactional information and text characteristics, then compare their predictive contribution.

Contribution

For my thesis, I analysed 50,000 Amazon baby-product reviews, created 17 variables and trained an XGBoost model to estimate perceived review helpfulness.

Impact and learning

Outcome

Verified purchases and images emerged as the strongest signals, followed by review length.

Key learning

Customer data becomes more useful when analysis connects observable behaviours to a concrete decision — here, the trust placed in a review.

“Results from academic research conducted with public data.”

Master’s thesis, EDHEC Business School

My approach

Prepared and explored a sample of 50,000 product reviews.
Created variables connected to content, credibility and publication context.
Trained and interpreted an XGBoost model.
Translated findings into implications for marketing and e-commerce teams.

Research question

What makes an online review genuinely helpful to another customer? This matters to e-commerce, marketing and customer-insights teams because review quality affects trust and decision-making.

Method

I used public Amazon-review data for baby products. The analysis combined structured signals — such as verified purchase or image presence — with characteristics of the review text.

Result and implication

The findings show that credibility signals and visual evidence matter more than the rating alone. For a business team, this suggests better highlighting verified, illustrated and sufficiently detailed reviews.