3. Identifying Fraud Buyers
● Preempting a fraud transaction is key to success for an
ecommerce business
● There are several ways in operations to detect frauds like
○ Two factor authentication for credit card frauds
○ Address parsing for COD/RTO frauds
Inspite of all this, machine learning can prove extremely
4. Labelled Data Generation
● Labelled Data is food for supervised learning problems.
● Generally human raters are employed to generate a
labelled data set.
● Platforms such as Amazon Mechanical turks are used in
this case.
5. Feature Generation
● For each human rated transaction, we generate features
which might be a good predictor of whether that
transaction is fraudulent or not
● Some examples of features are :
○ Buyer Rating
○ # Credit Cards used by buyer
○ # previous fraudulent purchases by buyer
6. Machine Learning
● Once you have labelled data and features, we can use
classification techniques like Logistic Regression,
Random Forests to detect fraudulent users.
● Issues:
○ Imbalanced Datasets
○ Evaluation Metric: Depends on application
7. Human in the loop approach
● As a result of machine learning, humans are not
eliminated but their job is reduced.
0 0.3 0.7 1
Human Evaluation Definitely FraudDefinitely Legitimate
Fraud Probability
8. Customer Feedback
● Customer Service is one of the integral part of customer
experience for ecommerce companies. A good customer
service contributes to the brand value of the company.
● Serves two purposes:
○ Address customer grievances
○ Serve as feedback loop for the product
9. Metrics on Customer Feedback
● Explicit
○ Net Promoter Score
■ Promoters : People rating the product 9 and 10
■ Detractors: People rating the product 6 or below
■ Passives: People rating the product 7 or 8
■ NPS above 0 are considered decent and above 30-40 are
considered great
10. Data Analytics on Feedbacks
● While giving feedback, a customer writes lot of stuff in
feedback form.
● Natural Language Processing (NLP) can be used to
identify the sentiment of reviews and understand the
frequent pain points of the customers
11. Bag of Words Model
● This model can be used to identify classify the reviews into positive and
negative using supervised classification techniques
● Again, the first step here is to generate labelled data using human raters
● Removal of english stop words from the reviews
● Filtering out only the adjectives which might correspond to positive or
negative words.
● Construct a feature saying whether a particular adjective appears in the
review or not