How will you identify fake reviews on Flipkart? [Flipkart PM Ques]

By -

Interviewer: How will you identify fake reviews on Flipkart? What will you do to decrease them?

Image source -  https://www.searchenginejournal.com/

Priyanka: To get a better understanding of the problem, I would like to seek some clarification.
Interviewer: Yes, please go ahead.

Priyanka: Firstly I would like to clarify my understanding of Flipkart, it is an e-commerce platform where sellers interact with users and sell their products through the platform. Flipkart gets a commission from each sale made through its platform. It offers other services like travel booking, grocery etc.
Interviewer: Yes.

Priyanka: Now talking about fake reviews. Reviews are users’ opinions for a particular product which help other shoppers get a better understanding of the product. It usually contains text and might contain images also. For example — If a user buys a dress from Flipkart and finds it good they post a picture of the dress and write a review to describe their experience. With e-commerce growing in popularity a lot of users rely on reviews of the product posted on the platform. Reviews are meant to build trust in the product and provide a first-hand experience for users.
Fake reviews are information stated by a user to falsely say a product is good or bad. Fake reviews can lead to a loss of trust in the product and reduce its sales. It can damage a brand’s reputation and create a negative image. It can be used to sabotage the competition.
Interviewer: Yes, that’s well defined.

Priyanka: It is important to mention that fake reviews can be positive, neutral, or negative. For example — A positive review about a bad product can make a buyer lose trust in the Flipkart platform whereas a negative review about a good product can reduce its sales.
Interviewer: That's a good observation.

Priyanka: As we know Flipkart has a lot of categories like electronics, accessories, home décor, personal care, etc. is there any particular category where we are facing this issue?
Interviewer: Yes, we are facing an issue in the electronics category.

Priyanka: Is there any particular metric that we are observing is getting impacted due to this? What I can think of is that if people see good reviews about a product they might purchase it but if it doesn’t meet their expectations, then they will return the product.
Interviewer: Yes, the return rate is high.

Priyanka: To ensure we are on the same page I would like to summarise what we have discussed till now. We talked about what fake reviews are, how they harm the platform and why is it important to do away with fake reviews. We have narrowed down our problem statement by identifying that there is an increase in fake reviews on Flipkart in the electronics category. Now we will discuss what we can do to solve this issue.
Interviewer: That’s well summarised.

Priyanka: Talking about the user journey of a buyer who would post review post-purchase:
User Journey

User journey

Ways to handle fake reviews

We can create a classification engine with if/else rules from heuristics to help us identify fake reviews.

  • Fake reviews might post links to other products
  • Fake reviews have either excellent or poor grammar
  • Fake reviews have extremely negative or positive emotion
  • Fake reviews arise from users who have not purchased the product. Currently, Flipkart does not allow users who have not purchased the product to post reviews about the product.

Image source - Flipkart review page

The moderation team will audit the reviews before they are posted on the platform. The reviews should not violate the policy laid down by the platform. Additionally, there should not be any derogatory remarks. Accounts that have previously shown some dubious behaviour will be kept in check.

Human moderation will also come into play when reviews are erroneously removed by ML algorithms due to bias present in the model.

We can build a machine learning model to classify reviews into fake and non-fake by giving a probability likelihood of each review.

Cleaning data: First we have to clean the data by removing any duplicate reviews or outliers. Duplicate reviews can be reviews posted from two different user IDs on the same product or exactly the same review posted by a user on two different products. Outlier reviews can be thought of as those whose rating deviates a lot from the average product rating.

Feature selection: We will take into account some of the important information related to the review like who is the reviewer, what is the content, and what is the product that is being reviewed.

Model building: As we need to classify the reviews into fake and genuine, so we can use logistic regression. We can set a probability score to reflect the likelihood of the review being fake/non-fake. We will use metrics like accuracy or sensitivity vs specificity to track the performance of our model.

To compare our logistic regression model, we can also build a decision tree model and see which one is performing better. If we have imbalanced data i.e. the ratio of fake vs non-fake reviews is not balanced, we should perform upsampling to balance the dataset.

We can use a combination of human moderation and machine learning to obtain the best results and remove bias arising out of machine learning models.

Metrics

After implementing the techniques that we discussed, we can track the following metrics:

  • % increase/decrease on the returns made by a user
  • % increase/decrease in the periodic sales for the products where fake reviews were identified
  • % increase/decrease in average revenue per user

To summarise, we discussed what fake reviews are and how they impact the Flipkart platform. We narrowed down our problem statement by saying that the fake reviews have increased on the Flipkart platform in the electronics category leading to an increase in the return rate. We discussed techniques like rule-based classification, human moderation and review analysis to solve the problem. We identified metrics that will help us track the impact on the platform once the fake reviews have decreased.


Interviewer: That was a great discussion. Thank you.

Comments

Post a Comment

Popular posts from this blog

Guesstimates for Product Interviews, Tips to ask clarification Questions

By -
Image
  Photo by  Antoine Dautry  on  Unsplash Guesstimates for Product Interviews, Tips to ask clarification Questions Estimate the number of Hotels | Estimate the number of users of Google Drive Estimate the number of hotels? Clarification Questions Understanding of Hotel Are we also including Airbnb, lodges or only hotels? [Only Hotels] Which geography are we talking about? [Bangalore] Overall Approach Define the area of Bangalore that will be around 600Km². Since there is a national park, a lot of lakes, forests, and wetlands so assuming that around ~50% of the area is there where a potential hotel can reside. So we are looking around an area of 300Km², we can divide this area into 3 parts according to the density, assuming a normal distribution of 25% highly dense for the hotel, 50% mid dense and the rest 25% is low dense. For highly dense that is closely around 75km², for a Km² we can have 5 hotels, that is around 375 hotels. For mid dense(150Km²), on an average, we have around 2.5Hote
Click here to Read »

System Design Concepts for Product Managers — Part 1

By -
Image
This is a system design series for Product Managers who should have a high-level understanding of System Design. We will first cover the basic concepts of System Design. This is part 1, and stay tuned for future updates. 1. Load Balancer What is a Load Balancer? Load Balancer:  Load Balancer is a common term used widely when you are working on an application tha t  serves enormous traffic. Let’s understand with an example: Assume that Netflix has deployed several servers to handle the number of client requests. When we hit “Netflix.com”, how does Netflix know which server it has routed my request to? This is where the load balancer comes into the picture; load Balancers distribute the incoming traffic from the client efficiently across different servers and make sure that no server is getting overheated or overloaded. Reliability and availability are maintained by redirecting the incoming request to the available servers. Credits:  https://codeburst.io/ Load Balancing Algorithms The lo
Click here to Read »

Design Uber for Kids. [Product Management Interview Question]

By -
Image
Image credits - https://gbksoft.com/blog Interviewer: Design Uber for kids Priyanka: Before we proceed, I would like to clarify my understanding of Uber. So, Uber is a digital ride-hailing platform that connects riders with drivers. The passengers can request a ride on the platform, and the driver can fulfil the demand by dropping them at the desired location. Uber acts as a facilitator between the two and takes a cut out of each ride that happens through its platform. Besides this, Uber also has business verticals like Uber Eats, Uber Freight, etc. Is my understanding correct? Interviewer: Yes. Priyanka: Alright, to understand the problem better, I would like to ask some clarifying questions. When we talk about kids, I am assuming we are referring to an age group of 10–16 years who can travel alone. Also, I am assuming that parents will be booking the rides for their kids. Interviewer: That’s a fair assumption Priyanka: Also, we are catering to the age group of 10–16, so there w
Click here to Read »