Personalization at Daraz

Introduction:

Before we formally begin with our topic on Personalization@Daraz, I would like to start by introducing Daraz as a platform first, and then we will dive right in!

Daraz (Alibaba Group) has come a long way since its inception. From being a fashion marketplace to an e-commerce giant in South Asia, Daraz has always been ambitious. Currently established in over 5 countries; Pakistan, Bangladesh, Srilanka, Nepal & Myanmar, we have over ~60M assortment available and over 4M users active daily(outside of campaign months, where we average up to ~9M active users/daily for campaign period). Suffice to say, personalization is a core element of functionality in Daraz.

Breaking down the parts of the whole:

Daraz is structured to have modules that can be best understood from the imagery patched together below. Most Popular, DarazMall, Global Collection, and more are modules on the home page that can then be entered into and explored further. The items you see here are not random. These are personalised differently for each and every user based on their many characteristics and preferences. Each has a different purpose, and different behavior, but we will not get into the details of that for this article. But as a rough introduction, you may happen to click on Gaming on Most Popular, you will be greeted by a curated list of products by our stack of algorithms.??

Campaign: Personalised just for you!

Now that we understand Daraz for the most part, let's deep dive into what is at the heart of every e-commerce business in the world; Campaign. In the west, we see Black Friday sales causing havoc in-store and online, and in China, South East Asia & now in South Asia, we see the hype for 11.11 & 12.12. During the first hour of 11.11 this year, we reached over 5K queries/second on our campaign deployment engine for machine learning. A query is a new request generated every time a user enters the banner, scrolls through some amount of items and so forth.?

Inside ‘Campaign Entry’ is where the magic happens. Note that some of the Banners once entered are also personalised, and all of “Explore More Deals” or akin to that are powered by ML.

The campaign in Daraz is broken down into two parts; the operator and the algorithm. The operator decides on what a particular floor inside the banner will contain i.e categories of data points, the number of products behind that floor, the name of the floor, and so forth.?

The operator is free to apply more filters or enable personalization. If personalization is not enabled, static items are displayed. You might wonder, I only see 6 items in say, Computing & Gaming, what is the point of the operator selecting the size of datapool any greater? You may press on ‘More’ to browse the assortment, you may revisit and will be welcomed with different items.?

Recommender System @ Daraz:

Mega Deals are many, only a few are shown at first glance, and the best ones are based on our algorithm's understanding of you. Similarly, we have category floors; Electronics, Fashion & Beauty belong to that. Computing & Gaming belong to what I referred to earlier as ‘Floor’.?

An operator may configure 100 floors on the platform, and then choose to show at max 5 from personalization and may choose some to be static. The 5 will be chosen by our algorithms from 100 made available and shown to each user uniquely, and static will be shown as is.?

It doesn’t end just here, the items inside the floor obviously are customised per user too!?

The three stages of the recommendation engine over at Daraz are as follows:

• Match (Candidate Generation: user behaviour and item features bring together the initial pool of items considered relevant based on scores from different algorithms based on user navigation in real-time)?
• Rank (Match items are ranked based on business goals: should the user click? Should the user add to the cart? Should the user buy? This is based on user, item and context features )
• ReRank (Receives Rank + Match scores to further rank the list of items to be displayed)

But to say that is all that goes on in any recommender system in production would not just be wrong but an injustice. Once the items are scored, brand diversity, category diversity, and store diversity comes into play, and real-time user behavior is recorded across the platform to provide for fresh recommendations. But the art of personalization is far from being a solved problem, and it is far from being solved only by algorithms. For any company to make the black box and the explainable algorithms or the heuristics work right, it is essential that the Product is centered around the goal, understands the challenges, the front end and the back end work towards that goal also. If the user experience lacks, the recommendations could be terrifyingly good but may not convert a user to a buyer. Everything matters.?