Recommendation Engine for Channel Partners at CARS24 — an Overview
During peak hours, we at CARS24 witness more than 200 car auctions running in parallel across India, all with a shelf life of less than 30 minutes and average ticket size of ~$5,000
Recommendation engines, as many of you would know, are essentially algorithms that leverage historical user level click-stream to predict likelihood of a user to engage with a given entity.
We are frequently exposed to recommendation engines across products we use all the time from Instagram to LinkedIn to YouTube, Netflix, Spotify and more.
Recommendation algorithms vary from relatively straight forward collaborative filtering to predictive models (classification models for propensity to engage) to more advanced custom approaches using ranking algorithms (factorization machines, RankNet, Lambda-MART etc) depending upon the size of the data and maturity of ML practice in the organization. Refer this really good post by Pavel Kordík for more details.
Also, recommendation engines are not just limited to online world but have been used extensively in offline world for quite some time now — albeit under diff name and of course data collection solved differently
e.g. dunnhumby introduced loyalty cards at Tesco way back in 90s which while rewarded customers with clubcard points (driving retention, repeat purchase etc) also provided immensely rich customer mapped item level shopping data that is being used extensively for personalized pricing & promotions and cross selling.
Use case of recommendation engine at CARS24
Unlike eCommerce, social media & content platforms where key challenge is to solve for discovery across huge catalog having millions of items, at CARS24 we have a somewhat different challenge.
CARS24 C2B transactions happen through auctions — and unlike regular transaction sites where selling price is fixed, in case of auctions it is established by the competitiveness of buyers.
So, apart from the design of auction itself (leveraging principles of game theory and, as importantly, series of A/B tests), it becomes absolutely critical for us to get as many ‘relevant’ channel partners as possible in an auction for true price discovery and higher conversion & margins.
Having established the need, i.e. higher participation of relevant channel partners, let’s talk about the biggest constraint — TIME
- At CARS24, we see about ~200 auctions (increasing ~15% MoM) running in parallel during peak hours. As a brand promise, we aim to complete seller journey from check-in to payment in 2 hours, and hence live auctions are run for just about 30 min right after inspection.
- Also, as explained in my previous blog explaining ML driven pricing models at CARS24, pre-owned car is a highly ephemeral entity with each car being an SKU in itself having its unique condition parameters and pricing.
- Last but not the least, average ticket size at CARS24 is ~$5,000 and almost 15% of the transactions are above $15,000. This implies strong need for proper assessment before buyers can start bidding, refer screen shot of inspection report for reference. This assessment requires no less than 5 min even for seasoned channel partners
This ephemeral nature of pre-owned car supply clubbed with short auction time window and need for detailed assessment of the car before bidding introduces huge time sensitivity in our auctions — channel partners have only 30 min to find, assess, bid/outbid and win a car. And once the auction is over, the car, if not won, in most likelihood is lost, forever!
Even the most engaged channel partner who is on app all through the day, can assess no more than 5% of all live auctions!
Auctoris — first generation of recommendation engine at CARS24
Auctoris’ job, in a nutshell, is to ensure that the channel partner is able to discover, assess and bid on the 5% cars most relevant to him/her
On an average, our channel partners spends 3 hours daily on the auction app and spend $5,000 every week buying cars through auctions — this has to be one of the very few products of this scale (~$2 million worth of transactions everyday) with such high degree of ‘time’ as well as ‘monetary’ commitment.
The user level click-stream collected on the app is both high on quantity (# of user actions per day) and quality ($ value of the user actions). We are leveraging this rich data to establish affinities of channel partners to different attributes of cars and relative weights of those affinities using advanced predictive modelling. Affinity attributes include make-model, price band, age, fuel type, car body type (SUV vs Sedan vs Hatchback etc), nature of auction, time of day/day of week etc
This enables us to have ‘buyer — auction’ level propensity scores which power personalized catalog sorting and notifications driving faster discovery of relevant auctions for the channel partners.
Auctoris (recommendation engine at CARS24) went live in Delhi-NCR in late Feb’19 for 25% of randomly selected base and has given us very impressive results — we plan to gradually scale this Pan India by end of April’19.
In Pre-Auctoris era, auctions were sorted by ascending end time, i.e. those ending sooner will be positioned higher on the catalog. Refer graph above capturing how the clicks have got skewed towards the top of the catalog after the introduction of recommendation engine.
Almost 90% of the total ‘first clicks on an auction’ are coming in top 20 positions for test group (with catalog sorting powered by recommendation), same stat was ~60% for time based sorting.
This superior buyer experience (faster, top of the catalog discovery of relevant auctions) is driving significant improvements in the overall funnel of the test group — with almost 20% higher views and 10% higher bids & bought cars vs control group — refer funnel graph
What Next?
Auctoris is trying to solve the tough problem of real-time matching of buyers with relevant auctions which is absolutely critical to our growth plans given the high time sensitivity of our process and strong need of driving higher relevant buyer participation for auction efficiency.
This engine when clubbed with our pricing engine Profecto (and other algorithms predicting default intent of sellers and buyer’s willingness to pay ‘$x’ more for any given car — more on that in future blogs!) gives us a massive control over the outcome of an auction.
We will continue to strengthen Auctoris with more features and newer algorithms e.g. moving from 24 hour lag to near real-time attributes, upgrading from XGBoost to ranking algorithms.
Acknowledgement
I take this opportunity to thank data scientists who developed Auctoris, Senior DS Shashank Kumar and Associate DS Himanshu Dhingra for the hard work and perseverance they have shown over the last few months to bring Auctoris to life. Also, must acknowledge the effort put in by the tech team towards seamless integration of Auctoris APIs with the main auction app code base under the guidance of Senior Director Rajesh Ajmera and Engineering Manager Devender; and the efforts of product team led by Nikhil Gupta which kept us right from UX/UI perspective and ensured smooth coordination between tech and DS teams. And as always, special thanks to our Co-founders CEO Vikram and COO Mehul for their unwavering support and feedback throughout this project.
We are always seeking outstanding people to join our data science team working across complex problems like Pricing, Recommendation Engines, Auction Scheduling/Strategy, Marketing Optimization, Retail/Sales efficiency projects and Risk models for our lending vertical.
Please reach out to me directly at naresh.mehta@cars24.com or drop a mail to datascience@cars24.com or hiring@cars24.com for more details.
📝 Read this story later in Journal.
🗞 Wake up every Sunday morning to the week’s most noteworthy Tech stories, opinions, and news waiting in your inbox: Get the noteworthy newsletter >
