GO-JEK: Improving Customer Experience with Machine Learning
Scaling machine learning in a super-app and its challenges.
GO-JEK is a technology start-up from Indonesia that offers 18+ services under its umbrella. Established in 2010 as a motorcycle ride-hailing phone services, GO-JEK had evolved to a “super-app” status with services that includes transportation, logistics, mobile payments, food delivery and other on-demand services[1].
GO-JEK has experienced a hyper-growth of 6600x in the past 3 years, managing over 3 million orders and 2.5 million customers daily today[2]. It has more than 1 million drivers on its GO-RIDE and over 200K merchants on its GO-FOOD platforms[3]. The sheer scale that the company is operating gives rise to the importance of utilizing machine learning to enhance business efficiency and improve customer experience.
Machine Learning at GO-JEK
Machine learning plays a key role in helping make decisions real-time at GO-JEK. The company also looks at machine learning to better understand its customers, especially in the context of its recent international expansion strategy; GO-JEK wants to understand the unique needs of customers in different markets in order to offer better products to its customers.
Driver Allocation
Taking driver allocation as an example, the ultimate goal is to improve customer experience when requesting for a ride using the GO-RIDE service. One indicator to track this is the rate of bookings completed. GO-JEK wants to minimize the time a customer has to wait until his requested ride arrived and maximize driver utilization at the same time. Thus, an algorithm is built based on a predetermined set of variables to spit out the best possible driver allocation for the requested ride. This is also coupled with the dynamic pricing model that informs drivers on which areas to service[4]. However, the current approach is “passive”, as it relies on real-time data inputs to churn out an outcome. A customer requesting for a ride in a bad weather condition could be waiting for more than 30 minutes for the ride to come, because of a lack of drivers in the area and the need to call on drivers from other areas to come and made it through traffic. The management action plan going forward is to keep refining the current model so that it could predict demand more accurately based on input features such as weather, time and region to inform drivers in advance on where the rides requests would be. This way, the customer can have a reduced waiting time because there is already enough supply of drivers in the area.
Home Screen Personalization
GO-JEK’s massive customer base of over 25 million monthly active users[5] uses the app for different purposes. Some used the app for their daily morning commute to work, some used it to order food on Friday nights, some are merchants who used its courier service, and the list goes on. Therefore, it is important for GO-JEK to personalize its home screen to improve customer experience by showing a list of customized recommendations for each customer, rather than having a long feed for everyone. The company applies machine learning in this scenario by taking recorded data that depicts unique user behavior as inputs to precompute a list of personalized recommendations and display them on the user’s home screen each time he opens the app.
Challenges and Recommendations
The context of operating in emerging markets like Indonesia provides its own unique sets of challenges in scaling up machine learning at GO-JEK. The lack of local talents in the fields of data science has forced the company to maximize its small team of data scientists’ time to take on more projects, build more models and spend less time in data sourcing. As the company expands internationally, I would recommend that the strategy going forward in terms of human resources in scaling up machine learning is to figure out how to increase productivity in its current data scientists, as opposed to hiring more data scientists. Furthermore, building a solid data foundation and ensuring standardization and consistency in data captured would be key for machine learning to work effectively at GO-JEK.
Open Question
- Successfully scaling machine learning in a company like GO-JEK requires getting much more than just the technology/algorithm right. What would be other important considerations in terms of organization and process design?
(Word Count: 800 words)
Sources
[1] https://www.go-jek.com/about/
[2] Rayi Noormega, “From 5000 Orders to 3M Orders Per Day”, August 1 2018, https://medium.com/life-at-go-jek/from-5-000-orders-to-3m-orders-per-day-65c65b52041a, accessed November 2018; Shobhit Srivastava, “The story of a team focused on eventually automating everything for GO-JEK”, October 29 2018, https://blog.gojekengineering.com/the-story-of-an-internal-team-focused-on-automating-everything-eventually-9036d37e203a, accessed November 2018.
[3] Adithya Venkatesan, “How GO-JEK manages 1 million drivers with 12 engineers (Part 1)”, May 23 2018, https://blog.gojekengineering.com/how-go-jek-manages-1-million-drivers-with-12-engineers-part-1-978af9ccfd3, accessed November 2018.
[4] Google Cloud Platform, “Lessons Learned Scaling Machine Learning at Go-Jek on Google Cloud (Cloud Next ’18)”, July 26 2018, https://www.youtube.com/watch?v=57Q07-SIcm4, accessed November 2018.
[5] Reuters, “Indonesia’s Go-Jek close to profits in all segments except transport”, August 17 2018, https://www.reuters.com/article/us-indonesia-gojek-interview/indonesias-go-jek-close-to-profits-in-all-segments-except-transport-ceo-idUSKBN1L20SI, accessed November 2018.
Great read, Melina. I think the major consideration for me is understanding the underlying data that the recommendations are being built on, and how that relates to new verticals or markets that GO-JEK might expand into. Currently, there is likely a lot of data that GO-JEK just doesn’t collect on its user that impacts their ride experience and their overall satisfaction – maybe they had a bad mood going into the ride, or maybe they just weren’t looking when the scooter drove by. None of these variables are factored into the algorithm, but impact the experience that a user has in a way that does impact what the algorithm is trying to solve for. As GO-JEK expands into new markets, they will always have the challenge of building that new data set from the ground up so that they can train their new algorithm.
Thank you Melina. So fascinating to read about a new “super app” and all the many things it can offer consumers (from pizza to mopeds!). With the huge growth in number of consumers it is clear that they have a lot of data they can use machine learning for to help improve things like consumer interface. I wonder what else machine learning could improve within the website? Perhaps more importantly, as someone who hasn’t used the app, I also wonder whether all 18+ activities are useful to consumers (great if so!) and also whether they could add any more and expand further to become the “jack of all trades”. How could machine learning help them make these decisions and priorities certain categories?
Thanks, Melina! Really interesting read. To your point, the algorithm itself might be quite scalable as the company expands, but it is the other factors of the business that could prove more tricky. Yes, there will be variations in preferences in different markets, but a good machine learning algorithm should be able to take a series of inputs and develop a similar pattern. The potentially more challenging issues could be things like business development, partner management (for the restaurants for example), and even regulatory hurdles depending on the market. Maintaining a lean development team, as you suggest, and potentially focusing additional resources on building strategic partnerships and momentum in new markets would likely be a wise growth strategy.
Great article Melina! I’m impressed how GO-JEK was able to leverage machine learning to drive its exponential growth. Being able to make real-time decisions allocating driver and routes certainly improves customer experience and allows the company to scale (6600x in three years is insane!). Your concern on the talent recruiting in machine learning in Indonesia reminds me of the same issue in China. Many machine learning or AI experts are highly sought-after and a lot of them were starting their own startups, given the bubble in the industry. I agree improving data scientists’ productivity could be a very good solutions! I wonder as the industry matures, would it be possible to purchase machine learning or AI as a service from other SaaS companies?