Today’s session opened with a fantastic presentation by Ravi Vijayaraghavan , VP at Flipkart.
It was amazing start to learn that Flipkart is “The Amazon of India”. It reminds me of the presentation made by hotels.com at the Databricks AI summit where there analytics and artificial intelligence are used to solve real world business problems ( eg: supply / demand forecasting, fraud detection , optimise conversion via automated image recognition).
Presentation by Ravi Vijayaraghavan , VP at Flipkart
Key takeaways :
- Flipkart is the “Amazon” of India
- Flipkart has nearly 120m customers, with over 80m product reviews
- They have a team of 200 analysts (comprising of data analysts, business analysts who build predictive models and a small core data scientists).
- Business analysts can build predictive models and have a good understanding of statistics.
- Scale of data matters and can given a lot of insightsOver 120m users they really have a detailed understanding of customers
- Mobile phones are transforming India (not laptops) – Penetration of mobile phones reigns supreme.
Scale of Flipart data platform
- Leveraging the data platform and machine learning platform at scale
- Distinguish between data assisted human decisions (eg: Operations planning/strategy) vs moData assisted machine decisions (listing quality, selection quality, fraud, personalisation)
A culture of continual experimentation : The need to continually experiment -> nearly 50+ launches a week which go through an A/B test.
Data translating to business decisions across nearly all areas including product insights, user insights, traffic, price elasticity modelling , supply chain optimisation etc,
Data science vs Measurement science
Measurement science equally as important as data science
- Understand the company strategy, identifying the areas of focus, frame it into Level 0 metrics.
- Articulate these L0 metrics then build out a statistical model for the underlying driver metrics
- Assign each of drivers to the relevant business unit head to ensure accountability
Business problem: Flipkart net promoter scores were dropping and they wanted to turn it around.
- Flipkart felt that they lost ground in net promoter scores compared to a few years ago.
- The execs and Ravi set an ambituous goal to improve net promoter scores by 10pts in a year.
- Ravi’s team created a statistical model that looked at the relative importance of each driver, its sub drivers and assigned accountability where ownership is directly attributed to each of the sub drivers.
- The key here is really assigning KPI accountabilities to the relevant business area and holding them to accountPricing index assigned to Business category
- Product category assigned to the consumer platform etc etc
Business problem : How can we help customers purchase better through the use of online reviews?
- Over 80m reviews so it cant be done manually.
- Initial simplistic algorithm was based on ‘proportion of up-votes’ and recency of review however they found this was not working well
Applying machine learning to incorporate natural language ‘sentiment’ together with other metrics to inform a more balanced view for customers
Business problem solved by machine learning : “How much do inventory should Flipkart stock of each item?
Leveraging deep learning and computer vision to Improve purchase conversion rates and reduce fraud
- Customers are sensitive to the quality of images and they dont seem to purchase/convert if its a poor quality image. Leverage computer vision to identify and classify images by extracting key features
- Auto categorisation of images that do not comply with standards (eg: images with a shadow)
- Using image recognition to correctly scan product items to find their “Real retail price”
Presentation by Yellowfin – Future of analytics departments
- As an industry, there is a need to shift to using BI to create value for the business
- Majority of companies still struggling with preparing data
- Focus on people with techniques and don’t hire for people with tools skills
- We need to run analytics like a business
Round table Panel discussion: Aligning your data strategy to support AI
- Focus on business value and not Hollywood AI.
- Need to assign a business owner to each analytics initiative.
- Focus on educating and applying the solution in the business.
- Most companies still struggling with data prep and acquisition. AI not the real problem
- Bring use cases, quick POCs , bring minimum viable product . Showcase to business and leadership teams
Presentation by Agustinus Nalwan, Head of AI at carsales.com
Data quality is critical for AI to work effectively
- Rubbish data will generate rubbish AI outcomes eg: see example of generation of baby names
Cyclops 2.0 is car-sales image recognition software built in-house using the tensorflow framework
- Incredible accuracy
- 1st version was built in a few months and 2nd version built in 5 months (eg: Spent 1 day a week for next 3 months and buildout AI )
- Key learnings:
- AI enables competition advantage however
- User interface design and the way customers use the product is critical for adoption to occur.
- Use of transfer learning enabled the model to saved a lot of computation time( 4 weeks vs 8 months)
using AI to do comparisons on cars
leveraging AI to build features – Tensorflow
remember to incorporate customer behaviour
Roundtable : Future proofing for GDPR discussion
Key learnings include:
- Impact of data breach can be significant eg: equifax., this can be used to educate your board to invest in data and analytics.
- Run a scenario with your board for data breaches and the potential costs and impact.
- Take the opportunity to re-architect IT to support the business outcomes.
- Make the effort to identify data owners within the company.
- Data Stewart/data owners can be segmented by:
- department :
- data domain (can be multiple ) – Don’t look for perfection
- Identify process owners, applications owners, data owners, and IT owners.
- Data Stewart’s also funnel things upwards via management reports. Analytics can be used to influence data Stewards. eg: inconsistent management reports
- Setup a support function- Leverage existing IT platform eg: servicenow to align tickets. Ticket can go to legal for GDPR deletion, only the end user can close the query.