How lean can Big Data be? Dat Tran explains how the market entry of (digital) products can be accelerated using the MVP approach

This article provides an overview of Dat Tran’s presentation at the Data Festival 2018: “What is the big in (Big) data? And how lean can your data be…”. The Head of Data Science at idealo presents lean solutions to release data products faster. His suggestions revolve around one key concept: MVP – Minimal Viable Product.

Major expectations and major disappointments

Dat Tran starts off claiming that many companies (in Germany) are not ready for the digital transformation. He explains that many corporations want to harness on the Data Science hype and therefore start implementing a lot of expensive tools, hire consultancies or build a data lab.

However, after two years of implementation, they have not actually created any value or released any product – and in the worst case, they do not even have any suitable data to transfer into their brand-new data lake.

What is the moral of the story? Three key learnings can be drawn from this situation:

  1. People at power often do not have the necessary expertise in data and therefore fail at making right decisions.
  2. Spending a lot of money on tools and consultancy does not necessarily help.
  3. These companies have a suboptimal culture that does not support agile organization and failure.

A solution to these three issues can be found in the implementation of a MVP culture.

What is a MVP culture?

MVP is the acronym for ‘minimum viable product’. Eric Ries defines it as follows:

A Minimum Viable Product (MVP) is essentially a

“version of a new product which allows a team to collect the maximum amount of validated learning about customers with the least effort.”

What does that mean in real life?

Dat Tran gives an example to illustrate the MVP approach: Imagine, you are new in Berlin and need a shelter. There are several approaches to solve this issue. On the one hand, you could indulge into the lengthy process of searching for a rental apartment.

On the other hand, you could go and buy a tent. Once you have that minimum shelter, you can expand from there on and get a container and ultimately a nice property. As you might have guessed, the second approach illustrates MVP – it is simple, minimalist and fast.

Why is MVP culture important?

Dat Tran underlines the importance of a MVP approach by presenting the case of Juicero: The company developed an (expensive) device to juice fruits and vegetables in combination with juice packages. After two years of development, they released the product to the market without previous testing.

Unfortunately, customers bought the juice packages only and simply used their hands to squeeze them. As a result, the pricy juicing device could not be sold and the company disappeared from the market.

The use of MVP would have helped to determine whether the product will be actually used by customers. The concept of MVP can be easily illustrated in a two-dimensional diagram: The ordinate axis shows the risk of market entry; the abscissa axis resembles the time span until the market entry of a product. While traditionally the product only gets released after taking up on high risk and a lot of time, in the MVP approach a lot of small experiments are being conducted in an iterative process. They lower risk and accelerate the market entry of a product.


Market entry MVP (red) versus traditional (light blue)

How does Machine Learning relate to MVP?

Dat Tran emphasizes that Machine Learning Products should be considered Data Products and respectively be treated like ‘normal’ products. ML Products are characterized by three dimensions:

Minimum Viable Modell: Start with the most basic model and expand upon it.

Example: Imagine you want to implement a classification model for churn prediction. Instead of beginning with complicated Deep Learning systems, Dat Tran suggests using the MVP approach and start with a simple statistical regression analysis. Upon this simple model, more complex ones such as a tree model can be built. Finally, it is possible to expand with simple neuronal networks and ultimately add Deep Learning to the product.

Minimum Viable Platform: Use on-premise services that are ready to use and buy use-case specific options only. On demand, add more layers.

Example: Traditionally, many companies start building huge platforms integrating a lot of tools such as Hadoop or Apache Spark. This implementation process can easily take up to three years. In that time window, neither value has been created, nor products have been released.

In contrast to this approach, Dat Tran uses on-premise services for idealo’s platforms, that offer several advantages: The services are ready to use and it is not necessary to hire additional staff to set up these systems. However, since these on-premise services are rather pricey, he suggests to buy required services for specific use-cases only. This way, the company only pays a fraction of the on-premise service price and can add additional layers on demand.

Minimum Viable (Data) Product: Release one basic product and test it. After receiving test-data, iterate and improve the product.

In terms of the Minimum Viable (Data) Product, Dat Tran shows several use-cases including chatbots and recommendation systems. For illustration, one of these use-cases is depicted: The ‘Wizard of Oz’ chatbot: It is basically a chatbot without a bot – instead of a bot, a person anticipates needs of the chatting partner. Upon these collected data, the actual, use-case specific chatbot is developed.

Digital Labs are Cargo Kult?

Recently, the integration of digital labs became very trendy. However, Dat Tran uncovers the hype around digital labs to be a Cargo Kult. Metaphorically, a Cargo Kult describes the superficial imitation of successful people’s outward actions in expectation of wealth and success. In the case of digital labs, a lot of companies build these labs to copy the behavior of successful Silicon Valley companies such as Google or Amazon. However, the success of these companies does not base on fancy digital labs. Instead, it is derived from a certain culture, that is characterized by three features:

  1. engineering culture (which means, they are built to create digital products)
  2. less hypothesis-driven experimentation (meaning they use the MVP approach) and
  3. Machine Learning is considered a product

Additionally, Dat Tran emphasizes the importance of hiring the right people.

In a nutshell: Tipps from Dat Tran

Closing with some Tipps, Dat Tran provides several pieces of advice: He emphasizes, that it is important for companies, to implement an individual (MVP) solution that works for them – there is no universally valid solution. Furthermore, he suggests to establish a real iterative MVP culture for data, in which it is allowed, to fail (fast!). Finally, he notes that hiring is a challenge – in order to get the best personnel on board, the company (sometimes) needs to change.

Dat Tran’s complete presentation about MVP culture and Machine Learning is available here..