When we established a machine learning team for our product, Lexio, at Narrative Science last year, none of us guessed that not only would we not have a dedicated ML (machine learning) team today, but would be actively evangelizing that fact. That’s exactly what I’m doing here, though, despite an increase in active and planned work employing ML techniques.
So why would we eliminate our ML team when we have plenty of ML work to do? To be clear, we still have a group that owns ML projects at Narrative Science, but we have rebranded ourselves as the “data intelligence team”. This isn’t doublespeak, but instead reflects a fundamental reevaluation of the team’s goals and the role of machine learning at NS.
Our error – and one I hope this post can help others avoid – was to confuse a tool, machine learning, with a business goal. Sexy new ML techniques are exciting and can often be powerful, but organizations must be cautious not to deploy them for their own sake. Everything starts to look like a nail, after all, when you have a shiny new hammer. By defining our team in terms of a tool it would employ, rather than the business goals it aimed to achieve, we were unnecessarily limiting ourselves to particular technical approaches and diverting our focus away from delivering customer value.
Reviewing roadmaps at the team and product level, it became clear that what tied our team together was not the use of particular ML techniques, but the goal of intelligently leveraging data to provide delightful customer experiences, with or without the use of ML. Hence the name “data intelligence” was born.
The shift from an ML team to a data intelligence team is, on one level, superficial. Nothing fundamental has changed about our short-term initiatives, and we still work on ML applications. What is different, however, is that we have an improved framework for building out our long-term roadmap and communicating the team’s goals to stakeholders. Instead of initiatives being defined by the technologies used, we can more closely couple them to customer value. Specifically, we focus on four areas in which intelligently leveraging data can improve the user experiences in our product Lexio:
- Content Discovery – Make it easier for customers to find the content most useful, interesting, and relevant at any given time.
- Surfacing Insights – Use advanced analysis on customer data to extract valuable insights and hidden gems.
- Improved Communication – Tailor the language and communication style to better fit with the readers’ preferences.
- Expanding Lexio’s Knowledge – Expand what Lexio “knows” to enable expanding into more domains easily, streamline user onboarding by providing better defaults and guesses, and give it a better understanding on how it should analyze customer data.
What ties these four areas together is not a particular class of technology, but instead that they all require leveraging customer data, interaction patterns, or external data sources to improve our product. In fact, we have worked on our planned development of both ML and non-ML approaches to all of these areas. Take surfacing insights, for example: Here, some of the most valuable content generated by our team’s efforts has been simple analysis surfacing movement of opportunities through a sales pipeline for our Salesforce customers. This involved zero ML work, but provided immediate value to customers (e.g. “You had three opportunities progress from stage X to stage Y last week”). At the time we implemented the feature, there was discussion of whether or not this was “really ML” (it wasn’t), but my point here is that that was the wrong question to be asking, and only comes up when building a team around a technology instead of directly delivering positive experiences to customers.
As a second example, consider content discovery. While we are in the design phase of a full-fledged, ML-backed recommendation engine for Lexio story content, it has become apparent that there is a great deal of low-hanging fruit using simple, popularity-based techniques (e.g. surfacing the most popular stories amongst an organization’s users). If we remained committed to being “the ML team”, there might be questions as to whether this simpler work was really in our purview, but for a data intelligence team the line between the simpler approaches and “real” ML is irrelevant. Our mandate is to leverage data to improve our product. Full stop.
The path forward
ML remains core to the work we’re doing at Narrative Science, and will only increase in importance as our user base grows and we have more data to work with. Forecasting and advanced time series analysis, recommendation engines, natural language understanding, and more are all on the horizon. But we must avoid being seduced by technology for technology’s sake. ML is a tool in our toolbox; a powerful and exciting tool, yes, but only a tool.