Amalgam Insights attended Enterprise Data World, a conference focused on data management, in late March. Though the conference tracks covered a wide variety of data practices, our primary interest was in the sessions on the AI and Machine Learning track. We came away with the impression that the data management world is starting to understand and support some of the challenges that organizations face when trying to get complex data initiatives off the ground, but that the learning process will continue to have growing pains.
Data Strategy Bootcamp
I began my time at Enterprise Data World with the Data Strategy Bootcamp on Monday. Often, organizations focus on getting smaller data projects done quickly in a tactical fashion at the expense of consciously developing their broader data strategy. The bootcamp addressed how to incorporate these “quick wins” into the bigger picture, and delved into the details of what a data strategy should include, and what does the process of building one look like. For people in data analytics and data scientist roles, understanding and contributing to your organization’s data strategy is important because well-documented and properly-managed data means data analysts and data scientists can spend more of their time doing analytics and building machine learning models. The “data scientists spend 80% of their time cleaning and preparing data” number continues to circulate without measurable improvement. To build a successful data strategy, organizations will need to identify business goals that are data-centric to align the organization’s data strategy with its business strategy, assess the organization’s maturity and capabilities across its data ecosystem, and determine long-term goals and “quick wins” that will provide measurable progress towards those goals.
Getting Started with Data Science, Machine Learning, and Artificial Intelligence Initiatives
Actually getting started on data science, machine learning, and artificial intelligence initiatives remains a point of confusion for many organizations looking to expand beyond the basic data analytics they’re currently doing. Both Kristin Serafin and Lizzie Westin of FINRA and Vinay Seth Mohta of Manifold led sessions discussing how to turn talk about machine learning and artificial intelligence into action in your organizations, and how to do so in a way that can scale up quickly. Key takeaways: your organization needs to understand its data to understand what questions it wants answered that require a machine learning approach; it needs to understand what tools are necessary to move forward; it needs to understand who already has pertinent data capabilities within the organization, and who is best positioned to improve their skills in the necessary manner; and you need to obtain buy-in from relevant stakeholders.
Data Job Roles
Data job roles were discussed in multiple sessions; I attended one from the perspective of how analytical jobs themselves are evolving, and one from the perspective of analytical career development. Despite the hype, not everyone is a data scientist, even if they may perform some tasks that are part of a data science pipeline! Data engineers are the difference between data scientists’ experiments sitting in silos and getting them into production where they can affect your company. Data analysts aren’t going anywhere – yet. (Though Michael Stonebraker, in his keynote Tuesday morning, stated that he believed data science would eventually replace BI, pending upskilling a sufficient number of data workers.) And data scientists spend 80% of their time doing data prep instead of building machine learning models; they’d like to do more of the latter, and because they’re an expensive asset, the business needs them to be doing less prep and more building as well.
By the same token, there are so many different specialties across the data environment, and the tool landscape is incredibly large. No one will know everything; even relatively low-level people will need to provide leadership in their particular roles to bridge the much-bemoaned gap between IT and Business. So how can data people do that? They’ll need to learn to talk about their initiatives and accomplishments in business terms – increasing revenue, decreasing cost, managing risk. By doing this, data strategy can be tied to business strategy, and this barrier to success can be surmounted.
Data Integration at Scale
Michael Stonebraker’s keynote highlighted the growing need for people with data science capabilities, but the real meat of his talk centered around how to support complex data science initiatives: doing data integration at scale. One example: General Electric’s procurement system problem. Clearly, the ideal number of procurement systems in any company is “one.” Given mergers and acquisitions, over time, GE had accumulated *75* procurement systems. They could save $100M if they could bring together all of these systems, with all of the information on the terms and conditions negotiated with each vendor via each of these systems. But this required a rather complex data integration process. Once that was done, the same process remained for dealing with their supplier databases, and their customer databases, and a whole host of other data. Machine learning can help with this – once there are sufficient people with machine learning skills to address these large problems. But doing data integration at scale will remain a significant challenge for enterprises for now, with machine learning skills being relatively costly and rare, data accumulation continuing to grow exponentially, and bringing in third-party data to supplement existing analyses..
Knowledge Graphs and Semantic AI
A number of sessions discussed knowledge graphs and their importance for supporting both data management and data science tasks. Knowledge graphs provide a “semantic” layer over standard relational databases – they prioritize documenting the relationships between entities, making it easier to understand how different parts of your organization’s data are interrelated. Because having a knowledge graph about your organization’s data provides natural-language context around data relationships, it can make machine learning models based on that data more “explainable” due to the additional human-legible information available for interpretation and understanding. Another example: if you’re trying to perform a search, most results rely on exact matches. Having a knowledge graph makes it simple to pull up “related” results based on the relationships documented in that knowledge graph.
Data Access, Control, and Usage
My big takeaway from Scott Taylor’s Data Architecture session: data should be a shared, centralized asset for your entire organization; it must be 1) accessible by its consumers 2) in the format they require 3) via the method they require 4) if they have permission to access it (security) 5) and they will use it in a way that abides by governance standards and laws. Data scientists care about this because they need data to do their job, and any hurdle in accessing usable data makes it more likely they’ll avoid using official methods to access the data. Nobody has three months to wait for a data requisition from IT’s data warehouses to be turned around anymore; instead, “I’ll just use this data copy on my desktop” – or more likely these days, in a cloud-hosted data silo. Making centralized access easy to use makes data users much more likely to comply with data usage and access policies, which helps secure data properly, govern its use appropriately, and prevent data silos from forming.
Digging a bit more into the security and governance aspects mentioned above, it’s surprisingly easy to identify individuals in a set of anonymized data. In separate presentations, Matt Vogt of Immuta demonstrated this with a dataset consisting of anonymized NYC taxi data, even as more and more information was redacted from it. Jeff Jonas of Senzing’s keynote took this further – as context accumulates around data, it gets easier to make inferences, even when your data is far from clean. With GDPR on the table, and CCPA coming into effect in nine months, how data workers can use data, ethically and legally, will shift, significantly affecting data workflows. Both the use of data and the results provided by black-box machine learning models will be challenged.
Recommendations
Data scientists and machine learning practitioners should familiarize themselves with the broader data management ecosystem. Said practitioners understand why dirty data is problematic, given that they spend most of their work hours cleaning that data so they can do the actual machine learning model-building, but there are numerous tools available to help with this process, and possibly obviate the need for a particular cleaning job that’s already been done once. As enterprise data catalogs become more common, this will prevent data scientists from spending hours on duplicative work when someone else has already cleaned the set they were planning to use and made it available for the organization’s use.
Data scientists and data science managers should also learn how to communicate the business value of their data initiatives when speaking to business stakeholders. From a technical point of view, making a model more accurate is an achievement in and of itself. But knowing what it means from a business standpoint builds understanding of what that improved accuracy or speed means for the business as a whole. Maybe your 1% improvement in model accuracy means you save your company tens of thousands of dollars by more accurately targeting potential customers who are ready to buy your product – that’s what will get the attention of your line-of-business partners.
Data science directors and Chief Data or Chief Analytics Officers should approach building their organization’s data strategy and culture with the long-term view in mind. Aligning your data strategy with the organization’s business strategy is crucial to your organization’s success. Rather than having both departments tugging on opposite ends of the rope going in different directions, develop an understanding of each others’ needs and capabilities and apply that knowledge to keep everyone focused on the same goal.
Chief Data Officers and Chief Analytics Officers should understand their organization’s capabilities by conducting an assessment both of their data capabilities and capacity available by individual, and to assess the general maturity in each data practice area (such as Master Data Management, Data Integration, Data Architecture, etc.). Knowing the availability of both technical and people-based resources is necessary to develop a scalable set of data processes for your organization with consistent results no matter who the data scientist or analyst is in charge of executing on the process for any given project.
As part of developing their organization’s data strategy, Chief Data Officers and Chief Analytics Officers must work with their legal department to develop rules and processes for accumulating, storing, accessing, and using data appropriately. As laws like GDPR and the California Privacy Act start being enforced, data access and usage will be much more scrutinized; companies not adhering to the letters of those laws will find themselves fined heavily. Data scientists and data science managers who are working on projects that involve sensitive or personal data should talk to their general counsel to ensure they remain on the right side of the law.