Introducing Tableau Business Science

AI projects need data expertise and domain expertise to succeed

Despite the fears and misconceptions that AI will replace people, there are too many business problems where the smartest machines are no match for human judgement. For example, a sales organisation might use predictive modelling to determine the most profitable upsell/cross-sell opportunities. An algorithm can offer predictions about the likelihood that a customer will buy, but wouldn’t have the critical knowledge of someone managing the business relationship. An account executive would understand the customer’s goals and which products might help them achieve those goals, or may have learned through past experiences that a certain solution is unlikely to reach a consensus to buy.

These kinds of nuances require human judgement beyond what an algorithm alone can offer. Together, human expertise, judgement, and contextual awareness combined with the rigour, automation and scalability of machine-generated insights drive better business outcomes. As another example, a retailer might want to explore which new products to add to its stores to increase profits in a certain region. A business professional understands how factors like supplier relationships, regional trends and other impactful, qualitative implications could affect the decision – details that a machine might never understand or apply successfully.

Organisations spend a lot of effort trying to recruit “unicorns” that have both data science and domain expertise, but people that fit this description are incredibly rare. There are a lot of mission-critical problems where data science techniques are the most important consideration, but Business Science brings many of these techniques to more people, so they too can leverage predictions in their analysis without requiring an advanced, technical degree.

Business Science is about matching the method and the expert to the problem. When data science teams take on new projects for the business, they often struggle from a lack of context and domain knowledge. Without the right context, data science teams spend a lot of time trying to identify the right data to address the problem, then curating, cleaning and preparing it for analysis. As AI and ML rely heavily on data, this is fundamentally a data problem, so it’s important to have a good match between the problem they’re trying to solve and the right data – in the right shape – to do so.

The proximity of domain experts to the unique problems their organisations face, and their specialised knowledge about their line of business data, gives them an advantage in developing, validating and deploying predictive models to be more timely, efficient and impactful.

An iterative revise-and-redeploy approach saves time, reduces costs and elevates unique skills

Traditional data science and ML cycles can be long and the “last mile” is often the most challenging and resource-heavy in terms of time, effort and cost. Deploying and integrating traditional custom models is complex, requiring statisticians and data scientists to make solutions consumable and actionable by end users. When time to market is a major factor, empowering people to create and iterate on predictive models in a matter of days and weeks – not months – maximises resources and helps to deliver a better return on investment.

For business experts who need to lean on data and predictions to inform decision-making, it’s incredibly valuable to be able to control various aspects throughout the model-building process and have the appropriate confidence in their data relative to each use case. This is especially worthwhile when the model-building approach is iterative and focused on continuous improvement of prediction accuracy. Such an approach helps people take action while the data is still relevant. And sometimes, even attempting to build a model can cause a team to prove the data to the point that they’ve achieved some value.

With Business Science, the objective is to move a KPI – not to perfect a model until it’s the most precise. By facilitating a more iterative, revise-and-redeploy process than that of traditional data science cycles, Business Science eliminates barriers for people with business context to quickly build models and apply predictions. More people can get value faster from these advanced analytical techniques and make smarter decisions when and where they need to – potentially leading to major cost savings or profits for a specific use case.

Think about assigning sales quotas: geographical assignments and targets change all the time. The same model for assigning quotas one year won’t work for the following year, because the inputs to their model are going to continuously evolve. And what if a regional sales lead vetoes the suggested model because they have a goal to do even more business in their geography? This makes the ability to quickly pivot or iterate on the model really valuable, and potentially result in new opportunities to generate more revenue in these sales territories.

By enabling analysts and business users with the ability to safely self-serve with machine learning, Business Science elevates the unique skills and domain expertise of more people. It also reinforces the integral role data science teams play by freeing them up to focus efforts on large-scale, mission-critical projects. This class of AI-powered solutions can also create an experimental environment for analysts and advanced business users to explore new use cases that aren’t priorities for data science teams. With more people examining larger numbers of use cases and analysing drivers of potential outcomes, the likelihood of delivering a successful model increases. Domain experts are empowered to build no-code, ML-driven models for their needs, producing excellent, business-specific analytics that don’t require the level of depth that data professionals go into.

It’s important to note that while Business Science solutions can help address the skills gap between analysts and data scientists, it is not a replacement for them. Data science professionals will continue to provide custom models, statistical analysis, etc., but more often, they will partner with business experts to validate the data being used in ML-powered models. This increased cross-team collaboration remains critical to the success and performance of these solutions.

Democratising data science responsibly requires guidance to identify and mitigate bias and unethical use

Expanding AI-powered analytics to more people has many benefits, but also some potential dangers. Bias is inherent in data and technologies, and unless mitigated, can perpetuate bias through predictions and recommendations – even causing harm. Facial recognition technology is a well-known example of a confluence of failure points, where biased training data, technology and predictions have created harmful outcomes for communities of colour. (This is one of the reasons  why we don’t allow facial recognition technology in our  acceptable use policy.)

The average business professional isn’t necessarily thinking about these things in their analysis. As a start, robust model documentation can provide more transparency behind predictions and recommendations and make things easier to trace. Understanding what contributed to the prediction, what columns in the data and where the drivers are can help identify potential bias in your data sets and models.

Further guidance might look like built-in bias detection, model monitoring and natural-language explanations of the variables influencing predictions, helping people create more ethical models. With  Einstein Discovery, the product asks users to select variables and sensitive fields that could possibly include bias – like race, age, marital status and residence location. Through proxy detection, it will also look for correlated fields in the data set, where there is potential for bias if one field, but not the other, were removed. Another important metric is disparate impact: for example, seeing if the outcome of the model differs across genders or different ZIP codes. For many regulated industries, this is one metric used to hold machine learning models accountable.

The demand for data skills is only growing, and with it, tremendous opportunity to help people and organisations learn and solve problems more effectively. While some approaches may focus on developing or implementing completely automated AI solutions, this limited approach will likely fail to deliver the value it’s intended to provide. Discounting and removing human ingenuity, contextual awareness and expertise when democratising AI and data science techniques isn’t just counterproductive, it’s irresponsible.

Tableau has always believed in the power of human potential. Unleashing human potential is a part of Tableau DNA, having been built into our industry-leading visual analytics platform. Now, with Tableau Business Science, this new class of innovation uses AI and ML to continue our tradition of empowering people with the right tools and functionality to help them ask and answer questions, uncover insights and solve problems. With the introduction of Einstein Discovery in Tableau, a Business Science solution, we’re safely putting machine learning into the hands of more people, democratising powerful data science techniques to help people make better, faster decisions.