The Translator & Interpreter Role in Analytics

Analytics has seen significant adoption and business sponsorship in the past decade. Data Scientists are being hired to build analytics models—largely machine learning and, increasingly, deep learning—capable of turning vast amounts of data into insights.

Business leaders continue to be invested in the concepts, but are not without concerns around cost and results. The need for integration of analytics solutions in business can’t be overstated as a way to drive end value.

Success with analytics requires data scientists and an upstream team of data engineers, data architects and downstream teams of software developers working in agile scrums to implement their frameworks and solutions.

As a result of this dynamic, an analytics initiative can easily take on its own life and become a significant cost center. It is therefore critical to ensure that analytics initiatives consistently show business value and business outcomes are attributed to these specific initiatives.

This is where a need for a more operational role emerges to help shepherd these initiatives to implementation — the analytics translator and interpreter.

Defining ATI

The distinction between translation and interpretation can be fuzzy in an analytics effort.

Loosely, translation is related to explaining the data science process, corresponding algorithms with an ability to break down questions from data scientists into business context,

Interpretation, however, entails explaining lifts, predictions, attributions and risks with context to business operations.

Whether these are separate roles based on employee interests and competencies or a Data Science leader who wears this hat is a question specific to organizations. For the scope of this article, we’ll refer to the role as an ‘Analytic Translator & Interpreter’ [ATI].

An ATI needs strong business acumen, some technical knowledge, and project management and delivery chops. The following skills are warranted for the role:

Understand use of analytics in the industry and the business value proposition
Be able to quickly prioritize use cases based on business understanding and end customer value
Define business problems, target variables, feature definitions and be able to explain the potential solution
Methods to evaluate outcome and dilution once solution is implemented
Test based development; embedding analytics solutions in the business and overcoming implementation difficulties

An Indicative Use Case

Let’s evaluate the need for an ATI through a indicative use case:

Envision the ask from a leader in your organization.

‘We need a 5-star rating system to improve transparency and ease of doing business for our customers.’

A team of data scientists plans to use historical transactions, customer interactions, ratings and payments data as predictors to build a boosted tree model with a multi-class classification output corresponding to the 5 star ratings. The 5 class ratings are extracted from the marketplace website for a small number of businesses who have received ratings from their associates.

At the onset of the project the ATI needs to provide direction and seek syndication of scoping questions such as:

What data to use and to exclude? What is the time value of data? How do we create samples and weigh them?
Who/ What are good test candidates and what is acceptable test case results?
What is the risk of misclassification? What type of misclassification trade offs are acceptable? What is acceptable- a 5 star misclassified as a 1 star or vice versa?
How do we choose labels? How should we functionally deal with data gaps and imbalance?
What quality metrics should be used? In a multiclass classification problem it is common to use log loss as a quality metric, but does business understand it?

Once the initial iteration of model development is complete the ATI then has to think about effective business integration:

Why is the machine learning model acceptable? What is the value of the quality metrics and what does it mean in layman’s term. For example how much better is 0.89 log loss compared to a flip of a coin or a random pick?
A model might be built with about 1-3% of all data available. Does this scale extrapolate well to the entire business? How do we handle business entities who are new and don’t have adequate data?
What dilution should we expect due to operational departures from the assumptions of the model and attribution to other factors? Data gaps and sufficiency, possible overfiting/ underfiting, time value of data, acquired versus organic data, reliability of data all have a great deal of role to play on this one.
How will the machine learning model be implemented in production? The ‘how’ usually determines production data capture for scoring, refresh rates of model and upstream instrumentation for data, API deployment of scoring models. For example, if new review/ ratings information is downloaded once a week, the data needs to be downloaded, model scored and tested with a tight turn around for refreshing the ratings.
Most important- what analysis and data needs to be shown to non-technical counterparts in sales, marketing, content development and product management for them to trust and champion the solution.

Ultimately a machine learning model is a math formula and how it behaves depends on the robustness of implementation and business acumen of the ATI.

Find ways of answering key questions while attributing to other environmental factors that may impact the outcome – How does the solution contribute to the top and bottom line?

ATI’s Versus Business Analysts

At the long last one might ask the question- ‘Isn’t this just a glorified business analyst?” While some business analysts may be uberskilled, this question may not be straight forward to answer. Usually business analysts are aligned with application development and more focused on UX/UI and customer specific functional requirements.

Unlike many business analytics, ATIs have a direct line of sight into the business and are embedded in the business process. In general the ATI needs to be skilled at obfuscating technical complexities around data, implementation pain points and algorithmic complexities while continuously delivering business value through analytics. ATI’s not only respect the functions of the data scientists, but also understand the needs of decision-makers; therefore, successful ATIs are typically respected by those entities in return.

About Chisel Analytics

Chisel provides end-to-end solutions for your data and analytics programs. Looking to escalate your digital and analytics transformation efforts? Contact us today and find the experts you need to get started.