Data Science Communication

• Effective communication within data science and machine learning is crucial

Importance of Communication

• Communication skills relate to job performance and is valid in industry and in academic settings.
• Your perceived value by an organization depends on communication.
• It is important to create a portfolio of analyses and products to showcase your expertise.

Why Communication Is Challenging

• Stakeholders often have varied expertise levels.
• Conflicting priorities or goals may exist among stakeholders.
• Stakeholders often have limited time to absorb detailed data analysis.
• Audiences may lack comfort interpreting complex datasets or visualizations.

Stakeholder Identification

• Primary stakeholders are directly impacted by your analysis
• Examples of primary stakeholders includes customers, internal teams, or decision-makers.
• Secondary stakeholders have less direct interest but are still impacted
• Examples of secondary stakeholders includes regulators, investors, or the community where your project has an impact.
• Non-obvious stakeholders may be less visible but still affected e.g., customer service teams when analyzing customer data.
• The stakeholder landscape evolves, requiring regular reassessment.

Managing Stakeholders

• Establish proactive communication to build relationships.
• Understand stakeholder priorities, concerns, and goals that data analysis can address.
• Adjust the detail level of reporting to match the needs of certain stakeholders.
• Involve stakeholders early to gather input and improve the data science process.
• You should be transparent about timelines, data limitations, and insight scope when managing expectations.

Domain Expertise

• Speaking the language of stakeholders requires domain expertise
• Domain expertise allows focuses on framing how project results solve the problems and opportunities
• Domain expertise helps build credibility and trust, thus making analysis more impactful.
• Asking informed questions comes from domain expertise.

Adapting to Stakeholders

• Communicating requires tailoring language at multiple levels
• This includes relevant vocabulary for specific stakeholders (technical vs. jargon).
• Considerations should be made whether to speak briefly or at length to the recipient of the information
• Adjust the level of explanation and facts based on the audience's familiarity with the data.

Common Forms of Communication

• Data science spans academia and industry, and central to both is communication.
• Common forms includes academic research papers, posters, executive summaries, and dashboards.

Academic Papers

• Academic papers aim to deliver information and have a structure of introduction, methods, results, and discussion
• IMRaD is an efficient method of structuring, but can stray from it to make the writing more exciting

Key areas for scientific papers

• Introduction is for setting up the problem, explaining the topic, and giving solution
• Methods explains a solution and how to evaluate it
• Results shows the results of an evaluation
• Discussion focuses on impact related to results

V Structure

• Introduction to introduce the problem and extract research questions
• Methods explains and evaluates a proposed solution.
• Results presents the findings and contextualizes these findings.
• Conclusion summarizes research questions and results; future applications can also be presented here

Paper Structure: Zooming In

• Level 1 headings should contain contain major sections, like the methods, results, etc.
• Level 2 headings should identify themes (e.g., Methods > Evaluation)
• Level 3 Headings contain topics (Methods > Evaluation > Evaluation Metrics.)
• Paragraphs should contain ideas (Methods > Evaluation > Metric > F1-Score)

Paragraph Construction

• Each paragraph should consist of:
• An introductory sentence setting the stage.
• A main body to elaborate on key points.
• A closing sentence to provides closure or transition

Academic Posters

• Academic posters are graphical, per-reviewed, and meant to inform within academic settings.
• The posters are non-linear, meant to explain a speaker with many plots present.

Executive Summaries

• These summaries should act as a single page linked to a technical report
• The target should be executives and decision makers, meant to assist in the decision making process.
• Key elements are problem statements, methodologies, key findings, recommendations, and visualizations.

Dashboards

• Interactive dashboards allow exploration of data analysis.
• These are often informed from live data and should assist with decision making.
• They often center around key performance indicators (KPIs) and metrics.

Dashboard Design Principles

• Dashboards should have an intuitive layout for immediate user understanding.
• Visual emphasis should be given to impactful data using appropriate charts.
• Interactivity should be supported.
• Dashboards should be clean and uncluttered.

Active vs Passive Dashboards

• Active dashboards allow to interact with and act upon the data, whereas passive are read only.
• Human involvement becomes important for active dashboards due to data, removing of observations and labelling

The Need for Human-in-the-Loop AI

• Various approaches exist, including interactive sense-making, explainable AI, adversarial training with human involvement, active learning, and meta-learning.
• There is an emphasis on full automation in AI and machine learning with large datasets and models.
• This process lends itself better to benchmark marking and specialization.
• Retaining humans in the process of improving beyond limitations, improving context

Active Learning

• Traditional supervised learning relies on randomly sampled and labeled training examples.
• Active learning enables algorithms to start with limited amount of labeled or unlabeled data.

General Active Learning

• A new batch of data points are selected based on specific criteria.
• You query human user to label that batch and retrain the model for the next batch.

Common Query Models

• Uncertainty sampling: Labelling is least certain among all instances; classifiers find documents closest to random chance
• Query by committee involves training multiple models, disagreeing on labelling based on votes and sends to user
• Expected outcomes involves documents where knowing would produce an estimate on the generalisation error.

Reasons to prefer active learning over supervised learning:

• Active learning allows for continuous labeling and allows to adapt to shifts in data.
• Active learning is cost effective.
• Quality is preferred over quantity.

Data Storytelling

• Data storytelling connects facts, transforming the analysis into a memorable narrative
• It uses a compelling question, builds tension, resolves with impacts, and utilizes visuals

Storytelling Frameworks

• A.I.D.A. captures attention, presents interest, shows how positive outcomes happen, and calls to action with information
• 5W is about using the structure to write a complete and through summary
• S.C.Q.A. defines a situation, followed with a complication, then a question, and the answering with the data given
• A Pyramid Structure gives conclusion first, followed with arguments and background data

Presenting an Analysis

• Presentations should begin with relevant hooks
• Focus on 2-3 major findings.
• Give recommendations with actionable steps
• Reiterate the importance of the work

Presentations

• Explain why results matter, and use impactful charts.
• Practice presenting, and anticipate questions.

Communication: BLUF

• BLUF communicates the bottom line up front early, sets clear expectations, and starts with the key conclusion.

Communication: Being Clear and Concise

• Cut unnecessary jargon and tangential details
• In a paper/dissertation use the active voice rather than passive, avoid subjective language, and be precise.

Communication: Visuals

• Utilize charts, as often visual are most memorable when comparing to text alone
• Common graphs are stacked bar charts, scatter plots, and line graph.

Communication: Abstracts

• Short versions of a paper, the abstract includes objectives to the work, ideas of methods used, results and important outcomes.
• Abstracts should aim to capture the reader

Communication: Audience

• Scientific papers should be adapted to relevant audience, discover/use vocabulary, and conventions as well should be used.

Communication: Formatting

• Figures should be correctly labeled
• Choose the right figure and be careful to avoid "chart junk"

Ethics

• The main ethical aspect should be focused on plagiarism prevention
• Data should not be manipulated
• Cherry picking and random seed optimisation is to be avoided
• It all comes down presenting work in a way where the audience understands