Data Science Mentoring

Every other page in this section is written from the seat of someone who learned the topic. This one is written from the other seat — the one where you have to make someone else understand it. I mentored data science students through the University of Melbourne's peer-mentoring program and the ANU Analytics Plus program, and it changed how I think about the whole field.

The biggest surprise of teaching is how much it teaches you. You can't explain what you only half-understand — the gaps in your own knowledge surface the moment a student asks "but why?" This page is the reflective counterpart to the rest of the section: not a topic I learned, but what learning to teach the topics taught me.

From the other side of the desk

Mentoring data science is different from tutoring a single subject, because the field is so broad and the learners arrive from everywhere — maths people scared of code, coders scared of stats, domain experts new to both. The job isn't to download facts; it's to help someone build a mental model they can extend on their own. A good mentor works themselves out of a job.

That reframes everything below. The goal of a session isn't to answer the question in front of you — it's to leave the student a little more able to answer the next question without you. Independence, not dependence, is the measure.

Where learners get stuck

After enough sessions you see the same walls again and again, and almost none of them are about intelligence:

The maths-anxiety wall — a learner convinced they're "not a maths person" freezes at a formula they could understand if it were unpacked in words first. The block is emotional before it's technical.
Tool-fixation — obsessing over which library or which model instead of asking what question they're actually trying to answer. They want the how before the what, which is backwards (the problem-first lesson).
Tutorial-following without understanding — they can run a notebook top to bottom and feel productive, but change one thing and it falls apart, because they followed steps rather than grasping why the steps work.

Recognising which wall someone is at matters more than knowing the material, because the response is completely different. The maths-anxious learner needs reassurance and intuition; the tool-fixated learner needs to be pulled back to the question; the tutorial-follower needs to be made to predict before they run.

Explaining hard things simply

The core craft of mentoring is making the hard thing simple without making it wrong. A few techniques do most of the work:

Intuition before formalism — give the picture first, the formula second. "Gradient descent is walking downhill in fog" lands before the equation does, and then the equation has somewhere to attach.
Analogy — connect the new idea to something they already know. The whole knowledge section is built on this: a model card, a foggy hillside, a shadow on a tabletop.
The "explain it back" test — the real check of understanding isn't whether they nod; it's whether they can explain it to you, in their own words. If they can't, the gap is exactly where their explanation breaks.

The mentor's playbook

A handful of principles, learned the hard way, that make a session work:

Meet them where they are — pitch to their actual level, not the level you wish they were at. Going over their head loses them; going under bores them.
Productive struggle — don't hand over the answer. The learning happens in the wrestling, so guide with questions and let them reach it. The help that feels most generous (just telling them) teaches the least.
Debug the thinking, not the code — when something's broken, the error is usually in the mental model, not the syntax. Fix the misunderstanding and the code fixes itself; fix only the code and the misunderstanding returns next week.

The mentoring loop. A learner attempts, gets stuck, and the mentor's job is to ask the question that unblocks their thinking — not hand over the answer — so they reach it themselves and can do it again next time.

Real, messy data

Textbooks teach with clean datasets; the world hands you mess. One of the most valuable things a mentor can do is move a learner off tidy toy problems and onto real, messy data as early as possible — because that's where the actual skills live. Wrestling with missing values, weird formats, and ambiguous questions teaches what no clean tutorial can: that most of the work is the data, and that judgement matters more than memorised steps.

It also builds the right relationship with being stuck. On real data everyone is stuck constantly; normalising that — "this is the job, not a sign you're failing" — is half of keeping a learner going.

Teaching deepens learning

The phenomenon has a name — the protégé effect: you learn material more deeply when you prepare to teach it and explain it to others. Teaching forces you to organise your knowledge, find the gaps, and build the clean explanations that only exist once you truly understand. I learned more data science by mentoring it than by sitting in some of the classes.

This is, frankly, the whole reason this knowledge section exists. Writing each page from scratch is teaching at scale — and the act of having to explain embeddings, or the bootstrap, or the CAP theorem clearly is exactly what keeps my own understanding sharp. The section is the protégé effect, applied to myself.

The skills nobody tests

Mentoring surfaced something the curriculum never grades: in real data science, communication and judgement matter as much as the maths. Beginners systematically underrate this — they think the job is the algorithm, when the job is framing the right question, working with people, and explaining the result so it gets used (the communication page). The best thing I could do for a mentee was widen their definition of "the skill" to include the parts no exam measures.

The through-line to my work

Refresh in 60 seconds

Reflects current writing on teaching data science and statistics (data-science education pieces, the apprenticeship/mentoring model) alongside hands-on mentoring experience.