Data Science Decoded

Currently writing about learning mathematics for machine learning via math academy | Senior data scientist with 5+ years of experience

[Data Science Decoded @DSDecoded]

Daily Math Academy

Working towards mathematics for machine learning

• Achieved 29/50 XP
• Scored 23/18 XP on lesson on diagonalizing a 2x2 matrix
• Scored 6/4 XP on review of limits and continuity of multi variable functions

[Data Science Decoded @DSDecoded]

Daily Math Academy

Working towards mathematics for machine learning

• Achieved 26/50 XP
• Scored 6/4 XP on review of log-likelihood functions for discrete probability distributions
• Scored 9/7 XP on lesson for elliptic cones
• Scored 11/9 XP on lesson for introduction to partial derivatives

[Data Science Decoded @DSDecoded]

Daily Math Academy

Working towards mathematics for machine learning

• Achieved 27/50 XP
• Scored 6/4 XP in review of likelihood functions for discrete probability distributions
• Scored 6/4 XP for review of defining abstract vector spaces
• Scored 15/15 XP for Quiz
• Got one review pending for log likelihood functions for discrete probability distributions

[Data Science Decoded @DSDecoded]

Daily Math Academy

Working towards mathematics for machine learning

• Achieved 19/50 XP
• One lesson today on limits and continuity of multi variable functions scoring 19/15 XP

[Data Science Decoded @DSDecoded]

Daily Math Academy

Working towards mathematics for machine learning

• Achieved 27/50 XP
• Scored 11/9 XP on Paraboloids
• Scored 6/4 XP on review of interiors and boundaries of sets
• Scored 4/4 XP on review of domain and range of quadratic functions
• Scored 6/4 XP on review of finding a basis of the column space of a matrix

[Data Science Decoded @DSDecoded]

Daily Math Academy

Working towards mathematics for machine learning

• Achieved 28/50 XP
• Scored 4/4 XP on review of type 1 and type 2 regions in 2 dimensional space
• Scored 6/4 XP on review of logical equivalence with conditional statements
• Scored 18 XP on lesson on characteristic equation of a matrix

[Data Science Decoded @DSDecoded]

Daily Math Academy

Working towards mathematics for machine learning

• Achieved 29/50 XP
• Scored 6/4 XP on review of dividing polynomials by manipulating rational expressions
• Scored 8/9 XP on lesson on domain of multi variable functions
• Scored 6/4 XP on review of invertible matrix theorem
• Scored 2/4 XP on dimension of the null space of a matrix
• Scored 7/5 XP on review of orthogonal complements

[Data Science Decoded @DSDecoded]

Daily Math Academy

Working towards mathematics for machine learning

• Achieved 19/50 XP
• Need to re-do lesson on domain of a multi-variable function scoring -2/9 XP as struggled with the graphs - I’m putting this down to rustiness from a few months off
• Completed lesson on conditional statements with predicates 12/12 XP
• Re-took Quiz 13 and scored 9/13 XP which I’m fairly happy with
• Right now I’m just focusing on getting the habits back rather than the pace of completion, previously I used to aim for 50XP as a stretch per day but now I’m scoring low 20’s at the moment

[Data Science Decoded @DSDecoded]

Daily Math Academy

Working towards mathematics for machine learning

• Achieved 25/50 XP
• Scored 1/15 XP on Quiz 13 - this was a lot of stuff that I had forgotten from my 8 week break
• Scored 7/5 XP on review of orthogonal complements
• Scored 6/4 XP on review of expressing the coordinates of a vector in a given basis
• Scored 8/6 XP on review of variance of sample means
• Scored 0/4 XP on domain of multi variable functions
• Scored -2/4 XP on conditional statements with predicates

[Data Science Decoded @DSDecoded]

Daily Math Academy

Working towards mathematics for machine learning
Quick session today just to maintain momentum

• Achieved 8/50 XP
• Reviewed sum of squares 4/4 XP
• Reviewed dimension of null space of a matrix 4/4XP

[Data Science Decoded @DSDecoded]

Daily Math Academy

Working towards mathematics for machine learning
Back after a break due to other committments

• Achieved 25/50 XP
• Scored 9/10 XP on lesson on hyperboloids
• Scored 6/4 XP on review of calculating the eigenvector of a 2x2 matrix
• Scored 4/4 XP on review of binconditional statements
• Scored 6/4 XP on review of permutations

[Data Science Decoded @DSDecoded]

3 things that decide whether an ML project succeeds, none of them modelling:

Whether the data you need actually exists

Whether the model is trusted enough to be used

Whether it keeps working after you ship it

The model is usually the easiest part.

[Data Science Decoded @DSDecoded]

3 more things Kaggle doesn't teach you:

How to tell a stakeholder the model isn't worth building

How to get a model out of a notebook and into production

How to keep it working six months after you ship it

Kaggle ends at the submission. In industry, that's where the real work begins.

[Data Science Decoded @DSDecoded]

Frame your model improvement in business impact, not model metrics.

Stakeholders don't work in log loss.

Tell them what the improvement means in revenue, conversions or cost, and you'll get their attention and their buy in.

[Data Science Decoded @DSDecoded]

Kaggle skips the first question entirely: does the data even exist?

In industry you'll spend weeks finding out what's logged, what's logged wrong, and what nobody thought to log at all.

That work often takes longer than the modelling does.

[Data Science Decoded @DSDecoded]

On Kaggle the problem is framed for you and the metric is handed over.

In industry neither is.

You have to understand the business problem first, then work out how to frame it as an ML problem.

That means time with stakeholders deciding which problem is actually worth solving.

[Data Science Decoded @DSDecoded]

3 problems worth solving with ML in ad tech:

Ads ranking: predict CTR and rank ads across the available slots

Delivery forecasting: predict impressions before a campaign commits budget

Creative optimisation: use bandits to find the best ad variant while still exploring

[Data Science Decoded @DSDecoded]

Forecasting ad campaign delivery is hard because the forecast is a snapshot and the ads market isn't.

You predict impressions or clicks at campaign start.

Then competitors enter, budgets shift, traffic moves.

The marketplace you deliver into is never the one you forecast for.

[Data Science Decoded @DSDecoded]

Not everything needs to be an agentic solution.

What matters is whether the selected solution solves the problem and adds value.

Whether that's an agent, an ML model or a SQL query is irrelevant.

[Data Science Decoded @DSDecoded]

3 ways data science in industry isn't like Kaggle:

You have to frame the problem yourself

You have to source the data yourself

You have to translate model metrics into something stakeholders actually care about

None of that shows up on a leaderboard.