[2023 kcdc] data leakage – why your ML model knows too much

Speaker: Leah Berg

For more, see theĀ table of contents.


Data Leakage

  • Also known as leakage or target leakage
  • Different meaning for information security (data leaking to outside organization)
  • Can be difficult to spot
  • Training data includes info about test.
  • Model trained on info not available in production

How models learn

  • Split data into training data and test data.
  • Test data – data model has never seen before and makes sure model gets is right
  • Can also have an optional validation set
  • Randomly pick whether data points are training or test data. – Called random train/test split
  • More training data than test data

Don’t include data from the future

  • Using a random split of time series data doesn’t work because model has learned about future data.
  • Better to use a sliding window. Use first few months to predict next month. Then add that next value and predict one after. And keep going. Adding up error gives you accuracy of model.
  • This works because model only knows about data before one asked to predict.
  • Create timeline for when events happen. That way you make sure you aren’t using data from before the prediction
  • Don’t always know where/when data was created. Important to understand business process

Don’t randomly split groups

  • Have some data from the group you are then predicting
  • Problem when new student shows up so prediction will be bad
  • scikit-learn has GroupShuffleSplit() to get full group in same set – testing or training

Don’t forget your data is a snapshot

  • In school, have pristine data set.
  • In real world, data is always changing.
  • Could tell model about data that occurred after prediction. Again think about data on timeline

Don’t randomly split data when retraining

  • Want to use same training/test data on production and challenger models to see which better.
  • One has already seen data points during training that you are testing so you don’t know if it is better.
  • Challenger model can get more data that wasn’t available originally. Ok to split new data into test/train as long as original data part is split same way.

Split data immediately

  • Risky to rescale before split because data isn’t represented same way. Min/max can vary if split after
  • Run normalization on different sets of data
  • Before split, do analysis with business, exploratory data analysis. Split data before start modeling

Use Cross Validation

  • KFold Validation – split training data into K parts
  • ex: 3 fold validation – two parts stay as training and one is validation. The test data remains as test data and is kept separate for final evaluation.
  • The validation set is for an initial test.
  • Gives more options to train model

Be Skeptical of High Performance

  • If validation much higher than train/test, suspicious.
  • If train/test/validation sets are all high/the same, suspicious.

Use scikit-learn pipeline

  • Helps avoid leaking test data into training data

Check for features correlated with target

  • If another attribute has a high match with what looking for, make sure not mixing up correlation/causation.
  • Also, avoid timeline errors for reverse causation. Ex: the thing you are looking for causes, something else

My take

Great talk. Almost all of this was new to me. It was understandable and I learned a lot.

[2023 kcdc] the elephant in your data set – avoid bias in machine learning

Speaker: Michelle Frost

For more, see theĀ table of contents.


  • Intersectionality wheel of privileged. Many spokes and range from power to erased to marginalized. Used the version posted here
  • Bias – inclination or prejudice for or against one person or group
  • ML Bias – systematic error in the model itself due to assumptions
  • Sometimes bias is necessary – inductive bias – assumptions combined with training examples to classify
  • Models with high bias oversimplify the model
  • Each stage has potential harmful bias
  • Bias feeds back into model
  • In ML, when something looks two good to be true, it probably is

Points of bias

  • Historical – prejudice in world as it exists today. Gave example from ChatGPT where assumed a nurse was female even when replaced pronouns. Full example here
  • Representation bias – Sample under-represents part of population. Can’t make effective predictions for that group. Article describing. “Solved” by dropping gorillas as a label
  • Measurement bias – using a proxy to represent a construct. Problem if oversimplifying or accuracy varies across groups. Compas (Correctional Offender Management Profiling for Alternative Sanctions) example. Data measures policing not just the offender.
  • Aggregation bias – one size fits all model assumes mapping inputs to labels is consistent. For example, could mean something different across cultures. Such as LSD being Lake Shore Drive in Chicago and not a drug. Or racial differences for HbA1c
  • Learning bias – modeling choice may prioritize one objective which damages another. Such as Amazon’s recruiting tool discriminating against women
  • Evaluation bias – benchmark data does not represent the population. Might make sense in some scenarios. Project Gender Shades analyzed differences in different tools.
  • Deployment bias – model attended to solve one problem, but used a different way. Make a hook for tuna and use it on a shark. Child abuse protection tool fails poor families.

Simpson’s paradox

  • Other attributes are a proxy for the thing leaving out
  • Association disappears, reappears or reverses when divide population


  • Protected class – category where bias is relevant
  • Sensitive characteristics – algorithmic decisions where bias could be factor
  • Disparate treatment
  • Disparate outcome/impact
  • Fairness – area of research to ensure biases and model inaccuracies do not lead to models that treat individuals unfavorable due to sensitive characteristics.


  • Demographic partiy – decisions/outcomes independent of protected attribute. Does not protect all unfairness
  • Equal odds – decision independent of protected attributes. True and false positive rates must be equal
  • Equal opportunity – like equal odds but only measures fairness for true positive rates


  • A popular (bad) data set is “adult data set”. I think i this one.
  • Not balanced by gender, race, country

Book recommendations

  • Weapons of math destruction
  • Biased
  • The alignment Ppoblem
  • Invisible Women
  • The Big Nine
  • Automating Inequality

My take

The types of bias and examples were interesting. Good end to the day. The demo graphs provided the point about biased data nicely.

[2023 kcdc] data science: zero to hero

Speaker: Gary Short

Twitter; @garyshort

Repo for presentation/samples: https://github.com/garyshort/kcdc2023

For more, see the table of contents.

Data science overview/rules

  • Applied data science – solving business problems
  • Curiosity is most important
  • The universe does random stuff so you haven’t discovered anything until you prove you’ve discovered something
  • Only qualitative and quantitative data – people lie, Can’t trust what you ask
  • Can only do math with numbers. Some things will pretend to be numbers when they are not. Also, can’t add different things (dollars vs killograms)
  • If you can’t explain it to a six year old, you don’t really understand it
  • Only have to be more than 51% accurate to do better than guessing
  • True random data has some clusters. The cluster will not last forever. Gambler’s paradox. 27 blacks doesn’t mean due a red.
  • If it’s not in production, it doesn’t exist. Can’t just be on your laptop. Most data scientists need to give to someone else to get it to prod. Cultural difference between data scientist and person who is building/deploying.
  • % chance of hypothesis being right or wrong doesn’t have to sum up to 100%. ex: grass is wet. Could be rain or a dog peeing or something else

Types of data


  • Relational data
  • Get connection, create cursor, fill cursor, close connection
  • Schema is important on data write.

Semi structured

  • ex: JSON/MongoDB.
  • Get connection, name collection, fill cursor, close connection
  • Schema important when read data


  • Blob (binary large object)
  • Stored in pages/blocks
  • Access via URL


  • Degrees of separation – can you deliver a message directly
  • People in room now more closely connected because in this session (and would stay so if shared contact info)
  • Wide network effect
  • Nodes tend to be nouns
  • Edges tend to be verbs. Can be unidirectional or bidirectional
  • Get connection, state query, fill cursor, close connection

AI/ML works on data types

  • Categorical – segregate data by category where category is not important (ex: blue eyes)
  • Ordinal – order is important but distance between is not important (ex: position in a race)
  • Numeric – order is important but distance is the same (ex: counting)
  • Ratio – numeric but with positive numbers

Can only do math with ordinal and ratio types. A survey on a scale of 1-5 (likert scale) is ordinal, not numeric/ratio. Can’t do average. This is categorical data (ex: very happy, pissed off). Can do math with counts of categorical data but not single items.

Exploratory Data analysis

  • Need to understand the variables. Ex: is it really a number
  • Handle missing values – depends on scenario. Ex: use mean or median (if not looking for that particular thing), delete row with incomplete data
  • Outlier detection – sometimes genuinely an outlier (ex: someone who is 8 feet tall), sometimes it is the important piece of data (ex: which exits people use in a fire; one person went the other way and want to know why). Need to determine why outlier and if care so don’t delete data need
  • Univariate analysis – ex: histogram for categorical data
  • Bivariate analysis – correlated data; could be hidden variable. Don’t need both of them since one predicts the other. Want minimal variables in model so chose the one that brings in the most info.

Feature Selection

  • Preprocess the data
  • Normalize data – units have to be the same. Using variance doesn’t help because unit is now original unit squared. Can use Z-score so everything on scale 0-1 using mean and divisor
  • Encode the categories – make so can do math
  • Booleans are numbers (0 and 1)
  • Word vector – can use math to represent a word. Complicated. Ok to have to look up every time.
  • Bi/multivariate analysis – high correlation means redundant info
  • Feature importance – check coefficients from regressions and scores from gradient boosting

Model Selection

  • People have a favorite model
  • Use one or more models. See which gives best result before making any changes to the model.
  • Good to use a linear and non linear one. Normal the linear model is enough because normally dealing with people (directly or indirectly). Linear equations work for a normal distrobution.
  • Make sure to find global minimum, not local/current one
  • Compete with yourself. Try to have your second best model beat your current best model. Once something in prod, start again

Train/test split

  • 80/20 split
  • 80% data for training
  • 20% data real
  • Model never sees training data because can’t grade own homework

Model evaluation

  • Outcome – model + error
  • Error is difference between predicted and observed values.
  • Sample of population can be model. Get error because of sampling bias

Hyper Parameter Tuning

  • Every models have parameters to govern how works.
  • Hyper param tuning is fiddling with these
  • Will be an optional value for each of these parameters for your particular use case

Model Validation

  • Need to make sure model doesn’t work by chance
  • K-Fold Cross Validation – after do 80/20 split, can feed data back in and do again
  • Stratified Cross Validation – same as K-Fold but unbalanced classes

Bayesian inference in Real Life

  • P(h|e) = P(e|h) * P(h) / P(e)
  • In English: current belief = new evidence


  • Important to be able to estimate values when have no data
  • Dumb questions like “how many piano tuners are there in Chicago” was testing this. So few people could do it that pulled question. [I suspect the ridicule and people memorizing the answer was a factor too]
  • Easier to estimate a range than an actual value
  • Pick a minimum that it couldn’t possibly be below. You’d be surprised and skeptical if less than that.
  • Pick a maximum that it couldn’t possible be above.
  • Pick value spits range in two so that the possibility of being above/below has equal probability. Call this the medium. Resist temptation to pick the mean.
  • Repeat finding the minimum to median. Call this Q1
  • Then repeat finding the median to maximum to get Q3.
  • This gets you a five point description of a distribution
  • Use sampling to get mean of distribution

Lab part

The lab was to predict something you want to predict and make a model and/or predict a probability. Can do individually or in groups. He also gave the option to leave. I chose leave because there was a little over an hour left when he finished explaining the lab. I need to go over the material for my own workshop so doing that instead of the lab.

My take

This was a good intro and Gary is a good, engaging speaker. I learned (and re-learned) a bunch of stuff. Both concepts and terms. Having a bunch of rules and getting into them made it fun. (ex: math needs numbers). I like that the concept part was longer (except for the lack of a break), but it would hav been better if it was advertised that way in the intro.

I disagree with Gary’s philosophy on not having a bathroom break. He started by saying there would be 60-90 minutes of lecture and then a lab. [wound up to being just over 2.5 hours] And that we are all adults and can go to the bathroom whenever. Someone asked at the 90 minute mark if there would be a bathroom break and he repeated the all adults thing expanding that you’ll catch up and the slides will be online later. He also said people feel compelled to hold it until break or go when told it is break. However, the tradeoff is that you don’t want to go to the bathroom lest you miss something that will wind up being important during the session. It’s super frustrating to miss stuff and then struggle to understand later. It may be that this workshop isn’t cumulative but there’s no way to know. Also, by not having a break, you aren’t giving people’s brain a break. It’s not just about the bathroom.

Gary stated he puts the materials online after so people don’t read it during the session. That I agree with!