A walk through Schottenfeldgasse 55/3 in the name of Machine Learning.
Us General Concepts
Bed Stands: Type 1 and Type 2 error:Seb: H0: Someone who has Covid tests positive.Barry: Recall/Type 1 error: Someone who does not have Covid tests positive (bad for the individual).- True Positive / True Positive + False Negative
Blue: Precision/Type 2 error: Someone who has Covid tests negative (bad for everyone).- True Positive / True Positive + False Positive
My Room Theories
Chair: Behavioral Change: operant (response is affected by consequences, e.g., gamification) vs classical conditioning (response occurs to stimuli, e.g., reminders)Box: Personal Thoughts: Precontemplation > Contemplation > Preparation > Action > Maintenance (actively set goals for users and encourage them to hit the goals).Blanket: Social Imitation: Consider collaboration features.Window: Fogg Model of Change: pleasure / pain, hope / fair and social acceptance get lazy people to do things.Shutters: ABA Model of Change: people engage in behavior to (a) achieve a desired result or (b) avoid an undesired resultRadiator: Randomized Variable Schedule: slot machines are addictive because the randomize administering awardsLaundry: Outsized positive rewards & mitigating losses: loss aversionRobe: Path dependence for least resistance: one step leads to the next step
Bath room ML Terminology
Bath Ward Robe: Supervised Learning: develop predictive models based on input and output (classification and regression).Contacts: K-Nearest Neighbors: birds of a feather will flock together.Gel: Linear regression: identify lines that best fits data points.Lip stick: Logistic regression: used for binary outcome, probs of accepting a friend request.Deo: Decision trees: best of 20 questions.Perfume: SVMs: perceptron machine divides two dimensional space.
Towels: Unsupervised Learning: group and interpret data based on input data.Small: K-means: bucket observations in k clusters.Large: PCA: converts correlated variables into linearly uncorrelated vectors.
Spanners: Dependent & independent variables: labels & key variable of interest.Pillow Cases: Confounding variables: variables that could drive spurious correlations.Blankets: Different Test StatisticsPaper rolls: z-test / z-score: two populations, normal distributionTissues: t-test / t-statistic: two small samples, normal distributionWindow cleaner: Anova / f-statistic: three or more samples, normal distributionToilet cleaner: chi-squared test / chi-squared: two samples, any distribution
Sink: Regression terms:Silvio Brush: Lasso (L1) Regularization: takes absolute value to avoid overfittingMy brush: Ridge (L2) Regularization: takes squared value to avoid overfittingToothpaste: Normalization: adjusting different scales to same scaleLights: One-hot encoding: when we have data on more than one level (fe/male)Mirror: Underfitting: too simple to explain varianceShower: Overfitting: force-fitting, too good to be true
Silvio's office Validation of Supervised Learning:
Massive picture: Train, test validation: Test different models on test data / val data.Coffee Machine: K-fold cross validation: Create k-folds and use these as test data.3D Printer: F1-score: 2 * 1/recall + 1/precision = (sensitivity = recall, specificity = 1-precision).3D Printer Workshop: Area under ROC: visualization of classification errorDesk: Ad-Hoc Experiments - A/B Testing:Chair: Approach: experiment run on two groups of random users (RCT), one retrieving the treatment and one that does not.Srreen: Premise: all conditions have to be the same for treated and control groups, except for the treatment.Silvio: ATE (average treatment effect): what is the average difference in our chosen metric between treated and control groups.iPad: Example: Sending out friend requests with and without a personalized message.
Hover Post-Hoc Experiments - Causal Inference Techniques
Window: Difference-in-Difference Modeling: find comparable counterfactual or control group to compare our treated users. (How do markets behave after airing or not airing a commercial in an area?)Wine: Placebo Testing: Test if there is no difference in two areas where there is no commercial aired.Book: Regression Discontinuity: Find comparable counterfactuals or control group members to compare treated users against.Towel box: Interrupted Time Series:Whiteboard: Autocorrelation: serial correlation between values of the process at different times.Wine shelve: Stationarity: Mean and variance are constant over time.Gin & Tonic: ARMA & ARIMA: autoregression models are linear regression models where we predict series values on the lagged value of the series itselfWinter Wine w/ rum: Seasonality Decomposition: known patterns that are repeated over fixed time intervals.
Light: Statistical Matching: achieve similar distributions on observed variables for treatment and control.Tado: Propensity Matching: classification algorithm to assign a treatment score or a probability of being treated.Table: General Heuristics: Educated guessing, dosage effect measuring, placebo tests, coherence, analogy.Salt: Uplift Modeling: targeting users that will modify their behavior based on treatmentPepper: Random Forests: collection of many decision trees to identify features and make predictions