The Unofficial Google Data Science Blog

OCTOBER 7, 2015

by HENNING HOHNHOLD, DEIRDRE O'BRIEN, and DIANE TANG In this post we discuss the challenges in measuring and modeling the long-term effect of ads on user behavior. Nevertheless, A/B testing has challenges and blind spots, such as: the difficulty of identifying suitable metrics that give "works well" a measurable meaning.

Modeling 61

Modeling Experimentation Knowledge Discovery Testing 61

The Unofficial Google Data Science Blog

JANUARY 14, 2016

And an LSOS is awash in data, right? Well, it turns out that depending on what it cares to measure, an LSOS might not have enough data. The practical consequence of this is that we can’t afford to be sloppy about measuring statistical significance and confidence intervals.

Experimentation 41

Experimentation Statistics Metrics Measurement 41

The Unofficial Google Data Science Blog

JULY 22, 2020

For this reason we don’t report uncertainty measures or statistical significance in the results of the simulation. Ramp-up solution: measure epoch and condition on its effect If one wants to do full traffic ramp-up and use data from all epochs, they must use an adjusted estimator to get an unbiased estimate of the average reward in each arm.

Experimentation 95

Experimentation Statistics Testing Strategy 95

Domino Data Lab

MAY 20, 2021

Working with highly imbalanced data can be problematic in several aspects: Distorted performance metrics — In a highly imbalanced dataset, say a binary dataset with a class ratio of 98:2, an algorithm that always predicts the majority class and completely ignores the minority class will still be 98% correct. Machine Learning, 57–78.

Machine Learning 136

Machine Learning Metrics Data mining Knowledge Discovery 136

The Unofficial Google Data Science Blog

NOVEMBER 4, 2015

Posteriors are useful to understand the system, measure accuracy, and make better decisions. Methods like the Poisson bootstrap can help us measure the variability of $t$, but don’t give us posteriors either, particularly since good high-dimensional estimators aren’t unbiased.

KDD 40

KDD Testing Machine Learning Modeling 40

The Unofficial Google Data Science Blog

FEBRUARY 29, 2016

And since the metric average is different in each hour of day, this is a source of variation in measuring the experimental effect. Let’s go back to our example of measuring the fraction of user sessions with purchase. Let $Y_i$ be the response measured on the $i$th user session.

Experimentation 40

Experimentation Metrics Statistics Measurement 40

Domino Data Lab

AUGUST 1, 2021

but it generally relies on measuring the entropy in the change of predictions given a perturbation of a feature. Conference on Knowledge Discovery and Data Mining, pp. The implementation of the attribute importance computation is based on Variable importance analysis (VIA). See Wei et al. Guestrin, C., Bahdanau, D.,

Modeling 139

Modeling Deep Learning Machine Learning Knowledge Discovery 139