Sanaz Bahargam

Sanaz Bahargam

Machine Learning Engineer @Twitter

Covariate Shift and Dataset Shift

2 minute read

Published:

Table of Contents

If you have productionized a new machine learning model, most probably you have faced the situation after a while the performance of your model has degraded. Let’s say train a model with 90% precision and you are happy with it. You ship the model into production and check the model performance over time. In a month, you see that the model precision has been decreasing over time and has reached 70% precision. Now you may wonder why this is happening. The chances are you the distribution of new data in production is significantly different than the distribution of the data you used for training. This phenomenon is known as the covariate shift or dataset shift.

Covariate shift:

This phenomenon happens when the distribution of inputs used as predictors (covariates) changes between the training and production stage, i.e. Ptrain(X) != Pproduction(X)

Dataset shift:

when the joint distribution of inputs and the outputs (the target being predicted) also changes, i.e. Ptrain(X, y) != Pproduction(X, y)

Causes of covariate and dataset shift:

The covariate and dataset shift happen due to two major factors

  • sample selection bias
  • non-stationary environments (when the predictions distribution changes due to different factors such as seasonality, new users, change of entities/products).

How to detect/treat covariate and dataser shift

In the real-world, the models are retrained with fresh data on a regular basis to resolved this issue this process can be automated with Airflow. In case, it is expensive to retraining the model regularly, you can check for a covariate shift in your data and when the distribution of data has changed dramatically, you can start the process of retraining the model with fresh data.

We can detect covariate shift by using data from the training stage and label it with “train” and get a sample from new data/fresh and label it “fresh”. We can train any ML model on this data and if the model predicts the label better than random, there is a covariate shift happening in the data (You can check if your model is doing better than random chance by checking Matthews correlation coefficient (MCC) or phi coefficient)

Read more here: DATASET SHIFT IN MACHINE LEARNING

You May Also Enjoy

Naacl2021

9 minute read

Published:

Here are the summary of some of talks from NAACL 2021. So far, I only summarized the following papers, I will be summarizing more. So, I will either append it here or a new blog post. Feel free to check back soon.

ICLR 2021

8 minute read

Published:

ICLR 2021

Edge Probing

3 minute read

Published:

In the past couple of years, Transformers has acheived state of art results in a variety of natural language tasks. In order to better understand Transformers and what they are learning in practice, researchers have done layer-wise analysis of Transformer’s hidden states to understand what the Transformer is learning in each layer. A wave of recent work has started to “prob” the state of the art Tranformers to inspect the structure of the network to assess whether there exist localizable regions associated with distinct types of linguistic decisions, both syntactic and semantic information. Researchers examine the hidden states between encoder layers directly and use those hidden states in a linear layer + softmax to predict what kind of information in encoded in each hidden state.

Transformers 2

5 minute read

Published:

This blog post is the continuation of my previous blog post, Transformers. In my previous blog post, I explained original Transformer paper, BERT, GPT, XLNet, RoBERTa, ALBERT, BART, and AMBER. In this blog post, I will explain MARGE, ConveRT, Generalization through Memorization, AdapterHub, and T5. Images and content used in this blogpost, otherwise mentioned, are all taken from the papers on each model.