AutoAI Tutorial: Univariate Time Series Experiment

Use sample data to train a univariate (single prediction column) time series experiment that predicts minimum daily temperatures.

Tech preview This is a technology preview and is not yet supported for use in production environments.

When you set up the experiment, you will load data that tracks daily minimum temperatures for the city of Melbourne, Australia. The experiment will generate a set of pipelines that use algorithms to predict future minimum daily temperatures. After generating the pipelines, AutoAI compares and tests them, chooses the best performers, and presents them in a leaderboard for you to review.

Overview of Data Set

The data set Daily_Min_Temperatures.csv describes the minimum daily temperatures over 10 years (1981-1990) in the city Melbourne, Australia. The units are in degrees Celsius and there are 3650 observations. The source of the data is the Australian Bureau of Meteorology. Note the following about the data set:

• You will use the Min_Temp column as the prediction column to build pipelines and forecast the future daily minimum temperatures. Before the pipeline training, the date column and Min_Temp column will be used together to figure out the appropriate lookback window.

The prediction column will forecast a prediction for the daily minimum temperature on a given day.

The sample data is structured in rows and columns, and saved as a .csv file.

Tasks Overview

This tutorial presents the basic steps for building and training a machine learning model using AutoAI:

1. Create a project
2. Create an AutoAI experiment
3. Configure experiment
4. Training the experiment

Task 1: Create a project

1. From the Gallery, download the zip locally and extract the CSV files: Daily_Min_Temperatures.csv

2. In the Projects page to create a new project, select New Project.
   a. Select Create an empty project.
   b. Include your project name.
   c. Click Create.
   

Task 2: Create an AutoAI experiment

1. In the project page, select Add to Project and choose AutoAI experiment.

2. Specify a name and optional description for your new experiment.

3. There are two ways to upload your file to the experiment:
   a. If you downloaded your file locally, upload the training data file, Daily_Min_Temperatures.csv, from your local computer by dragging the file onto the data panel or by clicking browse and then following the prompts.
   b. If you already uploaded your file to your project, click select from project, then select the data asset tab and choose Daily_Min_Temperatures.csv.

Task 3: Configure experiment

1. Click Yes for the option to create a Time Series Forecast.

2. Choose as prediction columns: Min_Temp.

3. Choose as the date/time column: Date.

   

4. Click Experiment settings to configure the experiment.
   a. In the Data Source page, select the Time series tab, specify 4 backtests for validating the model, and a holdout size of 20. You can see that the validation length changes accordingly.
   b. You can optionally review the remaining settings, but accept the defaults.
   c. Click Save settings to update the configuration.
   

   

5. Click Run experiment to begin the training.

Task 4: Training the experiment

1. As the experiment trains, a visualization shows the transformations used to create pipelines. Hover over any node to get details on the transformation.

   

2. You can then click Pipeline comparison to see how they differ. For example:

   

3. When the training completes, the top three best performing pipelines are saved to the leaderboard. You can also review the discarded pipelines.

   

4. Select the pipeline with Rank 1 and click Save as to create your model. Then select Create. This saves the pipeline under the Saved models section in the Assets tab.

Parent topic: Building a time series experiment (Tech preview)