Name: DataRobot
Uploaded: 2025-02-02T14:00:27.462Z
Duration: 2 min 33 s
Description: An AI platform offering tools to build, deploy, and manage predictive and generative AI solutions.

datarobot

An AI platform offering tools to build, deploy, and manage predictive and generative AI solutions.

DataRobot

Welcome. Here's a preview of what you'll do and learn in this walkthrough. You'll start off by creating a Use Case folder named Car Sales and uploading the data set we provide which has month-by-month sales for many makes and models of vehicles. In the graphical user interface you'll set up the initial Experiment settings to start modeling on all of the vehicles together. That will result in having the first Experiment here in the Experiment list. In this case there are 36 models inside this experiment. 

Then you'll import the Jupyter-formatted notebook that we provide into DataRobot Notebooks. When you execute this notebook, the other Experiments in this list are created for you segment by segment such as Cars, Luxury cars, Luxury SUVs etc. You can see that there's roughly 40 or 50 models inside each one of these Experiments. As your experiments are running you will be able to start investigating the insights. 

When forecasting all of the vehicles together, this Light Gradient Boosting on Elastic-Net Predictions is the winning blueprint. From the Feature impact we learn that the sales volume is the most important feature in the model, and specifically the log of the last 12 months median sales volume is the most important feature. But that isn't necessarily true for all segments. For example, for the sports car segment the month of the year is the most important feature. 

We can validate and probe that insight a bit further with Feature effects. Holding all other features constant, the partial dependence plot shows that predicted sales for sports cars are lowest in January and highest in August and December. Accuracy over time allows you to visually inspect how well the model fits your training data. Here you're seeing the actual values in orange and the predicted values in blue for each individual series of car. 

You can examine how well the model fits the entire 9 years of data provided or just zoom into the validation period. Series insights gives you easy access to the error for each of the individual makes and models of cars.

The Toyota Camry has the lowest error. And the Chevy Impala currently has the highest at 0.4. The Chevy Impala is in the 'Car' segment so let's see if the segmented model is any better. In this case the winning blueprint is a Ridge Regressor with Forecast Distance Modeling and Series Scaling. Let's check the insights for that Chevy Impala. The Chevy Impala still has the highest error, although it is lower than before. And that's where this walkthrough ends, but of course in real life you might now carve out the Chevy Impala and continue to model it separately. 

So to recap, if you follow this walkthrough you'll have one data set, many experiments, one notebook. And when all of this is completed you can easily share it with other team members. Take some time to explore all those insights and remember if you're going to buy a sports car you'll probably get a better price in January. From the homepage select Workbench, then create a new Use Case and name it 'Car Sales'. Upload the FAST version of the data set provided with this tutorial. DataRobot will advance automatically to this screen and you'll have a short wait while the data set registers. 

Once it's registered, click 'Explore' to see what's inside. This is time series and note that our time step is monthly. The target value to forecast is the monthly sales volume. The FAST version of the data set has 10 models of cars in two major segments. There are five columns of contextual information - the average price and average dealer incentive, and three economic indicators. When you're done exploring you can choose 'Start modeling' from this menu or return to the Use Case folder and start the modeling from the three dots. 

This table is great for a quick data quality check such as how many missing values and how many unique values. Fill in the name of the target feature which is 'Sales_Volume'. Change the modeling mode to Comprehensive, and view your options for optimization metrics. I'm going to go with the recommended metric. Then note the experiment summary information. Right now we're still in the configuration stage so all of this can be modified.

Once the experiment begins it will be locked and this will provide the metadata and governance for the experiment. Let's use 'Next' to advance to the next screen. Since this is time series select 'Time series modeling' and toggle it on. DataRobot has correctly detected the Date feature. You may need to set this manually if you have more than one date in your data set. DataRobot has also correctly detected the series identifier. I can click on this word 'Model' or I can type it into the drop-down list. 

Scroll down so that you can see the Feature derivation window, the Forecast window, and this illustration. Enter the numbers 13 and 1 for the Feature derivation window. Then enter 2 for the First forecast and 3 for the Number of values to forecast. DataRobot will create derived features using information from the last 13 months. The models in the Experiment will be able to predict 2, 3, and 4 months into the future. The winning model will be the one that minimizes the error for all makes and models of cars for all forecast distances. 

Turn on features known in advance and enter 'Brand' and 'Major_segment'. These features do not vary with time and are definitively known at prediction time. In the Experiment summary panel, click on Partitioning. DataRobot has configured multiple equally-spaced backtests for you. Explore the options available for precise configuration of backtests, but for simplicity accept the defaults. Then review all the details in the Experiment summary panel. These cannot be changed once you start modeling. 

When everything looks good hit 'Start modeling'. The automation is now in control. DataRobot will derive your time series features, apply other necessary pre-processing to the data, select the algorithms that should be tested, and then start testing them one by one. It's as if you've taken your experiment and handed it to a team of data scientists and you'll come back later to review all of their results. No further input is needed from you at this point and you can walk away and come back later.

The time it takes for the Experiment to complete depends on the amount of starting data and the number of concurrent jobs that your account allows. At this point two models are finished and two are building. You can view what's in the queue by expanding this menu on the right. Get in touch with your account representative if you want to increase the number of concurrent jobs running in your account. 

Anything on the Leaderboard that has a number next to it is complete and you can start to explore it. It's helpful to mark it with the star to find it easily again later. Start by expanding the Blueprint. Notice the pathway that your categorical and numerical data has traveled. Each of these boxes is clickable bringing up summary information. Click on DataRobot Model Docs for even more detail. 

Close the Blueprint view and open Feature impact. Click 'Compute' to request this computation. It will now be added to the queue, but may not be immediately visible if other jobs are ahead of it. Let's request a few other computations and then come back to explore. Open Feature effects and click 'Compute'. Then go to Accuracy Over Time, scroll down a bit and click 'Compute for training'. And lastly go to Series Insights, scroll down a bit and click the compute action above the table. 

While those computations happen in the background, click on 'View experiment info'. Here's your setup meta data, your original data, and the derived modeling data. Notice that DataRobot has created 107 new features. These include both lags and aggregations from your starting data. Here's the first lag of the Avg_Incentive meaning what was the value the month prior. And here's the 12-month maximum of the Avg_Incentive. 

The full derivation log is available for viewing and downloading. On the Feature list tab you'll see the automatically created feature lists. DataRobot knows which algorithms require differencing and which do not and creates all of these features for you automatically. In addition to these automatically generated feature lists you can also create your own.

Select the features that should be in the list, give it a name, and hit 'Save'. We won't do that right now because we don't need it. The Blueprint repository shows you all of the time series blueprints that are relevant for your use case. If there's a specific technique you know you want to try you can search for it by name. Let's hope that most of our requested calculations are done now, and I'm going to use the star to find the model that I was working on. 

Let's take a look at the Feature impact. This chart shows the relative contribution of the original features. You can also view the relative contribution of all of the derived features. This chart is for all 10 makes and models of cars in both segments, but you can create your own Slices to get a finer grained view. Click the menu next to 'Data slice' and then 'Create slice'. I'll create a slice for the pickup trucks. Select Major_Segment equals Pickup. When creating slices you can combine multiple logical operators. Let's save this slice. 

Now from the Slices menu I can request the chart for just pickup trucks. I do need to cue that computation and then I can come back and view it. Let's move on to Feature effects. Here you can view how a feature like the Commodity Price Index effects sales volume, or the Month of the year. I can put these in order by Effect size and see that the fewest cars are sold in January and the most are sold in August and December. Feature effects are also available by slice, but need to be requested. 

Let's go down to Accuracy Over Time. Here you can see the predicted values in blue and the actual values in orange for the training period and the validation period. Select any specific series from the drop-down menu and zoom in on any specific time period. Now let's look at the Series Insights. This column Backtest 1 shows you the error for each series. Currently the error is the lowest for the Toyota Camry and it's the highest for the Ford Focus. It's helpful to put errors in context with the actual sales volume. If the vehicles with very high sales volumes are among the ones with the lower errors you're in pretty good shape.

We spent some time exploring this particular model, but let's take a look at how the entire experiment is shaping up. At this point 14 models are complete and several are still building. More recent additions to the Leaderboard have improved upon the error for my starred model which is now near the bottom of the list. 

Keep exploring what's currently at the top by requesting those additional insights, or simply wait for the entire Experiment to finish. You can leave this view and go back to the Use Case and return to this Experiment at any time. You can also go do other work in DataRobot or exit the application entirely and your Experiment will keep going. 

While the automation in that first Experiment continues, let's move on to the Notebooks section to bring even more automation to this process. From the Notebooks tab of the Use Case folder, select Upload notebook and navigate to the notebook file that came along with this tutorial. Once you find the correct file, click Import. 

This is a lightweight notebook using basic Python functions, so the small resource type and the default environment will be fine. You can start the container from either of these two sliders. Once the container is started, go ahead and run the entire notebook. As this starts running let me orient you to what it contains. 

We'll connect to DataRobot. We'll import data as a pandas data frame and quickly examine the contents. Then we'll configure the time series experiment using code and we'll set it up in a way that matches what we did earlier in the GUI. We'll create and start a time series Experiment using all of the vehicles in the data frame. 

And then using a simple loop in Python, we'll start an Experiment for each segment of vehicle. This first cell imports datarobot as dr and a couple other common data science libraries. This is the easiest way to connect to DataRobot when you're working in a DataRobot notebook. 

This video uses the _FAST version of the dataset which has fewer vehicles and fewer segments. If you'd like to try it with all of the vehicles, use this data path instead.

This is a pandas data frame and we're viewing the top five rows. This video uses the smaller data set that has 10 unique vehicles in two major segments. The cell below selects five of the 10 at random and plots them so you can see roughly what the data looks like shapewise. Here we provide all of the Experiment settings. The target is Sales_Volume. The name of the date column is 'Date'. Known in advance is 'Brand' and 'Major_Segment'. The multi-series column ID is 'Model'. Feature derivation will be from -13 to -1. And the forecast window is from 2 to 4 months in the future. In addition to these there are other specifications possible as part of DatetimePartitioningSpecification(). 

Here we run the Experiment for all vehicles. This only creates a name, and then this one actually creates the Project. Note that Experiment and Project are interchangeable terms. In the GUI you'll see 'Experiment' and in the API you'll see 'Project'. This command analyze_and_model() is equivalent to hitting the start button. At this point your initial settings are locked and Autopilot takes over. This notebook uses Quick Autopilot with all available workers to make the process go as fast as possible. This process is still in progress so we can actually switch over to the GUI to see it happening. 

Here's the Experiment that was created 4 minutes ago, and we can click into it. This experiment has a few remaining jobs to complete. You can start to interact with any of these models and it will not interfere with the automation nor the notebook in any way. Let's go back to the Notebook. This data set only has two major segments - pickup and car. We create a dictionary named projects to save all of the projects one for each major segment. Then we execute this for loop for every unique value in vehicle segments. In this case we'll do it twice - once for cars and once for pickups. Here we subset the data by the Major_Segment then we give that new experiment a name. This creates the Experiment for that segment and this actually starts it using the specifications we had before in ts_specs.

There's a print statement here to tell us which project DataRobot is currently working on, how many vehicles there are, and how many vehicles in that segment. Here you see the time stamp for creating the pickup Experiment, and about 2 and a half minutes later, the cars Experiment is created. Now all there is to do is wait for all the models inside all the Experiments to complete. 

Again the automation is in charge and you can follow along in the GUI without interrupting any of the processes. Here's the newly created pickup and car Experiment. If I click inside pickups I can see exactly what's happening. Anything with a completed number is ready to interact with. 

Going back to the notebook, I hope you can see how the power of DataRobot's experiment automation and the power of python loops can allow you to accomplish a lot of exploratory work very quickly. In addition to looping by segment, you could loop for individual models, you could loop for individual forecast distances, and you could loop for different feature derivation windows. 

Visit docs.datarobot.com and search for 'model factory' for additional ideas and examples. We hope this walkthrough prepared you well to start on your own time series forecast and we would greatly appreciate you rating the quality of this learning content. Thank you and happy experimenting!

Note, that there may be slight differences in text and video. Watch video for the original content.

DataRobot Video