×
Community Blog Friday Blog - Week 34 - Detecting Diabetes With PAI Studio

Friday Blog - Week 34 - Detecting Diabetes With PAI Studio

This week we're taking a look at PAI Studio, a drag-and-drop machine learning tool that makes it easy for anybody to build and train machine learning models.

By Jeremy Pedersen

Welcome back! This week we're taking a look at PAI Studio, a drag-and-drop machine learning tool that makes it easy for anybody to build and train machine learning models.

Detecting Diabetes With PAI Studio

What Is PAI Studio?

PAI Studio is a drag-and-drop tool for building Machine Learning systems. It's a built-in part of Alibaba Cloud's Machine Learning Platform for AI (also called "PAI"), alongside other tools like PAI DSW (a Jupyter notebook service) and PAI EAS (a model deployment service).

The idea behind PAI Studio is to make it easy for people who are new to Machine Learning to get started training and deploying models, with little to no coding required.

PAI studio lets you build Machine Learning workflows by dragging-and-dropping components onto a dashboard, like this one:

01_workflow

Of course, the best way to understand PAI Studio is to try it for yourself. The rest of this blog post will take you through the steps needed to build and train your own Logistic Regression model to detect diabetes.

We'll use sample data from Kaggle, which I will provide links to later.

Ok, Let's Get Started!

First, a little background: PAI Studio runs its machine learning tasks on top of Alibaba Cloud's batch computing engine, MaxCompute.

MaxCompute itself is just a compute engine. To interact with MaxCompute, we use another tool called DataWorks.

Thus, to start using PAI Studio, we need to:

  1. Create a new DataWorks Workspace
  2. Choose MaxCompute as our Compute Engine
  3. Link DataWorks and MaxCompute to PAI
  4. Switch to the PAI Console, and enter PAI Studio

Although this might seem complicated, it's quite easy to do from the Alibaba Cloud web console. Follow along with the steps below to create your own PAI Studio project.

Note: You should use the "basic edition" of DataWorks, which is free, and the Pay-As-You-Go version of MaxCompute. This will ensure that you are only charged when your model is being actively trained, and even then the fees will be very low, since our example dataset is so small. Running all the steps in this blog should cost you well below 1.00 USD.

Creating A New Project

First, we need to create a DataWorks Workspace:

02_open_search

03_find_dataworks

04_choose_region

05_choose_region

06_create_workspace

07_create_workspace

08_create_workspace

09_create_workspace

That's it! We have now created our DataWorks Workspace and MaxCompute project. The next step is to switch over to the PAI Studio console and create a new Experiment:

10_open_pai

11_open_pai

12_open_pai

13_open_pai

14_open_pai

15_open_pai

Importing Data

We need some data on which to train our model. One great source of openly available Machine Learning data is Kaggle.

Just in case you aren't familiar with Kaggle: it's a machine learning community site where people post open datasets and take part in competitions to see who can train the most accurate machine learning models.

What we will be doing in this walkthrough is training a logistic regression model. Logistic Regression models are a kind of binary classifier, meaning they divide a dataset into two classes.

This makes them great for scenarios where we want to use input data to answer a yes or no question like "should we give this person a loan?" or "does this person have cancer?"

In our case, we'll be trying to predict whether or not a given patient has diabetes, using this diabetes dataset from Kaggle.

You may need to create a Kaggle account to download the dataset: don't worry, it's free.

Once you've downloaded the dataset, open it up in a spreadsheet program like Excel or Numbers (or a text editor if you prefer - it's a .csv file after all). You should see something like this:

16_table_data

Note that you will need to remove the header row before passing the data to PAI. Take a minute to save a copy of the file with the header row removed, before moving on to the next step.

Preparing The Data

With our .csv file ready to go, the next step is to create a new table in PAI Studio and upload our data. Use the SQL DDL statement below:

create table diabetes_data (
    pregnancies                    bigint,
    glucose                        bigint,
    bloodpressure                bigint,
    skinthickness                bigint,
    insulin                        bigint,
    bmi                            double,
    diabetespedigreefunction    double,
    age                            bigint,
    outcome                        bigint
);

Here are the steps you'll need to take in the console:

17_create_table

18_create_table

19_create_table

20_create_table

Now that we've created our table, we need to add some nodes to our Machine Learning workflow. Specifically, we want to:

  1. Convert all the input data into floating point values
  2. Normalize all the input data
  3. Split the data into separate test and training sets (we'll use an 80-20 split)

We can do this easily by dragging and dropping blocks onto the dashboard:

21_create_workflow

22_create_workflow

23_create_workflow

24_create_workflow

25_create_workflow

26_create_workflow

27_create_workflow

28_create_workflow

29_create_workflow

30_create_workflow

We can now run the workflow to make sure all these nodes are configured correctly:

31_run

32_run

Great! If a green "check" mark appears next to all four nodes, then everything has run successfully. If one of your nodes fails (a red "x" will appear next to it), left click on the node and double check its settings (or right click and choose View Log to take a look at the logs and see if you can figure out what went wrong from there).

Setting Model Parameters

Now, we can start setting up our model. We're going to use a built-in "Logistic Regression" model that's already part of PAI Studio.

33_model

34_model

35_model

36_model

Training And Testing The Model

With our model configured, we'll also want to run some tests and see how accurate the model is. Todo that, we need to add a few more nodes to the workflow, like so:

37_prediction

38_prediction

Next, we add "Confusion Matrix" and "Binary Classifier Evaluator" nodes:

39_confusion

40_binary_class_eval

For each of these nodes, we need to indicate which column in our dataset is the "label" (i.e. which column indicates whether or not a patient really has diabetes). In this case, that is the outcome column:

41_outcome

42_outcome

Running It All

Finally, we need to train our model and then test its performance using the Prediction node and the two evaluation nodes, Confusion Matrix and Binary Classification Evaluation. To do this, we right click on the Logistic Regression node and choose "Run From Here", like this:

43_run

If everything runs successfully, green check marks will appear next to each node, like so:

44_done

We can now take a look at the confusion matrix as well as some of the charts (such as ROC and precision/recall) generated by the "Confusion Matrix" and "Binary Classification Evaluation" nodes. First, the Confusion Matrix:

45_confusion

46_confusion

Next, let's look at the ROC curve, and evaluate the AUC (area under curve) which gives us a rough idea how accurate the model is:

47_auc

48_auc

Not bad! This model is OK for a first try, with an accuracy of around 70% on the test data. We could play around with PAI's AutoML feature to adjust the parameters of our Logistic Regression model and try for better accuracy. You can learn more about AutoML in this blog post. We could also experiment with different models such as GBDT (Gradient Boosting Decision Tree) to see if we can get higher accuracy.

I've Got A Question!

Great! Reach out to me at jierui.pjr@alibabacloud.com and I'll do my best to answer in a future Friday Q&A blog.

You can also follow the Alibaba Cloud Academy LinkedIn Page. We'll re-post these blogs there each Friday.

Not a LinkedIn person? We're also on Twitter and YouTube.

0 0 0
Share on

JDP

72 posts | 133 followers

You may also like

Comments

JDP

72 posts | 133 followers

Related Products