Build and Deploy a Machine Learning Model using Azure ML Studio

Create a Machine Learning workspace

1. Go to **https://portal.azure.com**
2. In the top search bar, type “Machine Learning” and select Azure Machine Learning.
3. Click + Create → New workspace.

4.Fill in:

• Subscription:Select your active Azure subscription (e.g., Azure for Students).
• Resource group:Choose an existing resource group or create a new one to logically organize your project assets.
• Workspace name:Provide a unique and recognizable name for your workspace environment.
• Region: pick the one closest to you(e.g: Cntral india)

5.Leave Storage account, Key Vault, App Insights as auto-generated defaults.

6.Click Review + create → Create.

7.Wait 1–3 minutes for deployment to finish → click Go to resource → click Launch studio

8. Accessing the Azure Machine Learning Designer

To build our visual machine learning workflow, navigate to the left authoring menu inside the Azure Machine Learning Studio and select the Designer section. This workspace provides an interactive, drag-and-drop canvas that allows us to build, train, and deploy machine learning models sequentially without writing any background code.

Click cerate a new pipline using classic prebuilt components

9.create the data asset

Give a Name -> select the Tabuer format

10.Select Data Source: Choose the From local files option from the file ingestion menu to upload data directly from your computer.

11.Upload Dataset: Select the local source file named red_wine.csv, which contains the comprehensive chemical profiles of various red wines.

12. Data Preview: Now we can see all the rows, continuous features, and target parameters beautifully populated within the tabular preview table schema.

13.Progress to Next: Verify the column attributes, formats, and structural layouts within the data settings, and then click Next

14.Finalize Asset Creation: Click the Create button to securely save and register the dataset configuration as a static reference asset.

15. Canvas Injection: After creating the data asset, drag and drop your newly registered red_wine dataset component directly onto the empty visual Designer canvas workspace.

16. Handle Missing Values: Search for the Clean Missing Data module in the component asset library, drag it into the canvas workspace, and connect the dataset output to its input port to clean structural null parameters.

17. Integrate Normalization: Find the Normalize Data module in the asset panel, drag it down, and connect the cleaned data output port directly into its training interface.

18.Data Scale Transformation (Normalize Data)

To prepare the dataset for model training, the Normalize Data module was integrated into the pipeline. Since different physicochemical features in the wine dataset have varying ranges and magnitudes, MinMax Normalization was selected as the transformation method to scale all numeric values into a strict standard range between 0 and 1.

• Transformation Method: MinMax
• Configured Columns: The normalization transformation was explicitly applied to all 11 continuous independent chemical attributes: fixed acidity, volatile acidity, citric acid, residual sugar, chlorides, free sulfur dioxide, total sulfur dioxide, density, pH, sulphates, and alcohol.
• Target Label Exclusion: The target dependent variable, quality, was intentionally excluded from this transformation to preserve its original categorical structure for classification.

19.Data Splitting (Split Data): To robustly train and validate our predictive model, a Split Data module was introduced right after the normalization phase. Splitting the dataset prevents data leakage and ensures we can thoroughly evaluate how well our model generalizes to unseen instances. The configuration details are as follows:

• Splitting Mode: Split Rows
• Fraction of Rows in the First Output Dataset: 0.7 (This allocates 70% of the preprocessed data directly for training the machine learning model, while reserving the remaining 30% as a completely separate validation set for model scoring and evaluation).
• Randomized Split: True (This shuffles the sample rows randomly before partitioning to eliminate structural ordering biases)

20.Model Selection and Training (Train Model)

For predicting the wine quality grades, a supervised classification approach was selected. The pipeline utilizes the Multiclass Decision Forest algorithm, which is an ensemble learning method that builds multiple decision trees and combines their outputs to improve classification accuracy and control over-fitting.

The training phase was configured as follows:

• Algorithm: Multiclass Decision Forest (Ideal for handling multi-value categorical targets like wine quality scores).
• Training Input: The Train Model module takes the untrained algorithm on its left input and receives the 70% training subset from the Split Data module on its right input.
• Label Column Selection: The target column was set to quality, defining it as the dependent variable that the algorithm must learn to predict based on the 11 input chemical features.

21. select the quality as target class

22.Score and Evaluate Setup: Select the Score Model and Evaluate Model blocks from the component panel and drag them onto the workspace canvas. Connect the trained output weights and the independent 30% testing dataset partition into the input ports of the Score Model node, then forward those predictions directly into the Evaluate Model block.

23.Configuring and Submitting the Pipeline Job: Click the blue Configure & Submit button located at the upper right corner of the authoring interface to trigger the operational deployment setup wizard.

24.Create compute cluster

25. select compute name

26. Job Tracking and Monitoring: Go to the Jobs section in the left navigation sidebar and select your active training experiment model block. The user interface opens the live running graph dashboard where we can see the model running successfully as each node sequentially executes and updates its tracking state on the cloud canvas.

we can see the model running

27.Pipeline Execution and Verification

As shown this image, the entire machine learning workflow completed its execution successfully without any structural errors or failure points.

• Run Status: The pipeline state officially transitioned to Completed, confirming that Azure ML successfully provisioned the compute cluster, processed the modules, and compiled the performance metrics chronologically.
• Visual Tracking: Every node layout on the canvas — ranging from data imputation, feature normalization, and dataset splitting, down to model scoring and final evaluation — displays a green checkmark indicator denoting successful validation and runtime processing.

With the complete infrastructure built and executed, the trained Multiclass Decision Forest model is now fully ready to have its classification performance metrics and confusion matrix inspected directly from the output of the evaluation module.

28.Machine Learning Model Evaluation Metrics Summary

Inside the Azure Machine Learning Designer workflow dashboard, the computed values were visualized directly from the evaluation node asset stream as shown in the image below. The output summarizes the core validation counts across the structured test split:

• Overall Accuracy: 0.6375
• Micro Precision: 0.6375
• Macro Precision: 0.415722

The snapshot confirms that the underlying tabular structure contains 1 row of aggregated summary data spanning across 5 core evaluation columns, ensuring proper model validation setup inside the workspace studio environment.

Step 27: Optimizing and Comparing with an Alternative Model

As illustrated in the image below, the machine learning pipeline was further optimized by establishing a dual-model comparison framework within the Azure ML Designer canvas. Alongside the baseline Multiclass Decision Forest algorithm, a Multiclass Neural Network module was integrated into the workflow to evaluate predictive capabilities under identical environmental constraints. The preprocessed training subset (70%) was branched simultaneously to feed both individual training blocks (train_model and train_model_1), while the validation subset (30%) was routed evenly to score both concurrent networks. By piping the finalized prediction streams from both Score Model nodes directly into the left and right input ports of a single, unified Evaluate Model asset, the platform is configured to generate a side-by-side performance visualization matrix, making it seamless to evaluate and select the absolute best performing classification model for our dataset.

Step 28: Model Comparison Results and Analysis

As shown in the evaluation results visualization below, Azure ML Studio successfully compiled the comparative metrics for both algorithms, displaying 2 rows of structured performance data:

• Multiclass Decision Forest (Row 1): Achieved an Overall Accuracy of 0.6375 (63.75%) and a Macro Precision of 0.415722.
• Multiclass Neural Network (Row 2): Achieved an Overall Accuracy of 0.554167 (55.41%) and a Macro Precision of 0.225272.

The direct comparison reveals that the Multiclass Decision Forest ensemble method outperformed the Neural Network framework on this specific red wine dataset, providing higher baseline accuracy and better precision across the quality classification categories.