Lab - Deploying a Predictive Model

July 6, 2015

383 views
100 downloads

Algorithms

Report Abuse
This lab explores unsupervised learning in Azure Machine Learning and how to deploy a predictive model as a web service. The lab will walk through copying an experiment from the Azure Machine Learning Gallery into the ML Studio, creating a scoring experiment, deploying a model as a web service, and interacting with the API using the included web interface.
“Where should I open my next restaurant location?” This question is often very difficult to answer. The right choice could lead to increased revenue and profit, but the wrong choice could lead to losing a major investment. Trying to make this decision by manually sifting through hundreds or even thousands of possible cities or neighborhoods can be almost impossible. Machine learning can help with this task by analyzing large volumes of data about different locations, finding common characteristics among locations, and grouping those like-attributed locations together. These groups can then be compared to previously successful restaurant locations to help narrow the choices for where to open next. In this lab, you will work with a dataset that includes geographic, economic, and demographic data about different US cities. The model you will explore uses a K-Means algorithm to cluster cities into distinctive buckets. K-Means Clustering is an unsupervised learning approach that uses iterative techniques to analyze cases and then group cases with similar characteristics together. These groups can then be used for labeling, exploration, identifying anomalies, or even further prediction. **Learning Objectives** Upon completing this lab, you will have hands-on experience with the following functions and concepts related to Azure Machine Learning: • Copying an experiment from the Azure Machine Learning Gallery to the ML Studio • Normalizing data • Creating a Scoring Experiment • Deploying a model as a web service o Modifying Web service inputs and outputs • Testing the web service via the web UI Lab Requirements/Prerequisites • A Microsoft account is required to access an Azure Machine Learning workspace. If you don’t already have a Microsoft account, you can obtain one for free by following the link below: https://www.microsoft.com/en-us/account/default.aspx  **Copy Experiment from Azure Machine Learning Gallery** Next, you will go to the Azure Machine Learning Gallery to find an experiment that has already been created for you, and copy it into your workspace. 1. Click the Gallery link at the top of the workspace. The Azure Machine Learning Gallery website will open. The gallery is Microsoft’s community portal for Azure Machine Learning. Here, users can upload experiments and share tutorials and comments with each other. 2. In the search box near the top of the website, type mslabs and hit Enter. 3. Find the experiment titled City-Market Clustering and click either the title or picture on the experiment. This will open the Experiment page in the gallery. Notice the user who contributed the experiment, a Summary and Description for the experiment, and some metadata about the experiment in the panel on the right. 4. Click the OPEN IN STUDIO button found in the panel on the right side of the screen. 5. In the Copy experiment from Gallery dialogue box, choose the workspace from the dropdown box you want to copy the experiment into (this should be the same workspace you logged into at the beginning of the lab). 6. Click the button on the Copy experiment from Gallery dialogue box. After a few moments, the experiment you selected in the Gallery will open in ML Studio under the workspace you selected. This is a copy of the experiment from the gallery, so you are free to run/modify it as you would like. It will also be saved as a training experiment in the workspace. 7. Click RUN at the bottom of the page to execute the training experiment. After the experiment finishes executing, a CREATE SCORING EXPERIMENT dialogue box might be displayed. 8. If the CREATE SCORING EXPERIMENT dialogue box is displayed, click the X in the top right corner of the CREATE SCORING EXPERIMENT dialogue box to close it. 9. Click the SAVE AS button at the bottom of the page. 10. In the Save As dialogue box, change the Experiment name to Lab – Deploying a Predictive Model, and click the button. The ML Studio canvas should now look similar to the below image.   **Explore the Training Experiment** Visualize the Input Dataset Next, you will take a quick look at the dataset used for the experiment. 1. Click RUN at the bottom of the page to execute the training experiment. 2. Once the experiment finishes running, click the output port on the Market Data dataset and select Visualize from the displayed menu Notice each row in the data set represents a location and there are a series of geographic, economic, and demographic features describing each location. 3. Click the X in the top right corner of the Visualize dialogue box to close it. 4. Click the left output port on the Normalize Data module and select Visualize from the displayed menu. Notice each of the features in the dataset has been scaled to a value between -1 and 1. Normalizing features is the process of setting all of the numeric features in a dataset to a similar scale (usually somewhere between -1 and 1). This often prevents features with larger scales from dominating influence in the model and typically makes features easier to visualize. Having normalized data is necessary for K-Means Clustering. 5. Click the X at the top right of the Visualize dialogue box to close it. **Explore the Clustering Model** The next series of modules trains the K-Means Clustering model. New input data is then assigned to the clusters using the trained model. 1. Click the K-Means Clustering module on the canvas. Notice the different parameters that can be configured in the Properties pane, including the number of clusters (centroids) to be created. The next module in the data flow, Train Clustering Model, will train the clustering model using the K-Means algorithm and the input dataset. Next, records are assigned to clusters using the Assign to Clusters module. 2. Click the output port on the Assign to Clusters module and select Visualize on the displayed menu. The visual that is displayed shows the size and relative position among features of each of the clusters. 3. Click the X in the top right corner of the Visualize dialogue box to close it. 4. Click the output port on the Convert to Dataset module and select Visualize from the displayed menu to view how each individual city-market was clustered. The Assignments column has the numeric value for the assigned cluster for each row in the data set. This numeric value does not provide any context into what type of cities/locations are in each cluster. At this point, you would typically download the assigned dataset to a tool like Excel to analyze and try to label each cluster. Upon doing this, you would likely find clusters for average/established cities, quickly growing cities, extremely large and dense cities, and cities that are very geographically spread out. 5. Click the X in the top right corner of the Visualize dialogue box to close it.   **Publish a Model as a Web Service** Create a Scoring Experiment from the Training Experiment Once you are satisfied with the results of a training experiment, you need to convert that training experiment into an experiment that is optimized for scoring/predicting new data. For this, you will create a Scoring Experiment. 1. Click RUN at the bottom of the page to execute the training experiment. 2. Once the experiment finishes executing, click CREATE SCORING EXPERIMENT at the bottom of the page. After you click CREATE SCORING EXPERIMENT, a message box will appear at the bottom of the screen that says Creating scoring experiment. Once the scoring experiment has been created, the NEW EXPERIMENT dialogue box will be displayed near the top of the canvas. 3. Click the X in the top right corner of the NEW EXPERIMENT dialogue box to close it. Notice a few things: a. There are now 2 tabs at the top of the ML Studio canvas: one for the Scoring experiment, and one for the Training experiment. b. The modules for training the model have all been replaced with a new module representing the trained clustering model. The new trained model module feeds the Assign to Clusters module. c. Modules for Web service input and Web service output have been added to the scoring experiment. These modules indicate where data would be passed into and out of an experiment that is published as a web service. 4. Click RUN at the bottom of the page to execute the experiment. The experiment will execute with the data set that is attached as the input data set (which is the same data as the Training experiment). It will score all of the rows using the trained clustering model. 5. If the PUBLISH WEB SERVICE dialogue box pops up, close it by clicking the X in the top right corner. **Publish the Scoring Experiment as a Web Service** Next, you will publish your scoring experiment as a web service so that predictions can be made from other applications outside of the ML Studio. 1. Click PUBLISH WEB SERVICE at the bottom of the page. The web service is created, and you now see the web service dashboard. Here you will find: a. Links to the latest experiment that was used to create the web service b. A description of the web service c. The API key used for authentication when other applications call the web service d. Other information about the Default Endpoint including: a. help pages for the REQUEST/RESPONSE and BATCH EXECUTION APIs b. a TEST option c. an Excel Workbook to download d. a data indicating when the endpoint was LAST UPDATED **Make Predictions using the Web Service** Test the API using the Web Interface Finally, we will test our API using the provided Web UI on the web service dashboard page. 1. Click the Test button. 2. Enter the following values in their corresponding text boxes on the Enter data to predict dialogue box: LOCATION: Houston, TX GINIINDEX: 0.4802 POP2014: 2239558 LANDAREA: 599.6 POPULTREND: 6.63 POPDENSITY: 3735 3. Click the at the bottom of the Enter data to predict dialogue box. Notice a status bar at the bottom of the page is displayed indicating the web service call is being made. Once the web service call is complete, the status bar will have a green checkmark next to the web service name along with the returned prediction from the API. 4. Click DETAILS on the status bar. The status bar will expand to show the entire result body that was returned from the API including the cluster number. 5. Click CLOSE on the status bar to return to the web service dashboard. 6. Sign out of your workspace by clicking the profile picture at the top right of the page and selecting Sign Out from the displayed menu. **Conclusion** This concludes the Deploying a Predictive Model with Azure Machine Learning lab. To recap, you have successfully copied an experiment from the Azure Machine Learning Gallery into ML Studio, created a scoring experiment, deployed a model as a web service, and interacted with the API via the included web interface. Now that you have a deployed web service, you can use the provided web UI or the provided Excel Worksheet to predict with new data. You can also create custom applications and data flows that access the API for automated prediction. To answer our earlier question of where to open your next restaurant location, you would start by downloading the cluster assignments into a tool like Excel to try to label the type of city-market each cluster represents. Then you could take those cluster labels and compare them to your restaurant locations that have done well in the past. This would give you a list of prime city-markets for future restaurant locations.