Knowledge Chat Bot (No Code)

Recently my customers was looking for a solution where on-field guys can search for answers related to their tasks and it was a perfect scenario to develop a Chat Bot. Microsoft Bot Framework has been available for quite some time now and its functionality is keep on getting enhanced with different aspect of Artificial Intelligence like speech recognition, face recognition etc. With yesterday's announcement of GA of Azure Bot Services, now we have one place to develop, deploy and expose our intelligent bots to channels like Web, Skype, FB Messenger etc.

In this post, I will walk you through Azure Bot Services(https://docs.microsoft.com/en-us/Bot-Framework) to build and connect a bot but first we need to build our bot's brain and fill it with the knowledge using QnAMaker( http://qnamaker.ai).

This post is divided into following three sections:

1. Architecture
2. Building the Knowledge Base (KB)
3. Building the Chat Bot

Note: You need Azure subscription to build this bot.

Architecture

---

Following is the architecture of our solution where we will build a service (bot's brain) in QnA Maker and then fill it with our Knowledge Base. Then we will create our bot in Azure Bot Services and link it to the bot's brain (MyBotService) and then we can publish our bot via different channels. We can use Channels to display our bot inside a SharePoint Page, Skype, Teams, Facebook Messenger etc.

Building the Knowledge Base (KB)

---

1. Log on to QnA Maker at https://qnamaker.ai.
2. Select Create New Service tab and provide name of the bot service e.g. MyBotService, leave remaining fields as it is and hit Create button to create the service which will serve as brain of the bot.
   
3. You should see the screen where you can provide the knowledge base for your bot. Add couple of questions and their answers by selecting + Add QnA Pair. Once you add few QnA pairs, hit Save and retrain to train your model. You can also provide the questions and answers in text, pdf and document format.
   
4. You can select Test tab to test the knowledge base in chat format.
5. Hit Publish once testing is done, and then hit publish again on the next screen to publish MyBotService to the web. We will consumed this service through our Bot.
6. Once the MyBotService is published, Save the two values highlighted in the screen shot below as QnAKnowledgebaseId and QnASubscriptionKey. We need this information in our Bot's settings.
   

Note: You can access further documentation about QnA Maker at https://qnamaker.ai/Documentation.

Building the Chat Bot 

---

1. Log on to Azure at http://portal.azure.com.
2. Select New and go to AI + Cognitive Services section and choose Web App Bot.
   
3. Provide a unique Bot Name (e.g. DemoBot101), select Subscription and then choose Bot template as Question and Answer. You can also choose between C# or Node.js SDK which doesn't matter in this case as we are not writing any code. Select Create button at the end to create the Bot.
   
4. Once the basic bot has been created you will be notified and select Go to resources from the notification section to see the Bot's details, settings etc.
   
5. Select Application Settings of the bot and go to the App settings section and provide the value of the keys QnAKnowledgebaseId and QnASubscriptionKey which we copied while creating the MyBotService and hit Save.
   
6. Now you can test the bot from Test in Web Chat tab.
   
7. Once the testing is completed, this Chat Bot can be published to web, Skype, Teams, MSN Messenger etc via Channels.
   

Note: You can access Azure Bot Services documentation here: https://azure.microsoft.com/en-us/services/bot-service

Predicting Parole Violators

Summary: In this experiment, we will build a model that will predict for an inmate that if he will commit crime when released on parole.

AzureML Gallery: https://gallery.azureml.net/Details/eae76138f1f84dfeb7bfb6e9ea7dfa26

Description:
Parole board in Criminal Justice System, analyze an inmate's application (along with behavior, history etc) and decide to release an inmate on parole or deny the application. This type of model can be very helpful for the parole board.

The dataset used in this experiment is taken from 15.071x The Analytics Edge, a course taught by MIT, which is a subset of the data provided by National Corrections Reporting Program. This model uses Logistic Regression (also known as Logit Regression) for prediction. We will try to predict "violator" column's value which can be 1 (if parolee violated the parole) or 0 (if parolee did not committed any crime while on parole). Since the outcome of our predictive model is either 0 or 1 we can apply "Two-Class Logistic Regression" algorithm to our dataset.

Dataset:
male: 1 if the parolee is male, 0 if female
race: 1 if the parolee is white, 2 otherwise
age: the parolee's age (in years) when he or she was released from prison
state: a code for the parolee's state. 2 is Kentucky, 3 is Louisiana, 4 is Virginia, and 1 is any other state.
time.served: the number of months the parolee served in prison (limited by the inclusion criteria to not exceed 6 months).
max.sentence: the maximum sentence length for all charges, in months (limited by the inclusion criteria to not exceed 18 months).
multiple.offenses: 1 if the parolee was incarcerated for multiple offenses, 0 otherwise.
crime: 2 is larceny, 3 is drug-related crime, 4 is driving-related crime, and 1 is any other crime.
violator: 1 if the parolee violated the parole, and 0 if the parolee completed the parole without violation.

Steps:

1. Drag the parole.csv from Saved Datasets. 
2. Split the datset into 70% Training data and 30% Test data. Make the random seed to 144.
3. Fill the 70% training data into Train Model and select the column "violator" as we are trying to predict this column's outcome.
4. Select "Two-Class Logistic Regression" from Machine Learning section and apply it to the Training Model. Leave default values as it is.
5. Connect the outcome of the Train Model to the Score Model that also takes 30% of the test data to apply the algorithm.
6. Lastly we have to evaluate the Score Model using Evaluate Model to determine the accuracy of our model.
7. Select Visualize from the output node of the Train Model and you will see the Accuracy of this model is determined as 0.906 which is pretty awesome. Remember perfect accuracy is 1. Area Under the Curve (AUC) has been calculated as 0.882 (with 0.5 threshold value) which also shows that this is a pretty strong model for parole prediction. 

Ref: http://www.icpsr.umich.edu/icpsrweb/NACJD/series/38/studies/26521?archive=NACJD&sortBy=7
Ref: https://www.edx.org/course/v2/analytics-edge-mitx-15-071x-0

Predictive Analytics with AzureML

I started exploring AzureML(Azure Machine Learning) few weeks back and quickly fell in love with its simplicity and robustness.

I grabbed the sample data of Down Jones Index from UC Irvine Machine Learning Repository and applied the Linear Regression algorithm to create a prediction model to estimate the future values of Microsoft stock's opening weekly price (so that I can be rich) and here how my model looks like in AzureML.

First I am removing the entire rows with missing values from the data. Then I am applying the filter for MSFT symbol in the first split and I am dividing the data to 80-20 ratio to train the actual model on 80% of the data with the help of Linear Regression algorithm. After that I am trying to predict price variable in Train Model and verifying it using 20% of remaining data. In the last, I am evaluating the model that how effective and reliable it is.

At this point I need to seriously improve my model using other algorithms, removing/adding new variables etc because the Coefficient of Determination is nowhere closer to 1 and Mean Absolute Error, Root Mean Squared Error, Relative Absolute Error & Relative Squared Error are very high. But that's how a prediction model (more or less) will eventually look like in AzureML. It can be published as a web service with few clicks.

I have published this experiment/source to AzureML gallery and can be accessed here:

https://gallery.azureml.net/Details/3f4d92649bfa4fa3bf4b0c93a3635227

My next step would be to grab data from SharePoint lists and apply some prediction algos on it.