Building a Production-ready Predictive App for Customer Service - Alex Ingerman @ PAPIs Connect

© 2015, Amazon Web Services, Inc. or its Affiliates. All rights reserved.

Alex IngermanSr. Manager, Tech. Product Management, Amazon Machine Learning

1/28/2016

Building a Production-Ready Predictive Customer Service Application with AWS

Agenda

• What is predictive customer service?

• Using machine learning to find important social media conversations

• Building an end-to-end application to act on these conversations

Application details

Goal: build a smart application for social media listening in the cloud

Full source code and documentation are on GitHub: http://bit.ly/AmazonMLCodeSample

Amazon Kinesis

AWSLambda

Amazon Machine Learning

AmazonSNS

Amazon Mechanical Turk

Motivation for listening to social media

Customer is reporting a possible service issue


Customer is making a feature request


Customer is angry or unhappy


Customer is asking a question

Why do we need machine learning for this?

The social media stream is high-volume, and most of the messages are not CS-actionable

Amazon Machine Learning in one slide

• Easy to use, managed machine learning service built for developers

• Robust, powerful machine learning technology based on Amazon’s internal systems

• Create models using your data already stored in the AWS cloud

• Deploy models to production in seconds

Formulating the Business Problem

Formulating the business problem

We would like to…

Instantly find new tweets mentioning @awscloud, ingest and analyze each one to predict whether a customer service agent should act on it, and, if so, send that tweet to customer service agents.

Establishing the Data Flow

Establishing the data flow

We would like to…


Twitter API


We would like to…


Twitter API Amazon Kinesis


We would like to…



AWSLambda


We would like to…



AWSLambda



We would like to…



AWSLambda


AmazonSNS

Picking the machine learning strategy

Picking the machine learning strategy

Question we want to answer: Is this tweet customer service-actionable, or not?

Our dataset: Text and metadata from past tweets mentioning @awscloud

Machine learning approach:Create a binary classification model to answer a yes/no question, and provide a confidence score

Obtaining the data

Retrieve past tweets

Twitter API can be used to search for tweets containing our company’s handle (e.g., @awscloud)

import twitter

twitter_api = twitter.Api(**twitter_credentials)

twitter_handle = ‘awscloud’

search_query = '@' + twitter_handle + ' -from:' + twitter_handle

results = twitter_api.GetSearch(term=search_query, count=100, result_type='recent’)

# We can go further back in time by issuing additional search requests

Retrieve past tweets

Twitter API can be used to search for tweets containing our company’s handle (e.g., @awscloud)

import twitter

twitter_api = twitter.Api(**twitter_credentials)

twitter_handle = ‘awscloud’

search_query = '@' + twitter_handle + ' -from:' + twitter_handle

results = twitter_api.GetSearch(term=search_query, count=100, result_type='recent')

# We can go further back in time by issuing additional search requests

Good news: data is well-structured and cleanBad news: tweets are not categorized (labeled) for us

Labeling past tweets

Why label tweets? (Many) machine learning algorithms work by discovering patterns connecting data points and labels

How many tweets need to be labeled? Several thousands to start with

Can I pay someone to do this? Yes! Amazon Mechanical Turk is a marketplace for tasks that require human intelligence

Setting up the machine learning model

Amazon ML process, in a nutshell

1. Create your datasourcesTwo API calls to create your training and evaluation dataSanity-check your data in service console

2. Create your ML modelOne API call to build a model, with smart default or custom setting

3. Evaluate your ML modelOne API call to compute your model’s quality metric

4. Adjust your ML modelUse console to align performance trade-offs to your business goals

Aligning the ML model with business requirements

Building the data flow

Reminder: Our data flow


AWSLambda


AmazonSNS

Create an Amazon ML endpoint for retrieving real-time predictions

import boto

ml = boto.connect_machinelearning()

ml.create_realtime_endpoint(“ml-tweets”)

# Endpoint information can be retrieved using the get_ml_model() method. Sample output: #"EndpointInfo": {

# "CreatedAt": 1424378682.266,

# "EndpointStatus": "READY",

# "EndpointUrl": ”https://realtime.machinelearning.us-east-1.amazonaws.com",

# "PeakRequestsPerSecond": 200}


AWSLambda


AmazonSNS

Create an Amazon Kinesis stream for receiving tweets

import boto

kinesis = boto.connect_kinesis()

kinesis.create_stream(stream_name = ‘tweetStream’, shard_count = 1)

# Each open shard can support up to 5 read transactions per second, up to a

# maximum total of 2 MB of data read per second. Each shard can support up to

# 1000 records written per second, up to a maximum total of 1 MB data written

# per second.


AWSLambda


AmazonSNS

Set up AWS Lambda to coordinate the data flow

The Lambda function is our application’s backbone. We will:

1. Write the code that will process and route tweets2. Configure the Lambda execution policy (what is it allowed to do?)3. Add the Kinesis stream as the data source for the Lambda function


AWSLambda


AmazonSNS

Create Lambda functions


AWSLambda


AmazonSNS

// These are our function’s signatures and globals only. See GitHub repository for full source.

var ml = new AWS.MachineLearning();

var endpointUrl = '';

var mlModelId = ’ml-tweets';

var snsTopicArn = 'arn:aws:sns:{region}:{awsAccountId}:{snsTopic}';

var snsMessageSubject = 'Respond to tweet';

var snsMessagePrefix = 'ML model '+mlModelId+': Respond to this tweet: https://twitter.com/0/status/';

var processRecords = function() {…} // Base64 decode the Kinesis payload and parse JSON

var callPredict = function(tweetData) {…} // Call Amazon ML real-time prediction API

var updateSns = function(tweetData) {…} // Publish CS-actionable tweets to SNS topic

var checkRealtimeEndpoint = function(err, data) {…} // Get Amazon ML endpoint URI

Create Lambda functions


AWSLambda


AmazonSNS

// These are our function’s signatures and globals only. See GitHub repository for full source.

var ml = new AWS.MachineLearning();

var endpointUrl = '';

var mlModelId = ’ml-tweets';

var snsTopicArn = 'arn:aws:sns:{region}:{awsAccountId}:{snsTopic}';

var snsMessageSubject = 'Respond to tweet';

var snsMessagePrefix = 'ML model '+mlModelId+': Respond to this tweet: https://twitter.com/0/status/';

var processRecords = function() {…} // Base64 decode the Kinesis payload and parse JSON

var callPredict = function(tweetData) {…} // Call Amazon ML real-time prediction API

var updateSns = function(tweetData) {…} // Publish CS-actionable tweets to SNS topic

var checkRealtimeEndpoint = function(err, data) {…} // Get Amazon ML endpoint URI

Configure Lambda execution policy


AWSLambda


AmazonSNS

{ "Statement": [

{ "Action": [ "logs:*” ],

"Effect": "Allow",

"Resource": "arn:aws:logs:{region}:{awsAccountId}:log-group:/aws/lambda/{lambdaFunctionName}:*"

},

{ "Action": [ "sns:publish” ],

"Effect": "Allow",

"Resource": "arn:aws:sns:{region}:{awsAccountId}:{snsTopic}"

},

{ "Action": [ "machinelearning:GetMLModel”, "machinelearning:Predict” ],

"Effect": "Allow",

"Resource": "arn:aws:machinelearning:{region}:{awsAccountId}:mlmodel/{mlModelId}”

},

{ "Action": [ "kinesis:ReadStream”, "kinesis:GetRecords”, "kinesis:GetShardIterator”, "kinesis:DescribeStream”,"kinesis:ListStreams” ],

"Effect": "Allow",

"Resource": "arn:aws:kinesis:{region}:{awsAccountId}:stream/{kinesisStream}"

}

] }



AWSLambda


AmazonSNS

{ "Statement": [


"Effect": "Allow",


},


"Effect": "Allow",


},


"Effect": "Allow",


},


"Effect": "Allow",


}

] }

Allow request logging in Amazon

CloudWatch



AWSLambda


AmazonSNS

{ "Statement": [


"Effect": "Allow",


},


"Effect": "Allow",


},


"Effect": "Allow",


},


"Effect": "Allow",


}

] }

Allow publication of notifications to

SNS topic



AWSLambda


AmazonSNS

{ "Statement": [


"Effect": "Allow",


},


"Effect": "Allow",


},


"Effect": "Allow",


},


"Effect": "Allow",


}

] }

Allow calls to Amazon ML

real-time prediction APIs



AWSLambda


AmazonSNS

{ "Statement": [


"Effect": "Allow",


},


"Effect": "Allow",


},


"Effect": "Allow",


},


"Effect": "Allow",


}

] }

Allow reading of data from

Kinesis stream

Connect Kinesis stream and Lambda function

import boto

aws_lambda = boto.connect_awslambda()

aws_lambda.add_event_source(

event_source = 'arn:aws:kinesis:' + region + ':' + aws_account_id + ':stream/' + “tweetStream”,

function_name = “process_tweets”,

role = 'arn:aws:iam::' + aws_account_id + ':role/' + lambda_execution_role)


AWSLambda


AmazonSNS

End-to-end flow

Amazon ML real-time predictions test

Here is a tweet:


Here is the same tweet…as a JSON blob:

{

"statuses_count": "8617",

"description": "Software Developer",

"friends_count": "96",

"text": "`scala-aws-s3` A Simple Amazon #S3 Wrapper for #Scala 1.10.20 available : https://t.co/q76PLTovFg",

"verified": "False",

"geo_enabled": "True",

"uid": "3800711",

"favourites_count": "36",

"screen_name": "turutosiya",

"followers_count": "640",

"user.name": "Toshiya TSURU",

"sid": "647222291672100864"

}


Let’s use the AWS Command Line Interface to request a prediction for this tweet:

aws machinelearning predict \

--predict-endpoint https://realtime.machinelearning.us-east-1.amazonaws.com \

--ml-model-id ml-tweets \

--record ‘<json_blob>’


Let’s use the AWS Command Line Interface to request a prediction for this tweet:

aws machinelearning predict \

--predict-endpoint https://realtime.machinelearning.us-east-1.amazonaws.com \

--ml-model-id ml-tweets \

--record ‘<json_blob>’

{"Prediction": {

"predictedLabel": "0", "predictedScores": {

"0": 0.012336540967226028}, "details": {

"PredictiveModelType": "BINARY", "Algorithm": "SGD"

}}

}

Generalizing to more feedback channels

Amazon Kinesis

AWSLambda

Model 1 AmazonSNS

Model 2

Model 3

What’s next?

Try the service:http://aws.amazon.com/machine-learning/

Download the Social Media Listening application code: http://bit.ly/AmazonMLCodeSample

Get in [email protected]

Thank you!

Building a Production-ready Predictive App for Customer Service - Alex Ingerman @ PAPIs Connect

Technology