Whats Next for Machine Learning
4 likes1,731 views
The document presents a webinar on machine learning (ML) and artificial intelligence (AI), discussing their definitions, significance, and practical applications such as chatbot utilities and customer service optimization. It emphasizes the impact of big data and advancements in computational power on the development of these technologies. Additionally, it showcases case studies that illustrate the deployment of machine learning for enhancing customer engagement and lifetime value predictions.
Whats Next for Machine Learning
- 2. Hello! Andrew Van Aken Consultant, OgilvyOne Worldwide Laurie Close Global Brand Partnerships, OgilvyRED Michael McCarthy Senior Consultant, OgilvyOne Worldwide
- 3. What’s the weather like in your city? Tell us where you’re dialing in from!
- 4. Want this deck? It will be available for download shortly after the webinar on: slideshare.net/socialogilvy Ogilvy staff: It’s also on The Market! themarket.ogilvy.com Are you on the go? You can join our webinars on mobile, too! Download the GoToWebinar app from the App Store or Google Play
- 5. This Talk • We will demystify machine learning (ML) and artificial intelligence (AI) • Why now for ML and AI? • Ogilvy case studies
- 6. What is Machine Learning Machine learning gives “computers the ability to learn without being explicitly programmed.” -Arthur Samuel, 1959
- 7. Any Type of Data
- 8. Machine Learning Concept • Machine learning takes an input • to an output: David Ogilvy
- 9. How does it do it? x1 x2 x3 y 23 146 1 91 x1 x2 x3 y 23 146 68 163
- 11. Panda or Gibbon?
- 12. Soccer/Football Example Visitor Goals Score Visitor Goals Allowed Home Goals Scored Home Goals Allowed Outcome 2 3 1 4 0 3 3 1 2 1 5 6 2 1 1
- 15. All Models are Wrong • After the tree has been built, a calculation is done to show how accurate your model is • The algorithm will try its best to minimize the error
- 17. New Example Visitor Goals Score Visitor Goals Allowed Home Goals Scored Home Goals Allowed Outcome 1 2 4 2 ? 4 3 1 4 ? 3 4 1 1 ?
- 18. What is Artificial Intelligence “Artificial intelligence is whatever hasn't been done yet” -Larry Tesler, 1970
- 19. Is This AI? • A program that can beat anyone in chess? • A software service that can tell you the answer to almost any question? • A digital assistant? • C3PO?
- 20. Is This AI?
- 21. Is This AI?
- 22. Is This AI?
- 23. Is This AI?
- 24. Is This AI? • While not a universal definition, at Ogilvy we consider a main differentiation of AI versus Machine Learning to be the ability to “self-learn” or “self- update” • This is in terms of analytics techniques, while a different criteria might be applied to interactive marketing tools like ChatBots, etc.
- 25. What is an Example of AI? • Example 1: Autonomous Media Buying
- 26. What is an Example of AI? • Example 2: AI Generated Content
- 27. What is AI
- 28. WHY NOW?
- 29. Why Now • Big Data • Compute
- 30. Google Trends - Machine Learning
- 31. Corporations
- 32. Why Now? “90% of the data in the world has been created in the past two years” -IBM, 2017
- 33. Big Data
- 35. Enormous Data
- 36. But CPU’s are Slowing
- 37. Enter GPUs
- 38. Enter GPUs
- 39. But at a Cost • A single GPU can cost up to $10,000 and uses tremendous amounts of power • Facebook recently used 256 GPUs to train 40,000 images a second • Can rent on the cloud for cheaper
- 40. Where Next? • Do we just keep adding data and power? • Do we need new methods?
- 41. What do we Think! • It’s complicated…
- 42. CASE STUDIES
- 43. Text Mining -> Chatbot Text mining analysis to provide insights into best use of Chatbot functionality
- 44. The Challenge - Utility Client Social media customer service is a significant cost expenditure and usage continues to rise Competitors and businesses are implementing Chatbots, which are crucial to scaling customer service and making brand engagement more interactive Existing data around customer service conversations was insufficient to examine cost-effectiveness and feasibility of a Chatbot Business Case Landscape Existing Data
- 45. The Ask Process Social Media Data Analyze Recommend Utilize Machine Learning to Extract Key Topics from Text Data Provide Recommendations on Deploying a Chatbot
- 46. The Data CONVERSATIONS BY TYPE CONVERSATIONS BY SENTIMENT AVERAGE CONVERSATION LENGTH AVERAGE WORDS PER CONVERSATION 4.5 messages ~50
- 47. The Solution Topic Modeling (Non-Negative Matrix Factorization) Programming Language Data Science Platform Machine Learning Package In-line Coding and Visualizations Data Science Toolkit Matrix Representation d1 d2 d3 bi1 1 0 1 bi2 0 2 0 bi3 0 1 4 Text Conversations ----- --- ------ Matrix Factorisation to Derive Topic Vectors ----- --- ------ Summarize Key Topics 1 2 .. 3
- 48. Identifying Viral Tweets Text mining analysis revealed 28% of conservation activity could be directed away from customer care, with 6% related to viral or marketing activity. Revealed an opportunity for a heuristic or machine learning model to flag these tweets algorithmically. # # # # #
- 49. Extracting Key Phrases by Sentiment Pulling out the top phrases by positive and negative customer service conversations gave insight into potential flags for a Chatbot to either continue chatting or divert a customer to a representative.
- 50. Summarizing Customer Service Topics customer service, poor customer, service today, excellent customer, shocking customer, service advisor, worst customer Customer Service Seekers email address, change email, old email, send email, address received, details follows, got right, technical error Contact Us power cut, post code, red triangle, pls help, Saturday night, fuse box, know long, tell long, gets sorted, getting address Help Seekers A total of 9 topics were generated from the data through unsupervised topic modeling. Three key topics (below) show a diversity of customer service conversations not previously categorized by agents.
- 51. Evaluating Chatbot Usage customer service… email address… power cut… Sentiment: 70% negative Complexity: ↑ average Recommendation: divert away from Chatbot Sentiment: 60% negative Complexity: ↑ average Recommendation: divert away from Chatbot Sentiment: 66% positive Complexity: ↓ average Recommendation: potential to utilize Chatbot Customer Service Seekers Contact Us Help Seekers
- 52. Client Recommendations 1. Brand and Viral comments could be diverted to a Chatbot with machine learning algorithm 2. Negative and positive sentiment are distinguishable by key phrases, allowing for direction to Chatbot or human where necessary 3. After applying non-negative matrix factorization, we can determine which conversation types are suitable for a Chatbot based on conversation complexity and sentiment
- 53. Customer Lifetime Value Scalable machine learning applied to millions of members
- 54. LTV Challenge • Build reproducible, production level lifetime value model which scales to millions of users • Writes to database and allows others to use • Refreshes every month
- 55. What did we Predict? • Revenue - a regression problem • Cost of goods sold - logistic problem • Coupons redeemed - Bayesian LTV = Revenue – COGS - Coupons
- 56. Prediction Error -$40 $160 $360 $560 $760 $960 $1,160 $1,360 1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85 89 93 97 Predicted CLV($) Actual Net Revenue($)
- 57. Data Pipeline Data- warehouse Stored Procedure Trains Model Trains Model Trains Model Stored Procedure Predicts Predicts PRedicts Writes Error Metrics Data- warehouse Writes Scores User User User*Process takes less than an hour
- 58. Going Forward • Develop a model to find what drives LTV • Will sending more emails affect LTV • What’s the optimal number of coupons to serve? • Segmenting users around LTV • What do we do with the most valuable • Do we do anything at all? • How do we engage users to spend more?
- 59. Want to Stay Present? • We write weekly on machine learning, artificial intelligence, cloud computing and other technology • Cloudy with a Chance of AI - Subscribe today!
- 60. Questions? Andrew Van Aken Consultant, OgilvyOne Worldwide Laurie Close Global Brand Partnerships, OgilvyRED Michael McCarthy Senior Consultant, OgilvyOne Worldwide