Business Intelligence 14

658 Part IV • Robotics, Social Networks, AI and IoT

Benefits to customers are:

According to ir.netflix.com, Netflix is (Spring 2018 data) the world’s leading Internet television network with more than 118 million members in over 190 countries enjoying more than 150 million hours of

TV shows and movies per day, including original series, documentaries, and feature films. Members can view unlimited shows without commercials for a monthly fee.

Application Case 12.3 Netflix Recommender: A Critical Success Factor

• Personalization. They receive recommendations that are very close to fulfilling what they like or need. This depends, of course, on the quality of the method used.

• Discovery. They may receive recommendations for products that they did not even know existed but were what they really need.

• Customer satisfaction. With repeated recommendations tends to increase. • Reports. Some recommenders provide reports and others provide explanations

about the selected products. • Increased dialog with sellers. Because recommendations may come with expla-

nations, buyers may want more interactions with the sellers.

Benefits to sellers are:

• Higher conversion rate. With personalized product recommendations, buyers tend to buy more.

• Increased cross-sell. Recommendation systems can suggest additional products. Amazon.com, for example, shows other products that “people bought together with the product you ordered.”

• Increased customer loyalty. As benefits to customers increase, their loyalty to the seller increases.

• Enabling of mass customization. This provides more information on potential customized orders.

Several methods are (or were) used for building recommendation systems. Two classic methods are collaborative filtering and content-based filtering.

COLLABORATIVE FILTERING This method builds a model that summarizes the past be- havior of shoppers, how they surf the Internet, what they were looking for, what they have purchased, and how much they like (rate) the products. Furthermore, collaborative filtering considers what shoppers with similar profiles bought and how they rated their purchases. From this, the method uses AI algorithms to predict the preference of both old and new customers. Then, the computer program makes a recommendation.

CONTENT-BASED FILTERING This technique allows vendors to identify preferences by the attributes of the product(s) that customers have bought or intend to buy. Knowing these preferences, the vendor recommends to customers products with similar attributes. For instance, the system may recommend a text-mining book to a customer who has shown interest in data mining, or action movies after a consumer has rented one in this category.

Each of these types has advantages and limitations (see example at en.wikipedia.org/ wiki/Recommender_system). Sometimes the two are combined into a unified method.

Several other filtering methods exist. Examples include rule-based filtering and activity-based filtering. Newer methods include machine learning and other AI technolo- gies, as illustrated in Application Case 12.3.

Chapter 12 • Knowledge Systems: Expert Systems, Recommenders, Chatbots, Virtual Personal Assistants 659

The Challenges

Netflix has several million titles and now produces its own shows. The large titles inventory often creates a problem for customers who have difficulty determin- ing which offerings they want to watch. An additional challenge is that Netflix expanded its business from the United States and Canada to 190 other countries. Netflix operates in a very competitive environment in which large players such as Apple, Amazon.com, and Google operate. Netflix was looking for a way to distinguish itself from the competition by making useful recommendations to its customers.

The Original Recommendation Engine

Netflix originally was solely a mail-order business for DVDs. At that time, it encountered inventory prob- lems due to its customers’ difficulties in determining which DVDs to rent. The solution was to develop a recommendation engine (called Cinematch) that told subscribers which titles they probably would like. Cinematch used data mining tools to sift through a database of billions of film ratings and customers’ rental histories. Using proprietary algorithms, it recom- mended rentals to customers. The recommendation was accomplished by comparing an individual’s likes, dislikes, and preferences against those of people with similar tastes, using a variant of collaborative filtering. Cinematch was like the geeky clerk at a small movie store who sets aside titles he knows you will like and suggests them to you when you visit the store.

To improve Cinematch’s accuracy, Netflix began a contest in October 2016, offering $1 million to the first person or team that will write a program that would increase Cinematch’s prediction accuracy by at least 10 percent. The company understood that this would take quite some time; therefore, it offered a $50,000 Progress Prize each year in which the contest was conducted. After more than two years of com- petition, the grand prize went to Bellkor’s Pragmatic Chaos team, a combination of two runner-up teams.

To learn how the movie recommendation algo- rithms work, see quora.com/How-does-the-Netflix- movie-recommendation-algorithm-work/.

The New Era

As time passed, Netflix moved to the streaming business and then to Internet TV. Also, the spread of cloud technology enabled improvement in the

recommendation system. The new system stopped making recommendations based on what people have seen in the past. Instead, it is using Amazon’s cloud to mimic the human brain in order to find what people really like in their favorite movies and shows. The system is based on AI and its technology of deep learning. The company can now visualize Big Data and draw insights for the recommenda- tions. The analysis is also used in creating the com- pany’s productions. Another major change dealt with the transformation to the global arena. In the past, recommendations had been based on information collected in the country (or region) where users live. The recommendations were based on what other people in the same country enjoyed. This approach did not work well in the global environment due to cultural, political, and social differences. The modi- fied system considers what people who live in many countries view and their viewing habits and likes.

Implementation of the new system was dif- ficult, especially when a new country or region was added. Recommendations were initially made without knowing much about the new customers. It took 70 engineers and a year of work to modify the recommendation system. For details, see Popper (2016).

The Results

As a result of implementing its recommender sys- tem, Netflix has seen very fast growth in sales and membership. The benefits include the following:

• Effective recommendations. Many Netflix members select their movies based on recom- mendations tailored to their individual tastes.

• Customer satisfaction. More than 90 per- cent of Netflix members say they are so satisfied with the Netflix service that they recommend it to family members and friends.

• Finance. The number of Netflix members has grown from 10 million in 2008 to 118 million in 2018. Its sales and profits are climbing steadily. In spring 2018, Netflix stock sold for over $400 per share compared with $140 a year earlier.

Sources: Based on Popper (2016), Arora (2016), and StartUp (2016).

(Continued )

660 Part IV • Robotics, Social Networks, AI and IoT

Questions for Case 12.3

1. Why is the recommender system useful? (Relate it to one-to-one targeted marketing.)

2. Explain how recommendations are generated.

3. Amazon disclosed its recommendation algo- rithms to the public but Netflix did not. Why?

4. Research the research activities that attempt to “mimic the human brain.”

5. Explain the changes due to the globalization of the company.

Application Case 12.3 (Continued)

u SECTION 12.2 REVIEW QUESTIONS

1. Define expert systems. 2. What is the major objective of ES? 3. Describe experts. 4. What is expertise? 5. List some areas especially amenable to ES. 6. List the major components of ES and describe each briefly. 7. Why is ES usage on the decline? 8. Define recommendation systems and describe their operations and benefits. 9. How do recommendation systems relate to AI?

12.3 CONCEPTS, DRIVERS, AND BENEFITS OF CHATBOTS

The world is now infested with chatbots. According to 2017 data (Knight, 2017c), 60 percent of millennials have already used chatbots and 53 percent of those who have not used them are interested in doing so. Millennials are not the only generation using chat- bots, although they may use them more than others. What chatbots are and what they do is the subject of this section.

What Is a Chatbot?

Short for chat robot, a chatbot, also known as a “bot” or “robo,” is a computerized service that enables easy conversations between humans and humanlike computerized robots or image characters, sometimes over the Internet. The conversations can be in writing, and more and more are by voice and images. The conversations frequently involve short questions and answers and are executed in a natural language. More intel- ligent chatbots are equipped with NLPs, so the computer can understand unstructured dialog. Interactions also can occur by taking or uploading images (e.g., as is done by Samsung Bixby on the Samsung S8 and 8). Some companies experiment with learning chatbots, which gain more knowledge with their accumulated experience. The ability of the computer to converse with a human is provided by a knowledge system (e.g., rule- based) and a natural language understanding capability. The service is often available on messaging services such as Facebook Messenger or WeChat, and on Twitter.

Chatbot Evolution

Chatbots originated decades ago. They were simple ES that enabled machines to answer questions posted by users. The first known such machine was Eliza (en.wikipedia. org/wiki/ELIZA). Eliza and similar machines were developed to work in Q&A mode.