In the previous video, I talked about one of the common ways to apply machine learning in a business context. That is to replace or simplify rule-based systems. In this video, I'll focus on the next two common ways you can discover ML use cases, automating processes, and understanding unstructured data. To help you think through the Automation opportunities within your organization, I'd like you to think back to an important point we covered in module two. The purpose of automating a business process is not to reduce costs, instead, it's to reduce drudgery, increase accuracy, and to gain more visibility into the business. Essentially, you're focuses on how to do more with less. Look for ways to achieve more by applying ML to your business. If this is done properly, the cost savings will then actually follow. Let's look at an example of a used car dealership. Whenever customers brought in cars into the dealership, the agents had to manually upload and label photos for each car and set a price. This process took them on average about 20 minutes per car. How could a car dealership, in this case do more with less by automating its processes. One automobile auction company decided to take on the challenge and is now using ML to predict the price value of a car based on the many car photos it's agents upload into accustomed trained ML model. The overall process to photograph and evaluate a car now has dropped from 20 minutes to two to three minutes per car. That's a 95 percent reduction in time and massive improvements in overall service. There are many more examples of process automation, of course, such as predicting weather forecasts, loan approvals, even sorting resumes to create a candidate pipeline. These should give you a few ideas to get you thinking about process automation which are now within your business. Now, before I dive into examples of how you can use ML to understand unstructured data let me refresh your memory on what unstructured data is. Unstructured data is data that can't be directly compared to other Data. For example, some characteristics of books are structured like their publisher, location of publishing, number of pages, again known as tabular data. But it's not easy to directly compare the content of two books or to precisely determine how they're really related or different. Even human experts might not agree on exactly how similar two books are. Open text or language is just one example of unstructured data. Other examples include pictures, videos, and audio. Let me use an example to illustrate how ML can be used to understand unstructured data to solve business problems. Suppose you have a global customer base. They're going to be speaking different languages. So how much you use the same source content to serve all of your customers in their own language. You can use ML to automatically translate your content. For example, Bloomberg aggregates financial news, data and analytics worldwide. But disseminating that data across all languages proved to be a challenge. Google Translate allow Bloomberg to bring all of the data to their customers regardless of language. The table on the right shows a subset of the language pairs available in translation API. There are 97 pairs in all. Translating content into multiple languages is a common ML use-case. Do you have customers all over the world? Do you have dynamic content like product reviews? Use automatic translation to serve them. All right. Now that you're familiar with some of the most common ways to uncover opportunities for using ML within your business let's review a few real-world cases for further inspiration. netmarble is a leading video game publishing company based in South Korea that develops multiplayer games. Video games are big business with a total revenue greater than a $100 billion worldwide in 2018. For gaming companies like netmarble, it is extremely critical to retain game players in a fair environment so that players feel comfortable enough to invest time and money into the games. Game abusers can destroy the trust of other users and netmarble was using a rule-based Engine to detect them. But this system wasn't meeting performance expectations. To understand why this was happening, we need to think about some of the ways that players cheat in video games. Bob players are users that can play consecutively for unrealistic times and are abnormal or repetitive patterns. Playtime hackers are users that can finish the game unrealistically fast and without defeating opponents. Damage hackers are users that modify the characteristics of their equipment such as damage or attack speed to unrealistic standards. Reward hackers are users that receive unrealistic rewards from the game. We cannot use static rules such as, if playtime is greater than 24 hours, it is [inaudible] to identify the abuse. We have to identify patterns of behavior. Google Cloud work together with netmarble to help them replace these rule-based systems with ML. The final ML model predicts each user's behavior by learning the probability of actions they would take in certain situations based on every game ever played. The model learns the probability of the next action given the prior actions. This is a lot like a computer predicting the next move in a game of chess. Players with behavior unpredictable by the model are flagged as suspicious and the information is sent for evaluation. That's not all, netmarble was able to uncover additional ML use cases by replacing the rule-based systems for user behavior. They were able to better understand the behaviors of their core users, which meant that they were able to better market their games and end game services. The model can also predict churn players who are sure to leave within a week, players who will remain, and players who should be managed. It can also analyze the factors causing churn and enabled their game designers to make adjustments as needed. Most surprising to the team is that they discovered that they could use machine learning to detect fraudulent behavior in the games. This is a great example. Having a successful pilot project will motivate other departments to carry out their own ML Projects. Let's move on to our next customer use case. Ocado is the world's largest online-only grocery supermarket. Previously, when Ocado received emails, all e-mails would go to a central mailbox for sorting and forwarding by a person. This process was time consuming and would lead to a poor customer experience. To improve and scale this process, Ocado used ML to automatically route emails to the department that needs to process them. This new process eliminated multiple rounds of reading and triaging. Just to be sure you're paying attention, which of these methods did Ocado use to triage it's customer emails? If you answered, automated process, you're correct. But they also used ML to understand unstructured data. Specifically, they used Natural language to identify the customers sentiment and the topic of each message so they could route it immediately to the relevant department. Here's another example of automating a business process and understanding unstructured data. Kewpie is a company that manufactures baby food. In Kewpie's case, quality is not necessarily a matter of life or death because the food itself is safe. But discoloration can concern parents. The company used ML to automate the process of having to manually inspect each food ingredient, such as a potato. They used video, a type of unstructured data to complete their automatic inspection. That's a clever solution. All right. We covered a lot. We talked about the three ways to discover ML use cases in day-to-day businesses and went through several customer examples. Hopefully you now feel more inspired and confident about finding opportunities to use ML in your own business. In the next video, we'll continue our journey and offer examples for how to personalize applications with ML.
