First step: Design data

1. Ask questions.
2. Break down questions.
3. Gather data
4. Analyze data

Second step: Tell your story with data

1. Put data puzzle together
2. Give a direction
3. Share the data
   customize for different shareholders.

Third step: Decide with data

1. Build an action plan
2. Get feedback from the set.
   Either from on-site workers’ feedback or website A/B testing.
3. Catch reaction from consumers.

It will be easier to understand when practicing it on cases.

Case: “My restaurant’s revenue decreases.”

First step: Design data

1. Ask questions:

My restaurant’s revenue decreased. What did I do wrong? Are there other reasons?

2. Break down questions.

• What does reducing revenue mean? Compare to yesterday? Last week? Last month? Last year?
• What did I do wrong:
  Is the restaurant’s setting the same as before? Consider changes in decoration, menu, billboard, tableware, ads, sales clerks, service, food, and product display.
• Other reason:
  Consider external reasons that reduce revenue. Maybe fewer people visit this neighborhood. Perhaps a new competitor emerges. Maybe the weather changed. Maybe the economy is in a recession?

3. Gather data

Gather data to answer the questions. Collect customer feedback from social media. Know your industry’s trend by searching keywords.

Gather sales data down to the hour.

4. Analyze data

Second step: Tell your story with data

1. Put data puzzle together
   Make a sales chart base on weekends, weekdays, hours, and months to analyze. We may figure out sales decreased much more on the weekend.
2. Give a direction
   We may separate our target audience into weekday customers, weekend customers, and customers that do not come on weekends. Improve sales by providing different tractions to them.
3. Share the data
   Share data with others and customize the presentation for different shareholders.

Third step: Decide with data

1. Build an action plan

• We figure out pizza sales better on weekends and sandwich sales better on weekdays.
• Our action plan can be to boost sales of sandwiches on weekdays by offering discounts when getting more than one sandwich. Or we can make a bundle with a sandwich to boost sales of other products.

2. Get feedback from the set.
Listen to shopkeepers, and they will give valuable insights. They might tell you that the sandwich is too challenging to prepare on-site.

3. Catch reaction from consumers.
The action plan may either work or not.

When a customer is not buying, it may be that the bundle price is too high or the quality of the product is not satisfying.

If we fail to boost the sales on weekdays, we can try to increase the sales on weekends. Suppose customers on weekends are mostly young generations. Do interviews to figure out what the young generation likes. Promote your product on social media, which targets young audiences.

Conclusion

Be curious and ask questions. Get the “why” behind the trend. Break down questions into measurable metrics. Then make action plans. Finally, get feedback from customers. Those are keys to a great data analysis.

E-commerce may often ask, whether should I change a button or layout of my website to increase sales? Usually, e-commerce does AB testing to answer these kinds of questions.
However, there are questions that AB testing can’t solve.

What AB testing is for

AB testing is often used for evaluating new features, UI, or products.
Take 2 sets of users, and show one set of users the control version which is the original version. And show the other set the testing version. See how they respond differently and conclude which version is better.
For example, Amazon uses AB testing to decide whether personalized product recommendation increases sales. Another use case is testing changes in the backend and see if it affects user behavior.

What AB testing is NOT capable of

• Testing new experiences
  Old users may either hate totally new experiences which is called ‘change aversion’ or they may be super excited about the new changes and try all of them which is called the ‘novelty effect’.
  These come with two problems. How much time do you need for the user to adapt to the experience and reflect on the true impact of your change? And also what is your baseline comparison.
• Long term effects
  The long-term effect such as will user recommend your site to others to buy a car is hard for AB testing. Or like changing a logo is quite more a long term effect. You just can’t get an answer from a short-term evaluation.
• Can’t tell if you are missing something

Other techniques rather than AB testing

AB testing is more like a “quantitive” measurement. There are other techniques to support Ab testing. More deep and “qualitative” measurements.

• Research on user logs perspectively
• Experiment research
• Focus group
• Surveys
• Human evaluation

Reference: Udacity A/B testing

# 偵測 啟動 TPU 
tpu = tf.distribute.cluster_resolver.TPUClusterResolver()
tf.config.experimental_connect_to_cluster(tpu)
tf.tpu.experimental.initialize_tpu_system(tpu)

# 啟動 distribution strategy
tpu_strategy = tf.distribute.experimental.TPUStrategy(tpu)

# 把 model 建在 strategy scope 會使得 model 在 TPU RUN 
with tpu_strategy.scope():
    model = tf.keras.Sequential( … ) # define your model normally
    model.compile( … )

# 正常 train model 
model.fit(training_dataset, epochs=EPOCHS, steps_per_epoch=…)

Batch size, learning rate

若想要TPU 跑得快，試著增加batch size. 習慣是 128 elements per core. (ex: batch size of 128*8=1024 for a TPU with 8 cores). Batch size of 8 per core 就可看到速度增加.

以下是怎麼 用cores定義 batch size:

BATCH_SIZE = 16 * tpu_strategy.num_replicas_in_sync

Reference: https://www.kaggle.com/docs/tpu

Data Science society 所舉辦的講座，邀請各領域的資料科學家談談資料科學如何被應用。有著global investment manager “Schroders”, Candy crush遊戲公司”King”, 資料技術顧問公司“Satalia”, 進修課程General Assembly和我這次最感興趣的醫療數據分析公司”QuintilesIMS”.

這些專家被問到請給在場的學生一些過來人意見。他們提到
1. 別侷限住自己專注的領域，有可能你的Team某些人請假而讓整個計畫停滯 不前，如果你能因為有些基本其他領域知識，適時站出來做點貢獻，整個TEAM才能繼續運轉。又或者你懂其他人在做什麼，讓你更能與人溝通合作。

2. 不要只談論用多快的速度，多小的記憶體解決問題，而是我們在解決什麼問題。

3. Try to understand why the algorithm works. 別只是建立運作方法，試著解釋其背後含義。 就算你的algorithm對，醫生若不知道其背後道理，還是不會輕易相信。

4.試著了解你是怎麼樣性格的人。你是喜歡幾個禮拜就見到成效的人，那你可能適合在公司工作。如果你喜歡長時間解決一個問題所帶來的成就感，那你應該在類似學術的環境工作。

5. 誠實的面對。對於你不知道的事要承認。這領域太廣泛，沒有人知道所有事情。

6. 多做些Side Projects, 哪怕只是運算自己的睡眠時間，參加hackathon，或是Kaggle competition. 多多累積自己的經驗，學習會更多，也才有東西秀給人家看。

7.當你對一個技術感到熟悉時，就該是時候嘗試學習新的東西。

試著記錄我在UCL 經歷到的事情。雖然跟Data Science 主題有點偏離。

學校 artificial intelligence 社團邀請Stuart Russell, University of California, Berkeley 教授來講述”Provably beneficial AI” (人工智慧發展的威脅，如何去面對設計AI.) Stuart 著作”Artificial Intelligence A modern approach” 一書可說是AI 的先驅。 在他的TED talk 也可看到Professor Russell在講述這項議題。
當Elon Musk , Stephen Hawking 和 Bill Gates 都擔憂 AI發展對人類生存造成的威脅。End of Human Race!!!!!! Professor Russell又如何看待這件事？ 他笑著說首先這些人都不是研究AI 的。但這些憂慮也絕非空穴來風。人工智慧有兩件事情必須要擔憂，第一件當你下達指令給他，他可能解讀錯誤產生錯誤的結果。第二，對他而言指令是第一，不論是用什麼方式達成指令。

那怎麼辦？ 那就應該從一開始就設計好AI，而不是假裝這些事情不會發生。 所以Professor Russell 致力開發 Human -compatible AI. 有三個目標。
1. AI只能更加實現人類的目的(沒有自己存活的意識)
2. AI並不知道 人類目的的意義
3. AI藉由觀察人類行為 了解人類目的。

例子： 一個機器人被下指令去買咖啡，他會想我因為要買咖啡所以我不能被關掉，我會把所有在Starbuck擋在我面前排隊的人除掉。

如果他Follow Human -compatible AI 原則
他會想，我可能會被關掉，但那是因為我有做錯事，我並不想做錯事。在被人類關掉後，他會學習什麼是錯事。

透過文獻與對人類的學習，可能就不會發生，機器人在家要幫小孩煮東西吃，但發現冰箱沒食物，這時看到一隻貓，所以就把貓抓來煮個小孩吃的狀況。