Opening

Processor Alice Wong
Associate Dean of Science, HKU

• WiDS 2017 is a collaboration among Stanford University, SAP, Google, Microsoft and Walmart Labs.
• 50th Anniversary for HKU Department of Statistic and Actuarial Science
• the big data research cluster @HKU

Talk 1: Women in Data Science

Speaker:
Anita Varshney
Global Strategy Transformation Lead, SAP Hong Kong

1. WiDS

• held by Stanford every February (March in Asia)
• keynote speakers from various industries that are doing data science now
• having largest attending number actually in middle east

2. SAP

• the world's largest provider of enterprise application software
• HQ in Germany; founded in 1972
• career suggestion: look for a good mentor
• present in 26 industries
• Real time processes, Prediction and simulation, great User experience, Agility and TCO

3. SAP next-gen

• Providing platform for college students to present their ideas directly to business customers.
• Technologies
  - Machine learning
  - IoT

*Amazing time management of presentation

Talk 2: Big Data Decision Analysis

"Big data is something that breaks Microsoft Excel" (lol)

Research project - Machine Learning for Chinese Suicide Newspaper Articles Classification

Analysis how the media report suicide incidence, and to figure out how to prevent suicide.

• WiseNews database: over 220K search result for the keyword "suicide", containing 84 million terms
• Big data challenges
  • Noisy dataset: e.g. "suicide car booming attack"
  • Data classification
• Supervised Machine Learning (use labeled articles to train)
• Web Interface for manually label
• Article features extraction for ML
  • Text Segmentation: Sentence -> Words -> N-grams
    • Tool: Jieba(结巴) - functionalities like MP & HMM(Hidden Markov Model)
      • State Transition Matrix: P(M|B) >> P(E|B)
  • Document Representation
    • Word to Document Matrix (not very efficient)
    • Chosen approach - Word Embedding (Word2Vec)
      • each word is represented by a vector of fixed number of dimensions (usually 30-500d)
      • Neural network: to determine the dimensions of the document vector, CBOW and Skip-Gram Model
      • Cosine similarity
• Classification (Training)
  • labeled dataset: 70% for training and 30% for testing
  • P(Suicide = Yes) 85.9% accuracy, P(Student = No), P(HK = Yes), ...
• Future work
  • Identify any pattern of misclassification
  • Increase dimensions of the word vectors
  • Deep learning approach for other NLP tasks with this dataset
    • Predict the method used for suicide
    • Predict the reasons used for suicide

Talk 3: Predictive Analytics

Vanessa Ko
Head of Presales SAP Hong Kong

1. SAP HK

• Customers: I.T., Cathay Pacific, PizzaHut, etc.
• Biggest competitor: overall no, only in some sub-areas.

2. Predictive Analytics

• How to make use of digitalized historical data
• Case: Obama for America 2012
  • Data source: Historical voting data, Census, Volunteer collected data, Facebook, etc;
  • Segments of voters, Found raising prediction, who's persuadable?
  • Data Modeling: VOTING RATE MODEL, SUPPORT RATE MODEL, Persuasive Rating, Overall score;
  • Goal: Target Voters, Donators and Volunteers -> especially swing voters (not too supportive or too opposing)