Big Data For Data Science

Module 6: Data Analysis for Big Data

Xingang (Ian) Fang

Outline

  • Products of data analysis

  • Data Analysis in Big Data Context

  • Machine learning

  • Deep learning

Products of data analysis

  • Reports and visualizations

  • Models

    • Statistical

    • Machine learning (including deep learning)

  • Optimized process/algorithm

  • New datasets

Reports and Dashboard

Data Analysis in Big Data Context

  • Two scenarios

    • Start big, finish small

    • Start big, finish big

  • Most data analysis methods only needs small data except for deep learning models

  • Statistical models

    • find trends and metrics

    • cannot model complex patterns

  • Machine learning: a subset of artificial intelligence that allows computers to learn from data and make predictions or decisions without being explicitly programmed.

  • Deep learning models (the only group of large models)

Most models are not “big”

Model performance vs data size

Credit: Andrew Ng, CC BY-SA 4.0 <https://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons

Machine Learning vs Deep Learning

  • Machine Learning

    • Evolution: from rule-based to data-driven

    • Type of learning tasks

    • Traditional ML models

    • Libraries and frameworks

  • Deep Learning

    • Characteristics of DL

    • Comparison with traditional ML

    • DL models

      • Trending DL models

    • Libraries and frameworks

AI Hierarchy

Credit: Lollixzc, CC BY-SA 4.0 <https://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons

Machine Learning

  • Evolution: from rule-based to data-driven

    • Rule-based: expert knowledge

    • Data-driven: data mining

    • Future: big data driven DL models

Rule Based and Machine Learning

Credit: https://www.led-professional.com/resources-1/articles/ai-lighting

Type of learning tasks

  • Supervised learning: labeled data, learn to predict

  • Unsupervised learning: unlabeled data, learn to analyze and generate

  • Reinforcement learning: learn to make better decision with reward/punishment

  • Semi-supervised learning: a mix of labeled and unlabeled data

  • Self-supervised learning: unlabeled data, learn to predict itself

  • Ensemble learning: combine multiple models to improve performance

  • Transfer learning: learn from a related task and apply to a new task

Supervised Learning Example

Credit: EpochFail, CC BY-SA 4.0 <https://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons

Traditional ML Models

  • Supervised learning

    • Linear Regression

    • Logistic Regression

    • Support Vector Machine

  • Unsupervised learning

    • K-means clustering

    • Principal Component Analysis (PCA)

  • Reinforcement learning

    • Q-learning

Scikit Learning Choosing the Right Estimator

ML Libraries and Frameworks

  • Traditional ML

    • Scikit learning

    • R-studio

    • Spark MLLib

  • Visualization

    • Matplotlib

    • Seaborn

    • Plotly

    • Bokeh

Machine Learning (include deep learning) libraries

Deep Learning

  • Characteristics of DL

    • Ability to learn complex representations

    • Automatic feature extraction

    • Large amounts of data

    • Computationally intensive

  • Comparison with traditional ML

    • Data representation: learned feature vs hand-crafted feature

    • Complexity: huge number of parameters in DL models

    • Performance: Much better as data size increases

    • Data requirements: large amount of data

Difference between ML and DL

Credit: Asher, Clint, et al. “The Role of AI in Characterizing the DCM Phenotype.” Frontiers in Cardiovascular Medicine 8 (2021): 1986.

Deep Leaning Models

By learning task:

  • Supervised learning

    • Convolutional Neural Network (CNN)

    • Recurrent Neural Network (RNN)

    • Generative Adversarial Network (GAN)

  • Unsupervised learning

    • Autoencoder

    • Diffusion models

    • Generative Adversarial Network (GAN)

  • Reinforcement learning

    • Deep Q-Network (DQN)

By data type:

  • Computer vision

    • Convolutional Neural Network (CNN)

    • Recurrent Neural Network (RNN)

    • Generative Adversarial Network (GAN)

    • Diffusion models

  • Natural language processing

    • Transformer models

    • Matrix Factorization

    • Recurrent Neural Network (RNN)

  • Time-series data

    • Example: sound, speech, video, stock market data, etc.

    • Recurrent Neural Network (RNN)

    • Convolutional Neural Network (CNN)

Neural Network Zoo

Credit: https://www.asimovinstitute.org/neural-network-zoo/

Trending DL Models

  • Large language models (LLM)

    • GPT family from OpenAI: GPT-3, GPT-4, ChatGPT

    • Google BERT, Palm, T5, Bard

    • Facebook RoBERTa, BigBird

    • BigScience BLOOM

  • Image generation models

    • OpenAI DALL-E

    • Stable Diffusion

    • MidJourney (possibly based-on stable diffusion)

  • Voice/Speech models

    • Google WaveNet, Universal Speech Model (USM)

    • OpenAI Whisper

DL Libraries and Frameworks

  • TensorFlow

  • PyTorch

  • Theano

  • Caffe 2

  • Chainer

Top Deep Learning libraries

Credit: https://analyticsdrift.com/top-deep-learning-libraries/

Case Studies on Data Analysis for Big Data