Introduction¶

Data science is turning raw data into understanding, insight, and knowledge to drive data-driven decisions and solve problems - Collecting data - Analyzing data - Statistics - Machine learning - Deep learning - Communicating analysis

We cannot move away from domain knowledge and solely depend on algorithms

Data Professionals¶

Decision-Making¶

Iterative process with feedback loops; not linear

Knowledge Hierarchy¶

Chain of increasing value

flowchart LR
Data -->
Information -->
Knowledge -->
Decision

Process¶

• Preparation: Plan to turn data into information, with a specific model and decision in mind
• Testing/Experimenting/Measurement
• Analysis, with uncertainty: Use model to turn information to knowledge
• Decision
• Using uncertainty
• Risk/benefit analysis
• Post-mortem: Learnings to improve things

Notes¶

• All models are wrong, some are useful
• Applying data mining algorithms on data that you don’t understand may result in false conclusions
• Always keep track of performed tests & analyses, to factor in data snooping

Questions¶

Fields Overview¶

Objectives¶

Use ML models to discover structural models, and then let the structural models to make the predictions, not the ML models

• Why: Black swans can be predicted by theory, even if they cannot be predicted by ML
• How: Use a non-parametric ML to identify important variables and then develop a parametric structural form model.

Types of Analysis¶

The complexity increases as we go down the above list, but the value obtained increases as well

Project Lifecycle¶

flowchart TB

subgraph Scoping
    dp[Define<br/>Project] -->
    me["Define Metrics<br/>(Accuracy, Recall)"] -->
    re[Resources<br/>Budget] -->
    ba["Establish<br />Baseline"]
end

subgraph Data
    d[(Data Source)] -->
    l[Label &<br />Organize Data]
end

subgraph Modelling
  pre[Preprocessing] -->
    s[Modelling] -->
    train[Training] -->
  pp[Post<br />Processing] -->
    vt[Validation &<br />Testing] -->
    e[Error Analysis] -->
    pre
end

subgraph Deploy
    dep[Deploy in<br />Production] -->
    m[Monitor &<br />Maintain] & dss[Decision<br />Support System]
end

Scoping --> Data --> Modelling --> Deploy

https://www.youtube.com/watch?v=UyEtTyeahus&list=PLkDaE6sCZn6GMoA0wbpJLi3t34Gd8l0aK&index=5

Data Mining¶

Generate Decision Support Systems

Non-trivial extraction of implicit, previously-unknown and potentially useful information from data

Automatic/Semi-automatic means of discovering meaningful patterns from large quantities of data

Predictive Tasks¶

Predict value of target/independent variable using values of independent variables

• Regression - Continuous
• Classification - Discrete

Descriptive Tasks¶

Goal is to find

• Patterns
• Associations/Relationships

Association Analysis¶

Find hidden assocations and patterns, using association rules

Applications¶

• Gene Discovery
• Market Baset Data Analysis Find items that are bought together

Clustering/Cluster Analysis¶

Grouping similar customers

Metrics¶

• Similarity
• Dissimilarity/Distance Metrics

Applications¶

• Grouping similar documents

• Clustering documents

• Vocabulary - All terms(key words) from all docs

• Generate document-term frequency matrix

  Document \vert Term	T1	T2	…	Tn
  D1
  D2
  …
  Dm

Deviation/Outlier/Anomaly Detection¶

Outlier is a data point that does not follow the norms.

Don’t mistake outlier for noise.

Application¶

• Credit Card Fraud Detection

  • Collect user profile such as Name, Age, Location
  • Collect user behavior data

• Network Intrusion Detection

• Identify anomalous behavior from surveillance camera videos

Misconceptions¶

• All forecasts will be inaccurate, so no point
• If we had the latest forecasting technology, all problems would be solved

IDK¶

Ensure you are looking at the correct scale

Model \(y_t/y_0\) instead of \(y_t\) to standardize all time series

Common problems with analysis¶

• Poorly-defined goals
• Data doesn’t meet needs of analysis objectives
• Analysis makes unwarranted assumptions
• Model is wong
• Data doesn’t support conclusion

Learning Process

1. Model building: Functional form
2. Identify parameter weights
3. Distribution of random errors

Each of them can have different levels of generalizability

For eg: Ohm’s Law

1. \(V=IR\) remains constant for all materials (under certain conditions)
2. \(R\) Changes for different materials
3. Errors are dependent on measurement and experimental methods, and are independent of materials

Communication of Results¶

Storytelling is important

Pitfalls¶

• Manipulation
  • Don't lie
  • Emphasize the story, but make the full data available
• Misrepresentation
  • Control expectations
  • Exaggerate your work, but do not over-exaggerate
• Ethos
  • Ethos \(\ne\) Credentials
  • Credentials do not make an expert
  • Credentials do not not make an expert
• Equity
  • Don't dumb down: Just translate in a simpler way
  • Amplify underrepresented voices

Comments