#29 Introduction to Data Science | Data Science for Engineers

Main Ideas and Lessons Conveyed

Goal of the course/lecture series

The video begins a series of lectures introducing data science, covering:

• Various data science techniques (selected ones)
• Small illustrative examples showing how each technique applies to typical problems
• An end-of-course case study for participants to practice

This lecture is the first introduction, meant to help learners understand:

• What data science techniques do at a high level
• How to think about data science problems, especially problem formulation (turning unclear problems into solvable ones)

Common “laundry list” of techniques (and why there are so many)

The speaker notes that curricula/books often present many unrelated techniques (e.g., regression, clustering, SVMs, random forests, deep nets, etc.). This lecture challenges the idea of memorizing methods as a disconnected set and reframes learning around:

• What types of problems those techniques solve
• Why multiple techniques exist for similar problem types

Two fundamental engineering problem categories

From an engineering perspective, data science primarily solves two broad categories:

1. Classification problems
2. Function approximation problems

---

Concepts Explained in Detail

1) Classification problems

Core definition

• You have labeled data
• For a new input data point (with attributes/features), you assign a class/label
• Often the goal is to compute the likelihood/probability that a new point belongs to each class, then choose the most likely class

Binary classification

Example setup:

• Data points described by attributes: (x = [x_1, x_2, …, x_n])
• Two classes: (c_1) and (c_2)

Classification task:

• Given a new point (x^*), decide whether it is likely from (c_1) or (c_2)
• Example decision using likelihoods: 0.9 vs 0.1 ⇒ choose class (c_1)

Real-world engineering examples

• Fraud detection (binary classification)

  • Transactions have measurable attributes (amount, time of day, location, product type, etc.)
  • Historical transactions are labeled:
    • “illegal/fraudulent” vs “legal/legitimate”
  • For a new transaction:
    • Run it through a classifier to get fraud likelihood
    • If likelihood is very high:
      • Contact the cardholder to verify and potentially stop payment if confirmed fraudulent

• Fault diagnosis / failure prediction (multi-class classification)

  • Equipment state is described by attributes (power draw, performance, vibration, noise, temperature, etc.)
  • Historical labeled blocks correspond to states:
    • Normal ((n))
    • Fault mode 1 ((f_1))
    • Fault mode 2 ((f_2))
  • Classification of new operating data:
    • If normal: do nothing
    • If (f_1): stop the pump if severe, or schedule maintenance depending on severity

Linear vs non-linear classification

• Linear classification
  • Decision boundary can be a line/plane/hyperplane
  • In 2D, a straight line can separate classes well
• Non-linear classification
  • Classes may not be separable with a simple line/hyperplane
  • A non-linear decision function (curved boundary) is needed

Key question introduced:

• In the non-linear case, there are infinitely many possible functional forms, so you must decide which non-linear decision function to use.

---

2) Function approximation problems

Core definition

• Learn a function mapping inputs (attributes) to outputs
• The function is typically parameterized (it has parameters you must learn)

Data and objective

Given samples of:

• Inputs/attributes (e.g., (x_1, x_2, …, x_n))
• Corresponding outputs (observations/labels in a regression-like sense)

You must:

1. Choose the functional form (f(\cdot))
2. Estimate the parameters within that form

Examples

• Linear function form

  • Example: (y = a_0 x + b_0)
  • Parameters: (a_0, b_0)

• Quadratic function form

  • Example: (y = a_0 x^2 + a_1 x + a_2)
  • Parameters: (a_0, a_1, a_2)

Relation to regression

The speaker notes the course will cover linear regression as a linear function approximation approach.

Linear vs non-linear function approximation

• Linear case: straight line/hyperplane form
• Non-linear case: curve/surface that fits points (often involving clustering/approximation ideas)

---

Methodology / “Thinking Framework” Emphasized

The lecture’s main operational lesson is: select techniques based on assumptions, then validate them.

Assumption-validation cycle (core methodology described)

Thought experiment: unseen microorganisms

You can “see” only what is visible; unseen elements require a testing method. You generate hypotheses/assumptions about what exists (e.g., which microorganisms are present), then apply a chemical test known to react to a specific microorganism.

• If results match expectations, the assumption is supported.
• If results don’t match, the assumption is wrong (for the tested case), and you try the next hypothesis.

Through repeated assumption testing, you infer the unseen composition.

Connection to data science

• In high-dimensional data, you can’t directly “visualize” relationships.
• Data analytic tools act like a microscope:
  • You assume structure (e.g., randomness, Gaussian distribution, linear separability)
  • You choose a technique proven to work under those assumptions
  • You check whether the result “makes sense” (mathematically/empirically)
  • If it fails, the issue is typically that assumptions are incorrect, so you revise assumptions and try again

Testing and evaluation

• Results should be evaluated using test data
• Different methods may use different metrics and thresholds for deciding whether something “makes sense,” introducing subjectivity—but the overall process is still validation-driven

---

Why So Many Techniques Exist (Reframed Answer)

There are many techniques because:

• There are many possible assumptions about how data is structured
• For each assumption set, you can have techniques that perform well when those assumptions hold
• The combinations of assumptions are numerous, so technique diversity follows

Therefore, blindly comparing “which is best” is less important than:

• Understanding the assumptions each technique makes
• Matching a technique (or family of techniques) to the structure likely present in the specific problem

---

Course Transition / Next Lecture Preview

The speaker concludes:

• Takeaway 1: Most engineering data science problems are classification or function approximation
• Takeaway 2: Many techniques exist due to the assumptions they rely on and their ability to help “see” structure in multi-dimensional data

Next lecture planned:

• Introduce a data science problem-solving framework
• Use data imputation as the example technique/activity
• Show how the assumption-validation cycle is applied inside that framework

---

Speakers / Sources Featured

• Speaker: Unspecified (single lecturer presenting the material; no name provided in the subtitles)
• Sources referenced: None external (no named authors, institutions, or studies mentioned)