Video summary

1. МО-1 ФКН: вводный семинар :)

Main summary

Key takeaways

Educational

Main ideas / lessons

1) Seminar format and teaching philosophy

  • Machine learning seminars focus on understanding concepts through a “story” in notebooks, rather than heavy manual coding.
  • Typical teaching flow across the semester:
    • Early semester
      • Little mathematics
      • Some starter code
      • Focus on understanding what the notebook is doing and the overall approach
    • As the semester progresses
      • More mathematics (e.g., derivatives, optimization concepts like gradients/gradient descent)
      • More deeper understanding of how algorithms work
      • More problem solving and homework that requires implementing/using those ideas
  • Students are expected to already know basic programming habits:
    • Using documentation
    • Searching for functions (e.g., via Google)
    • Asking how to solve tasks

2) What machine learning tasks are (object/target/features)

Exercises are structured around identifying:

  • Object: what each training example corresponds to
  • Target: what we predict
  • Features / predictors: inputs used by the model
  • Problem type: regression, classification, ranking, clustering, etc.

3) Example: food delivery ETA (regression)

  • Scenario: A user orders food; the app shows how many minutes it will take for delivery.
  • ML framing:
    • Target: delivery time (a real-valued number)
    • Task type: regression
    • Object: the order
      • Not the courier, because a courier has many orders and we need per-order predictions.
  • Features examples:
    • Courier-related historical attributes (e.g., an age proxy, whether they have a bicycle)
    • Environment/state variables (weather, time of day, etc.)
    • Area and contextual variables of pickup/dropoff
    • Restaurant-related attributes
    • Speed/history-type variables related to the restaurant/courier

4) Example: spam detection in YouTube comments (binary classification)

  • Scenario: Many comments appear under YouTube videos; want to detect spam.
  • ML framing:
    • Target/labels: spam vs not spam
    • Task type: binary classification
    • Object: a comment
  • Why not classify the video or user:
    • Each comment is a unique sample
    • Users can write both spam and legitimate comments
  • Features examples:
    • Word frequency / presence (e.g., how often certain words occur)
    • Comment statistics / ratings / metadata-like counts (as discussed in the conversation)

5) Other ML task types (with examples)

  • Ranking: search results / ordered lists (query returns items ranked by predicted relevance/position)
  • Classification (multi-class examples):
    • Exam grade buckets (e.g., 4 classes)
    • Pass vs fail
  • Regression:
    • Taxi cost, numeric price/ETA-type predictions
  • Clustering:
    • Group similar items when labels aren’t provided (e.g., image/object grouping)
  • Computer vision:
    • Image classification / segmentation (zones/pixels)
  • Text recommendation examples:
    • YouTube recommendations as a ranking problem
    • Possible reformulation using pairwise comparisons (ordering correctness)

6) Train/test splitting: avoiding data leakage and time effects

  • Normally split the dataset into:
    • Train: learn model parameters
    • Test: evaluate generalization on unseen data
  • Main goal: approximate performance on future/unseen data.
  • Common splitting approaches:
    • Random split: sometimes misleading when there are time-dependent correlations
    • Time-based split: keeps later periods for testing

Why random split can fail (data leakage / future information leaking):

  • Example described: a news/tragedy classifier.
  • If you randomly split articles across time, the model can effectively “see” that a tragedy is ongoing/has been discussed, because the target signal correlates with time or event propagation.
  • Result:
    • Test performance may look artificially high
    • Future real data shows a dramatic drop

7) Instability of evaluation and a solution: multi-splitting / cross-validation-like idea

The lecturer discusses cases where:

  • A model gets high metrics on one split but drops on real future data.
  • Causes:
    • Structural shifts (new entities/conditions appear)
    • Unlucky random splits (test set too “easy” or too “hard”)

Proposed robust evaluation idea (rotating splits):

  • Split the data into multiple parts (“folds”).
  • For each fold:
    • Train on some parts
    • Test on the remaining part
  • Aggregate results:
    • Average metrics for a more stable estimate
    • Produces multiple trained models and reduces sensitivity to one unlucky split

8) Data modalities and which model families tend to work

  • Tabular data
    • Often best with linear models and gradient boosting / classical ML
  • Images
    • Use neural networks to exploit continuity/spatial structure
    • Segmentation vs classification
    • Mentions CNN-style ideas implicitly; references to Transformers hype
  • Audio/time series
    • Treated as continuous signals; classical methods and neural nets can work
  • Text
    • Meaningful sequences occupy a tiny subset of all possible strings; neural networks work well
    • Classical alternative: feature engineering like word frequency (works initially)
  • Video
    • Sequence of frames/pictures
  • Graph data
    • Used in biology (proteins/molecules) and social networks (fake account farms)
  • Mentions multimodal models as a direction (combining text+image+video, etc.)

9) Seminar tooling: Pandas and workflow for tabular data

Main practical notebook guidance (first week):

  • Use Pandas for student/course assignment tabular data.
  • Steps emphasized:
    • Understand how Pandas works:
      • read the notebook tutorial
      • search the documentation
    • Inspect the dataset:
      • column names
      • types
      • table size
      • distributions
  • Typical Pandas operations highlighted:
    • Access columns and inspect types
    • Slice/filter rows (e.g., select students belonging to a specific group)
    • Apply functions to columns (including lambda functions)
  • Strong guidance:
    • Avoid manual loops when possible; use vectorized operations / built-ins
    • Homework will penalize inefficient looping
  • Data quality / preprocessing issues to watch for:
    • Wrong delimiter/format (e.g., decimal comma vs dot causing numeric columns to become objects)
    • Inconsistent scaling between years (necessity of normalization)
    • Missing values (“gaps”) need handling
  • Additional tasks:
    • Grouping/counting distributions (e.g., how many choose each course priority)
    • Visualization using a matplotlib-like workflow (lookup the needed commands)

10) Logistics, participation, and homework/tests

  • Tests:
    • Lecturer mentions “3 or 4” planned tests (not fully finalized yet)
  • Participation:
    • Students may interrupt if unclear:
      • use the microphone or write questions in chat
  • Homework:
    • Described as “big assignments”
    • Includes using Pandas and visualization
    • Includes searching for correct library commands and applying them

Speakers / sources featured

  • Philip: main lecturer; works at Yandex; also described practical work in taxi safety/security systems
  • Kolya: mentioned as a target audience reference
  • Zhenya: mentioned as someone who “paid well”/as a relevant organizer or fellow lecturer; referenced in earlier discussions
  • Yulia: appears in chat/questions during the lecture
  • Kruse/Cruise: referenced in the phrase “Cruise Not because…”, likely referring to a company/system; not a speaker
  • Yandex: source of work/practical examples
  • YouTube: platform used for the spam example; referenced for comments
  • Minority Report: mentioned as an analogy
  • Grey / students / groups course materials: mentioned indirectly as existing course/notebooks
  • Pandas documentation / Google: recommended sources for function lookup
  • Lecture notebooks / “ML SE / first notebook about date”: referenced as existing course materials

Original video