📚 Data Science Riddle
You're training a hiring model. What's the biggest ethical risk?
You're training a hiring model. What's the biggest ethical risk?
Anonymous Quiz
19%
High Variance
16%
Algorithm Choice
7%
Large dataset size
57%
Biased training data
📚 Data Science Riddle
In Naive Bayes, what's the "naive" assumption?
In Naive Bayes, what's the "naive" assumption?
Anonymous Quiz
21%
Features are Gaussian distributed
51%
Features are conditionally independent given the class
15%
Classes are equally probable
13%
Noisy data is ignored
Parameters vs Hyperparameters
People confuse these all the time.
Parameters: learned by the model during training. (e.g., weights in a neural network, coefficients in regression).
Hyperparameters: set before training. They control how the model learns. (e.g., learning rate, number of layers, batch size).
✔️ Parameters = the student’s knowledge (changes as they study).
✔️ Hyperparameters = the teacher’s instructions (fixed rules of how to study).
Tuning hyperparameters is often the difference between a good model and a useless one.
People confuse these all the time.
Parameters: learned by the model during training. (e.g., weights in a neural network, coefficients in regression).
Hyperparameters: set before training. They control how the model learns. (e.g., learning rate, number of layers, batch size).
✔️ Parameters = the student’s knowledge (changes as they study).
✔️ Hyperparameters = the teacher’s instructions (fixed rules of how to study).
Tuning hyperparameters is often the difference between a good model and a useless one.
❤3🔥3
📚 Data Science Riddle
You're classifying product reviews (positive/negative). Which feature method is more effective for capturing context?
You're classifying product reviews (positive/negative). Which feature method is more effective for capturing context?
Anonymous Quiz
20%
Bag of Words
25%
TF-IDF
26%
Word2Vec
28%
One-Hot Encoding
Data Drift: The reason Good Models Go Bad
You built a model that performed amazingly last month.
Now? Accuracy tanked. Confusion Matrix looks like a crime scene.
Welcome to Data Drift. The silent model killer.
📉 What Is Data Drift?
It’s when the data your model sees today is different from the data it was trained on.
Imagine you trained a model on pre-COVID shopping data then you tried to predict online purchases in 2021.
People’s behavior changed. Your model didn’t.
That’s drift. Reality shifted, but your math stayed still.
🧠 The Core Types
➡️ Covariate Drift: Input features change (e.g., user age distribution shifts).
➡️ Prior Drift: The target variable’s frequency changes (e.g., fewer defaults now).
➡️ Concept Drift: The relationship between input and output changes entirely.
The last one is deadly. your model’s logic literally stops making sense.
🚨 Why It’s Dangerous
Models decay quietly.
By the time you notice lower performance, the damage( business or otherwise ) is already done.
That’s why top teams monitor models like systems, not code.
🧩 The Fix
1. Track feature distributions over time (use KS test, PSI, or histograms).
2. Monitor prediction confidence — sudden uncertainty = red flag.
3. Retrain models periodically with fresh data.
AI isn’t “build once.” It’s “maintain forever.”
You built a model that performed amazingly last month.
Now? Accuracy tanked. Confusion Matrix looks like a crime scene.
Welcome to Data Drift. The silent model killer.
📉 What Is Data Drift?
It’s when the data your model sees today is different from the data it was trained on.
Imagine you trained a model on pre-COVID shopping data then you tried to predict online purchases in 2021.
People’s behavior changed. Your model didn’t.
That’s drift. Reality shifted, but your math stayed still.
🧠 The Core Types
➡️ Covariate Drift: Input features change (e.g., user age distribution shifts).
➡️ Prior Drift: The target variable’s frequency changes (e.g., fewer defaults now).
➡️ Concept Drift: The relationship between input and output changes entirely.
The last one is deadly. your model’s logic literally stops making sense.
🚨 Why It’s Dangerous
Models decay quietly.
By the time you notice lower performance, the damage( business or otherwise ) is already done.
That’s why top teams monitor models like systems, not code.
🧩 The Fix
1. Track feature distributions over time (use KS test, PSI, or histograms).
2. Monitor prediction confidence — sudden uncertainty = red flag.
3. Retrain models periodically with fresh data.
AI isn’t “build once.” It’s “maintain forever.”
A model is only as good as the world it was trained in
and the world never stops changing.
❤6
📚 Data Science Riddle
You're building a chatbot but it gives generic answers. What's the root issue?
You're building a chatbot but it gives generic answers. What's the root issue?
Anonymous Quiz
11%
Model is too deep
65%
Training data lacks context
11%
Wrong loss function
14%
Poor tokenization
📚 Data Science Riddle
Model Accuracy improves after dropping half the features. Why?
Model Accuracy improves after dropping half the features. Why?
Anonymous Quiz
12%
Model became smaller
69%
Overfitting reduced
12%
Data size shrank
7%
Training faster
❤3
Understanding the Forecast Statistics and Four Moments (4P).pdf
181.8 KB
Statistical Moments (M1, M2) for Data Analysis
Here are 5 curated PDFs diving into the mean (M1), variance (M2), and their applications in crafting research questions and sourcing data.
A channel member requested resources on this topic and we delivered.
If you have a topic you want resources on let us know, and we’ll make it happen!
@datascience_bds
Here are 5 curated PDFs diving into the mean (M1), variance (M2), and their applications in crafting research questions and sourcing data.
A channel member requested resources on this topic and we delivered.
If you have a topic you want resources on let us know, and we’ll make it happen!
@datascience_bds
❤7