Collaborative filtering in ML Python - Model Pipeline Trace

Model Pipeline - Collaborative filtering

Collaborative filtering helps recommend items to users by learning from their past preferences and the preferences of similar users.

Data Flow - 5 Stages

1Raw user-item ratings

1000 users x 500 items→Collect user ratings for items (e.g., movies, products)→1000 users x 500 items

User 1 rated Movie A: 4 stars, Movie B: 5 stars; User 2 rated Movie A: 3 stars

↓

2Data preprocessing

1000 users x 500 items→Fill missing ratings with zeros or average ratings→1000 users x 500 items

Missing rating for User 3 on Movie C replaced with 0

↓

3Matrix factorization

1000 users x 500 items→Decompose rating matrix into user and item feature matrices→User features: 1000 x 20, Item features: 500 x 20

User 1 features: [0.2, 0.5, ..., 0.1], Movie A features: [0.3, 0.7, ..., 0.4]

↓

4Model training

User features: 1000 x 20, Item features: 500 x 20→Optimize features to minimize difference between predicted and actual ratings→Trained user and item feature matrices

User 1 and Movie A features adjusted to predict rating close to 4 stars

↓

5Prediction

User features: 1 x 20, Item features: 1 x 20→Calculate predicted rating by dot product of user and item features→Predicted rating scalar

Predicted rating for User 1 on Movie C: 3.8 stars

Training Trace - Epoch by Epoch

Loss
0.9 |*       
0.8 |**      
0.7 |**      
0.6 |***     
0.5 |****    
0.4 |*****   
0.3 |******  
0.2 |******* 
    +--------
     1 5 10 15 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.85	N/A	Initial loss is high as model starts learning user-item patterns
5	0.45	N/A	Loss decreases as model improves rating predictions
10	0.30	N/A	Loss continues to decrease, model converging
15	0.25	N/A	Loss stabilizes, model has learned meaningful features

Prediction Trace - 4 Layers

Layer 1: User feature vector extraction

Layer 2: Item feature vector extraction

Layer 3: Dot product calculation

Layer 4: Rating prediction output

Model Quiz - 3 Questions

What does the matrix factorization step do in collaborative filtering?

A It breaks down the user-item rating matrix into smaller feature matrices

B It removes missing ratings from the dataset

C It directly predicts the final ratings without training

D It collects raw user ratings from the database

Key Insight

Collaborative filtering uses patterns in user-item interactions to predict preferences, learning hidden features that represent users and items. This helps recommend items users might like based on similar users' tastes.