---
tags: COTAI LHP
---

# Lecture Notes ML4AI 2021

## SESSION 4: RecSys: Recommender Systems
- widely use as core engine for user platforms (amazon,spotify,etc)

((recsys success story insert))

- can learn to serve humans: Virtual assistant, cobot

### Generalization counts
- Observe items or users and recommend other items to the users
- **vital for a good RecSys**

### Personalization counts
- going beyond common statistics of the mass
- have a unique differentiated stats between users
- **vital for a good RecSys**

### Long-tail issues
(insert image here)

### Problem formulation

- $Input: users' rating \rightarrow Output: prediction(rating, stars,etc.)$
- Two common methods: **User-based filtering** and **Item-based filtering**

#### Input Data

#### Taxonomy

### Collaborative filtering

- Require:
- $\overrightarrow{Z}$ : embeddings
- similarity

##### Item-based CF
- Key idea: users as features → compute item-item similarity matrix.
- Advantage: items don’t change much as users.

##### User-based CF
- Key idea: items as features → compute user-user similarity matrix.
###### Note: Transposing the utility matrix in user-based CF before computing basically gives item-based CF result (more efficient since $n^o$ of users > $n^o$ of items).

##### Matrix factorization (model-based) CF
- Key idea: user-item weighted-sum/similarity score ≈ rating score

- SVD: singular vector decomposition.

- Matrix factorization solved by Regularized Alternating Least Squares
- Solved by truncated SVD ***after*** CF

### Personalized RecSys

#### Content-based filtering
- Key idea: $\theta_k⋅z_i+b_k= user_k–item_i$, weighted sum score ≈ rating (could be formulated as a regression or a classification problem)

#### Context-aware & KB personalized RecSys!
[](https://i.imgur.com/7tlsUUH.png)

### Hybrid RecSys