Please enable JavaScript.
Coggle requires JavaScript to display documents.
Recommendation System (Collaborative Filtering (CF) (Categories (Memory…
Recommendation System
Collaborative Filtering (CF)
Categories
Model-based CF
methods
models
Non-deep methods
Conventional Methods
Matrix factorization
Latent semantic model
Bayesian network model
Mixture models
Rich side information
For User/Item
Social Network
1 more item...
User-contributed info
1 more item...
For interaction
Timestamp, location, hunger status...
3 more items...
Deep learning based methods
Purpose
Interaction Function Learning
User/Item Feature Representation Learning
Advantage
Ability to process sequence data
Capture Non-linearity for modelling interaction/features
Underlying Models
MLP
Difficult to model low-rank interaction
AE
Powerful feature extractor, can also model interaction in reconstruction
CNN
Good for feature extraction (image and text)
RNN
Good for sequential data processing
RBM
NADE
tractable distribution estimator better than RBM
Attention based
Good interpretability
Adversary network
DRL
Real-time recommendation
Hybrid
CNN-RNN
CNN-AE
RNN-AE
RNN-DRL
Disadvantages
Interpretability
Data-hungry
Hyperparameter tuning
Future Research Direction
Joint Representation Learning & Automatic feature crafting
Explainable Recommendation
Cross domain recommendation
Multitask learning
Better evaluation datasets and metrics
Learning
Point-wise learning (Predict score)
Pair-wise learning (Total Ordering)
Bayesian Personalized Ranking
Memory-based CF
methods
User-based: aggregate similar user ratings using similarity metrics
Item-based: recommend items based on other items user has previously rated
drawbacks
Computationally expensive
Similarity measures are not optimal relations between users/items
Challenges
New Conditions and Tasks
Group Recommendation
Social Recommendation
Long Tail Recommendation
Cross-Domain Recommendation
New Perspectives and Models
Search and Recommendation
Interaction and Recommendation
Economics and Recommendation
Note: Interesting point actually, using economics to build better objective
Hybrid Methods
Combination of CF and content-based
Content-based methods
Use user/item auxillary data to predict