Introduction to Recommender System
- 1. Deep Learning for Recommendations: Fundamentals and Advances Wenqi Fan1, Xiangyu Zhao2, Dawei Yin3, Jiliang Tang4 1The Hong Kong Polytechnic University, 2City University of Hong Kong, 3Baidu Inc., 4Michigan State University 1 Data Science and Engineering Lab Tutorial website: https://advanced-recommender-systems.github.io/ijcai2021-tutorial/
- 2. 2 Recommender Systems Recommender Systems Information overload Age of Information Explosion Recommend item X to user Items can be: Products, News, Movies, Videos, Friends, etc.
- 3. 3 Recommender Systems A B C Recommendation has been widely applied in online services: - E-commerce, Content Sharing, Social Networking ... Product Recommendation
- 4. 4 Recommender Systems Recommendation has been widely applied in online services: - E-commerce, Content Sharing, Social Networking ... News/Video/Image Recommendation
- 5. 5 Recommender Systems Recommendation has been widely applied in online services: - E-commerce, Content Sharing, Social Networking ... Friend Recommendation
- 6. Problem Formulation Historical user-item interactions or additional side information (e.g., social relations, item’s knowledge, etc.) INPUT Predict how likely a user would interact with a target Item (e.g., click, view, or purchase) OUTPUT 6 Item set User set social relations, age, gender, occupation, etc. year, genre, actor, reviews, etc. Side information Side information ! users ! items (movies) … … User-item Interaction History 5 5 5 5 3 5 1 4 Spider Man Iron Man Toy Story Minions Captain America Lily Lala Peter David
- 7. 7 Recommender Systems • Collaborative Filtering (CF) is the most well-known technique for recommendation. - Similar users (with respect to their historical interactions) have similar preferences. - Modelling users’ preference on items based on their past interactions (e.g., ratings and clicks). • Learning representations of users and items is the key of CF. items … users … ! users ! items (movies) … … User-item Interaction History 5 5 5 5 3 5 1 4 Spider Man Iron Man Toy Story Minions Captain America Lily Lala Peter David Task: predicting missing movie ratings in Netflix. 5 4 … 5 … 5 … 5 1 2 5 ! users " items (movies) Lily … ? … ? ? ? ? ? ? … ? ? … ? ? ? Spider Man Iron Man Toy Story Minions Captain America … User-item Rating Matrix 𝐑
- 8. 8 Matrix Factorization ≈ 𝒒! 𝒑" # 𝑑 𝑑 Predicted rating of item 𝒋 for user 𝒊 : ̂ 𝑟"! ≈ 𝒑" # 𝒒! = * $%& ' 𝑝"$ 𝑞!$ User representations 𝐏 ∈ ℝ!×# Items representations QT ∈ ℝ#×$ 𝐑 ≈ 𝐏 × 𝐐T × ≪ 𝒎𝒊𝒏(𝒏, 𝒎) User-item Rating Matrix 𝐑 5 4 … 5 … 5 … 5 1 2 5 𝑛 users 𝑚 items (movies) Lily Lala Peter David … Spider Man Iron Man Toy Story Minions Captain America … Lily Lala Peter David … … Spider Man Iron Man Toy Story Minions Captain America ? … ? ? ? ? ? ? … ? ? … ? ? ? … … 4 𝒓𝒊𝒋 Ø Learn representations to describe users and items based on user-item rating matrix 𝐑.
- 9. … … … Objective with rating reconstruction error: 𝑚𝑖𝑛 𝐏, 𝐐 * ",!∈. (𝑟"! − ̂ 𝑟"!)/ = * ",!∈. (𝑟"! − 𝒑" # 𝒒!)/ Given 𝑛×𝑚 matrix 𝐑, the goal is to learn: Users/Items representations: 𝐏 ∈ ℝ0×', 𝐐 ∈ ℝ2×' Observed user-item interactions (known): 𝑺 observed rating score predicted rating score ≈ 𝒒! 𝒑" # 𝒅 𝒅 User representations 𝐏 ∈ ℝ!×# Item representations QT ∈ ℝ#×$ × ≪ 𝒎𝒊𝒏(𝒏, 𝒎) User-item Rating Matrix 𝐑 5 4 … 5 … 5 … 5 1 2 5 𝑛 users 𝒎 items (movies) … … Task: rating prediction in Netflix 9 Matrix Factorization
- 10. 10 Deep Learning is Changing Our Lives
- 11. 11 Reinforcement Learning (RL) Graph Neural Networks (GNNs) Automated Machine Learning (AutoML) Deep Recommender Systems Fundamentals of Deep Recommender Systems Graph-structured Data § Information isolated island Issue: ignore implicit/explicit relationships among instances Items Users !! "! !! "" !! "# !" "! … … 0 0 1 Field 1 Field m Field M 0 1 0 1 0 0 Manually Deisgned Architectures § Expert knowledge § Time and engineering efforts Recommendation Policies § Offline optimization § Short-term reward Adversarial Attacks ... ... ... ... ... ... ... ... ... ... Normal Data Fake User Profiles Generating Fake User Profiles Attacker promote/demote Poisoning attacks: § Promote/demote items § White/grey/black-box attacks
- 12. 12 Agenda • Introduction to Recommender Systems (Jiliang Tang) • Fundamentals of Deep Recommender Systems (Wenqi Fan) • Reinforcement Learning for Recommendations (Xiangyu Zhao) • Coffee Break (10 mins) • Graph Neural Network for Recommendations (Wenqi Fan) • AutoML for Recommendations (Xiangyu Zhao) • Adversarial Attacks for Recommendations (Wenqi Fan) • Future