Deep Learning Based Recommendation Systems: Part 2 Prof. Srijan - PowerPoint PPT Presentation

CSE 6240: Web Search and Text Mining. Spring 2020 Deep Learning Based Recommendation Systems: Part 2 Prof. Srijan Kumar http://cc.gatech.edu/~srijan 1 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Announcements • HW3: – Deadline extended to Wednesday upon multiple requests. – Sample output released on Piazza. Please check your submissions. • Project: – Milestone grades released. Grade queries to be submitted by Tuesday night. • Very nicely written reports! – Good luck for the final presentation and report! 2 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Today’s Lecture • Introduction • Neural Collaborative Filtering • RRN • JODIE 3 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

RRN • RRN = Recurrent Recommender Networks • One of the first methods to model the temporal evolution of user and item behavior • Reference paper: Recurrent Recommender Networks. CY Wu, A Ahmed, A Beutel, A Smola, H Jing. WSDM 2017 4 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Traditional Methods • Existing models assume user and item states are stationary – States = embeddings, hidden factors, representations • However, user preferences and item states change over time • How to model this? • Key idea: use of RNNs to learn evolution of user embeddings 5 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

User Preferences • User preference changes over time 10 years ago ? now 6 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Item States • Movie reception changes over time So bad that it’s great to watch Bad movie 7 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Exogenous Effects “La La Land” won big at Golden Globes 8 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Seasonal Effects Only watch during Christmas 9 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Traditional Methods • Traditional matrix factorization, including NCF, assumes user state u i and item state m j are fixed and independent of each other • Use both to make predictions about the rating score r ij • Right figure: latent variable block diagram of traditional MF 10 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

RRN Framework • RRN innovates by modeling temporal dynamics within each user state u i and movie state m j • u it depends on u it- and influences u it+ – Same for movies • User and item states are independent of each other 11 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Model Learning Setting • Actions are happening over time • How to split training and testing data to respect the time dependency? 12 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Traditional Random Split: N/A • Random train/test split violates the temporal dependency – Future actions can be in train, while past actions can be in test ? ? ? 13 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Realistic Learning Setting • Train on first K% data and test in the last data points ? ? ? ? 14 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

RRN Model • Train two RNNs: one for all users and other for all movies – User RNN parameters are shared across all users; same for movies User RNN Movie RNN 15 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

RRN Process • Initialization: User and movie embeddings are initialized – Initialization can be one-hot • Embedding update • Prediction: To predict the rating a user gives to a movie, the user’s embedding is multiplied with the movie’s embedding • Loss: User-movie rating score prediction error is used to update the RNN parameters 16 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

User RNN • User RNN takes a user’s (movie, rating) sequence – Each input: concatenation of movie embedding and one-hot vector of rating score – RNN initialization: special ‘new’ vector to indicate a new user • For the next user , the process is repeated, starting from initialization 17 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Movie RNN • Movie RNN takes the movie’s (user, rating) sequence – Each input: concatenation of user embedding and one-hot vector of rating score – RNN initialization: special ‘new’ vector to indicate a new movie • For the next movie , the process is repeated, starting from initialization 18 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Rating Prediction • What is the rating by a u i to m j at time t? • Take the user and movie embedding till time t and output rating ) • Output function: MLP, Hadammard product, etc. 19 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Model Training • Learn the model parameters 𝛴 such that the predicted rating is close to the actual rating • R( 𝛴 ) is a regularization term to avoid overfitting 20 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Experiments • Three datasets, several baselines – PMF: Salakhutdinov & Mnih NIPS ’07 – T-STD: Koren KDD ’09 – U-AR & I-AR: Sedhain et al. WWW ‘15 • Metric = RMSE (Root Mean Square Error) (RMSE) 21 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Temporal Effects • How well does the model capture the temporal effects? ? 22 22 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Exogenous Effects • RRN automatically captures the exogenous effects Oscar & Golden globe 23 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

System Effects • RRN automatically learns the system effects Netflix changed the Likert scale 24 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Movie Age Effect • RRN automatically learns effects that we typically capture via hand-crafted features Movie age effects 25 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

RRN Summary Future prediction Novel model ? ? ? ? Temporal dynamics Accurate prediction 26 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Today’s Lecture • Introduction • Neural Collaborative Filtering • RRN • JODIE 27 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

Predicting Dynamic Embedding Trajectory in Temporal Interaction Networks Xikun Zhang Jure Leskovec Srijan Kumar Stanford University Stanford University UIUC Georgia Institute of Technology Code and Data: https://snap.stanford.edu/jodie 28 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

[KDD’19] Temporal Interaction Networks Flexible way to represent time-evolving relations Represented as a sequence of interactions, Time sorted by time: Feature interaction user item time features Users Items 29 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

[KDD’19] Temporal Interaction Networks E-commerce Social media …... Web Education Finance IoT Accounts Application domains Posts 30 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

[KDD’19] Temporal Interaction Networks E-commerce Social media …... Web Education Finance IoT Students Application domains Courses 31 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

[KDD’19] Problem Setup Given a temporal interaction network interaction user item time features where generate an embedding trajectory of every user and an embedding trajectory of every item 32 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

[KDD’19] Goal: Generate Dynamic Trajectory 2 4 1 5 6 3 Input: Temporal Output: Dynamic trajectory interaction network in embedding space 33 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

[KDD’19] Challenges Challenges in modeling: • C1: How to learn inter-dependent user and item embeddings? • C2: How to generate embedding for every point in time? Challenges in scalability: • C3: How to scalably train models on temporal networks? 34 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

[KDD’19] Existing Methods C1 C2 C3 Co- Embed Train in influence any time batches Deep recommender systems • Time-LSTM (IJCAI 2017) • Recurrent Recommender Networks (WSDM 2017) • Latent Cross (WSDM 2018) Dynamic co-evolution • Deep Coevolve (DLRS, 2016) Temporal network embedding • CTDNE (BigNet, 2018) Our model: JODIE 35 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining

[KDD’19] Our Model: JODIE JODIE: Joint Dynamic Interaction Embedding • Mutually-recursive recurrent neural network framework Update User RNN Item RNN Component Project Projection Component Operator 36 Srijan Kumar, Georgia Tech, CSE6240 Spring 2020: Web Search and Text Mining