SLIDE 1
Dynamics in Recommenders
- Users and Items are constantly in flux.
… …
Recurrent Recommendation with Local Coherence Jianling Wang and - - PowerPoint PPT Presentation
Recurrent Recommendation with Local Coherence Jianling Wang and - - PowerPoint PPT Presentation
Recurrent Recommendation with Local Coherence Jianling Wang and James Caverlee Dynamics in Recommenders Users and Items are constantly in flux. Local Coherence Within a short-term sequence, the neighboring items or users is
SLIDE 2
SLIDE 3
Local Coherence
- Within a short-term sequence, the neighboring items or
users is likely to be coherent.
… … …
SLIDE 4
Local Coherence
- Within a short-term sequence, the neighboring items or
users is likely to be coherent.
… … …
Coherent Items:
SLIDE 5
Problem
- Balance the local coherence with long-term evolution
among both implicit and explicit feedback.
SLIDE 6
Our Goal
- Predict how users will rate items in the future.
?
SLIDE 7
Our Approach
Recurrent Recommendation with Local Coherence - RRLC
SLIDE 8
Challenge 1: Feedback Sparsity
- We want to model the consistent latent factors of users
and items with their interactions.
Embedding Embedding Consistent Consistent
SLIDE 9
Challenge 1: Feedback Sparsity
- The interactions between users and items are sparse.
Embedding Embedding Consistent Consistent
Sparse Data !
SLIDE 10
- Coherence-based Neighbors.
{
{
Embedding
Coherence-based Neighboring Users
Embedding Consistent Consistent
Sparse Data !
Coherence-based Neighboring Items
Solution
SLIDE 11
Challenge 2: Representations
- We use Recurrent Neural Networks (RNN) to model the
dynamics of users and items.
∆T
Dynamic Dynamic … …
∆T
User-RNN Item-RNN
SLIDE 12
∆T
User-RNN Item-RNN
∆T
Dynamic Dynamic … …
? ?
Challenge 2: Representations
- We want effective representations of rating events.
SLIDE 13
∆T ∆T
Dynamic Dynamic … …
Coherence-based Rating Event Embedding Coherence-based Rating Event Embedding
Solution
- Rating Event Embedding utilizing local coherence
User-RNN Item-RNN
SLIDE 14
RRLC vs. Neural Models
Time-independent Neural Models:
- ~6% Improvement
SLIDE 15
RRLC vs. Time-Dependent Models
Time-dependent Models:
- ~5% Improvement
SLIDE 16
Conclusion
- Propose a novel RNN-based recommender
- Enhanced with local coherence on both implicit and
explicit feedback sequences
- Outperform state-of-art models in ratings prediction
- Help to alleviate the cold start problem
Texas A&M University, USA