Recsys 2018 論文読み会
Item recommendation on monotonic
behavior chains
Item recommendation on monotonic behavior chains
Mengting Wan, Julian McAuley (University of California, San Diego)
• Treat multiple types of feedback as monotonic behavior chains.
• Multiple types of user feedback (e.g., click, purchase, review, recommend)
• Several studies seek to connect implicit and explicit signals [8, 15, 18-20, survey 9]
• Propose chainRec, which models multiple types of interactions as monotonic behavior
chain. (Also releases new dataset, Goodreads, for this task.)
• Factorization by CP/PARAFAC tensor decomposition.
• Additive scoring function with a parametric rectifier
• Edgewise optimization
• Adding PMI to objectives for negative samples
• Comparison of AUC and NDCG for Steam, Yoochoose, Yelp, GoogleLocal, Goodreads
• Especially with variants or ablation model of proposed models.
In Short
Problem
Related
Work
Contrib
ution
Evalu
ation
Method
Monotonic behavior chain and its tensor representation
Proposed Model
parametric rectifier
Monotonic scoring function
CP/PARAFAC tensor decomposition
Optimization
Adding PMI to objectives for negative samples
(to adjust confidence for unobserved negatives)
Edgewise optimization:
training sampling focused on edges
(edge = two consecutive stages where
users exhibit different responses)
Other Optimization
Optimize for each stage Optimize conditional
probability
Dataset and Performance Comparison
edgeOptsliceOptcondOpt
Additional Comparison
Comparison on
each stage
Sensitivity to
dimensionality
Visualization
Separation of genres on each interaction’s embedding

Recsys2018 item recommendation on monotonic behavior chains

  • 1.
    Recsys 2018 論文読み会 Itemrecommendation on monotonic behavior chains
  • 2.
    Item recommendation onmonotonic behavior chains Mengting Wan, Julian McAuley (University of California, San Diego) • Treat multiple types of feedback as monotonic behavior chains. • Multiple types of user feedback (e.g., click, purchase, review, recommend) • Several studies seek to connect implicit and explicit signals [8, 15, 18-20, survey 9] • Propose chainRec, which models multiple types of interactions as monotonic behavior chain. (Also releases new dataset, Goodreads, for this task.) • Factorization by CP/PARAFAC tensor decomposition. • Additive scoring function with a parametric rectifier • Edgewise optimization • Adding PMI to objectives for negative samples • Comparison of AUC and NDCG for Steam, Yoochoose, Yelp, GoogleLocal, Goodreads • Especially with variants or ablation model of proposed models. In Short Problem Related Work Contrib ution Evalu ation Method
  • 3.
    Monotonic behavior chainand its tensor representation
  • 4.
    Proposed Model parametric rectifier Monotonicscoring function CP/PARAFAC tensor decomposition
  • 5.
    Optimization Adding PMI toobjectives for negative samples (to adjust confidence for unobserved negatives) Edgewise optimization: training sampling focused on edges (edge = two consecutive stages where users exhibit different responses)
  • 6.
    Other Optimization Optimize foreach stage Optimize conditional probability
  • 7.
    Dataset and PerformanceComparison edgeOptsliceOptcondOpt
  • 8.
    Additional Comparison Comparison on eachstage Sensitivity to dimensionality
  • 9.
    Visualization Separation of genreson each interaction’s embedding