WWW 2014 Notes

| Tags Conference  Notes 

The WWW2014 Conference Notes

08th April, 2014

Mining User Trails in Critiquing Based Recommenders

** Dataset (restaurant critiquing history) **


** useful reference: **

  • S. Gupta and S. Chakraborti. Utilsim: Iterativelyhelping users discover their preferences. InE-Commerce and Web Technologies, pages 113–124.Springer, 2013
  • Y. Salem and J. Hong. History-aware critiquing-basedconversational recommendation. WWW2013

Towards a scalable Social REcommender Engine for Online Marketplaces

  • real-time

  • scalability

  • social-network

Dating matching

  • reciprotocal revealed preference

    use LDA to infer the revealed preference emerged in the users’ behaviors

09th April, 2014

Personalized Collaborative Clustering

LCR: Local Collaborative Ranking

  • Ordering Problem

  • the rating matrix is low-rank only locally

  • PREA toolkit is opensource.

COBAFI: Collabarative Bayesian Filtering

  • Limitation: Gaussian assumption

  • Goals

    • Fit the recommender distribution

    • cold-start/cool-start/few ratings

    • spam and fraud

  • model

    • co-cluster users and items (extend probabilistic matrix factorization)

    • [great!]the recommender distribution [Tan et al, 2013]


  • Switching detection Challenge 2012

Repeat consumtption

  • BrightKite: location checkins
  • G+: ppublic location checkins

  • MapClicks: clicks on Google maps business

  • Youtube: last 10K video waches

  • Shakespeare http://www.cs.cornel.com/~shuochen/

  • popularity more frequently consumed items are morel likely to be reconsumed

  • recency How does the recency of consumptions affect thlieklihood of reconsumption? to answer this question, we usa cacehe-based analysis technique

the hit ratio is an indication of the degress to which recency is displayed in a consumption history

Real consumption sequences display a significant amount of recency

user-leel item popularity generally postive xxx

  • model *
  1. quality model: item quality dictates consumption behavior simply th eemprical fraction of occurences

  2. recency model: we formulate a copying model based on recency additive in weights item was consumed i steps ago and j steps ago maximize likelihood with SGD

  3. hybrid model

log-likelihood per item of models, normlizec by log-likelihood of hrybird model

we learn a wieght for each possible previous position weights follow power law with exponential cutoff

Demograhics, Weather and Online Reviews: A study of restaurant recommendations

Endogenous: restauants can control Exogenous: restauant cannot control

Weather & offline bahavior Weather & season

840K restaurant 1.1M reviews (text, time, locaiton) demographics and weather condition (using weather station)

  • Modeling reviews: Negative binomial regression count variable

  • Modeling rating: cumulative linke model (orderred logistic regression)

Monetary effects:

  • Low-priced restaurants: fewer reviews lower ratings

  • Online promotions: more reviews not necessarily higher ratings

Higher population, higher education and higher racial -> higher likelihood reviewed

Dataset is available http://tinyurl.com/k26sf5g

Twitter timeline generation

  • odesk platform

Growing and death of membership website prediction



How do peaople adopt new behavior ? Mansfiedld’61, ROgers’03




Myspace data

** Adoption Models: popuarity of successufl website: bass model

died from competition diedfrom incompeten xxx

Positive & negative Attention Loops

Modeled as reaction-decay process


clustering items using comments

finding progresssion stages in Time-Evolving Event sequences


Burst prediction

Previous     Next