WWW 2013

Related work

User and product features change over time

Recommending products to users

Movies may be preferred once they are considered 'classics' (Koren, 2010); new products may cause users to change their focus (Kolter & Malouf, 2007)

• Users should be compared to others reviewing the product at the same time

From amateurs to connoisseurs:

modeling the evolution of user expertise

through online reviews

Goal is to predict users' ratings of products

Special effects that were good in 2003

may not be good in 2013

• Age of the product

We need to understand the preferences of the user, and the properties of the product

e.g. How much will Julian like Harry Potter?

Julian McAuley, Jure Leskovec

Stanford University

this paper!

• Users should be compared to others with the same 'age'

• Age ('development') of the user

A child who likes Harry Potter in 2003

may have outgrown it in 2013

• Age ('zeitgeist') of the community

Even though today's children like Harry Potter,

the children of 2023 may not

• Users should be compared to members of their community or social network

Users may influence each other (Ma et al., 2011; Moe & Schweidel, 2012), and communities may shift (Xiong et al., 2010)

Each model makes different assumptions about which users are 'similar'

Our goal is to compare and model these effects

An example

How do users 'age'?

Models of temporal evolution

'Global' versus 'Personal' clocks

We don't model physical age, but rather development, or 'expertise'

Start with the 'standard' latent factor model

Should users/products/communities be modeled in terms of their age or in terms of the date?

'expert preference'

• A user should watch many movies before they will enjoy Citizen Kane
• A user should drink many wines before they will enjoy a Romanee Conti
• A user won't enjoy a bitter IPA, a smokey whiskey, or a strong cheese, until they have acquired a taste for them

Replace it by a series of models, indexed by a user's 'life stage'

• Are 13 year old wines good, or are wines from 2000 good?
• Did I start to enjoy Games of Thrones because it became popular, or because I became interested in fantasy?
• Is Harry Potter popular in 2013, or amongst 13 year olds?

Users need to be ready for the products they consume

The very act of consuming products will cause users' tastes to change

rating

If we know each user's experience level, fitting the parameters is easy

Products/communities evolve over time:

time

Experiments

Users evolve over time:

Learn the rate at which each user evolves:

We need to fit the parameters of each recommender system

never became an expert

What are we really learning?

started out (somewhat) expert

Experts agree with each other

How do experts differ from beginners?

Users who fail to become

experts are likely to quit

Thanks!

time

1) Experts give similar ratings when reviewing the same products

time

Conclusion

Datasets

Results

The two steps repeat

until convergence

Model fitting

(mean squared error of rating predictions on test set

users with at least 50 ratings)

Bibliography:

• The constraint that users 'progress' between successive models means that all users evolve in the same 'direction'
• Really, we're learning some property of evolution that's common to all users, regardless of when they arrive in the community
• 'Expertise' is the name we give to this property

• Temporal evolution occurs on the level of products, users, and communities
• We studied how users evolve as they review more products - or as they become more expert
• Modeling expertise leads to more accurate performance, and leads to novel insights about how beginners and experts behave

• J. Kolter and M. Maloof. Dynamic weighted majority: An ensemble method for drifting concepts. JMLR, 2007.
• Y. Koren. Collaborative filtering with temporal dynamics. Commun. ACM, 2010.
• H. Ma, D. Zhou, C. Liu, M. Lyu, and I. King. Recommender systems with social regularization. In WSDM, 2011.
• W. Moe and D. Schweidel. Online product opinions: Incidence, evaluation, and evolution. Marketing Science, 2012.
• H. Einhorn. Expert judgment: Some necessary conditions and an example. Journal of Applied Psychology, 1974.

• Beloved products are most beloved by experts
• Hated products are most hated by experts

Data are available on http://snap.stanford.edu/data/

2) They also are individually more predictable (see paper)

see also Danescu Niculescu-Mizil et al. (Thursday session 5)

for train/test splits, regularization etc., see our paper

• Some users start out experienced
• Others never become experienced
• Each user evolves at their own rate

(1) and (2) are arguably necessary conditions for users to be considered 'experts' (Einhorn, 1974)

We also need to fit users' experience progression

• Users gradually become expert over time
• All 'beginners' use the same model, regardless of when they arrive in the community
• But! Not all users should evolve at the same rate!

Since users gain experience monotonically, we can fit experience using Dynamic Programming