23
FOLLOW MY FRIENDS THIS FRIDAY! Ruth García Gavilanes Universitat Pompeu Fabra Neil O´Hare, Yahoo! Research Luca Maria Aiello, Yahoo! Research Alejandro Jaimes, Yahoo! Research An Analysis of Human-Generated Friendship Recommendations

Follow My Friends This Friday! An Analysis of Human-generated Friendship Recommendations

Embed Size (px)

Citation preview

FOLLOW MY FRIENDS THIS FRIDAY!

Ruth García Gavilanes Universitat Pompeu Fabra

Neil O´Hare, Yahoo! Research

Luca Maria Aiello, Yahoo! Research Alejandro Jaimes, Yahoo! Research

An Analysis of Human-Generated Friendship Recommendations

Twitter: Users tweet recommendations

Advantages

Recommendations from friends • Trust • More acceptances( Amin et al., Recsys’12)

•  Features unknown for machines

Disadvantages

•  Noise •  Commercial purpose and Spam •  Not all recommendations are delivered to or seen by target users

We propose a way to measure the effect of recommendations from humans only

Who-to-Follow Recommendations

Humans

In 2010 •  The most popular hashtag in 2010 was #followfriday

or #ff in several countries •  Recommending people became trendy • No personalized recommendations of who to follow

What is Follow Friday?

SOCIAL NETWORK RECOMMENDS PEOPLE TO

FOLLOW ON FRIDAYS.

TARGET USER

follows

Tie Strength

to follow or not to follow?

Objective

• Analyze the dynamics of Follow Friday : impact, effect in time, repetitions and longevity.

•  Identify important features in each recommendation by using a classifier

Method

FRIDAY SATURDAY SUNDAY MONDAYTHURSDAY

Follows Who is new?Follows ?

#followfriday

RECEIVER

RECOMMENDER RECOMMENDED USERS

ACCEPTED RECOMMENDATION

93% of follow friday Tweets

48 snapshots 24 weeks

snapshot snapshot

Acceptance Total

Initial set of users 55,000

Receivers 21,270

Recommenders 589,844

Recommended Users 3,261,133

Recommendation Instances

59,055,205

Accepted Recommendation Instances

354,687

Most Follow Friday Recommendations are not taken into account

right away

0.60% instance acceptance

Interactions

Most Follow Friday Recommendations are not taken into account

right away

Mentions Acceptance Rate

Recommender -> Recommendation 0.006 Recommendation <-> Recommender 0.009 Receiver -> Recommender 0.010 Recommender -> Receiver 0.011 Recommender <-> Recommendation 0.012 Receiver -> Recommendation 0.095 Recommendation -> Receiver 0.097 Receiver <-> Recommendation 0.145

IMPACT

Impact

• We need to compare Follow Friday recommendations to other models:

•  Implicit : Mentions that were not Follow Friday recommendations

• Unobserved : Follow Friday recommendations of the future only

#FF

Acceptance after n weeks

Repetitions & Recommenders

Features USER-BASED (per user) •  Attention

•  Followers vs Followees •  Mentions by other users •  Recommendations

•  Activity •  Average tweets per

day •  New followees •  Accepted

recommendations and recommenders

•  Mentions

RECOMMENDATION-BASED (per recommendation) •  Repetitions

–  Repeated recommendations

–  Different recommenders •  Format

–  Day of the week –  Re-tweet or not –  Same tweet

recommendations –  Urls

RELATION-BASED (per pair)

•  Tie Strength –  Mentionss –  Folllow Friday

recommendations –  Previous

acceptances –  Friendship longevity

•  Similarity –  Words, mentions,

hashtags and urls –  Geolocation

Methodology •  Three methods: Rotation Forest, Linear combination and

random •  Training : week 1 to 16 •  Test : week 17 to 23

•  Up to 2 weeks to calculate acceptance rate •  Recommendations accepted after two weeks were not

considered in the classifier. •  Balanced set for training •  Goal : accepted recommendations towards the top of

the ranking •  Evaluation with Mean Average Precision

Results

Ranking MAP

Rotation Forest 0.496

Linear Combination

0.057

Random 0.037

Features MAP

All 0.496

User-based 0.074

Relation-based 0.398

Recommendation-based

0.062

User + Relation 0.518

User + Format 0.079

Relation + Format 0.379

Lessons Learned •  Recommendations derived from Social Networks have an

impact on users decisions

•  Social accepted recommendations seems to last longer/more relevant

•  Many broadcasted recommendations are not seen

•  Not accepted recommendations can be followed in the future

•  Relation and user based features are better predictors of tie formation

Future Work • Can we rate recommendations according to

permanence/tenure?

• When should we consider an accepted recommendation? (never vs. some day)

• User study: Can we build an online recommender of social recommendations (and so promote recommendations not seen)?

• Add cultural differences in features, is there an improvement?

THANK YOU @ruthygarcia