Recommending Twitter Users to Follow Using Content and Collaborative Filtering Approaches

John Hannon, Mike Bennett, Barry Smyth CLARITY Centre for Sensor Web Technologies

University College Dublin

Outline

Problem1

Method & Experiment3

Result & Analysis4

Related work & Innovation2

Problem

• The paper solves an important

recommendation problem— for a given user,

UT which other users might be recommended

as followers/followees, based on a large

dataset of Twitter users and their tweets.

• The motivation of the paper is to demonstrate

the potential for effective and efficient followee

recommendation.

Related Work

• Analysis of Twitter’s real-time data.– Kwak et al : reciprocity and homophily among Twitter users,

information diffusesion.

• User-generated content like review as an additional source is used in recommender system.– The use of user-generated movie reviews from IMDb as part of

a movie recommender system.

• Research to help users find and contact with people online.– The information such as co-authorships are used to identify

similar users.– Freyne and Geyer et al have done much work about

relationship building.• Make recommendations to new users during their sign-up

process. • Recommend Topics for Self-Descriptions in Online User Profiles

Innovation

• Twitter’s potential as a powerful source of

profiling data. This is a novel take on

profiling and recommendation in itself.

• Focus on noisy, unstructured micro-

blogging data.

• Novel contribution of the paper is that

noisy as Twitter data is, it can still provide

a useful recommendation signal.

twittomender

Approach

• How users are profiled

– Content-based techniques which rely on the

content of tweets.

– Collaborative filtering approaches based on

the followees and followers of users.

• How these profiles can be used to

suggest interesting users to follow.

– Lucene platform are used to develop the

framework.

Profiling Users on Twitter

• 5 basic profiling strategies:

(1) Representing users by their own tweets

(tweets(UT));

(2) By the tweets of their followees

(followeetweets(UT));

(3) By the tweets of their followers

(followertweets(UT));

(4) By the ids of their followees (followees(UT));

(5) By the ids of their followers (followers(UT)).

Indexing & Recommendation

• Using Lucene’s indexing features we can represent

each, UT , as a weighted term-vector, profile (UT,

source).

• profile (UT ,source) = {w1,…,wn}

• Term weighting function: TF-IDF

• Query-based retrieval and profile-based

recommendation are then implemented using Lucene's

standard retrieval function, with the target user's profile

document serving as the search query in the case of

the latter.

Experiment—dataset

• Imported 20,000 users directly using the

Twitter API as dataset. The dataset is split into

two sets of users –one containing 1000 users

to act as test users, and a larger training-set of

19,000 users;

9 different profile information

• S1: tweets(UT)

• S2: followeestweets(UT)

• S3: followerstweets(UT)

• S4: tweets(UT), followeestweets(UT), followerstweets(UT)

• S5: followee(UT)

• S6: follower(UT)

• S7: followee(UT), follower(UT)

• S8: the scoring function is based on a combination of

content and collaborative strategies S1 and S6;

• S9: the scoring function is based on the position of the

user in each of the recommendation lists.

Recommendation Precision

• Our basic measure of recommendation performance is

the average percentage overlap between a given

recommendation list and the target user's actual

followees-list;

• We can also see that relevant recommendations tend to

be clustered towards the top of recommendation lists

since the precision of all strategies is seen to decline

within increasing recommendation-list size. Interestingly,

the collaborative strategies perform better than the

content strategies;

Ranking Effectiveness

• The position of relevant recommendations is also an important consideration, especially since we know that users focus the lion's share of their attention on items at the top of results or recommendation-lists.

A live-user trial

• Shortage of the off-line evaluation ?• It’s unwise to discount the non-overlapping

recommendations as definitively not relevant to the target user.

User Recommendation: an average of 6.9 users per recommendation-list. User Search: an average of 4.9 of the suggested users per search.

Conclusion

Advantage:

• User-generated contents are used as a source of profiling data.

• Tweet doesn’t been preprocessed.

My idea:

• User’s tweet should be preprocessed, such as extracting tag

from tweet. The tag may be more important than content.

• Besides, other information such as the group user join in is also

worthy to take into account.

• Users can be divided into celebrity and people. For different kind

of users, different strategy should be take into account.

• Barry Smyth ： Centre Director • his research interests include personalization,

recommender systems, case-based reasoning, machine learning, and information retrieval. 

• Mr. John HannonPh.D. Student

• Mike Bennett is a postdoctoral researcher and interaction designer