Multidimensional Social Network in the Social Recommender System

Przemyslaw Kazienko; Katarzyna Musial; Tomasz Kajdanowicz

Source-linked AI summary

Multidimensional Social Network in the Social Recommender System

Przemyslaw Kazienko, Katarzyna Musial, Tomasz Kajdanowicz

arXiv:1303.0093v1Mar 1, 2013cs.SIcs.IRphysics.soc-ph

TL;DR

The paper asks how user activities in multimedia sharing systems can reveal multiple social and semantic relationships for analysis and recommendation. It constructs a layered Flickr network, measures its relations, and adapts recommendation weights to users’ behaviour. The experiments report denser, increasingly multidimensional connections and higher-rated recommendations after adaptation.

Problem
Existing analyses of Flickr did not fully represent the multiple direct and object-based relationships available in multimedia-sharing activity data.
Method
The paper extracts direct and object-based relation layers from Flickr, assigns layer-specific strengths, and uses adaptively recalculated personal weights to recommend users.
Results
Recommendations after adaptation were rated 8% better on average, while the number of ties linking the same users increased almost four times year by year.
Takeaways & Limitations
Layer-specific and adaptive weighting makes human recommendations personalized to users’ activities and preferences within the multidimensional network.

Abstract

from arXiv · show

All online sharing systems gather data that reflects users' collective behaviour and their shared activities. This data can be used to extract different kinds of relationships, which can be grouped into layers, and which are basic components of the multidimensional social network proposed in the paper. The layers are created on the basis of two types of relations between humans, i.e. direct and object-based ones which respectively correspond to either social or semantic links between individuals. For better understanding of the complexity of the social network structure, layers and their profiles were identified and studied on two, spanned in time, snapshots of the Flickr population. Additionally, for each layer, a separate strength measure was proposed. The experiments on the Flickr photo sharing system revealed that the relationships between users result either from semantic links between objects they operate on or from social connections of these users. Moreover, the density of the social network increases in time. The second part of the study is devoted to building a social recommender system that supports the creation of new relations between users in a multimedia sharing system. Its main goal is to generate personalized suggestions that are continuously adapted to users' needs depending on the personal weights assigned to each layer in the multidimensional social network. The conducted experiments confirmed the usefulness of the proposed model.

I. INTRODUCTION

The paper treats multimedia sharing systems as sources of layered social networks, where users connect through direct relationships and shared activities involving multimedia objects. It analyzes these layers in Flickr and uses them to personalize human recommendations.

Motivation: Multimedia sharing systems record tagging, commenting, favouriting, grouping, and other activities that reveal relationships among users.Comments can indicate similar interests, while direct contact lists represent explicit relationships.
Motivation: Users interact through both semantic links involving multimedia objects and social links involving people they know or like.These two relationship types motivate a multidimensional social-network representation.
Contribution: The proposed recommender suggests potentially interesting users by exploiting relationships that are often indirect rather than explicitly visible.Its purpose is to help users establish new interpersonal relationships in multimedia sharing systems.
Related work: Flickr had commonly been studied as a social-tagging system, but the paper extends analysis to multiple relation types extracted from available user activities.Earlier work also examined direct contacts, common tags, groups, favourites, comments, and authorship relations.
Contribution: The paper analyzes layer profiles across Flickr snapshots from 2007 and 2008 and compares the resulting social layers.It identifies two additional layers beyond earlier studies and develops a deeper human-recommendation analysis.

III. MULTIDIMENSIONAL SOCIAL NETWORK

The multidimensional social network aggregates direct intentional links and object-based relations into layered ties between Flickr users. Its layers distinguish social connections from shared activities and assign relation strengths based on user behaviour.

Relation types: Direct relations represent explicit user links, whereas object-based relations connect users through shared multimedia objects.Object-based relations may have equal roles, such as commenting on the same picture, or different roles, such as commentator and author.
Network definition: A user tie exists when users have a direct intentional link or share an activity, and the multidimensional network aggregates all relation layers.For Flickr, the layers include contacts, tags, groups, favourites, and opinions about pictures.
Flickr layers: Flickr contains eleven identified relation types spanning contact lists, shared tags, groups, favourites, and opinions about pictures.These relations are classified by whether users have direct links, equal roles, or different roles toward an object.
Relation strength: The strength of each relation is layer-specific and generally increases with the amount of activity by one user toward another.Relations may use unary or real-valued strengths rather than a single common weighting scheme.
Contact-derived layers: Contact-of-contact relations connect users when an intermediary appears on both users’ contact lists.The relation strength is based on the number of intermediary users who simultaneously list the target user.

C. Relations Based on Tags

Tag-based relations connect Flickr users through shared tags applied to their photos. The relation is an equal-role object-based link whose strength reflects the proportion of one user’s shared tags that overlap with another user.

Relation extraction: A tag-based relation exists when two users have used at least one common tag to describe their photos.Only tags used by at least two users enter the set of shared tags.
Relation strength: The tag-relation strength is calculated as the number of tags common to two users divided by the number of shared tags used by the first user.The measure counts whether a tag was shared at least once, not how many photos received that tag.
Relation type: Tag-based relations are object-based relations with equal roles because all users have the same role toward the pictures they tag.The shared object is the tagging activity associated with users’ photos.

D. Relations Based on Groups

Group-based relations connect users through shared groups containing photos from both users, while favourite-based relations connect users through shared favourites or author–favourite roles. Their strengths are normalized by users’ related activity counts.

D. Relations Based on Groups: A group-based relation exists when users’ photos occur in at least one common group.The group contains photos authored by both users.
E. Relations Based on List of Favourites: Favourite–favourite relations connect users who marked at least one common photo as a favourite.The corresponding strength uses jointly favourited photos relative to the first user’s favourited photos.
E. Relations Based on List of Favourites: Author–favourite relations connect a photo author with a user who marked that author’s photo as a favourite, with reverse directions represented separately.The paper distinguishes author-to-favourite and favourite-to-author relations.
E. Relations Based on List of Favourites: Favourite-based relations may have equal roles or different roles, depending on whether users share favourites or connect through authorship.The three favourite layers are Rff, Raf, and Rfa.

F. Relations Based on Opinions

Opinion-based relations are extracted from users commenting on common photos or from author–commentator interactions. The paper distinguishes equal-role and different-role opinion layers and assigns each a normalized strength.

F. Relations Based on Opinions: An opinion–opinion relation exists when two users comment on at least one common photo.Its strength is based on jointly commented photos relative to the first user’s total commented photos.
F. Relations Based on Opinions: An author–opinion relation connects a photo author to a user who commented on that author’s photo.The reverse opinion–author relation connects the commentator to the author.
F. Relations Based on Opinions: The three opinion-layer strengths are computed from shared comments or author–commentator counts normalized by users’ commenting or authorship activity.Equations (9)–(11) define the corresponding strength values.
F. Relations Based on Opinions: Opinion-based relations comprise equal-role Roo and different-role Rao and Roa layers.The existence of an author–opinion relation produces the corresponding reverse opinion–author relation.

G. Aggregation of Layers

The multidimensional social network aggregates direct links and shared-activity relations into combined ties between users. A tie exists whenever at least one relation layer connects the two users.

G. Aggregation of Layers: MSN=(U,L) contains users and ties derived from direct intentional links or shared activities.The tie set represents relations extracted from the system’s user activities.
G. Aggregation of Layers: A combined tie lij from ui to uj exists if any relation of any kind connects ui to uj.The aggregation therefore records whether at least one layer supplies a connection.
G. Aggregation of Layers: The combined network is formed by summing the relation sets identified across the system.Individual layers are aggregated into the multidimensional network’s tie set L.

L= RcRrcRcocRtRgRffRfaRafRooRaoRoa, (12)

The model assigns strengths to component relations and aggregates them into linkage strengths across layers. Layer importance weights can control the contribution of each relation type, while historical activity can be time-adjusted.

L= RcRrcRcocRtRgRffRfaRafRooRaoRoa, (12): Each existing tie receives a linkage strength based on the strengths of its component relations.The paper allows different formulas for relation strengths.
L= RcRrcRcocRtRgRffRfaRafRooRaoRoa, (12): αk represents the static importance coefficient of the kth relation layer.For Flickr, k indexes eleven relation layers, including contact, tag, group, favourite, and opinion relations.
L= RcRrcRcocRtRgRffRfaRafRooRaoRoa, (12): Relation and tie strengths are constrained to the range [0;1].The aggregation combines strength values from all discovered relation levels.
L= RcRrcRcocRtRgRffRfaRafRooRaoRoa, (12): Historical activity can be time-weighted using tp^λ, where tp is the number of fixed periods since the activity.This discounts or otherwise adjusts past activity according to λ.

IV. COMPARISON OF THE LAYERS IN MULTIDIMENSIONAL SOCIAL NETWORK

The paper compares nine Flickr relation layers across two datasets using binary and correlation-based similarity measures, after extracting layer-specific ties from user activity data. The comparison uses snapshots from 2007 and 2008 and excludes technically impractical contact relations.

Data and layer construction: Nine Flickr relation layers were extracted and compared across datasets from 2007 and 2008.The analyzed layers were Rc, Rt, Rg, Roo, Roa, Rao, Rff, Rfa, and Raf.
Similarity measures: Pearson correlation compares directed, real-valued layers from -1 for different corresponding relations to 1 for identical links with fully correlated strengths.The coefficient is applied to any two Flickr relation layers R1 and R2.
Similarity measures: Layer similarity was assessed with binary cosine M1, binary Jaccard M2, Pearson correlation M3, and additional binary graph-density measures.M1 and M2 operate on binary relations, while Pearson correlation incorporates real-valued relation strength; binary measures range from 0 to 1.
Data and layer construction: The experiments used the same users’ activity data collected in January 2007 and February 2008, selecting the 1,000 most tag-active users from an initial population of almost 2 million.After removing users without relations, the processed user sets contained 745 users in 2007 and 945 in 2008.
Layer statistics: The study calculated layer graph density and relation strength using the processed user sets and a common assumption αk=1 for every considered layer.These statistics were reported alongside layer-specific counts of ties, users, and contribution to the combined network.

U card U card

Across the two Flickr snapshots, the dominant relation layer shifted from groups in 2007 to tags in 2008, while favourite- and opinion-based relations showed stronger social profiles. Layer overlap and similarity patterns also changed substantially over time.

Layer profiles over time: 99.5% of all 2007 ties belonged to the group layer Rg, whereas Rg contributed 30% of ties in 2008.Rg covered 91% of users in 2007 and 77% in 2008; its average group size grew from 5.6 to 93.2 members.
Layer profiles over time: 99% of 2008 ties included the tag layer Rt, whose density approached 71% after nearly 200-fold growth in relations and a 280-fold increase in used tags.The number of users participating in Rt rose from 48% to 97%, and the number of tags exceeded 480 thousand.
Layer profiles over time: Rc remained relatively strong despite few contacts, averaging 1.4 relations per user at strength 0.73 in 2007 and 3.6 at strength 0.25 in 2008.For several layers, increases in relations per user accompanied decreases in average strength, unlike Rt, where both increased.
Social and semantic profiles: Favourite-author and author-favourite layers had average strengths above 0.9 in 2007 and above 0.4 in 2008, consistent with users favouriting material from relatively few close users.The paper interprets these layers as socially oriented, while Rt reflects semantic relationships.
Layer similarity: The highest Pearson correlation was 0.1 between Roa and Rao in 2007 and 0.682 between Rao and Roo in 2008.The strongest binary similarities also indicated mutual social involvement between complementary favourite and opinion relations.
Layer overlap: In 2007, only 4,026 relations, under 2.4% of total ties, were common to two or more layers; in 2008, overlap increased because Rt expanded and met Rc.Other layers remained comparatively independent in 2008.

V. SOCIAL RECOMMENDATIONS WITHIN MULTIDIMENSIONAL SOCIAL NETWORK

The system recommends users by aggregating similarities across multidimensional-network layers, using system and personal weights to personalize and adapt rankings from user feedback.

Recommendations aggregate similarities across all network layers into values used to rank candidate users.The ranking excludes existing contacts and blocked users, while previously viewed users receive reduced recommendation values.
System and personal weights separately encode the general and individual importance of each network layer.System weights aggregate personal weights across users, while personal weights reflect a layer's current usefulness for an individual.
User activities and ratings provide feedback for adapting personal layer weights after recommendations are used.Activities range from profile viewing to adding a recommended person to the contact list, with different importance values.
The process creates new network relations from interactions with recommended users across one or more layers.The system monitors profile views, comments, favourites, and contact-list additions as evidence of subsequent relationships.
For efficiency, weight updates and ranking-list revisions can run offline and periodically rather than after every activity.The paper gives once-a-day execution as an example schedule.

A. Data Preparation

The FlickrFront experiment compared equal-weight recommendations with recommendations adapted to user feedback. Adaptation improved ratings and shifted layer importance toward selected social and tag-based relations.

21,640 Flickr profiles supported recommendations for eight volunteers using all eleven network layers.The experiment used the online FlickrFront framework.
The initial recommendation list assigned each of the eleven layers an equal personal weight of 1/11.Users then rated the initial suggestions before adaptation.
After adaptation, users rated recommendations 8% better on average than the initial list.The second list was recalculated from adapted personal weights and presented two weeks later, excluding first-stage recommendations.
Rcoc and Rao personal-weight contributions increased by 220% and 65%, respectively, after adaptation.The tag-based layer Rt increased by 8%, while Roa, Rff, and Raf decreased by -59% to -66% on average.
The proposed framework uses layer-specific system and personal weights to personalize and adapt user recommendations.Personal weights are recalculated as users interact with recommendations.
Online calculation volume creates efficiency problems, motivating offline periodic processing of some tasks.The paper specifically mentions list creation and retaining only the n most similar users as tasks that can be moved offline.