Recsys14 int rs_vassileva

Interac(ve Visualiza(ons of Recommenda(ons for Social Streams

Julita Vassileva University of Saskatchewan

Plan

•  Problem •  Early works

– Open learner modeling – Social naviga(on

•  Visualiza(on and user control of recommenda(ons in streams – 6 approaches

•  Analysis and general challenges

Problem •  Recommender systems serve to reduce user’s overload •  Most recommender systems are evaluated for analy(cal metrics, such as

accuracy, coverage, diversity, novelty •  User sa(sfac(on with recommenda(ons is harder to evaluate

–  Depends on user trust (Felfering, 2007, Cramer et al, 2008) –  Trust can be gained with explana(on of how the system works (Tintrarev &

Masthoff, 2007) (Herlocker, et al. 2000) –  Depends on user understanding and control (Knijenburg et al, 2012): “…

inspectability and control indeed increase users’ perceived understanding of and control over the system, their ra8ng of the recommenda8on quality, and their sa8sfac8on of the system”

–  But explana(ons and control should not cause addi(onal overload on the user •  Research Ques(on: how to present the recommena(ons and

–  Explain the reasons / mechanism for adapta(on and –  Involve and Give control to the user –  Without causing extra work and overload on the user

General solu(on

•  Open or explain how the recommenda(ons are generated – Open the user model – Explain the mechanism of genera(ng recommenda(ons

•  Involve the user to tune the model and the recommender

•  Use a visualiza(on (less cogni(ve load)

Historical overview of interac(ve/visual recommenders

•  Intelligent tutoring systems and learning environments

•  Open learner models •  Adap(ve naviga(on support •  Social naviga(on support •  Interac(ve Recommender systems

Give user control

To change rec. process

To change model

Visualize other user models – open social

Content (Compare with others)

Footprints (Naviga(on support)

Visualize user model To s(mulate reflec(on

on content (Kay, Bull, Dimitrova,

Zapata Rivera & Greer…)

To explain recommenda(on (Brusilovsky)

The idea of “Skillometers”

•  Corbeg and Anderson, 1995 -‐ APTList Tutor •  … •  …. •  ……. •  ……… •  ………… •  Finally now it shows up in the HCI community: Malacria, S., Scarr J, Cockburn, A., Gutwin C., Grossman, T. (2013) Skillometers: reflec(ve Widgets that mo(vate and help users to improve performance, ACM UIST’13 , 321-‐333.

The ideas of opening up the learner model 1994-‐1995

•  Judy Kay with students -‐ 1994 – scrutability of learner /user models –  To verify it and eventually fix errors –  But can we trust the learner to change the model?

•  Susan Bull – 1995 – focus on open learner models as a tool for reflec(on – Don’t trust learner correc(ons, store both the learner’s and system’s model of the learner

– Discuss the differences, involve learner in argument

ELM: Visualiza(on of Rec’s for Naviga(on in a Course

Brusilovsky, Eklund (1998) A Study of User Model Based Link Annota(on in Educa(onal Hypermedia, JUCS4(4),429-‐448.

VisMod – open model to teachers & parents

JD Zapata-‐Rivera, JE Greer (2000) Inspec(ng and visualizing distributed Bayesian student models Intelligent Tutoring Systems, 544-‐553

CourseVis: presen(ng mul(ple models of students to teachers

R Mazza, V Dimitrova (2003) CourseVis: A graphical student monitoring tool for suppor(ng instructors in web-‐based distance courses, Interna(onal Journal of Human-‐Computer Studies 65 (2), 125-‐139

Comtella: Visualizing peer contribu(ons: to mo(vate par(cipa(on

Bretzke H., Vassileva J. (2003) Mo(va(ng Coopera(on in Peer to Peer Networks, Proceedings User Modeling UM03, Johnstown, PA, June 22-‐26, Springer Verlag LNCS 2702, 2003, 218-‐227.

Cheng R., Vassileva J. (2006) Design and Evalua(on of an Adap(ve Incen(ve Mechanism for Sustained Educa(onal Online Communi(es. User Modelling and User-‐Adapted Interac(on, 16 (2/3), 321-‐348.

Comparison with peers

Hsiao, Bakalov, Brusilovsky, Konig-‐Reiss (2011) Open Social Student Modeling: visualizing student Modes with parallel introspec(ve views, Proc. UMAP 2011, 171-‐182.

Comparison to peers on a scale: Social naviga(on

Brusilovsky & Sosnovsky, 2003

Social Naviga(on with QuizMap

Peter Brusilovsky, I-‐Han Hsiao, Yetunde Folajimi (2011) QuizMap: Open Social Student Modeling and Adap(ve Naviga(on Support with Tree Maps. Towards Ubiquitous Learning, Springer

Bostandjev’s layered approach for visualizing hybrid recommenda(ons

Control to the user: Interac(ve hybrid recommender visualiza(on

•  hgp://bit.ly/TasteWeights

Bostandjiev S., O’Donnovan, J., Hoellerer T. (2012) TasteWeights: a Visual Interac(ve Hybrid Recommender System. RecSys’12

•  Un(l 10 yrs ago -‐ recommending in a fairly sta(c context (educa(onal content, music, music, products)

•  However the social web (with user-‐generated content) has brought both opportuni(es and problems

hgp://www.citravel.com/blog/meet/?p=94

hgp://www.ilead.co.za/blog/the-‐age-‐of-‐informa(on-‐overload-‐and-‐ how-‐to-‐simplify-‐communica(on.html

What did we do 2006-‐now? •  Focused on social environments à recommending items of interest in (me-‐based streams of social (user-‐generated) data

•  Different requirements from recommending products or people à User sa(sfac(on is crucial (otherwise they won’t use the tool)

à You don’t want to miss something important; but if something unimportant gets recommended – no problem à Recall is more important than precision

Comtella-‐D •  Andrew Webster, 2006 •  Recommends posts in discussion forum, and threads based on their popularity and ra(ng

•  Visualizes highly rated content with visual emphasis (rec. by popularity) – Hierarchical propaga(on of visual emphasis from post to thread

•  Visualizes the impact of a single ra(ng on rec. For threaded discussion forums Popularity based, not personalized

Comtella-‐D (2006) Webster A.S., Vassileva J. (2006) Visualizing Personal Relations in Online Communities, (2006) in Adaptive Hypermedia and Adaptive Web-Based Systems, Dublin, Ireland, June 21-23, 2006, Springer LNCS 4018, 223-233.

The quick red fox jumped over the lazy red dog. By Andrew

Immediate visual feedback on user ac(ons of ra(ng posts

@work Energy Stored Energy

The quick red fox jumped over the lazy brown dog. By Andrew



All generaliza(ons are false, including this one. By Mark Twain

All generaliza(ons are false, including this one. By Mark Twain

Evalua(on While subtle, the visual “hea(ng up” or “cooling down” effect that followed the user’s act of ra(ng was fun, figed the energy metaphor, and didn’t distract the users’ agen(on or pose an addi(onal cogni(ve load. The energy system acted as an implicit recommender func(on

hgp://ecommons.usask.ca/bitstream/handle/10388/etd-‐09192007-‐204935/thesis_webster.pdf

KeepUP

•  Andrew Webster, 2007 •  Recommends RSS feeds from different streams, seman(cally annotated (tags, LSA)

•  Visualizes the user’s neighbourhood in CF and allows user to manually change the impact of other users over him

Personalized – Collab. Filtering within Channels, Content-‐Based for Channels

KeepUP Vis – Ist version

KeepUP (2007)

Webster A.S., Vassileva J. (2007) The KeepUp Recommender System. ACM RecSys 2007, 173-177

Evalua(on

•  13 par(cipants, ques(onnaire auer using KeepUp

IBlogVis •  Indratmo 2008-‐2010 •  Recommenda(on by the social interac(on history around content (#commenters, # digs) – like social naviga(on

•  For streams, archives, (meline-‐based data -‐ Visualiza(on of a blog archive

•  Interac(ve visualiza(on allows browsing, zooming (me periods…

Popularity-‐based rec (not personalized)

IBlogVis-‐1 (2008) Indratmo, Vassileva J. (2008) Exploring Blog Archives with Interac(ve Visualiza(on, in AVI'2008

Evalua(on

•  Lab study – 19 students, predefined exploratory tasks (overview, browsing, social naviga(onal), followed by ques(ons (closed and open-‐ended)

•  Results: – Error rate 0-‐3 out of 16 tasks, avg. 6.25% – User sa(sfac(on

iBlogVis can be used as a collabora(ve filter

IBlogVis-‐2 (2009) Indratmo, Vassileva J. (2012) The Role of Social Interac(on Filter and Visualiza(on in Casual Browsing, Proc. IEEE HICSS-‐45

Evalua(on •  Summa(ve – within subjects design

–  Independent variable: type of vis – 3 values: List, TimeVis, and SocialVis

–  Dependent variables: interest score, percep(on of support, wasted effort, user sa(sfac(on, and overall preference

–  30 par(cipants, different ages and background

SocConnect •  Yuan Wang and Jie Zhang, 2009-‐2010 •  Recommends social updates in social network aggregator – content based (LSA), machine learning

•  Visualiza(on of recommenda(ons via emphasis in the stream

•  User can filter -‐ from which SSN and which friends to see updates, and what tags/topics

Personalized – content based

SocConnect (2010)

Wang Y., Zhang J., Vassileva J. (2010). SocConnect: Intelligent Social Networks Aggregator, UMAP’2010, Hawaii, June 20-‐25, 2010, Adjunct proceedings of UMAP'2010. .

SocConnect Evalua(on

•  Field study – 11 valid par(cipants (out of 21 recruited on FB and Twiger); using SocConnect for 2 weeks

•  Usage data analysis and final ques(onnaire to evaluate user sa(sfac(on with the new func(ons

-‐  Only half of the par(cipants found the use of color to highlight recommenda(ons intui(ve

-‐  Ouen recommended items buried down in the stream -‐  Par(cipants split in their awareness and sa(sfac(on with

the new func(onali(es to control the filtering (blend, group) -‐

Details at: hgp://ecommons.usask.ca/bitstream/handle/10388/etd-‐11292010-‐112405/yuan_thesis.pdf

Rings

•  Shi Shi, 2011 (hgp://rings.usask.ca) •  Recommends what to read on Facebook – overview first – details on demand; “Big-‐Bang” (meline vis.

•  Prolific and reputable users and popular posts are emphasized visually

•  User can interact by filtering by users, by (me period (but no influence on the recommenda(ons)

•  Mostly used to avoid the filter bubble of Facebook

Popularity-‐based rec (not personalized)

Rings (2011)

Hgp://rings.usask.ca

Evalua(on

•  Field study – 21 users, need to use Rings 5 (mes over 3 weeks, auer 5 min usage prompted to answer 5 ques(ons .

5 par(cipants completed 5 daily ques(onnaires 1 par(cipant finished 4 daily ques(onnaires 2 completed 2, 6 filled just one ques(onnaire, 1 completed 9! And 6 didn’t fill even one…

hgp://ecommons.usask.ca/handle/10388/ETD-‐2011-‐09-‐139

MADMICA •  Sayooran Nagulendra (2013, 2014) •  Filters away posts from friends about things which are not of common interest with the user (Content-‐based overlaid over Social network)

•  Visualiza(on creates awareness of what is being filtered away from each friend (in what category of interest)

•  User can interac(vely change the way the recommender works

Personalized (content-‐based, social)

MADMICA (2013)

hgp://prezi.com -‐ Video

hgp://madmica.usask.ca

Evalua(on

•  Crowd-‐sourced lab study – M-‐turk with 326 par(cipants (2 equal-‐size groups, one with embedded help and one without help)

– Working bubble visualiza(on incorporated in a ques(onnaire with tasks and 25 ques(ons

– Metrics for visualiza(on undestandability based on Interna(onal Standards for Souware Quality Evalua(on

Analysis of our approaches-‐1 •  Applica(on area: recommending stream data

–  Threaded discussion forums: Comtella-‐D –  RSS feed aggregators: KeepUP –  Digg (archives): IBlogVis –  Social network aggregators: SocConnect –  Social network streams

•  Centralized (FB): Rings •  Decentralized: Madmica

•  Purposes of using visualiza(on –  Support exploratory browsing: Comtella-‐D, iBlogVis, Rings –  Explaining the recommenda(on process to increase user trust: Comtella-‐D, KeepUP, SocConnect

–  Avoiding the filter bubble: Rings, Madmica

Analysis of our approaches-‐2 •  Two main kinds of recommenders exploited

–  Popularity-‐based / social interac(on history (not personalized): Comtella-‐D, iBlogVis, Rings

–  Personalized: •  Collabora(ve Filtering + Content-‐based + Social: KeepUP, •  Content-‐Based + Social NW-‐based: SocConnect •  Content-‐Based + Social NW-‐based: Madmica

•  Two types of adapta(ons exploited –  Visual emphasis: Comtella-‐D, iBlogVis, Rings, Madmica –  Filtering of non-‐recommended items: KeepUp, SocConnect, Madmica

Analysis of our approaches-‐3

•  Interac(vity dimensions – provide immediate feedback to user ac(ons

•  User ra(ng ac(ons: Comtella-‐D à Visual effects on increasing/decreasing “hotness” of item

•  User adjus(ng strength of influence from other users: KeepUP, SocConnect, Madmica à Changing the recommended set of items in the user’s stream

•  Allow user to explore the data interac(vely : iBlogVis, Rings

– allow user to change the recommenda(on algorithm variables: KeepUP, Madmica

General Challenges •  Unconstrained design space for visualiza(on -‐ finding good pagerns?

•  Evalua(on challenges in field studies – Hard to isolate the effect of the visualiza(on and interac(ve func(ons from the recommender algorithm

–  Rec. sys. for social streams need many users and long (me to work well …

•  Difficulty in recrui(ng large number of par(cipants •  Difficulty ensuring par(cipa(on in loner-‐term studies •  Bias towards studies on large social networks (need insider connec(ons)

•  Hard for academics to publish small scale studies

References •  Comtella-‐D:

•  KeepUP:

•  IBlogVis-‐1

•  IBlogVis-‐2

•  SocConnect

•  Rings

•  MADMICA

•  Webster A.S., Vassileva J. (2006) Visualizing Personal Rela(ons in Online Communi(es, Proc. AH’06, Dublin, Ireland, LNCS 4018, 223-‐233.

•  Webster A.S., Vassileva L. (2007) Push-‐Poll Recommender: Suppor(ng Word of Mouth, Proc. UM’07, Corfu, Greece

•  Webster A.S, Vassileva J. (2007) The KeepUP Recommender System, Proc. ACM RecSys’07, 173-‐177.

•  Indratmo, Vassileva J. (2008) Exploring Blog Archives with Interac(ve Visualiza(on. Proc. AVI’2008, Napoli.

•  Indratmo, Vassileva J. (2012) The Role of Social Interac(on Filter and Visualiza(on in Casual Browsing. Proc. IEEE HICSS-‐45.

•  Wang Y, Zhang J, Vassileva J. (2010) SocConnct: A User-‐Centric Approach to Social Networking Sites Aggrega(on. Proc. WS on Social Seman(c Web, with IUI’2010, HK

•  Shi S., Largillier T., Vassileva J. (2012) Keeping Up with Friends’ Updates on Facebook. Proc. CRIWG’2012, LNCS 7493.

•  Nagulendra S, Vassileva J (2013) Providing Awareness, Understanding and Control of Personalized Stream Filtering in a P2P Social Network, CRIWG’2013.

•  Nagulendra S., Vassileva J (2013) Minimizing Social Data Overload through Interest-‐Based Stream Filtering in a P2P Social Network. Proc. ASE/IEEE SocialCom’2013.

•  Nagilendra S, Vassileva J (2013) Understanding and Controlling the Filter Bubble through Interac(ve Visualiza(on. CRIWG’2014.

Recsys14 int rs_vassileva

Social Media

Transcript of Recsys14 int rs_vassileva