220 likes | 359 Views
Flickr Tag Analysis. Ahmet Iscen. Outline. Social Media What is Flickr? Flickr Photos Association Rule Latent Semantic Analysis Latent Dirichlet Allocation Conclusions. Social Media. Important part of our daily lives today Twitter 12th largest country in the world
E N D
Flickr Tag Analysis Ahmet Iscen
Outline • Social Media • What is Flickr? • Flickr Photos • Association Rule • Latent Semantic Analysis • Latent Dirichlet Allocation • Conclusions
Social Media • Important part of our daily lives today • Twitter 12th largest country in the world • Two new members sign up every second to LinkedIn
What is Flickr? • Image and video hosting • Acquired by Yahoo! in 2005 • 51 million registered members and 80 million unique visitors as of June 2011 • 6 million photos • Widely used by researchers
Dataset • Xirong Li's Flickr-3.5M Dataset • 3,500,000 images • 570,000 unique tags • 270,000 unique user-ids • Randomly selected 250,000 images with their tags http://staff.science.uva.nl/~xirong/index.php?n=DataSet.Flickr3m
Challenges • Tags totally depend on the user • Can be extremely noisy • Huge range of possible words • Examples: milos tasic milosevrodjendan verjaardagmilos desember 2005 tmo
Preprocessing • Eliminate stopwords (a,for,the etc.) • Eliminate extreme words (those that appear less than 20 photos and more than 80% of the photos. • Porter Stemmer (only for association rule) • Convert everything to lowercase • Eliminate tags with less than 2 letters and more than 20 letters • Eliminate numerical tags
Association Rules Mining • Rapid Miner [york] --> [new] (confidence: 0.910) Support: 0.04 [geolat, geolon] --> [geotag] (confidence: 0.986) Support: 0.03 [hors, lotharlez] --> [caballo, cheval, hestur] (confidence: 0.846) Support: 0.03 [paard] --> [hors, lotharlenz, zirg] (confidence: 0.802) Support: 0.03 [hors, paard] --> [lotharlenz, zirg] (confidence: 0.802) Support: 0.03
Association Rules Mining • Poor results. • Probably due to noise and variance in data. • Takes too much time to process the words and find rules. • Need find alternative methods
Latent Semantic Analysis • Same as LSI (LSI used in IR field) • SVD on document-term matrix to reduce dimensionality • Words are compared by taking the cosine of the angle between two vectors by any two rows.
Implementation • Gensim – topic modeling toolkit • Python • Tested different corpus and topic sizes
Latent Semantic Analysis • 250000 photos, 20 topics topic #0: 0.997*"wedding" + 0.047*"family" + 0.023*"friends" + 0.022*"party" + 0.019*"reception" + 0.013*"california" + 0.011*"ceremony" + 0.009*"india" + 0.008*"church" + 0.008*"sanfrancisco" topic #11: 0.491*"newyork" + -0.463*"china" + 0.448*"nyc" + -0.233*"beach" + 0.174*"newyorkcity" + 0.146*"italy" + -0.132*"friends" + -0.123*"flowers" + 0.119*"new" + -0.117*"beijing" topic #4: 0.586*"paris" + -0.524*"family" + 0.417*"france" + 0.186*"london" + 0.178*"party" + -0.169*"halloween" + 0.156*"europe" + -0.121*"japan" + 0.103*"travel" + 0.063*"birthday" topic #1: 0.701*"halloween" + 0.588*"party" + 0.169*"friends" + 0.165*"family" + 0.157*"birthday" + 0.126*"japan" + 0.071*"christmas" + 0.059*"london" + 0.058*"travel" + 0.055*"beach"
Latent Semantic Analysis • 250000 photos, 50 topics topic #10: -0.655*"friends" + 0.633*"china" + 0.221*"travel" + 0.166*"beijing" + 0.136*"party" + -0.088*"beach" + 0.075*"vacation" + 0.071*"greatwall" + 0.070*"shanghai" + -0.066*"flowers" topic #28: -0.580*"india" + -0.323*"trip" + 0.279*"nature" + 0.262*"snow" + -0.258*"dog" + -0.224*"sunset" + 0.200*"winter" . topic #20: -0.527*"cat" + 0.511*"sunset" + 0.266*"sky" + -0.242*"california" + -0.209*"sanfrancisco" + 0.198*"clouds" + -0.167*"beach" + -0.156*"flower" + -0.149*"cats" + -0.132*"dog" topic #17: -0.323*"california" + -0.272*"sanfrancisco" + 0.269*"cat" + 0.254*"horse" + 0.211*"pferd" + 0.207*"cheval" + 0.205*"caballo" + 0.205*"paard" + 0.204*"hest" + 0.204*"cavalo"
Latent Semantic Analysis • 250000 photos, 100 topics topic #29: 0.689*"australia" + 0.279*"sydney" + -0.233*"nature" + 0.220*"trip" + -0.209*"france" + -0.187*"india" + -0.175*"snow" + 0.157*"new" + 0.144*"paris" + -0.134*"winter" topic #58: 0.401*"geotagged" + 0.385*"geolat" + 0.380*"geolon" + -0.261*"people" + 0.259*"day" + 0.198*"england" + 0.191*"newzealand" + -0.178*"canada" + 0.168*"water" + -0.144*"portrait". topic #45: 0.406*"fall" + 0.398*"park" + 0.315*"october" + -0.291*"animals" + 0.289*"autumn" + -0.262*"art" + 0.182*"leaves" + -0.175*"zoo" + -0.163*"sky" + 0.132*"garden" topic #85: -0.673*"hongkong" + 0.221*"florida" + 0.221*"singapore" + 0.209*"winter" + 0.174*"museum" + -0.170*"boston" + -0.165*"scotland" + -0.153*"prague" + 0.153*"cats" + -0.136*"island"
Latent Semantic Analysis • Notice the negative weights. • Hard to interpret • Probabilistic methods are not used
Latent Dirichlet Allocation • Expectation- Maximization • Each document is a mixture of topics • Find the posterior for topics in the E-Step p(topic t | document d) • Then update the assignment of the current word in the M-Step p(word w | topic t)
Latent Dirichlet Allocation • 250000 photos, 20 topics topic #13: 0.088*party + 0.072*halloween + 0.027*lake + 0.024*boat + 0.022*home + 0.019*park + 0.018*river + 0.016*ice + 0.015*spring + 0.014*birthday topic #3: 0.046*trip + 0.044*vacation + 0.044*sanfrancisco + 0.040*california + 0.026*road + 0.024*cats + 0.018*school + 0.018*cruise + 0.014*ca + 0.014*old topic #8: 0.051*paris + 0.042*france + 0.027*july + 0.027*4th + 0.025*music + 0.022*car + 0.021*rock + 0.020*dogs + 0.020*concert + 0.016*geotagged
Latent Dirichlet Allocation • 250000 photos, 50 topics topic #7: 0.111*sunset + 0.108*beach + 0.089*holiday + 0.047*fun + 0.029*smile + 0.028*forest + 0.023*rose + 0.020*wood + 0.019*disneyland + 0.019*costarica topic #14: 0.141*vacation + 0.046*san + 0.037*francisco + 0.034*sports + 0.020*hockey + 0.020*top + 0.019*cake + 0.014*cafe + 0.013*biking + 0.013*ruins topic #23: 0.112*trip + 0.070*bridge + 0.057*road + 0.048*blue + 0.048*building + 0.042*film + 0.035*orange + 0.022*university + 0.021*telephone + 0.018*sky topic #29: 0.124*party + 0.110*friends + 0.085*christmas + 0.045*rock + 0.038*lake + 0.038*ireland + 0.031*castle + 0.026*africa + 0.025*live + 0.025*music
Latent Dirichlet Allocation • 250000 photos, 100 topics topic #10: 0.109*hawaii + 0.093*island + 0.060*la + 0.030*photoshop + 0.027*walk + 0.026*hdr + 0.024*maui + 0.023*us + 0.019*fountain + 0.018*beach topic #24: 0.172*house + 0.106*architecture + 0.077*festival + 0.068*airplane + 0.038*flying + 0.029*flight + 0.026*air + 0.025*aircraft + 0.021*aviation + 0.020*airshow topic #34: 0.231*vacation + 0.159*trip + 0.136*lake + 0.095*florida + 0.088*birds + 0.062*san + 0.051*francisco + 0.015*yellowstone + 0.015*kayak + 0.015*maltay topic #70: 0.114*november + 0.074*thanksgiving + 0.050*soccer + 0.048*polarbear + 0.048*ski + 0.041*basketball + 0.035*safari + 0.034*bear + 0.023*wien + 0.021*flood
Conclusions • LSA and LDA are more useful for analyzing tags than Association Rule Mining • There is no “best” number of topics • Human interpretation still might be required
Future Works • Increase the corpus size to 1000000 documents • Analyze Flickr groups as well