User Profiling by Social Curation By GENG Xue Supervised by Prof. Chua Tat-Seng A Crowd of Social Network Platforms Changes in Concerts & Pope Inauguration 1990 2010 Exponential Multimedia The Visual nature of the web increases exponentially 2013 Internet TrendsοΌ http://www.kpcb.com/insights/2013-internet-trends One of Kind Big Multimedia How to Deliver Meaningful Contents to the Right Person ? User Profiling User Profiling • Definition – A process to establish user profiles by extracting & representing the characteristics and preferences of users. Better Service Better Experience Recommendation A B Similar Recommendation A B Similar • Basic info & Social relationships So, User Profiling by Multimedia Analysis However, • Multimedia data are very diverse & unorganized. Traditional approaches fail. Solutions • Structured multimedia & Social Intelligence. Flickr Galleries Facebook Like Social Curation Modern Funnel: Social Curation. People select organize keep track of items they like. Social Curation Services (Pinterest) Social Curation Services (Pinterest) Boards to which image are re-pinned Board name and holder Why SCSs good? • Organized Contents Better Content Models • Content-centric network Refine Content Models Why SCSs good? • Organized Contents Better Content Models Why SCSs good? • Content-centric network Social intelligence to refine content models Framework Profile Structure (Ontology) Learning Profile Structure Refinement User Profiles Ontology-based user profiles (e.g., Fashion Domain) Profile Ontology Construction • Idea: Pruning Wikipedia ontology. – Remove out-dated words which are not often used currently. – Insert new words. • Find the most probable position in the existing ontology. • Each new word β is represented by ππππππ | π − π½πΆ |ππ + π||πΆ||π π πΆ is the sparse coefficient. An example (‘pegged pants’) Profile Ontology Learning • Idea: sibling samples are more visually similar, classifiers should be more distinct. ‘V dresses’, ‘Strapless dresses’, ‘Halter dress’, ‘One Shoulder dress’ Failure Cases • Features may be Wrong Refinement of user profiles • Organized contents without social intelligence (content-centric network). • Social intelligence to refine user profiles. User/Board-level Connection • If images are shared by more users/boards simultaneously, they more likely belong to the same preference. Observation: Shopping Sites Recommendation T shirt People also see these Observation: Movie Recommendation Recommendation User/Board-level Connection • If images are shared by more users/boards simultaneously, they more likely belong to the same preference. Content-level Connection • Similar images share similar visual cues and semantics. More Similar Mathematical Social Intelligence • User-level: – π’ πππ π, if there are π users sharing images π and π. = 0, if there are no users sharing them. • Bundle-level: π – πππ = π, if there are π boards sharing images π and π. 0, if there are no boards sharing them. • Content-level: π£ – πππ = exp(− ππ −ππ π2 β – πππ = π©ππ ππ©π = and j. 2 2 ) visual similarities of images i and j. π π π,π ππ π»ππ ππ hierarchical similarities of image i Refinement of User Profiles • Multi-level connections are incorporated into the low-rank method After refinement Bundle-level Semantic-level connection connection π¦π’π§ π± πΉ = | π· − πΉ |ππ + πΆ| πΉ |∗ + π·πππππ{πΉπ» π³π + π³π + π³π + π³π πΉ} πΉ Before refinement User-level Visual-level connection connection π·, πΉ : hierarchical representations of all training images before and after refinement. π³π οΌπ³π οΌπ³π οΌπ³π : the graph Laplacians of corresponding graphs of social curation. Visualization of User Profiles It is a vector: (…, 0.13, …, 0.23, …, 0.3, …) Experimental Data • Data collection – 1,239 users, 1,538,658 images. • Profile learning and refinement – We split labeled images equally into training/testing sets as the ground truths. • Image recommendation – We split the dataset by pin-time for training/testing – We added half noisy data out of fashion domain to simulate real world recommendation system. Evaluation of Profile Learning Failure Cases: a) Some image samples are too fine-grained. b) Some concepts tend to co-occur in the same image frequently. Evaluation of Profile Refinement Failure Cases: a) Sparse & noisy connections from some outdated items. b) Some items are co-repined leading to similar multi-level connections Evaluation of Image Recommendation Conclusion • Social Curation is NEW! • It has –Well-organized Contents –Social Intelligence • We test it on Pinterest (fashion domain). Thank You