Personalized News and Video Recomendation System at LinkSure
- 2. LinkSure: a Mobile Internet Company
- 3. 3
- 4. THE WORLD’S LARGEST SHARING APP % Tech giants which passed 900M user milestones 4
- 5. WIFI MASTER KEY APP FUNCTIONS Connect to the strongest and nearest hotspot nearby 5
- 6. WIFI MASTER KEY APP FUNCTIONS: FEEDS RECOMMENDATION WiFi Master Key Feeds A media portal for users to discover interesting broadcast, film, publishing, animation, interactive media, games International version Chinese version 6
- 7. 0 7
- 8. Big Data Changes the World Asia Big Data Association (ABDA) Non-Profit Organisation ABDA is registered in 2016 as a Society with the Registry of Societies, Government of Singapore. 8
- 9. Entreprise Academic Government Great Ideas Research papers Ideas Deep Learning Annual Conference Meetups FinTech Data Mining A.I. Great Ideas Big Data Standard API & Services Deep Learning, A.I., FinTech, Data Mining, etc. Specialisation Technology Promoting Big Data research and standard adoption For Academia: Annual conferences, Publish research papers Service Bridging Academia & Industry in Big Data Research Translation For Industry: Meet-ups, Standard Data API & Services
- 10. Partners 10
- 11. Evolution of recommender system at LinkSure • 1st generation, 2015 • Connected data stream • Multiple types of recommendation • 2nd generation, 2016 • More user features, item features, WiFi hotspot features. • Multiple types of recall • Fusion of multi-algorithms • Video recommender system • 3rd generation, 2017 • Deep Learning based recommender system 11
- 12. Challengesin News and video Recommendation • New items: how to recommend items that has not been viewed by any one? • New users: how to recommend when the user has not viewed any? • Hot gets hotter: how to balance hot news and the less hot ones? News that is viewed by many users are more likely to be recommended to even more users, while those that are not viewed by many are less likely to get recommended. • Implicit data has no real negatives • Sparsity: how to effectively use hundreds of thousands of features and recommend to tens of millions of users? • Serendipity: how to allow the user extend his/her interest space? • Scalability: how to efficiently compute for millions or even billions of users? Computational costs increases drastically with huge number of users. 12
- 13. The first generation Before After • Connected Data stream • News tags • Multiple types of recommendation • Headline • Hot • Personalized • Latest 13
- 14. Online A/B Testing Dates Dates Improvement rateOther channels C A/B Test result User conversion News conversion User conversion News conversion News conversion User conversion A/B Testing Comparison of different recommendation types User conversion News conversion 14 headline headline
- 15. The second generation • More user, item and Wi-Fi hotspot features • Multiple types of recall methods • Fusion of multi-algorithms • FTRL • LibFM • FTRL-FM • CF • Long term interest • Short term interest • Negative feedback processing • … AlgsReal-time CTR based Real-time CTR with hot news Topic-based Long-term interest based Timeline-based Number of clicks 15
- 17. User Profiling • Demographics • Sex (Predictive model) • Age (Predictive model) • Marital Status (Rule) • Consumption Power (Rule) • Student(Rule) • Business Traveler (Rule) • Car Owner(Rule) • Office worker (online time, position stability) • Blue collar worker (active in industry area) • Position • Resident city • Active cities • Hometown (online in Spring Festival) • Active coordinates • Active biz areas • Active biz types • Home price 17
- 18. User Profiling • Interest • Biz interest • Game • Edu • … • Installed Apps • Tags • Types • List • Traffic of apps • Reading preferences • Channels • Topics • Keywords • Brand preference • Phone • Phone brand • Types • Prices 18
- 19. Wi-Fi hotspot profiling Chang Xu et al. WWW2017 Classify 200+ million Wi-Fi hotspots using only Connection and User counts via • Convolution Neural Networks (CNN) to capture spatio-temporal relationship between connection/user counts across 24 × 7 matrix, and • FFT (Fast Fourier Transforms) to extract user/connection frequencies 4T SPARK 2.0 classifies 239 m hotspots in 12 hours with ~80% F1 for 2-class. 1. Dining (restaurants) 2. Entertainment (KTV, amusement) 3. Living (convenience stores, barbers) 4. Hotel 5. Residential 6. Office 7. Shopping 8. Health (clinics, gyms, hospitals) 9. Tourist (parks, attractions) 10. Education (schools & colleges) 11. Transportation (railway stations, airports) Challenges • No information about HS, other than SSID and connection info, • Only 8.9% HS have SSID’s that matches a keyword dictionary, • HS MAC address can reveal brand/make, but not category in general 5-Fold CV F1 scores comparing CNN, FFT, CNN+FFT 19 Future work: incorporate additional features like the geo-information of POI to further improve classification accuracy.
- 20. Ku6 Media Ku6 Media (NASDAQ:KUTV) is one of the largest online video companies. Most of the videos posted on the Ku6 website are short-form videos submitted via user uploads. The company provides some customized services, such as user-navigating channels for video sharing and commenting tools; search history for frequent visitors; chatting box alongside videos. • Multiple types of recommendation • Related Videos • Popular Videos • Selected Video • Guess you like • Multiple channels • Mobile site • PC site • Home page • Playing page • APP 20
- 21. Features • Basic info • Title, description, user tags • Types, resolution, length, upload time, user id • Click data • Number of clicks • Number of comments • Pre- and post-fix of titles • Eg. LEGO 01 New Girl in Town • Prefix: Lego • Postfix: 01 • Hot video info to redis 21
- 22. Ranking and re-ranking Ranking • Filtering • Similar titles • Keep videos with different Length • Keep videos with same prefix, but with different postfix • Keep the high resolution video or more clicks • Relevance score • From Same search results; CF score(channel related) • Similar channel; same playlist; Same uploader • Same play list; same prefix but different postfix • User score • Number of clicks, Number of comments • Hot channels, higher view count • Video quality score • Resolution, upload time, vcu upload, long title Re-ranking • Guarantee the quality and variety of top videos • Long videos first • Variety of video sources • CF results, search results, positive feedback results • Similar but higher quality ones not in top • Variety of recommendation reasons • Scatter videos from different combination of queries • Put backward videos without dynamic reasons 22
- 23. Recommendation models Collaborative Filtering • Relevance score ü Relate(A,B)= CoOccur(A,B)/(Freq(A)*Freq(B)) ü CoOccur(A,B)= His_CoOccur(A,B)+ New_CoOccur(A,B) ü Freq(A)= His_Freq(A)+ New_Freq(A) ü Freq(B)= His_Freq(B)+ New_Freq(B) • Time factor ü Relate(A,B)= (His_CoOccur(A,B)+ New_CoOccur(A,B))/(( His_Freq(A)+ New_Freq(A))*( His_Freq(B)+ New_Freq(B))) ü Relate(A,B)=(Alfa*His_CoOccur(A,B)+ New_CoOccur(A,B))/(( Alfa*His_Freq(A)+ New_Freq(A))*( Alfa*His_Freq(B)+ New_Freq(B))) 23
- 24. Ku6 Video Recommendation at KU6 : : :: 24
- 25. Before After Online Result for KU6 25
- 26. Video Recommendation for WiFi Master Key Total Number of video views Total Number of devices Single Chanel All Chanels Number of video views and devices 26
- 27. Video Recommendation for Wifi Master Key Video conversion rate Device conversion rate Single Chanel All Chanels Conversion rate 27
- 28. Existing problems and potential solutions • Attention models • Hot gets hotter problem • Context-aware models • Context is indirect/implicit feedback • Explore/exploit approaches and learning across time • Temporal dynamics • Session modelling 28
- 29. The 3rd generation Deep Learning Based Recommender System
- 30. Neural Collaborative Filtering He et al., WWW 2017 30 NCF uses a multi-layer model to learn the user-item interaction function - Input: sparse feature vector for user u (vu) and item i (vi) - Output: predicted score ŷui Interaction function Note: Input feature vector can include any categorical variables other than user/item ID, such as attributes, contexts and content. Promising Results • Deeper models are helpful. • Combining deep models with MF in the latent space leads to better results.
- 31. Potential research directions • The objective is to explore the application of deep learning for news and video recommender system. • to retrieve or design useful features for deep recommender system from various explicit or implicit feedbacks of users • to design or propose special deep networks which is suitable for news and video recommendation • to make the deep network applicable in real-time recommender system either by optimizing the network structure or implementing it in novel hardware or software system • Potential Research Directions: • deep convolutional neural networks to model the latent factor vector for user or item separately or auxiliary information from user’s implicit feedbacks • reinforcement learning to CF to optimize the prediction of a user’s clicks for an unseen news by taking into account the sequence of news he has seen • Online learning and incremental learning • Temporal correlation modeling • Attention mechanism • Minimizing model complexity and increasing interpretability of models 31
- 32. Conclusion • Approaches and systems have evolved a lot in the past years. • Industry and academia working together has advanced the field since beginning • Welcome to collaborate with us and join Asia Big Data Association 32