Video recommendations and Machine Learning by Jerónimo Macanas at Big Data Spain 2017

A good content recommendation system is key for any video content provider. Machine Learning video recommendations provide a unique opportunity for broadcasters, Pay-TV operators, TV Networks, and any content distributor to increase engagement and reduce churn through content personalization.

https://www.bigdataspain.org/2017/talk/video-recommendations-and-machine-learning

Big Data Spain 2017
16th - 17th Kinépolis Madrid

Transcript

2. JUMP Actionable video insights Video recommendations and machine learning
3. WHO AM I? ☺ • MSIE in Industrial Engineering (Universidad Politécnica Madrid) • Executive Master’s in Growing Companies @ Stanford University • Data Science studies @ MIT • Helping video service providers since 2000 • Currently CEO at JUMP (Data Driven Video) • Data Lover!
4. JUMP offers a unified tool that in addition to providing fast insights into “what has happened”, “what is happening” and “what will happen” on video services, it makes an instant impact on business results. WHAT AM I DOING? ☺
5. 4 The Video Industry
6. 5 Video Industry
7. “By 2019, 80% of global Internet consumption will be video content” The Video Streaming Market • The global video streaming market is projected to grow from USD 30.29 Billion in 2016 to USD 70 Billion by 2021
8. THREE SUCCESS FACTORS IN A VIDEO SERVICE 1
11. 2 THREE SUCCESS FACTORS IN A VIDEO SERVICE
14. 3 THREE SUCCESS FACTORS IN A VIDEO SERVICE
15. 14 Video recommendations
16. • Households are watching up to 120 minutes more TV per week. • Services are able to increase user traffic by a factor of up to 10. • 25% increase of buy-rates in VoD • 8% increase in total ARPU WHY USE MACHINE LEARNING FOR VIDEO RECOMMENDATIONS? Business performance
17. User engagement & satisfaction • Users create on average 7 to 14 profiles per household. • Users enjoy previously unknown content. • Users want even more content to feed their personal profiles. • After less than 3 weeks, up to 80-90% of the viewing time is served by video recommendations. WHY USE MACHINE LEARNING FOR VIDEO RECOMMENDATIONS?
18. MACHINE LEARNING RECOMMENDATION ALGORITHM EXAMPLE 1. Considers the user’s view history, and optionally, the time of day. 2. Creates a "user preference vector“ to represent how much of each genre the user has seen; if the time of day is provided as a parameter, more weight is given to those movies watched at the given time of day. 3. Compiles a list of similar movies, selecting the 10 movies most similar to those watched by the user. Movies that have already been seen are discarded. The remaining are the candidate recommendations.
19. 4. Each candidate movie is assigned a score, computed by taking into account the user's genre preference and the movie’s genre. The higher the score, the closer it will be to the user’s preferences. 5. A recommendation is produced by taking the top x movies that scored the highest in (4) 6. The popularity of the movie within the video service, in the form of total number of views, is used for tie-breaking purposes. MACHINE LEARNING RECOMMENDATION ALGORITHM EXAMPLE
20. R&D NEW OPPORTUNITIES IN MACHINE LEARNING RECOMMENDATIONS 1. AUDIO/VIDEO RECOGNITION FOR METADATA ENRICHMENT 1. SOCIAL RECOMMENDATIONS 1. LINEAR TV/SMART EPG
21. • The creation of a neuronal network to enrich the content metadata, carrying out a comparative analysis with the tagging of audio, video, and audio and video together with the goal of a more efficient solution. • Additionally, the analysis of audio only (dimensionality reduction) could yield results similar to those resulting from video or combined video and audio analysis, thereby drastically improving the recommendation process. • The approach could tag content attributes such as: - Image velocity - colours and luminosity - localization - times of day within the video - emotion detection - objects that appear in the video 1. AUDIO/VIDEO RECOGNITION FOR METADATA ENRICHMENT E.G: Movie image recognition using AI for automatic movie tag generation
22. 2. SOCIAL RECOMMENDATIONS • Friends are the most powerful source of information • Friends have a high level of trust • Very likely to follow friends’ recommendations • Creating a recommendation model that includes the user’s social network when recommending content expands the limits of personalization getting closer and closer to a personalized discovery experience.
23. • No personalization in linear TV interface • Difficult to decide what to watch from +100 channels • Creating a recommendation model that factors in data regarding the channels the user watches at a given time, the programs that are the most interesting at a given time of day, together with an EPG display personalized according to this data, will create a much more personalized linear TV watching experience. 3. LINEAR TV/SMART EPG
24. • There is huge potential for Machine Learning in video recommendation/personalization with a proven impact on business results • We are just in the early stages of what might be done to personalize entertainment • Data Driven video technology is disrupting the video market - Recommendations/personalization - Acquisition through trial conversion predictions - Retention though churn analysis - Audience clustering - Advertisement pressure optimization - and much more … SUMMARY
25. info@jumptvs.com
26. “Thanks!” Jerónimo Macanás CEO jero@jumptvs.com +34 605 938 091 www.jumptvs.com