Welcome to RS_c, the central platform for the RecSys community. We provide curated lists of recommender-systems datasets, algorithms, books, conferences and many resources more. Maybe most importantly, we publish the latest recommender-system news. If you want your news to be reported on RS_c, read here.
Netflix announced that it is now personalizing the sizzles in the user interface. A sizzle reel is “a montage of video clips from different titles strung together into a seamless A/V asset that gets members excited about upcoming launches”. Netflix is reportedly creating the sizzles in real-time and dynamically based on the user who is looking at it. More precisely, the included titles and order of them are personalized. Netflix’s motivation for this was a better user experience and cost savings, as previously sizzles were created manually by employees.
Netflix’s blog post outlines the development of their sizzle reels, moving from manual creation to a more efficient and personalized approach called “Dynamic Sizzles.” Traditional production was costly and time-consuming, limiting scalability and personalization. In contrast, Dynamic Sizzles are a system-based approach, stitching together many video clips with a synchronized audio track to create personalized “Mega Assets”. These Mega Assets, large audio/video (A/V) assets compiled from various title clips, serve as a library for generating personalized sizzle reels, offering notable time and cost savings.
Creating Mega Assets was the initial challenge. The process required the selection and arrangement of video clips, done by human editors for creative and technical accuracy. The clips are laid out in a specific order in a timeline, each marked with an index for location. Netflix addressed the difficulty of this process by developing an Adobe Premiere plug-in to automate the indexing, combined with programmatic verifications via timecode data ingestion. This allows the generation of millions Dynamic Sizzles from a single Mega Asset.
The construction of a Dynamic Sizzle includes personalization and a strategic approach. When a sizzle reel request is made by the system, the system identifies titles in the Mega Asset and creates a personalized list based on member preferences. Higher-ranked titles receive more emphasis in the Dynamic Sizzle. The process involves identifying specific timecodes for each clip in the Mega Asset, enabled by Netflix’s Hollow technology. This technology enables fast timecode searches and use, ensuring efficient access during runtime. The clips are then ordered according to a predefined cadence, dictating the final layout and seamless assembly of the Dynamic Sizzle.
The delivery and playback of Dynamic Sizzles present their own set of challenges and innovations. The player uses timecodes to determine the start and stop points for each clip, playing them sequentially for a seamless experience. To support this, the API used by devices to fetch trailers was modified, requiring devices to indicate their capability to support Dynamic Sizzles. Additionally, the player had to be adapted to handle asymmetrical segment streaming and optimized for streaming shorter, discontiguous segments. Despite the challenges, the implementation of Dynamic Sizzles represents a significant advancement in delivering personalized and engaging content experiences, demonstrating Netflix’s commitment to innovation and collaboration in engineering solutions.
I am the founder of Recommender-Systems.com and head of the Intelligent Systems Group (ISG) at the University of Siegen, Germany https://isg.beel.org. We conduct research in recommender-systems (RecSys), personalization and information retrieval (IR) as well as on automated machine learning (AutoML), meta-learning and algorithm selection. Domains we are particularly interested in include smart places, eHealth, manufacturing (industry 4.0), mobility, visual computing, and digital libraries.
We founded or maintain, among others, LensKit-Auto, Darwin & Goliath, Mr. DLib, and Docear, each with thousand of users; we contributed to TensorFlow, JabRef and others; and we developed the first prototypes of automated recommender systems (AutoSurprise and Auto-CaseRec) and Federated Meta Learning (FMLearn Server and Client).