ByteDance compresses 100,000 watch history into a sketch for Douyin recommendations

ByteDance unveiled SequenceO1, a recommendation method that compresses a 100,000-interaction history into a fixed-size sketch read in O(1) time. It is live at full traffic on Douyin and Douyin Lite, with 49.9x cheaper training and 63.9x cheaper serving than naive scaling.

• A sketch of k prototypes compresses 100K items: at k=256 it holds 390x fewer rows
• Training is 49.9x cheaper and serving 63.9x cheaper than direct 100K scaling
• It retains 83% of the quality gain: +1.07% Finish AUC vs +1.29%
• The last 10,000 interactions are read in full; long-term taste comes from the sketch