CLOCK: Online Temporal Hierarchical Framework for Multi-scale Multi-granularity Forecasting of User Impression

Abstract

User impression forecasting underpins various commercial activities, from long-term strategic decisions to short-term automated operations. As a representative that involves both kinds, the highly profitable Guaranteed Delivery (GD) advertising focuses mainly on promoting brand effect by allowing advertisers to order target impressions weeks in advance and get allocated online at the scheduled time. Such a business mode naturally incurs three issues making existing solutions inferior: 1) Timescale-granularity dilemma of coherently supporting the sales of day-level impressions of the distant future and the corresponding fine-grained allocation in real-time. 2) High dimensionality due to the Cartesian product of user attribute combinations. 3) Stability-plasticity dilemma of instant adaptation to emerging patterns of temporal dependency without catastrophic forgetting of repeated ones facing the nonstationary traffic. To overcome the obstacles, we propose an online temporal hierarchical framework that functions analogously to a CLOCK and hence its name. Long-timescale, coarse-grained temporal data (e.g., the daily impression of one quarter) and short-timescale but fine-grained ones are handled separately by dedicated models, just like the hour/minute/second hands. Each tier in the hierarchy is triggered for forecasting and updating by need at different frequencies, thus saving the maintenance overhead. Furthermore, we devise a reconciliation mechanism to coordinate tiers by aggregating the separately learned local variance and global trends tier by tier. CLOCK solves the dimensionality dilemma by subsuming the autoencoder design to achieve an end-to-end, nonlinear factorization of streaming data into a low-dimension latent space, where a neural predictor produces predictions for the decoder to project them back to the high dimension. Lastly, we regulate the CLOCK's continual refinement by combining the complementary Experience Replay (ER) and Knowledge Distillation (KD) techniques to consolidate and recall previously learned temporal patterns. We conduct extensive evaluations on three public datasets and the real-life user impression log from the Tencent advertising system, and the results demonstrate CLOCK's efficacy.