Scalable Time-Decaying Adaptive Prediction Algorithm
Yinyan Tan*, Huawei; Zhe Fan, ; Guilin Li, ; Fangshan Wang, ; Zhengbing Li, ; Shikai Liu, ; Qiuling Pan, ; Eric Xing, CMU; Qirong Ho,
Online learning is used in a wide range of real applications, e.g., predicting ad click-through rates (CTR) and personalized recommendations. Based on the analysis of users’ behaviors in Video-On-Demand (VoD) recommender system-s, we discover that the most recent users’ actions can better reﬂect users’ current intentions and preferences. Under this observation, we thereby propose a novel time-decaying online learning algorithm derived from the state-of-the-art FTRL-proximal algorithm, called Time-Decaying Adaptive Prediction (TDAP) algorithm.
To scale Big Data, we further parallelize our algorithm following the data parallel scheme under both BSP and SSP consistency model. We experimentally evaluate our TDAP algorithm on real IPTV VoD datasets using two state-of-the-art distributed computing platforms, i.e., Spark and Petuum. TDAP achieves good accuracy: it improves at least 5.6% in terms of prediction accuracy, compared to FTRL-proximal algorithm; and TDAP scales well: it runs 4 times faster when the number of machines increases from 2 to 10. In our real running business cases, TDAP signiﬁcantly in-creases the degree of user activity, which brings more revenue than existing ones for our customers.
Filed under: Recommender Systems