Modeling Occasion Evolution in Frequency Domain for Promotion-Aware Click-Through Rate Prediction

Promotions are becoming very important and frequent in e-commerce platforms to attract customers and boost sales, resulting in various occasions which drive users to behave differently. Due to the frequent changes of occasions, existing Click-Through Rate (CTR) prediction methods are not able to generalize well to online serving because the data distribution is uncertain. Besides, with training data collected from different occasions, the assumption of identical distribution does not hold, imposing extra difficulties on model learning. In this paper, we propose a novel CTR model named MOEF for recommendation under frequent changes of occasions. Firstly, we generate occasion signals from the online business scenario with a proper sampling interval. For each occasion signal, we obtain a sequence of frequency spectrum via Fast Fourier Transformation applied on sliding time windows. Occasion signals are more discriminative in frequency domain, so we can model occasion evolution with sequences of frequency spectrum via LSTM more easily to learn a better occasion representation, helping tackle the online distribution uncertainty. To ease the difficulties of model learning introduced by non-identically distributed training data, we adopt multiple experts to learn feature representations from multiple aspects, which are guided by the occasion representation via an attention mechanism. Accordingly, a mixture of feature representations is obtained adaptively for different occasions and used for the final CTR prediction. Experimental results on real-world datasets validate the superiority of our MOEF model. Online A/B tests also show MOEF achieves significant gains of 4.23% on CTR and 6.47% on IPV during promotion periods as well as 4.61% and 6.96% in normal days, respectively. The code will be made publicly available.