Existing review-based recommendation methods learn a latent representation of user and item from user-generated reviews by a static strategy, which are unable to capture the dynamic evolution of users' interests and the dynamic attraction of items. Here, we propose a dynamic and static representation learning network (DSRLN) to improve the rating prediction accuracy by exploring fine-grained representations of users and items. Specifically, we built DSRLN with a dynamic representation extractor to model the dynamic evolution of users' interests by exploring the inner relations of an interaction sequence, and with a static representation extractor to model the users' intrinsic preferences by learning the semantic coherence and feature strength information from reviews. To identify the different influences of dynamic and static features for different users, a personalized adaptive fusion module was designed using a weighted attention mechanism. Extensive experiments on five real-world datasets from Amazon demonstrated the superiority of the proposed model, and the additional ablation studies verified the effectiveness of the components designed in the DSRLN model.

As the self-attention mechanism offers powerful capabilities for capturing sequential relationships, it has become increasingly popular to use it for modeling user behavior sequences in recommender systems. However, the self-attention mechanism has a quadratic computational complexity of O(n^2), as it conducts interactions among all item pairs in the sequence. This can lead to expensive model training and slow inference speeds, which may hinder practical deployment. To this end, we pursue to develop alternative approaches to improve the efficiency of the self-attention mechanism. We observe that the attention scores calculated from each item interacting with other items (including itself) are sparse, indicating that there are limited valuable item pairs (with non-zero attention weight) that contribute to the final output. This motivates us to develop effective strategies for discerning valuable items and computing attention scores solely for these items, thereby minimizing the consumption of unnecessary computations. Herein, we present a novel Progressive Sampling-based Self-Attention (PS-SA) mechanism, which utilizes a learnable progressive sampling strategy to identify the most valuable items. Subsequently, we solely utilize these selected items to produce the final output. Experiments on academic and production datasets demonstrate PS-SA could still achieve promising results while reducing computational costs. It is notable that we have successfully deployed it on Alibaba display advertising system, resulting in a 2.6% CTR and 1.3% RPM increase.

Cascading architecture has been widely adopted in large-scale advertising systems to balance efficiency and effectiveness. In this architecture, the pre-ranking model is expected to be a lightweight approximation of the ranking model, which handles more candidates with strict latency requirements. Due to the gap in model capacity, the pre-ranking and ranking models usually generate inconsistent ranked results, thus hurting the overall system effectiveness. The paradigm of score alignment is proposed to regularize their raw scores to be consistent. However, it suffers from inevitable alignment errors and error amplification by bids when applied in online advertising. To this end, we introduce a consistency-oriented pre-ranking framework for online advertising, which employs a chunk-based sampling module and a plug-and-play rank alignment module to explicitly optimize consistency of ECPM-ranked results. A ΔNDCG-based weighting mechanism is adopted to better distinguish the importance of inter-chunk samples in optimization. Both online and offline experiments have validated the superiority of our framework. When deployed in Taobao display advertising system, it achieves an improvement of up to +12.3% CTR and +5.6% RPM.

Conversion rate (CVR) prediction is one of the core components in online recommender systems, and various approaches have been proposed to obtain accurate and well-calibrated CVR estimation. However, we observe that a well-trained CVR prediction model often performs sub-optimally during sales promotions. This can be largely ascribed to the problem of the data distribution shift, in which the conventional methods no longer work. To this end, we seek to develop alternative modeling techniques for CVR prediction. Observing similar purchase patterns across different promotions, we propose reusing the historical promotion data to capture the promotional conversion patterns. Herein, we propose a novel Historical Data Reuse (HDR) approach that first retrieves historically similar promotion data and then fine-tunes the CVR prediction model with the acquired data for better adaptation to the promotion mode. HDR consists of three components: an automated data retrieval module that seeks similar data from historical promotions, a distribution shift correction module that re-weights the retrieved data for better aligning with the target promotion, and a TransBlock module that quickly fine-tunes the original model for better adaptation to the promotion mode. Experiments conducted with real-world data demonstrate the effectiveness of HDR, as it improves both ranking and calibration metrics to a large extent. HDR has also been deployed on the display advertising system in Alibaba, bringing a lift of 9% RPM and 16% CVR during Double 11 Sales in 2022.

Typically, existing works utilize static methods to extract the latent feature representation of user and item reviews, neglecting the time signals and behavior patterns hidden in the user-item interaction history, which may fail to capture users' instant interests and items' temporal attributes. Moreover, there is no framework that unifies recent behavior sequences and reviews. Therefore, in this paper, we first define dynamic and static features to describe users' short- and long-term preferences and items' temporal and inherent attributes. We then design feature extractors to capture these latent factors simultaneously from recent behavior sequences and reviews. Then, we propose a novel unified framework to extract and fuse these fine-grained characteristics, named unified model with dynamic-static features (UMDSF). Specifically, the proposed model extracts both temporal sequence and review features by two parallel feature extractors based on self-attention and a multi-head attention mechanism. Subsequently, an adaptive fusion module is utilized to combine the fine-grained representations for the downstream recommendation tasks. Extensive experiments on four real-world datasets demonstrate the superiority of UMDSF and additional ablation studies verify the effectiveness of the components designed in the proposed model.