Research on sequence-based ranking models has been a popular field in recommendation systems. In recent years, numerous researchers have devoted themselves to expanding the content of user behavior sequences, such as longer sequences, more types of sequences, and more variable sequence periods. These studies have achieved promising results, especially in the direction of longer sequences, where a large number of industrial recommendation systems have demonstrated that longer sequences can lead to better performance. However, as the sequence length approaches the upper limit of user behavior occurrences, the marginal benefit of increasing sequence length is gradually diminishing. Against this backdrop, this paper proposes a sequence expansion framework based on Side Information Memory Network (SIMN). Based on SIMN, theoretically all item-side features can be incorporated into the sequence, while avoiding additional sample development costs and storage costs. Furthermore, considering the application of this framework in small to medium-sized recommendation systems, this paper proposes a Feature Auto Encoder-Decoder (FAED) module, which further reduces the storage cost of SIMN. This paper integrates SIMN and FAED into a unified multitask training framework for modeling and validates it on two industrial datasets. Experimental results demonstrate that SIMN-FAED can be integrated with most mainstream sequence modeling methods and achieve better performance, with broad application prospects.

The Multi-stage Cascading Architecture (MCA), widely adopted in industrial advertising systems to balance efficiency and effectiveness, suffers from critical limitations: 1) ranking inconsistency caused by conflicting modeling objectives and capacity gaps across stages, and 2) the inability to model externalities-mutual influences among candidate ads in ranking stages. These issues degrade system performance and lead to suboptimal platform revenue. In this paper, we present UniROM, an end-to-end generative architecture that Unifies online advertising Ranking as One Model. UniROM replaces cascaded stages with a single model to directly generate optimal ad sequences from the full candidate ad corpus in location-based services (LBS). The primary challenges associated with this approach stem from high costs of feature processing and computational bottlenecks in modeling externalities of large-scale candidate pools. To address these challenges, UniROM introduces an algorithm and engine co-designed hybrid feature service to decouple user and ad feature processing, reducing latency while preserving expressiveness. To efficiently extract intra- and cross-sequence mutual information, we propose RecFormer with an innovative cluster-attention mechanism as its core architectural component. Furthermore, we propose a bi-stage training strategy that integrates pre-training with reinforcement learning-based post-training to meet sophisticated platform and advertising objectives. Extensive offline evaluations on public benchmarks and large-scale online A/B testing on industrial advertising platform have demonstrated the superior performance of UniROM over state-of-the-art MCAs.

Recommender system (RS) is widely applied in a multitude of scenarios to aid individuals obtaining the information they require efficiently. At the same time, the prevalence of popularity bias in such systems has become a widely acknowledged issue. To address this challenge, we propose a novel method named Model-Agnostic Popularity Debias Training Framework (MDTF). It consists of two basic modules including 1) General Ranking Model (GRM), which is model-agnostic and can be implemented as any ranking models; and 2) Popularity Debias Module (PDM), which estimates the impact of the competitiveness and popularity of candidate items on the CTR, by utilizing the feedback of cold-start users to re-weigh the loss in GRM. MDTF seamlessly integrates these two modules in an end-to-end multi-task learning framework. Extensive experiments on both real-world offline dataset and online A/B test demonstrate its superiority over state-of-the-art methods.

Sequential matching, which aims to predict the item a user will next interact with in the sequential context of the user's historical behaviors, is widely adopted in recommender systems. Existing works mainly characterize the sequential context as the dependencies of user interactions, which is less effective for online travel recommendation where users' behaviors are highly correlated with theirstages in the travel life cycle. Specifically, users on an online travel platform (OTP) usually go through different stages (e.g., exploring a destination, planning an itinerary), and make several correlated interactions (e.g., booking a flight, reserving a hotel, renting a car) at each stage. In this paper, we propose to capture the deep sequential context by modeling the evolving of user stages, and develop a novel stage-aware deep sequential matching network (SASNet) that incorporates inter-stage and intra-stage dependencies over stage-augmented interaction sequence for more accurate and interpretable recommendation. Extensive experiments on real-world datasets validate the superiority of our model for both online travel recommendation and general next-item recommendation. Our model has been successfully deployed at Fliggy, one of the most popular OTPs in China, and shows good performance in serving online traffic.