Shared-account usage is common on short-video platforms, especially on mobile and tablet devices, where a single device is accessed by multiple users. While existing industrial solutions generally focus on behavior sequence purification to disentangle mixed user preferences, such approaches inherently depend on behavior accumulation and therefore lack the capability for real-time user identification. To adapt to online recommendation, utilizing gesture interaction features is a natural and promising option, as they (1) are instantaneous without behavior collection and (2) naturally encode fine-grained user operation habits. Nevertheless, we empirically observe that directly incorporating raw gesture features into recommendation models yields limited gains. Identity-discriminative patterns embedded in gesture signals are largely entangled during the main model training, preventing them from being leveraged as explicit and reliable identity cues. As a result, efficiently utilizing gesture information to provide more distinct identity signals for recommendation models remains a critical challenge. To address this issue, we propose G-CORE (Gesture Clustering for Real-time REcommendation), an unsupervised framework that disentangles gesture representations via clustering before integrating them into the main recommendation model. By providing clearer and more identity-aware signals, G-CORE enables the main model with faster user switching without relying on a volume of behavior accumulation. Through extensive offline experiments and online A/B tests on Kuaishou platform, G-CORE demonstrates its effectiveness in various shared-account scenarios, and has been successfully deployed in the Mobile and Tablet system of the platform.

Short-video recommender systems typically optimize ranking models using dense user behavioral signals, such as clicks and watch time. However, these signals are only indirect proxies of user satisfaction and often suffer from noise and bias. Recently, explicit satisfaction feedback collected through questionnaires has emerged as a high-quality direct alignment supervision, but is extremely sparse and easily overwhelmed by abundant behavioral data, making it difficult to incorporate into online recommendation models. To address these challenges, we propose a novel framework which is towards End-to-End Alignment of user Satisfaction via Questionnaire, named EASQ, to enable real-time alignment of ranking models with true user satisfaction. Specifically, we first construct an independent parameter pathway for sparse questionnaire signals by combining a multi-task architecture and a lightweight LoRA module. The multi-task design separates sparse satisfaction supervision from dense behavioral signals, preventing the former from being overwhelmed. The LoRA module pre-inject these preferences in a parameter-isolated manner, ensuring stability in the backbone while optimizing user satisfaction. Furthermore, we employ a DPO-based optimization objective tailored for online learning, which aligns the main model outputs with sparse satisfaction signals in real time. This design enables end-to-end online learning, allowing the model to continuously adapt to new questionnaire feedback while maintaining the stability and effectiveness of the backbone. Extensive offline experiments and large-scale online A/B tests demonstrate that EASQ consistently improves user satisfaction metrics across multiple scenarios. EASQ has been successfully deployed in a production short-video recommendation system, delivering significant and stable business gains.

Short-video recommendation systems typically optimize for user satisfaction. However, allocating exposure to creators at critical growth stages incentivizes long-term content supply despite compromising immediate user engagement. Existing efforts concerning creator interests aim at either improving creator exposure fairness, matching creators with suitable audiences, or leveraging creator behavior to enhance user welfare. Directly maximizing the joint value of user satisfaction and creator incentive at recommendation time, however, remains largely unaddressed. This presents two challenges. First, the two objectives are heterogeneous in nature, making it non-trivial to formulate this joint optimization as a tractable problem. Second, optimizing creator incentive requires globally coordinated decisions across requests, making real-time serving infeasible. To address these challenges, we formulate the joint maximization as a constrained optimization problem that unifies the two heterogeneous objectives. We further derive an efficient online algorithm based on the primal-dual method, which decouples global incentive constraints into real-time decisions with theoretical guarantees. Experiments on a large-scale short-video platform demonstrate consistent improvements in joint user-creator value over existing baselines.

We present PushGen, an automated framework for generating high-quality push notifications comparable to human-crafted content. With the rise of generative models, there is growing interest in leveraging LLMs for push content generation. Although LLMs make content generation straightforward and cost-effective, maintaining stylistic control and reliable quality assessment remains challenging, as both directly impact user engagement. To address these issues, PushGen combines two key components: (1) a controllable category prompt technique to guide LLM outputs toward desired styles, and (2) a reward model that ranks and selects generated candidates. Extensive offline and online experiments demonstrate its effectiveness, which has been deployed in large-scale industrial applications, serving hundreds of millions of users daily.

How to preserve the data privacy during the training of deep neural network (DNN) is a key security concern in the artificial intelligence era. However, most existing solutions based on homomorphic encryption and Trusted Execution Environment (TEE) are incompatible with heterogeneous Neural Network Accelerators (NNAs), leading to significant performance loss. We propose a novel method based on vector decomposition to allocate operators across different NNAs, ensuring both throughput and privacy simultaneously. Furthermore, based on this approach, we have designed a compiler that automatically converts front-end model descriptions into backend encrypted computation graphs, which is running securely over trusted and untrusted hardware. This compiler heuristically determines the allocation scheme based on hardware affinity and cross-hardware communication costs, significantly reducing additional overhead. Experimental results demonstrate that our method does not incur extra accuracy costs and achieves a throughput significantly higher than existing methods. Deploying our approach at scale on a platform with a billion users, we have verified its negligible impact on real-world operations while ensuring the privacy protection capability for cross-domain data.

Recommender systems need to optimize various types of user feedback, e.g., clicks, likes, and shares. A typical recommender system handling multiple types of feedback has two components: a multi-task learning (MTL) module, predicting feedback such as click-through rate and like rate; and a multi-task fusion (MTF) module, integrating these predictions into a single score for item ranking. MTF is essential for ensuring user satisfaction, as it directly influences recommendation outcomes. Recently, reinforcement learning (RL) has been applied to MTF tasks to improve long-term user satisfaction. However, existing RL-based MTF methods are formula-based methods, which only adjust limited coefficients within pre-defined formulas. The pre-defined formulas restrict the RL search space and become a bottleneck for MTF. To overcome this, we propose a formula-free MTF framework. We demonstrate that any suitable fusion function can be expressed as a composition of single-variable monotonic functions, as per the Sprecher Representation Theorem. Leveraging this, we introduce a novel learnable monotonic fusion cell (MFC) to replace pre-defined formulas. We call this new MFC-based model eXtreme MTF (xMTF). Furthermore, we employ a two-stage hybrid (TSH) learning strategy to train xMTF effectively. By expanding the MTF search space, xMTF outperforms existing methods in extensive offline and online experiments.

Two-tower models are widely adopted in the industrial-scale matching stage across a broad range of application domains, such as content recommendations, advertisement systems, and search engines. This model efficiently handles large-scale candidate item screening by separating user and item representations. However, the decoupling network also leads to a neglect of potential information interaction between the user and item representations. Current state-of-the-art (SOTA) approaches include adding a shallow fully connected layer(i.e., COLD), which is limited by performance and can only be used in the ranking stage. For performance considerations, another approach attempts to capture historical positive interaction information from the other tower by regarding them as the input features(i.e., DAT). Later research showed that the gains achieved by this method are still limited because of lacking the guidance on the next user intent. To address the aforementioned challenges, we propose a "cross-interaction decoupling architecture" within our matching paradigm. This user-tower architecture leverages a diffusion module to reconstruct the next positive intention representation and employs a mixed-attention module to facilitate comprehensive cross-interaction. During the next positive intention generation, we further enhance the accuracy of its reconstruction by explicitly extracting the temporal drift within user behavior sequences. Experiments on two real-world datasets and one industrial dataset demonstrate that our method outperforms the SOTA two-tower models significantly, and our diffusion approach outperforms other generative models in reconstructing item representations.

The watch time is a significant indicator of user satisfaction in video recommender systems. However, the prediction of watch time as a target variable is often hindered by its highly imbalanced distribution with a scarcity of observations for larger target values and over-populated samples for small values. State-of-the-art watch time prediction models discretize the continuous watch time into a set of buckets in order to consider the distribution of watch time. However, it is highly uninvestigated how these discrete buckets should be created from the continuous watch time distribution, and existing discretization approaches suffer from either a large learning error or a large restoration error. To address this challenge, we propose a Classification-Restoration framework with Error-Adaptive-Discretization (CREAD) to accurately predict the watch time. The proposed framework contains a discretization module, a classification module, and a restoration module. It predicts the watch time through multiple classification problems. The discretization process is a key contribution of the CREAD framework. We theoretically analyze the impacts of the discretization on the learning error and the restoration error, and then propose the error-adaptive discretization (EAD) technique to better balance the two errors, which achieves better performance over traditional discretization approaches. We conduct detailed offline evaluations on a public dataset and an industrial dataset, both showing performance gains through the proposed approach. Moreover, We have fully launched our framework to an online video platform, which resulted in a significant increase in users' video watch time by 0.29% through A/B testing. These results highlight the effectiveness of the CREAD framework in watch time prediction in video recommender systems.

In recent years, there has been a growing interest in utilizing reinforcement learning (RL) to optimize long-term rewards in recommender systems. Since industrial recommender systems are typically designed as multi-stage systems, RL methods with a single agent face challenges when optimizing multiple stages simultaneously. The reason is that different stages have different observation spaces, and thus cannot be modeled by a single agent. To address this issue, we propose a novel UNidirectional-EXecution-based multi-agent Reinforcement Learning (UNEX-RL) framework to reinforce the long-term rewards in multi-stage recommender systems. We show that the unidirectional execution is a key feature of multi-stage recommender systems, bringing new challenges to the applications of multi-agent reinforcement learning (MARL), namely the observation dependency and the cascading effect. To tackle these challenges, we provide a cascading information chain (CIC) method to separate the independent observations from action-dependent observations and use CIC to train UNEX-RL effectively. We also discuss practical variance reduction techniques for UNEX-RL. Finally, we show the effectiveness of UNEX-RL on both public datasets and an online recommender system with over 100 million users. Specifically, UNEX-RL reveals a 0.558% increase in users' usage time compared with single-agent RL algorithms in online A/B experiments, highlighting the effectiveness of UNEX-RL in industrial recommender systems.

Modern recommender systems are built upon computation-intensive infrastructure, and it is challenging to perform real-time computation for each request, especially in peak periods, due to the limited computational resources. Recommending by user-wise result caches is widely used when the system cannot afford a real-time recommendation. However, it is challenging to allocate real-time and cached recommendations to maximize the users’ overall engagement. This paper shows two key challenges to cache allocation, i.e., the value-strategy dependency and the streaming allocation. Then, we propose a reinforcement prediction-allocation framework (RPAF) to address these issues. RPAF is a reinforcement-learning-based two-stage framework containing prediction and allocation stages. The prediction stage estimates the values of the cache choices considering the value-strategy dependency, and the allocation stage determines the cache choices for each individual request while satisfying the global budget constraint. We show that the challenge of training RPAF includes globality and the strictness of budget constraints, and a relaxed local allocator (RLA) is proposed to address this issue. Moreover, a PoolRank algorithm is used in the allocation stage to deal with the streaming allocation problem. Experiments show that RPAF significantly improves users’ engagement under computational budget constraints.