arXiv:2604.25834v1 [cs.AI] 28 Apr 2026
Action-Aware Generative Sequence Modeling for Short Video Recommendation Wenhao Li
Zihan Lin
ZhengXiao Guo
Kuaishou Inc. Beijing, China [email protected]
Kuaishou Inc. Beijing, China [email protected]
Kuaishou Inc. Beijing, China [email protected]
Jie Zhou
Shukai Liu
Yongqi Liu
Kuaishou Inc. Beijing, China [email protected]
Unaffiliated China [email protected]
Kuaishou Inc. Beijing, China [email protected]
Chuan Luo
Chaoyi Ma
Ruiming Tang
Beihang University Beijing, China [email protected]
Kuaishou Inc. Beijing, China [email protected]
Kuaishou Inc. Beijing, China [email protected]
Han Li Kuaishou Inc. Beijing, China [email protected]
Abstract With the rapid development of the Internet, users have increasingly higher expectations for the recommendation accuracy of online content consumption platforms (e.g., short video platforms). However, short videos often contain diverse segments, and users may not hold the same attitude toward all of them (e.g., music enthusiasts may not enjoy all songs in a medley). Traditional binary-classification recommendation models, which treat a video as a single holistic entity, face limitations in accurately capturing such nuanced preferences. Considering that user consumption is a temporal process, this paper demonstrates that the timing of user actions can represent diverse intentions through statistical analysis and examination of action patterns. Based on this insight, we propose a novel modeling paradigm: Action-Aware Generative Sequence Network (A2 Gen), which refines user actions (e.g., 𝐿𝑖𝑘𝑒 and 𝐹𝑜𝑙𝑙𝑜𝑤, etc.) along the temporal dimension and chains them into sequences for unified processing and prediction. First, we introduce the Context-aware Attention Module (CAM) to model action sequences enriched with item-specific contextual features. Building upon this, we develop the Hierarchical Sequence Encoder (HSE) to learn temporal action patterns from users’ historical actions. Finally, through leveraging CAM, we design a module for action sequence generation: the Action-seq Autoregressive Generator (AAG). Extensive offline experiments on the Kuaishou’s dataset and the Tmall public dataset Corresponding authors.
This work is licensed under a Creative Commons Attribution 4.0 International License. SIGIR ’26, Melbourne, VIC, Australia © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2599-9/2026/07 https://doi.org/10.1145/3805712.3809728
demonstrate the superiority of our proposed model. Furthermore, through large-scale online A/B testing deployed on Kuaishou’s platform, our model achieves significant improvements over baseline methods in multi-task prediction by leveraging sequential information. Specifically, it yields increases of 0.34% in user watch time, 8.1% in interaction rate, and 0.162% in overall user retention (LifeTime-7), leading to successful deployment across all traffic, serving over 400 million users every day.
CCS Concepts • Information systems → Recommender systems.
Keywords User Action Sequence, Autoregressive Generative Model, MultiTask Recommendation ACM Reference Format: Wenhao Li, Zihan Lin, ZhengXiao Guo, Jie Zhou, Shukai Liu, Yongqi Liu, Chuan Luo, Chaoyi Ma, Ruiming Tang, and Han Li. 2026. Action-Aware Generative Sequence Modeling for Short Video Recommendation. In Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’26), July 20–24, 2026, Melbourne, VIC, Australia. ACM, New York, NY, USA, 11 pages. https://doi.org/10.1145/ 3805712.3809728
1
Introduction
The rapid evolution of Internet platforms has led to increasingly diverse user emotional expressions. This presents a critical bottleneck for modern real-time recommendation systems, particularly in accurately detecting users’ true preferences [7, 32, 42]. As an indispensable part of entertainment, online video platforms (e.g.,
SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia
TikTok, YouTube Shorts) are committed to delivering precise recommendations to maximize user retention and boost corporate revenue. Traditional multi-task recommendation systems typically treat each video as a single entity, recommending items through IDs and associated features. They employ task-specific architectures to process user, item, and contextual information, to optimize the binary prediction accuracy across multiple user actions [14, 22, 27]. However, videos often consist of multiple segments of different content. While traditional recommendation models simplify training and inference by modeling videos as holistic units, they fail to capture the dynamic process of user consumption. As a result, they cannot accurately distinguish users’ varying interests across different segments, allowing irrelevant segments to contaminate the learned user interest embeddings and ultimately degrade recommendation performance (a typical case is illustrated in Figure 1). To address the above challenge (accurately capture users’ finegrained interests within specific video segments), we analyze the distribution of user actions along the video timeline1 (Figure 2) and uncover additional insights into user intent: (1) Temporal patterns of actions exhibit a multi-peak distribution, with peaks closely aligned with video highlights (using 𝐿𝑖𝑘𝑒 as an example, see Figure 3). Most 𝐿𝑖𝑘𝑒s near these peaks are content-driven responses. Samples corresponding to these peaks yield a Follow rate 3.3 times higher and a Collect rate 1.52 times higher than others, suggesting they better reflect users’ genuine interest in those segments. (2) Action sequences reveal differences in user attitude. For instance, the 𝐹𝑜𝑙𝑙𝑜𝑤-𝐿𝑖𝑘𝑒 2 sequence indicates users are interested in the author even without fully evaluating the content. Samples with this sequence exhibit 1.28 times longer watch time and a 1.66 times higher comment rate compared to those following the 𝐿𝑖𝑘𝑒-𝐹𝑜𝑙𝑙𝑜𝑤 sequence. Based on the above statistics and empirical observations, if we decompose the viewing process of a short video along its timeline, users are more likely to provide positive feedback (e.g., 𝐿𝑖𝑘𝑒) on the segments they find most engaging and negative feedback (e.g., 𝐸𝑥𝑖𝑡) on the segments they dislike. By capturing the timing of these signals, the model can more precisely infer users’ true interests toward different segments, enabling more fine-grained user modeling. Consequently, we incorporate action timing into the modeling process by connecting all user actions into a time-ordered sequence for each item. Both the input and output of the proposed model consist of the complete sequence, including the actions and corresponding timings. During prediction, an autoregressive generative paradigm is employed to leverage the context of preceding actions in the sequence to assist in predicting subsequent ones. Achieving the aforementioned functionality presents two major problems: effectively mining users’ segment-level interests from their historical action sequences, and leveraging these insights to accurately predict the action sequence for a target item. To this end, we propose a novel modeling framework: the Action-Aware Generative Sequence Network (A2 Gen). Unlike textual sequences, the contextual information embedded in action sequences (e.g., item 1 Based on online data analysis from Kuaishou. 2 𝐹𝑜𝑙𝑙𝑜𝑤 occurs before 𝐿𝑖𝑘𝑒 .
Wenhao Li et al.
features) directly influences both the occurrence of user actions and the interdependencies among them. Accordingly, we first propose a Context-aware Attention Module (CAM) that effectively models the input sequence by incorporating contextual information. Furthermore, since user action sequences are strongly correlated with their historical habits, we further develop a Hierarchical Sequence Encoder (HSE) based on CAM to model historical action sequences and extract information. Finally, by combining the extracted user representations with additional contextual factors (e.g., the target item, including video content features), we design an autoregressive action sequence generation module: the Action-seq Autoregressive Generator (AAG). Through this approach, A2 Gen incrementally generates the user action sequence based on the user’s historical interests and the content of the target item, enabling simultaneous prediction of both the occurrence probability and the timing of each action. Furthermore, the prediction of subsequent actions incorporates preceding actions, effectively leveraging the additional information provided by the temporal structure of the user action sequence. A2 Gen was deployed in online recommendation environment, where it optimized multi-task estimation and leveraged additional temporal information to better capture users’ true preferences. A/B tests demonstrated significant improvements in recommendation performance, ultimately driving increased user retention. Based on these experimental results, the model was scaled to full traffic, showcasing its effectiveness and reliability in real-world applications. In summary, the contributions of this paper are as follows: • We identify the limitation of binary classification models and first present the importance of action timing in distinguishing and capturing genuine user preferences. • We introduce the paradigm of chaining user actions into a sequence for unified modeling, and propose A2 Gen (including CAM, HSE, and AAG), which enables the generation of complete target action sequences through implicit autoregression. • By improving prediction accuracy and introducing the action timing ranking factor, A2 Gen achieves significant gains in Kuaishou (with hundreds of millions of daily active users) online experiments, including a 0.34% increase in user watch time, an 8.1% improvement in interaction rate, and a 0.162% enhancement in user retention value.
2 Related Work 2.1 Multi-Task Recommendation Multi-task recommendation has been developed and applied in both academia and industry for decades [22, 27]. To effectively model the dependency and relation between relevant recommendation tasks, several multi-task learning techniques are adopted to design sophisticated architecture [1, 16, 20, 34] or build comprehensive learning objectives [5, 17, 29]. As a classic approach, ShareBottom shares part of the parameters across the prediction tasks. Afterward, soft-sharing methods [28] and expert-sharing methods [16] become the mainstream to upgrade the structure of the network. MMoE [16] designs the gating mechanism on several shared expert networks to automatically
Action-Aware Generative Sequence Modeling for Short Video Recommendation
SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia
Figure 1: Consider a short video titled “Interview: Do you like Messi or Ronaldo?” A user gives the video a 𝐿𝑖𝑘𝑒 after seeing the interviewee choose “Messi,” as the user shares the same preference. However, the remaining 80% of the video showcases Ronaldo’s highlights. Traditional recommendation models, which treat the video as a holistic unit, may incorrectly infer that the user is interested in Ronaldo and subsequently recommend Ronaldo-related videos. An ideal recommendation model should recognize that the user’s interest is more strongly associated with the segment corresponding to the 𝐿𝑖𝑘𝑒 action (i.e., “Messi”) and recommend more Messi-related short videos instead. means samples nears the peaks indicates the video content at that time
The singer takes the stage.
Three judges press the button
The last judge presses the button.
Figure 3: The distribution of 𝐿𝑖𝑘𝑒 action timing (the x-axis represents time (s) and the y-axis represents the occurrence rate). Figure 2: A user’s actions with timing during short video consumption. The user 𝑆𝑡𝑎𝑟𝑡 watching at 0s (second), and performs 𝐿𝑖𝑘𝑒, 𝐶𝑜𝑚𝑚𝑒𝑛𝑡, and 𝐹𝑜𝑙𝑙𝑜𝑤 actions at 2s, 9s, 16s, respectively, and 𝐿𝑒𝑎𝑣𝑒 at 18s.
learn different combinations for each task. PLE [20] further incorporates task-specific experts to extract specific knowledge for each task. HMoE [8], STAR [18], HiNet [41] and STEM [19] solve the multi-domain recommendation task as the same as multi-task recommendation with common and specific structures. Among optimization-based methods, MetaBalance [5] proposes to balance the optimization between the target task and the auxiliary task in the magnitude of the gradient. AdaTask [29] proposes to calculate task-specific accumulative gradients to alleviate negative transfer on shared parameters. ForeMerge [17] designs an algorithm that
probabilistically drops specific tasks during optimization to dynamically selecting beneficial knowledge to transfer. Wang Xu et al. proposed the HoME [26] model to address expert network collapse. Scene-wise adaptive networks [9] further tackle dynamic cold-start optimization in CTR prediction. However, these methods do not consider how to effectively fuse information from action sequences and item sequences.
2.2
Sequence Modeling
In recommendation systems, sequence modeling plays a crucial role in capturing user preferences, significantly enhancing the performance of recommendation models. Among the extensive research about sequence modeling, a milestone research work is Deep Interest Network (DIN) [40]. DIN leverages attention architecture to learn representations of users’ historical behaviors for target items.
SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia
Wenhao Li et al.
Deep Interest Evolution Network (DIEN) [39] employs auxiliary losses to adjust representations from current to next behaviors, then uses AUGRU to model specific interest evolution processes across different target items. Deep Session Interest Network (DSIN) [4] constructs distinct interest preferences for each session. Behavior Sequence Transformer (BST) [2] utilizes multi-head attention network to learn characterizations of users’ historical behavior sequences. However, existing methods cannot handle the timing sequence of users’ different actions under a single item in short video recommendation.
2.3
Generative Recommendation
In recent years, numerous works in recommendation introduce behavior sequence and multifarious sequential module to refine the performance [2, 40]. Unlike traditional algorithms that retrieve and rank items from fixed candidates, generative recommendation methods have gained much development based on generative adversarial networks (GANs), variational autoencoders (VAEs), and Generative AI technology [3, 24]. Few years ago, many ID-based generative recommendation methods followed the paradigm of image generation to model user interests with VAE [11, 37] or diffusion models [13, 25]. Recently, with the success of the Transformers architecture and scaling law in language generation [35], LLM-based recommendation provides a new direction to encode history interaction and generate recommendations with natural language [15, 43]. These methods usually transform the recommendation task into text generation with designed prompt and instruction-tuning procedure [10, 31]. For instance, LLaRA [12] introduces a hybrid prompting method to combine item embeddings and textual item features. And LC-Rec [36] transforms the recommendation task into a language generation paradigm. Although these methods show great potential of LLM-based generative recommendation, they all rely on abundant text data. However, these approaches underperform in real-world scenarios due to data scarcity. Additionally, Meta researchers proposed HSTU [30] to replace the traditional DLRM. HSTU unifies and sequentializes all heterogeneous features in DLRM, achieving promising performance via sequential generation modeling with Transformers. Yet HSTU integrates features at the item level, lacking the ability to perform fine-grained modeling of user behavior toward single item in the temporal dimension. Nevertheless, in this work, we construct user behavior sequences (both the item level and the temporal dimension within individual items) to directly generate next behavior of user, which is both efficient and effective.
3 Problem Formulation 3.1 Problem Identification As illustrated in Figure 1, short videos contain highly diverse information, with the exposed content changing dynamically as users progress through the video. In most cases, users are not interested in every segment. However, traditional recommendation models do not explicitly model the user’s viewing process, making it difficult to identify which segments actually attracted the user’s interest. As a result, as shown in the figure, a user who liked the "Messi" segment may still be recommended many Ronaldo-related videos. In addition, correlations also exist between different user actions.
Figure 4: The modeling process of A2 Gen. For instance, if a user follows the author while watching a video, they are more likely to subsequently 𝐿𝑖𝑘𝑒 that video. Building on the above analysis, a model should precisely capture the segments that truly reflect users’ interests and exploit the hidden information contained in the sequence of actions during video consumption. Since watching a short video is inherently a temporal process, we aim to incorporate time into the modeling. A heuristic solution (as shown in Figure 4) is to represent each user’s history as a sequence of actions, where each action is associated with both its type and its position on the video timeline. The model then learns from these sequences, leveraging action timing to identify users’ genuine interest segments. During inference, a generative approach is adopted: the model combines segment-level video features with the learned representation to incrementally predict the actions a user is likely to take and their corresponding timeline positions, thereby providing richer auxiliary signals for downstream prediction. In this way, we decompose the conventional video-level prediction task into predicting when a user is most likely to perform target actions during the viewing process, achieving more fine-grained modeling.
3.2
Solution Design
During browsing, a user sends a request to the online service, which includes user information (e.g., user ID, gender, age, and other profile details), recent historical items, and 𝑛 candidate items to be ranked, denoted as X = {𝑥 1, · · · , 𝑥 𝑗 , · · · , 𝑥𝑛 }, where 𝑛 is the number of candidate items, and 𝑗 represents the 𝑗-th item. Given the candidate set X, the goal of the user action sequence generation is to generate the complete sequence: Act_seq 𝑗 = {(A1, T1 ), · · · , (A𝑖 , T𝑖 ), · · · , (A𝑚 𝑗 , T𝑚 𝑗 )}
(1)
for each item 𝑥 𝑗 , where A𝑖 represents a user action (e.g., 𝐿𝑖𝑘𝑒, 𝐹𝑜𝑙𝑙𝑜𝑤, 𝐶𝑜𝑙𝑙𝑒𝑐𝑡, etc., collectively forming set A), T𝑖 denotes the timing of each action, 𝑚 𝑗 is the length of the action sequence for 𝑥 𝑗 , and 𝑀 is the maximum sequence length, where each action occurs once, determined by the dataset. The action sequence generation model is F (𝑢, 𝑥; 𝜃 ), where 𝜃 denotes the model parameters. The computation logic of F includes
Action-Aware Generative Sequence Modeling for Short Video Recommendation
SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia
multi-class classification for action types and regression prediction for action timings. The primary objective of training F (𝑢, 𝑥; 𝜃 ) is to minimize the target function, defined as: ∑︁ b𝑖 ) 𝑎𝑟𝑔𝑚𝑖𝑛( Classification_loss(A𝑖 , A ∑︁ (2) + Regression_loss(T𝑖 , Tb𝑖 )), b𝑖 and Tb𝑖 represent the ground truth action type and timing. where A Sequence generation introduces a factorially growing permutation space, denoted as 𝑂 (𝑛 · 𝑚 𝑗 !). Typically, 𝑚 𝑗 is less than 5, and 𝑛 ranges from several tens to hundreds. Each permutation represents a potential action sequence, but only one matches the ground truth. Consequently, the key challenge lies in efficiently identifying the most accurate sequence within a vast solution space. To tackle the challenges, it’s essential to decompose the problem into manageable sub-tasks. Considering the inherent correlation between consecutive actions, a natural approach is to leverage autoregressive modeling [21], which predicts each action step by step based on previously generated actions. This paradigm simplifies the problem and aligns well with the sequential nature of user action.
3.3
Figure 5: Context-aware Attention Module (CAM)
Autoregressive Generation
Given a set of candidate items set X and an action set A, autoregressive models decompose the distribution over potential generated sequences 𝐴𝑐𝑡𝑠𝑒𝑞 into a series of conditional probabilities: 𝑝 (𝐴𝑐𝑡_𝑠𝑒𝑞 | 𝑥; 𝜃 ) =
𝑚 Ö
𝑝 (𝐴𝑖 ,𝑇𝑖 | 𝐴1:𝑖 −1,𝑇1:𝑖 −1, 𝑥; 𝜃 ),
(3)
𝑖=2
where the unique actions 𝐴1 and 𝐴𝑚 denote the beginning (𝑆𝑡𝑎𝑟𝑡) and end (𝐿𝑒𝑎𝑣𝑒) of the target sequences, respectively. Since 𝑚 is relatively small and constant, the length of the generated sequence can be set as predetermined and fixed, unlike the variable lengths typically seen in text. The logic of sequence generation is autoregressively decomposed, resulting in maximum likelihood training that involves an action type multi-class cross-entropy loss and a time regression loss at each time step: L𝑐𝑙𝑠 = − log 𝑝𝐴 (𝐴𝑐𝑡_𝑠𝑒𝑞 | 𝑥; 𝜃 ), L𝑟𝑒𝑔 =
1 ∑︁
2
(4)
𝑇 − 𝑇b . (5) 2 𝑚 The training objective aims to optimize the conditional classification probability and the regression prediction error at each step, where the computation depends on the outcomes of the preceding steps. During training, the target sequence is given, and the model conditions on the ground-truth labels of previous steps rather than the predictions generated by the model itself. This enables efficient parallelization following the teacher-forcing paradigm. In the inference phase, the autoregressive model generates the target sequence step by step, effectively capturing the influence of the preceding target sequence on the distribution of subsequent steps and improving sequence prediction. Given the strong correlations often observed between users’ sequential actions, the autoregressive paradigm is well-suited for predicting user action sequences, especially considering the vast space of possible permutations.
4 Approach 4.1 Context-aware Attention Module To effectively model users’ historical item sequences, action sequences, and the target action sequence, a general-purpose sequence processing module is required. Our sequences share similarities with text sequences: (1) Each position in the sequence involves multi-class predictions with a finite set of categories; (2) There is continuity and contextual relevance between adjacent positions. Leveraging these shared characteristics, this paper adopts mature sequence modeling techniques from the field of Natural Language Processing (NLP) [21] and employs a multi-head self-attentionbased module to model user action sequences effectively. However, our sequences differ from text sequences in important ways: (1) Each user action occurs within the context of a specific item. When calculating self-attention, the similarity between two actions should depend on the specific content of the associated items rather than being uniform. (2) Sequence generation during the prediction phase involves multi-task estimation, which requires optimizing the balance between shared and task-specific information. Based on these considerations, we developed the Contextaware Attention Module (CAM) to process both item and action sequences, shown as Figure 5. CAM incorporates two types of feature inputs: sequence features and contextual features. First, we integrate contextual information into the query to enable context-specific self-attention calculations within the sequence. It ensures that the similarity calculation between the query and the key reflects context-specific characteristics, resulting in differentiated similarity scores tailored to the unique content of different contexts. Next, in self-attention, multiple attention heads act as distinct information extraction modules, each capturing various aspects of the data. Since the importance of the information extracted by each head varies across tasks, we introduce a gating network that learns the relative importance of the
SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia
Figure 6: Hierarchical Sequence Encoder (HSE)
Wenhao Li et al.
vary significantly. Simply providing the model with historical items consumed by the user, as is typical in traditional approaches, is insufficient to accurately generate the user action sequence. First, ensuring that the model comprehensively understands the user’s historical interaction habits is crucial. It requires accurately capturing the specific interaction timings of each historical item and extracting meaningful insights from them. Furthermore, in most cases, the number of items a user has interacted with in their history is significantly larger than one. These items also form a sequential relationship along the temporal dimension, resulting in higher-dimensional sequence information. Since these items exhibit varying levels of relevance to the target item and contribute differently to the final prediction, the model must not only effectively capture the patterns within each one-dimensional sequence but also integrate and extract information across these higher-dimensional sequences in conjunction with the target item. Based on the above analysis and considerations, this paper proposes the Hierarchical Sequence Encoder (HSE, shown in Figure 6) to extract information from user historical action sequences. In our implementation, the model is designed to operate on two dimensions to capture user-item interactions effectively. Action Dimension: The sequential input of CAM is the user action sequence for each historical item, while the contextual input consists of the corresponding item features. Due to the significant content differences between items, the similarity between the same pair of actions varies across items. Finally, the model extracts and aggregates information from user historical action sequences. In summary, the formula for the action sequence of 𝑥 𝑗 is as follows: Seq 𝑗 = FCAMAct (Act_seq 𝑗 , 𝐹𝑥 𝑗 ; 𝜃 ),
Figure 7: Action-seq Autoregressive Generator (AAG) information from different heads, enabling task-specific optimization. Finally, a Multi-Layer Perceptron (MLP) structure is applied to further extract and refine the information, completing the modeling of the action sequence for a single item within the user’s interaction history. The specific formulation is as follows: ∑︁ FCAM (Act_seq, 𝑥; 𝜃 ) = 𝐺 (𝑥) · MHA(Act_seq, 𝑥), (6) where MHA(·) is multi-head attention layer, 𝐺 (·) means the gating network. Through this process, we achieve effective sequence information extraction.
4.2
Hierarchical Sequence Encoder
From users’ perspective, each individual exhibits unique action patterns, and their emotional triggers during content consumption
(7)
where 𝐹𝑥 𝑗 represents the features of the item 𝑥 𝑗 . Item Dimension: The sequential input of CAM is the concatenation of the embedding of each historical item with the output of the action-dimensional sequence extraction, while the contextual input consists of the target item features. First, the model integrates the target item’s feature information with the user’s historical sequence information to represent each historical item. Similar to the processing in the action dimension, the similarity between historical items should depend on their relevance to the target item. Therefore, the target item information is incorporated into the selfattention calculations to produce personalized similarity weights. Finally, the model captures the complete item-dimensional representation of user historical action sequences. The equation is as follows: 𝑉 𝑒𝑐 hist = FCAMItem ({Concat(Seq 𝑗 , 𝐹𝑥 𝑗 )}seq, 𝐹𝑥 target ; 𝜃 ),
(8)
where {𝜁 }seq denotes the sequence composed of 𝜁 , 𝐹𝑥 target means features of the target item.
4.3
Autoregressive Generator
The generative model leverages information from preceding actions to assist in the prediction of subsequent ones, generating the complete user action sequence step by step. Following this computation logic, we propose the Action-seq Autoregressive Generator (AAG), which is built upon CAM (Figure 7). During training, the input of AAG consists of three components: (1) Context features, including user, item, and other attributes; (2)
Action-Aware Generative Sequence Modeling for Short Video Recommendation
SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia
Figure 8: The overall architecture of A2 Gen. 𝑉 𝑒𝑐 hist , the representation of user historical action sequences; (3) the ground-truth action sequence on the target item. Parallelized training is achieved by adopting the masking mechanism as Transformer [21], allowing the model to simultaneously compute the action types and occurrence times at all positions in the sequence: 𝐹 context = Concat(𝐹𝑢 , 𝐹𝑥 target , 𝑉 𝑒𝑐 hist ),
(9)
1:𝑖 −1, 𝐹 context ; 𝜃 ), logits𝑖 = FCAMGen ( Act_seq
(10)
1:𝑖 −1 represents the ground-truth action sequence on where Act_seq the target item, and 𝐹𝑢 denotes user features. During generation, AAG leverages preceding actions to iteratively generate the types and timings of subsequent actions, ultimately producing the complete sequence: logits𝑖 = FCAMGen (Act_seq1:𝑖 −1, 𝐹 context ; 𝜃 ),
(11)
where Act_seq1:𝑖 −1 represents the preceding sequence predicted.
4.4
Decision Layer
To generate a complete user action sequence, the output layer of AAG utilizes two MLP towers specifically designed to predict user actions’ type (multi-class classification) and timing (regression). Both towers take the preceding output 𝑙𝑜𝑔𝑖𝑡𝑠𝑖 as input, with the specific formulas detailed as follows: Act𝑖 = Softmax(MLPcls (logits𝑖 ; 𝜃 )) (12) T𝑖 = MLPT (logits𝑖 ; 𝜃 ), where the number of classes is 𝑀 + 1, with the additional class representing padding, indicating that no action occurs at the position. For action timing regression, to ensure robustness and mitigate the significant variation caused by video length, the model normalizes the predicted time concerning the video duration as follows: 𝑇b𝑖 =
𝑇d 𝑖𝑠𝑒𝑐 , 𝐷𝑢𝑟𝑎𝑡𝑖𝑜𝑛𝑠𝑒𝑐
(13)
where 𝐷𝑢𝑟𝑎𝑡𝑖𝑜𝑛𝑠𝑒𝑐 denotes the total video duration (in seconds), and 𝑇d 𝑖𝑠𝑒𝑐 represents the ground-truth timing (in seconds) relative to the start of playback.
4.5
Loss
Generating action sequences requires simultaneously predicting the action type and corresponding timing: Lcls = − log 𝑝 Act (Act_seq | 𝑥; 𝜃 ), (14) ∑︁ 2 1 T−b T . (15) Lreg = 2 𝑚 Additionally, a sequential temporal order constraint loss is introduced to ensure that the generated sequences are meaningful by preserving the correct chronological order of actions: ∑︁ 2 Lorder = 𝑚𝑎𝑥 (T𝑝 − T𝑞 , 0) 2 . (16) 0<𝑝<𝑞<𝑚
To control the impact of different losses, the total loss is obtained by the weighted sum of the losses above: L = 𝛼 Lcls + 𝛽Lreg + 𝛾 Lorder,
(17)
where 𝛼, 𝛽, and 𝛾 are the weighting hyper-parameters. In summary, the overall architecture of A2 Gen, which integrates all the modules, is illustrated in Figure 8.
5 Experiments 5.1 Experimental Settings Public Dataset. The E-commerce dataset is from Tmall [33, 38], which includes features such as user-side (ID) and item-side (ID, title, et al.). The user actions include 𝐶𝑙𝑖𝑐𝑘, 𝐶𝑜𝑙𝑙𝑒𝑐𝑡, 𝐶𝑎𝑟𝑡, 𝑃𝑎𝑦, along with their corresponding timestamps. In this study, we concatenate the actions performed by user ID and item ID in chronological order to form action sequences. The data from September 22, 2014, and earlier is the training set, while data from later dates is the test set. Industrial Dataset. The industrial dataset of short videos is collected from Kuaishou (with hundreds of millions of daily active users) over one month. The dataset comprises nearly one billion users, tens of billions of items, and trillions of samples. For each behavior type, the number of positive instances ranges from billions to tens of billions. Each sample consists of user (user profile attributes and IDs), video (metrics, categories, IDs), and contextual features. The dataset includes six types of actions: 𝑆𝑡𝑎𝑟𝑡, 𝐿𝑖𝑘𝑒, 𝐹𝑜𝑙𝑙𝑜𝑤, 𝐹𝑜𝑟𝑤𝑎𝑟𝑑, 𝐶𝑜𝑙𝑙𝑒𝑐𝑡, and 𝐿𝑒𝑎𝑣𝑒, along with their respective timings. This dataset is used for offline and ablation experiments.
SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia
Wenhao Li et al.
Settings. The softmax cross-entropy and Mean Squared Error (MSE) loss functions are used, along with the Adam optimizer [6]. The number of heads in the multi-head attention is set to 4. In HSE, the user’s historical item count is 100. And 𝛼, 𝛽, and 𝛾 are set to 1, 1, and 0.1. To ensure fairness, the baseline model and A2 Gen share the same features and hyperparameters in the experiments. The comparative experiments in this paper were repeated 10 independent times for validation, and the results were statistically significant at the 0.05 level (Friedman test), indicating confidence in the effectiveness of A2 Gen. Metrics. In recommendation systems, items are typically classified as either relevant or irrelevant for a given user. To evaluate the model’s ability to distinguish between these two classes, we use the AUC (Area Under the Curve) score. A higher AUC score reflects the model’s effectiveness in differentiating relevant items from irrelevant ones, showcasing its stronger capability to recommend items that align with user preferences. In regression prediction tasks, we aim to minimize the numerical discrepancy between predicted values and ground truth. Therefore, we adopt the MAE (Mean Absolute Error) loss to evaluate model performance. For online experiments, our objectives focus on enhancing three key metrics: user watch time, interaction rate (the proportion of users performing actions such as 𝐿𝑖𝑘𝑒 and 𝐹𝑜𝑙𝑙𝑜𝑤, et al.), and the user retention metric LifeTime-7 (LT7) [23]: ∑︁7 DAU𝑖 , (18) LT7 = 𝑖=1 Numuser where DAU𝑖 denotes the Daily Active Users on the 𝑖-th day, and Numuser represents the number of active users over a 7-day period. Generally, higher AUC scores, lower loss values, and better online experiment metrics indicate that the recommended short videos better align with user preferences and expectations, demonstrating the superior ranking capability of the model.
5.2
Experimental Results Table 1: Experimental results on the public dataset.
A2 Gen Action /AUC
Time /MAE
MMoE [16] PLE [20] AdaTask [29] ForeMerge [17] STEM [19] HoME [26] A2 Gen MMoE [16] PLE [20] AdaTask [29] ForeMerge [17] STEM [19] HoME [26]
𝐶𝑜𝑙𝑙𝑒𝑐𝑡
𝐶𝑎𝑟𝑡
𝑃𝑎𝑦
0.7198 0.6891 0.6957 0.7107 0.7148 0.7151 0.7156 0.8829 0.9817 0.9372 0.9053 0.8989 0.8976 0.8928
0.7962 0.7594 0.7682 0.7889 0.7858 0.7872 0.7897 0.8781 0.9875 0.9682 0.8868 0.8904 0.8882 0.8841
0.7351 0.6948 0.7120 0.7268 0.7254 0.7285 0.7319 0.8292 0.9081 0.8810 0.8353 0.8382 0.8371 0.8344
This section presents the offline experiments in predicting the occurrence and timing of user actions on the Tmall and Kuaishou’s datasets. Since all 𝐿𝑒𝑎𝑣𝑒 in this dataset are positive (users inevitably exit after viewing), we only evaluate the time prediction. In addition,
Table 2: Action experiments on the industrial dataset (AUC).
A2 Gen MMoE [16] PLE [20] AdaTask [29] ForeMerge [17] STEM [19] HoME [26]
𝐿𝑖𝑘𝑒
𝐹𝑜𝑙𝑙𝑜𝑤
𝐹𝑜𝑟 𝑤𝑎𝑟𝑑
𝐶𝑜𝑙𝑙𝑒𝑐𝑡
0.7137 0.6814 0.6994 0.7043 0.7059 0.7068 0.7079
0.7303 0.6981 0.7131 0.7204 0.7231 0.7229 0.7242
0.7298 0.6899 0.7201 0.7211 0.7203 0.7223 0.7244
0.7286 0.6989 0.7179 0.7172 0.7198 0.7221 0.7245
Table 3: Timing experiments on the industrial dataset (MAE).
A2 Gen MMoE [16] PLE [20] AdaTask [29] ForeMerge [17] STEM [19] HoME [26]
𝐿𝑖𝑘𝑒
𝐹𝑜𝑙𝑙𝑜𝑤
𝐹𝑜𝑟 𝑤𝑎𝑟𝑑
𝐶𝑜𝑙𝑙𝑒𝑐𝑡
𝐿𝑒𝑎𝑣𝑒
0.5353 0.6013 0.5829 0.5637 0.5594 0.5553 0.5496
0.5042 0.5632 0.5519 0.5329 0.5435 0.5318 0.5278
0.4991 0.5599 0.5493 0.5248 0.5299 0.5217 0.5164
0.4979 0.5537 0.5459 0.5106 0.5148 0.5115 0.5079
0.5594 0.6041 0.5967 0.5747 0.5793 0.5789 0.5708
all model comparisons were conducted using an identical set of features to ensure fairness. Detailed experimental results are shown in Table 1, 2 and 3. First, we assess action prediction using 𝐴𝑈𝐶, where A2 Gen predicts each action by taking the maximum predicted probability across all positions in the sequence: 𝑃𝐴𝑐𝑡 = 𝑀𝑎𝑥 (𝑃𝐴𝑐𝑡1 , · · · , 𝑃𝐴𝑐𝑡𝑚 ). Compared to baseline models, A2 Gen achieves significant 𝐴𝑈 𝐶 improvements for all action types (e.g., achieves a 58 bp improvement in 𝐴𝑈𝐶 for 𝐿𝑖𝑘𝑒 action prediction compared to the SOTA HoME [26]). It also achieves a 143 bp improvement over the online production model PLE. In general, a higher 𝐴𝑈𝐶 indicates better ranking accuracy for items, enabling the system to provide users with videos that better match their interests. This, in turn, increases user engagement with the platform and drives economic benefits for the business. These results demonstrate A2 Gen ’s ability to capture richer latent patterns from sequential data and translate them into superior prediction performance. Second, we evaluate timing prediction via 𝑀𝐴𝐸 loss. A2 Gen estimates timings by computing the weighted sum of predicted probabilities at each sequence position multiplied by their corresponding Í timing: 𝑃𝐴𝑐𝑡𝑇 = 𝑆𝑒𝑞 𝑃𝐴𝑐𝑡 · 𝑃𝑇 . The results indicate that A2 Gen achieves lower 𝑀𝐴𝐸 loss for all action types (e.g., A2 Gen achieves a 2.60% reduction in 𝐿𝑖𝑘𝑒 timing loss compared to HoME [26]), suggesting its time predictions are closer to the ground truth.
5.3
Ablation Study and Analysis
We further conduct ablation studies on the Kuaishou dataset, with results in Table 4 and 5. First, we examine removing HSE (by replacing 𝑉 𝑒𝑐ℎ𝑖𝑠𝑡 in Eq. 9 with the sum-pooling embedding of user’s historical items and actions) and of retaining the item dimension in HSE (by replacing only 𝑆𝑒𝑞 𝑗 in Eq. 8 with the sum pooling embedding of the user’s historical actions), and observe a consistent decrease in accuracy across all action types. This highlights the importance of HSE in effectively extracting information from users’
Action-Aware Generative Sequence Modeling for Short Video Recommendation
(a) 𝛼 (𝛽 = 1, 𝛾 = 0.1)
SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia
(b) 𝛽 (𝛼 = 1, 𝛾 = 0.1)
(c) 𝛾 (𝛼 = 1, 𝛽 = 1)
Figure 9: Hyper-parameter analysis of the Loss function on the 𝐿𝑖𝑘𝑒 action. Two hyper-parameters are fixed while the remaining one varies. The corresponding AUC of 𝐿𝑖𝑘𝑒 and loss curves of 𝐿𝑖𝑘𝑒 Timing under different hyper-parameter settings are reported. Table 4: Action ablation study without (¬) structures (AUC).
A2 Gen ¬ HSE ¬ Action-seq ¬ CAM ¬ Lorder
𝐿𝑖𝑘𝑒
𝐹𝑜𝑙𝑙𝑜𝑤
𝐹𝑜𝑟 𝑤𝑎𝑟𝑑
𝐶𝑜𝑙𝑙𝑒𝑐𝑡
0.7137 0.7038 0.7082 0.7058 0.7121
0.7303 0.7159 0.7175 0.7204 0.7281
0.7298 0.7231 0.7251 0.7247 0.7282
0.7286 0.7146 0.7192 0.7184 0.7271
Figure 9b exhibits a similar trend for 𝛽, where both small and large values adversely affect AUC, further corroborating the interaction between classification objectives and timing estimation. As shown in Figure 9c, 𝛾 controls the Lorder term, which is designed to enforce the temporal ordering of actions in sequence generation. Consequently, a relatively small value of 𝛾 is sufficient to serve as an auxiliary training signal.
5.5 Table 5: Timing ablation study without (¬) structures (MAE).
A2 Gen ¬ HSE ¬ Action-seq ¬ CAM ¬ Lorder
𝐿𝑖𝑘𝑒
𝐹𝑜𝑙𝑙𝑜𝑤
𝐹𝑜𝑟 𝑤𝑎𝑟𝑑
𝐶𝑜𝑙𝑙𝑒𝑐𝑡
𝐿𝑒𝑎𝑣𝑒
0.5353 0.5543 0.5419 0.5523 0.5431
0.5042 0.5255 0.5164 0.5202 0.5114
0.4991 0.5148 0.5079 0.5129 0.5033
0.4979 0.5207 0.5139 0.5197 0.5059
0.5594 0.5712 0.5669 0.5691 0.5641
historical action sequences. Next, we compare CAM with the original attention mechanism module [21]. Since historical sequences in recommendation systems are typically strongly correlated with contextual information and the target item [40], the results demonstrate that CAM is more effective in leveraging such background information. Finally, we assess the effect of L𝑜𝑟𝑑𝑒𝑟 and find that this loss function helps the model generate logically coherent sequences, contributing to overall performance improvements.
5.4
Hyper-parameters Experiments
We investigate the impact of the loss-function hyper-parameters 𝛼, 𝛽, and 𝛾 on both classification performance and action timing regression, as illustrated in Figure 9. As shown in Figure 9a, when 𝛼 is small, increasing 𝛼 leads to simultaneous improvements in AUC and timing loss, indicating that the classification loss plays a critical role in optimizing AUC and can also cooperatively benefit action timing prediction. However, when 𝛼 becomes excessively large, the model tends to overemphasize classification while neglecting timing estimation, resulting in a degradation in AUC. These observations demonstrate a strong correlation between user action timing and action prediction.
Application in Practice
To demonstrate the effectiveness of A2 Gen, we deploy large-scale online A/B tests, where each group serves tens of millions of DAU over a one-month period. Performance gains are achieved through the following optimizations: Model Replacement. Tables 2 and 3 show that A2 Gen outperforms the baseline model (PLE), enabling more precise online predictions. Following the model replacement, we achieved a 0.11% increase in watch time and a 2.1% improvement in interaction rate. Action Timing Aware. As described in Section 1, the timing of actions reflects user intentions. Consequently, we optimize the item ranking mechanism by incorporating temporal predictions generated by AAG, informed by empirical data analysis and domain expertise. For example, we boost the ranking scores with lateroccurring 𝐿𝑖𝑘𝑒 (indicating that a user continues to appreciate a video after a deeper engagement). This strategy also filters out noisy data (e.g., downweights indiscriminate 𝐿𝑖𝑘𝑒𝑠 that typically occur early). This optimization resulted in a 0.13% increase in watch time, a 3.5% boost in interaction rate, and a 0.12% improvement in 𝐿𝑇 7. Action Sequence Aware. The order in which actions occur can reveal users’ underlying intentions. To capture this, we adjust ranking scores based on action sequences; for instance, we elevate the scores featuring a 𝐹𝑜𝑙𝑙𝑜𝑤-𝐿𝑖𝑘𝑒 (implying a particularly strong interest in the content author, and the proportion of action sequence samples in the experimental group is shown in Table 7). This modification yielded a 0.10% growth in watch time, a 1.4% enhancement in interaction rate. Action Timing Distribution Aware. Through an in-depth examination of aggregate data, we have uncovered that the temporal distribution of occurrence rates for various action types manifests as a multi-peak distribution. Each peak typically corresponds to an emotional highlight within a video, and actions triggered near
SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia
Table 6: The proportion of action timing distribution.
Exp / Base
Off Peak
Near peak
-1.0415%
+1.6982%
Table 7: The proportion of 𝐿𝑖𝑘𝑒 and 𝐹𝑜𝑙𝑙𝑜𝑤 action sequences.
Exp / Base
𝐿𝑖𝑘𝑒-𝐹𝑜𝑙𝑙𝑜𝑤
𝐹𝑜𝑙𝑙𝑜𝑤-𝐿𝑖𝑘𝑒
-0.8912%
+1.8247%
these peaks tend to align more closely with users’ genuine interests. Consequently, we boost the ranking scores whose action timings fall near these peaks. The proportion change of samples in the experimental group is shown in Table 6, ultimately achieving a 0.042% improvement in user 𝐿𝑇 7 and a 1.1% increase in interaction rate. All the above performance improvements have been verified through significance testing. Through the above online deployment optimizations, A2 Gen achieved an approximate 0.162% improvement in user retention 𝐿𝑇 7, equivalent to roughly one million additional DAU per day.
6
Conclusion
This paper identifies the limitations of binary classification models in capturing true user interests, due to the implicit intent variance behind user actions. Through in-depth data exploration, we further demonstrate the critical role of action timing in enhancing interest modeling. Motivated by these observations, we propose an ActionAware Generative Sequence Network (A2 Gen), which chains user actions into sequences for unified modeling and prediction, effectively leveraging the temporal dependencies among actions. By introducing a Context-aware Attention Module (CAM) and a Hierarchical Sequence Encoder (HSE), A2 Gen captures the intricate timing-aware relationships between different user actions. Additionally, we design AAG to produce complete user action sequences. A2 Gen outperforms baseline models on both Kuaishou and Tmall offline datasets. In online deployment, A2 Gen improves interest discovery and ranking performance through four strategies: Model Replacement, Action Timing Awareness, Action Sequence Awareness, and Action Timing Distribution Aware. Overall, it achieves a 0.34% increase in user watch time, an 8.1% improvement in interaction rate, and a 0.162% gain in overall user Lifetime-7 value, equivalent to roughly one million additional DAU per day.
Acknowledgments This work was supported in part by the National Key Research and Development Program of China under Grant 2023YFB3307500, in part by the National Natural Science Foundation of China under Grants 62522201 and 62202025, in part by the Young Elite Scientist Sponsorship Program by CAST under Grant YESS20230566, in part by Beijing Natural Science Foundation under Grant L241050, and in part by the Fundamental Research Fund Project of Beihang University.
References [1] Jianxin Chang, Chenbin Zhang, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, and Kun Gai. 2023. Pepnet: Parameter and embedding personalized network for infusing with personalized prior information. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3795–3804.
Wenhao Li et al.
[2] Qiwei Chen, Huan Zhao, Wei Li, Pipei Huang, and Wenwu Ou. 2019. Behavior sequence transformer for e-commerce recommendation in alibaba. In Proceedings of the 1st international workshop on deep learning practice for high-dimensional sparse data. 1–4. [3] Yashar Deldjoo, Zhankui He, Julian McAuley, Anton Korikov, Scott Sanner, Arnau Ramisa, René Vidal, Maheswaran Sathiamoorthy, Atoosa Kasirzadeh, and Silvia Milano. 2024. A review of modern recommender systems using generative models (gen-recsys). In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 6448–6458. [4] Yufei Feng, Fuyu Lv, Weichen Shen, Menghan Wang, Fei Sun, Yu Zhu, and Keping Yang. 2019. Deep session interest network for click-through rate prediction. In Proceedings of the 28th International Joint Conference on Artificial Intelligence. 2301–2307. [5] Yun He, Xue Feng, Cheng Cheng, Geng Ji, Yunsong Guo, and James Caverlee. 2022. Metabalance: improving multi-task recommendations via adapting gradient magnitudes of auxiliary tasks. In Proceedings of the ACM Web Conference 2022. 2205–2215. [6] Diederik P Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations (ICLR). [7] Hyeyoung Ko, Suyeon Lee, Yoonseo Park, and Anna Choi. 2022. A survey of recommendation systems: recommendation models, techniques, and application fields. Electronics 11, 1 (2022), 141. [8] Pengcheng Li, Runze Li, Qing Da, An-Xiang Zeng, and Lijun Zhang. 2020. Improving multi-scenario learning to rank in e-commerce by exploiting task relationships in the label space. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management. 2605–2612. [9] Wenhao Li, Jie Zhou, Chuan Luo, Chao Tang, Kun Zhang, and Shixiong Zhao. 2024. Scene-wise adaptive network for dynamic cold-start scenes optimization in ctr prediction. In Proceedings of the 18th ACM Conference on Recommender Systems. 370–379. [10] Yongqi Li, Xinyu Lin, Wenjie Wang, Fuli Feng, Liang Pang, Wenjie Li, Liqiang Nie, Xiangnan He, and Tat-Seng Chua. 2024. A survey of generative search and recommendation in the era of large language models. arXiv preprint arXiv:2404.16924 (2024). [11] Shangsong Liang, Zhou Pan, wei liu, Jian Yin, and Maarten de Rijke. 2024. A Survey on Variational Autoencoders in Recommender Systems. Comput. Surveys (2024). [12] Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2023. Llara: Aligning large language models with sequential recommenders. arXiv preprint arXiv:2312.02445 (2023). [13] Jianghao Lin, Jiaqi Liu, Jiachen Zhu, Yunjia Xi, Chengkai Liu, Yangtian Zhang, Yong Yu, and Weinan Zhang. 2024. A Survey on Diffusion Models for Recommender Systems. arXiv preprint arXiv:2409.05033 (2024). [14] Xiaofan Liu, Qinglin Jia, Chuhan Wu, Jingjie Li, Dai Quanyu, Lin Bo, Rui Zhang, and Ruiming Tang. 2023. Task adaptive multi-learner network for joint CTR and CVR estimation. In Companion Proceedings of the ACM Web Conference 2023. 490–494. [15] Sichun Luo, Yuxuan Yao, Bowei He, Yinya Huang, Aojun Zhou, Xinyi Zhang, Yuanzhang Xiao, Mingjie Zhan, and Linqi Song. 2024. Integrating large language models into recommendation via mutual augmentation and adaptive aggregation. arXiv preprint arXiv:2401.13870 (2024). [16] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. 2018. Modeling task relationships in multi-task learning with multi-gate mixture-ofexperts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1930–1939. [17] Aakarsh Malhotra, Mayank Vatsa, and Richa Singh. 2022. Dropped scheduled task: Mitigating negative transfer in multi-task learning using dynamic task dropping. Transactions on Machine Learning Research (2022). [18] Xiang-Rong Sheng, Liqin Zhao, Guorui Zhou, Xinyao Ding, Binding Dai, Qiang Luo, Siran Yang, Jingshan Lv, Chi Zhang, Hongbo Deng, et al. 2021. One model to serve all: Star topology adaptive recommender for multi-domain ctr prediction. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management. 4104–4113. [19] Liangcai Su, Junwei Pan, Ximei Wang, Xi Xiao, Shijie Quan, Xihua Chen, and Jie Jiang. 2024. STEM: unleashing the power of embeddings for multi-task recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 9002–9010. [20] Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. 2020. Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recommendations. In Proceedings of the 14th ACM Conference on Recommender Systems. 269–278. [21] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, Vol. 30. [22] Nelson Vithayathil Varghese and Qusay H Mahmoud. 2020. A survey of multi-task deep reinforcement learning. Electronics 9, 9 (2020), 1363.
Action-Aware Generative Sequence Modeling for Short Video Recommendation
[23] Ruize Wang, Hui Xu, Ying Cheng, Qi He, Xing Zhou, Rui Feng, Wei Xu, Lei Huang, and Jie Jiang. 2024. ADSNet: Cross-Domain LTV Prediction with an Adaptive Siamese Network in Advertising. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 5872–5881. [24] Wenjie Wang, Xinyu Lin, Fuli Feng, Xiangnan He, and Tat-Seng Chua. 2023. Generative recommendation: Towards next-generation recommender paradigm. arXiv preprint arXiv:2304.03516 (2023). [25] Wenjie Wang, Yiyan Xu, Fuli Feng, Xinyu Lin, Xiangnan He, and Tat-Seng Chua. 2023. Diffusion recommender model. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 832–841. [26] Xu Wang, Jiangxia Cao, Zhiyi Fu, Kun Gai, and Guorui Zhou. 2025. Home: Hierarchy of multi-gate experts for multi-task learning at kuaishou. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1. 2638–2647. [27] Yuhao Wang, Ha Tsz Lam, Yi Wong, Ziru Liu, Xiangyu Zhao, Yichao Wang, Bo Chen, Huifeng Guo, and Ruiming Tang. 2024. Multi-task deep recommender systems: A survey. IEEE Transactions on Knowledge and Data Engineering 36, 5 (2024), 2038–2057. [28] Shen Xin, Martin Ester, Jiajun Bu, Chengwei Yao, Zhao Li, Xun Zhou, Yizhou Ye, and Can Wang. 2019. Multi-task based sales predictions for online promotions. In Proceedings of the 28th ACM international conference on information and knowledge management. 2823–2831. [29] Enneng Yang, Junwei Pan, Ximei Wang, Haibin Yu, Li Shen, Xihua Chen, Lei Xiao, Jie Jiang, and Guibing Guo. 2023. Adatask: A task-aware adaptive learning rate approach to multi-task learning. In Proceedings of the AAAI conference on artificial intelligence, Vol. 37. 10745–10753. [30] Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Jiayuan He, et al. 2024. Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations. In International Conference on Machine Learning. PMLR, 58484–58509. [31] Junjie Zhang, Ruobing Xie, Yupeng Hou, Xin Zhao, Leyu Lin, and Ji-Rong Wen. 2023. Recommendation as instruction following: A large language model empowered recommendation approach. ACM Transactions on Information Systems (2023). [32] Shuai Zhang, Lina Yao, Aixin Sun, and Yi Tay. 2019. Deep learning based recommender system: A survey and new perspectives. ACM computing surveys (CSUR) 52, 1 (2019), 1–38. [33] Yuyu Zhang, Liang Pang, Lei Shi, and Bin Wang. 2014. Large scale purchase prediction with historical user actions on B2C online retail platform. arXiv
SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia
preprint arXiv:1408.6515 (2014). [34] Zijian Zhang, Shuchang Liu, Jiaao Yu, Qingpeng Cai, Xiangyu Zhao, Chunxu Zhang, Ziru Liu, Qidong Liu, Hongwei Zhao, Lantao Hu, et al. 2024. M3oe: Multi-domain multi-task mixture-of experts recommendation framework. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 893–902. [35] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223 (2023). [36] Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting large language models by integrating collaborative semantics for recommendation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 1435–1448. [37] Kai Zheng, Xianjun Yang, Yilei Wang, Yingjie Wu, and Xianghan Zheng. 2020. Collaborative filtering recommendation algorithm based on variational inference. International Journal of Crowd Science 4, 1 (2020), 31–44. [38] Wenliang Zhong, Rong Jin, Cheng Yang, Xiaowei Yan, Qi Zhang, and Qiang Li. 2015. Stock constrained recommendation in tmall. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2287–2296. [39] Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Deep interest evolution network for click-through rate prediction. In Proceedings of the AAAI conference on artificial intelligence, Vol. 33. 5941–5948. [40] Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep interest network for click-through rate prediction. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1059–1068. [41] Jie Zhou, Xianshuai Cao, Wenhao Li, Lin Bo, Kun Zhang, Chuan Luo, and Qian Yu. 2023. Hinet: Novel multi-scenario & multi-task learning with hierarchical information extraction. In 2023 IEEE 39th International Conference on Data Engineering (ICDE). IEEE, 2969–2975. [42] Jieming Zhu, Qinglin Jia, Guohao Cai, Quanyu Dai, Jingjie Li, Zhenhua Dong, Ruiming Tang, and Rui Zhang. 2023. Final: Factorized interaction layer for ctr prediction. In Proceedings of the 46th International ACM SIGIR conference on research and development in information retrieval. 2006–2010. [43] Yaochen Zhu, Liang Wu, Qi Guo, Liangjie Hong, and Jundong Li. 2024. Collaborative large language model for recommender systems. In Proceedings of the ACM on Web Conference 2024. 3162–3172.