Source-linked AI summary
Lifelong Sequential Modeling with Personalized Memorization for User Response Prediction
Kan Ren, Jiarui Qin, Yuchen Fang, Weinan Zhang, Lei Zheng, Weijie Bian, Guorui Zhou, Jian Xu, Yong Yu, Xiaoqiang Zhu, Kun Gai
TL;DR
Long-lived user histories contain changing interests and multi-scale dependencies that recent-behavior sequential models do not fully capture. The paper introduces HPMN, which maintains personalized hierarchical memories with periodic updates, and reports significant improvements over strong baselines across three real-world datasets.
Problem
Existing sequential user-modeling methods mainly use recent behaviors, leaving lifelong histories and their long-term, multi-scale dependencies insufficiently modeled.
Method
HPMN maintains personalized hierarchical memories and periodically updates them at different layers to retain multi-scale patterns from evolving user behavior.
Results
HPMN significantly improves user response prediction over strong baselines, including state-of-the-art methods, across three large-scale real-world datasets.
Takeaways & Limitations
Lifelong sequential modeling can improve user modeling and response prediction when supported by a memory architecture designed for long-term, multi-scale behavior patterns.
Takeaways & Limitations
The paper leaves multi-task user modeling and dynamically learned update periods for future work.
Abstract
from arXiv · showhide
User response prediction, which models the user preference w.r.t. the presented items, plays a key role in online services. With two-decade rapid development, nowadays the cumulated user behavior sequences on mature Internet service platforms have become extremely long since the user's first registration. Each user not only has intrinsic tastes, but also keeps changing her personal interests during lifetime. Hence, it is challenging to handle such lifelong sequential modeling for each individual user. Existing methodologies for sequential modeling are only capable of dealing with relatively recent user behaviors, which leaves huge space for modeling long-term especially lifelong sequential patterns to facilitate user modeling. Moreover, one user's behavior may be accounted for various previous behaviors within her whole online activity history, i.e., long-term dependency with multi-scale sequential patterns. In order to tackle these challenges, in this paper, we propose a Hierarchical Periodic Memory Network for lifelong sequential modeling with personalized memorization of sequential patterns for each user. The model also adopts a hierarchical and periodical updating mechanism to capture multi-scale sequential patterns of user interests while supporting the evolving user behavior logs. The experimental results over three large-scale real-world datasets have demonstrated the advantages of our proposed model with significant improvement in user response prediction performance against the state-of-the-arts.
1 INTRODUCTION
User response prediction must model both evolving interests and dependencies across users’ increasingly long behavior histories. The paper proposes HPMN, a personalized hierarchical memory network with periodic updates for lifelong, multi-scale sequential modeling.
- User behavior sequences vary in length, contain long-term dependencies, and exhibit multi-scale temporal patterns reflecting changing interests.Early behaviors may contribute to later decisions, while users differ in activity and registration histories.
- Existing user-modeling methods either aggregate histories while ignoring temporal dynamics or focus mainly on recent behaviors.These limitations leave lifelong sequential patterns insufficiently modeled.
- HPMN maintains personalized memories for intrinsic tastes and multi-facet interests, while hierarchical memories retain long-term behavioral knowledge.The model is designed to represent each user’s evolving lifelong history.
- HPMN updates memory at different periods across layers to capture multi-scale sequential patterns as new behaviors arrive.The incremental, hierarchical updating mechanism supports evolving behavior logs.
- Experiments on three large-scale real-world datasets show significant improvement over strong baselines, including state-of-the-art methods.The paper frames this as evidence for the proposed lifelong sequential modeling approach.
2 RELATED WORKS
Related work establishes user modeling as central to response prediction but identifies limitations in methods that ignore temporal drift or restrict attention to recent behavior. Memory-based approaches also have not adequately addressed lifelong, multi-scale dependencies in this setting.
- User modeling derives adaptive representations of latent user interests and is a key component of user response prediction.Existing approaches range from latent-factor methods to deep representation learning.
- Many conventional models aggregate historical behaviors as a whole, thereby ignoring temporal and drifting user interests.This motivates sequential approaches that preserve behavior order and evolution.
- Sequential user-modeling research includes temporal matrix factorization, Markov-chain methods, and other sequence-based approaches for changing preferences.The literature has increasingly focused on behavioral sequences because they contain information about drifting interests.
- Most sequential methods focus on short-term behavior, while long-term methods capture only simple patterns and do not fully model long-term, multi-scale dependencies.Static intrinsic-interest representations and hierarchical attention extend coverage but remain limited in sequential complexity.
- Memory networks have been applied to recommendation, but existing adaptations do not consider practical user-response issues such as multi-scale knowledge memorization and long-term dependencies.A recurrent multi-scale model from NLP targets fixed-length sentence modeling rather than lifelong user histories.
3 METHODOLOGY
The methodology section defines lifelong sequential modeling and presents its overall architecture and data flow using the Hierarchical Periodic Memory Network. It also introduces the notation used throughout the model description.
- The paper first establishes notation and preliminaries for user response prediction before defining lifelong sequential modeling.These foundations precede the model architecture.
- The methodology presents HPMN as the architecture for lifelong sequential modeling, including its overall data flow.Detailed motivation and model design are developed within this framework.
- Table 1 summarizes the notation used in the subsequent formulation and model description.
3.1 Preliminaries
The paper formulates response prediction from user, item, context, and historical interaction features, with variable-length behavior sequences. Conventional sequential modeling typically summarizes only a recent prefix of each user’s history.
- Each training instance contains a user, an item, and a binary behavior label indicating whether the user interacted with that item.The formulation treats clicks as the representative user behavior.
- User response prediction estimates the probability of a user action, such as a click, for a given item at a specified time.The task is represented through a learned prediction function over user, item, and context information.
- The paper focuses on CTR estimation, while other response types can be handled using the same formulation.
- User features combine side information with a sequence of previously interacted items whose length varies across users.The historical sequence length is denoted by T and is not fixed across the population.
- Sequential user modeling learns a comprehensive user representation from a recent window of s behaviors, where s is typically smaller than the full history length T.Examples in prior work use fixed windows such as s = 5 or s = 50.
3.2 Lifelong Sequential Modeling
Lifelong Sequential Modeling continuously mines sequential patterns from a user's entire behavior history while preserving efficient online updates. It targets comprehensive modeling of intrinsic interests, temporal tastes, and continual adaptation to new behavior.
- Existing sequential methods typically use only the recent s behaviors, where s may be much smaller than the full sequence length T.
- Lifelong Sequential Modeling continuously performs online user modeling and sequential pattern mining over lifelong behavior histories.
- LSM requires efficient preservation of lifelong behavior patterns because real-time inference cannot maintain each user's complete history.
- LSM combines intrinsic interests, temporal user tastes, and continuous adaptation to up-to-date behaviors.
- For online inference, each request queries the target user's maintained personalized memory using the target item's vector.
3.3 Hierarchical Periodic Memory Network
HPMN maintains personalized hierarchical memories that update incrementally at different periods to represent long-term, multi-scale user behavior patterns. It reads these memories attentively for response prediction while regularizing them to preserve diverse user-interest knowledge.
- 3.3 Hierarchical Periodic Memory Network: HPMN uses hierarchical memory layers with different update periods to retain long-term and multi-scale sequential patterns.Each layer maintains a memory slot, and layer states are transferred across time steps and layers.
- 3.3.2 Continuous Memory Update: The model incrementally updates each layer after user behaviors, while periodic updates leave a layer's memory unchanged between its scheduled update steps.Each update uses a GRU-based memory-writing operation governed by the layer-specific period t_j.
- 3.3.2 Continuous Memory Update: Upper layers update less frequently, helping model longer sequences and preserve longer-term dependencies than lower layers.The periodic mechanism is also described as reducing gradient vanishing or explosion.
- 3.3.2 Continuous Memory Update: HPMN extends clockwork-style updating by connecting layers through state transfer and adding external memory to preserve user-interest knowledge.
- 3.3.3 Attentional Memory Reading: For prediction, the model combines maintained memory slots into a user representation using learned contribution weights and an energy-based attention mechanism.The target query is matched against long-term memory to obtain a comprehensive representation.
- 3.3.4 Covariance Regularization: Covariance regularization penalizes covariance among memory slots so different slots preserve diverse user-interest knowledge.
3.4 Prediction Function and Losses
The model queries each user’s maintained personalized memory to form a comprehensive representation for response prediction, while periodic memory maintenance supports lifelong modeling.
- HPMN queries the target user’s personalized memory to obtain a comprehensive representation for estimating the user response probability.The maintained memory is queried with the target item rather than recomputing user modeling from the entire behavior sequence.
- Periodic hierarchical memory updates are designed to retain long-term and multi-scale sequential patterns in evolving user behavior.Upper layers update less frequently, helping model long sequences and retain longer-term dependencies.
- HPMN differs from lifelong machine learning because it retains user-specific knowledge for user modeling while behavior patterns drift over time.Lifelong machine learning instead retains model-specific knowledge for continuously changing tasks and data samples.
- The model maintains user-specific memory throughout lifelong modeling, expanding the memory as the user’s behavior history grows.The number of memory slots is a hyperparameter that depends on the practical setting.
4 EXPERIMENTS
The experiments evaluate lifelong sequential modeling through three research questions concerning lifelong data, comparative performance, and the sequential patterns captured by HPMN.
- The study presents experimental setups, corresponding results, and extended discussions to assess the proposed method’s effectiveness.The paper also states that its code has been published.
- The experiments ask whether lifelong behavior sequences improve response prediction and whether HPMN achieves the best performance under comparable settings.They also investigate whether HPMN captures long-term, short-term, and multi-scale sequential patterns.
- The third research question examines whether HPMN captures long-term, short-term, and multi-scale sequential patterns from user behavior sequences.
4.1 Experimental Setups
The study evaluates HPMN against aggregation-based, short-term sequential, and long-term sequential baselines on three real-world datasets using AUC and Log-loss.
- Datasets: The evaluation uses Amazon, Taobao, and XLong, covering approximately fifteen years of sparse behavior, nine days of active behavior, and one thousand clicks over half a year, respectively.XLong is described as containing longer historical behavior sequences than the other two datasets.
- Evaluation Metrics: AUC measures pairwise ranking between clicked and non-clicked samples, while Log-loss measures the overall likelihood of the test data.Higher AUC and lower Log-loss indicate better performance.
- Experimental Protocol: All models receive the same contextual features and side information, and training and test data are split chronologically at a timestamp cut.The resulting split is approximately 70% training data and 30% test data.
- Compared Settings: HPMN is the first work identified here for lifelong sequential modeling in user response prediction, whereas existing baselines target aggregation, short-term, or long-term sequential modeling.The compared methods include RNN-, CNN-, attention-, and memory-based approaches.
- Compared Settings: For online inference, HPMN and RUM query personalized memory, whereas other baselines load the whole behavior sequence for response prediction.The memory-based approach is described as more space-efficient for online sequential modeling.
- Compared Settings: HPMN uses hierarchical periodic updates to capture multi-scale sequential patterns, distinguishing it from RUM’s memory architecture.The HPMN structure includes memory-slot sizes and layer-specific update periods reported by dataset.
4.2 Experimental Results and Analysis
HPMN achieves state-of-the-art response-prediction performance against all baselines, while lifelong sequences generally improve modeling but require suitable memory mechanisms.
- HPMN significantly outperforms all baselines and achieves state-of-the-art performance.
- Comparison between HPMN and baselines: Aggregation-based DNN and SVD++ models underperform sequential methods, indicating that user behavior contains sequential patterns.
- Comparison between HPMN and baselines: HPMN outperforms sequential baselines across behavior-sequence lengths, while some baselines lose long-term dependencies or multi-scale patterns.
- Analysis about Lifelong Sequential Modeling: Lifelong sequences improve most models on Amazon and Taobao, but performance drops for several short-term models on XLong.
- Model Convergence: HPMN converges quickly, with Log-loss stabilizing after about one iteration over each training dataset.
4.3 Extended Investigation
Extended analyses show that HPMN captures long-term, short-term, and multi-scale dependencies through hierarchical memories, while its useful capacity depends on sequence length.
- Sequential Patterns with Multi-scale Dependency: HPMN captures long-term, short-term, and multi-scale sequential patterns in XLong behavior sequences.
- Sequential Patterns with Multi-scale Dependency: Higher HPMN layers attend to distant lotion clicks, demonstrating their role in capturing long-term dependencies.
- Sequential Patterns with Multi-scale Dependency: Lower HPMN layers model recent furniture-related behavior because they update memory more frequently.
- Sequential Patterns with Multi-scale Dependency: Different HPMN layers jointly capture recent and distant dependencies when both contribute to a target-item response.
- Memory Capacity: AUC rises sharply below five memory slots but drops slightly above five on XLong, indicating sequence-dependent capacity constraints.
5 CONCLUSION
The paper presents lifelong sequential modeling with personalized hierarchical memory and periodic updates for user response prediction, and reports significant gains over strong baselines.
- HPMN maintains personalized hierarchical memories and periodically updates them to retain multi-scale sequential knowledge for response prediction.
- Future work will extend lifelong sequential modeling to multi-task prediction of user clicks and conversions.