Uniform Sequence Better: Time Interval Aware Data Augmentation for Sequential Recommendation

Yizhou DangEnneng YangGuibing GuoLinying JiangXingwei WangXiaoxiao XuQinghui SunHong Liu

article2023AAAI105 citations

Proposes time-interval-aware data augmentation operators that transform irregular interaction sequences into uniform sequences to mitigate user preference drift and substantially improve sequential recommendation accuracy.

Listen

Modern online platforms rely heavily on sequential recommendation systems to predict what item a user will interact with next. Existing methods primarily focus on the ordering of interactions while ignoring the elapsed time between them. However, real-world user activity is often bursty or irregular, causing significant gaps in interaction histories. When long time gaps occur, user preferences frequently shift—a phenomenon known as preference drift—which degrades recommendation accuracy and diminishes user engagement.

The article evaluates the impact of time interval distributions on recommendation performance and introduces a novel framework that improves prediction accuracy by restructuring interaction histories through time-aware data augmentation.

To address this challenge, the authors first conducted an empirical benchmark showing that training models on sequences with uniformly spaced time intervals delivers substantial accuracy gains over sequences with irregular gaps. Building on this insight, they developed five targeted data modification techniques—insertion, cropping, masking, substitution, and reordering—specifically designed to regularize irregular time intervals or preserve core interaction patterns. These operators were integrated with contrastive learning, a self-supervised method that ensures modified data retains high similarity to the original behaviors, forming a comprehensive model termed TiCoSeRec. The approach was evaluated across four large real-world benchmarks encompassing e-commerce and consumer reviews.

The study yielded several key findings. First, existing recommendation algorithms achieved 15% to over 30% higher ranking accuracy when trained on uniform interaction sequences compared to irregular sequences. Second, across four diverse public datasets, the proposed TiCoSeRec framework consistently outperformed nine baseline recommendation models, yielding relative performance gains ranging from 5% to 18%. Third, ablation experiments revealed that the time-aware substitution operator had the most critical positive impact on accuracy. Finally, parameter sensitivity analysis demonstrated optimal results when leaving the top 20% most uniform user sequences intact while applying time-aware augmentation to the remaining 80%.

These findings demonstrate that improving data quality prior to model training is more effective than solely relying on complex temporal architectures. Rather than feeding noisy, irregular raw interaction histories directly into models, standardizing sequence timing reduces noise and prevents model disorientation caused by user preference drift. For organizations managing digital marketplaces or content platforms, adopting time-aware sequence refinement provides a practical, high-impact mechanism to enhance recommendation relevance and conversion metrics.

Organizations seeking to enhance their recommendation pipelines should implement time-interval-aware augmentation across irregular user histories, specifically targeting the least uniform 70% to 80% of sequence data. In terms of implementation options, teams should prioritize time-guided item substitution, insertion, and masking, while applying length-dependent rules to prevent over-modifying very short user histories. Next steps should explore incorporating product categories and contextual attributes directly into the augmentation logic.

The findings are supported by consistent results across four large real-world datasets spanning multiple product and service domains. However, certain limitations remain: the framework relies on tuned thresholds to classify uniform sequences, and evaluation was centered on standard e-commerce rating datasets. Overall, decision-makers can have high confidence in applying time-interval-aware augmentation to sequential recommendation workflows.

Alibaba Group Northeastern University data augmentation personalized recommendation recommendation sequential recommendation time interval aware data augmentation data augmentation

arXiv: 2212.08262

Paper: Self-Attentive Sequential Recommendation, Wang-Cheng Kang et al. (2018). SASRec establishes the foundational self-attentive sequential recommendation architecture that contrastive and data-augmented sequential recommendation models build upon.
Paper: BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer, Fei Sun et al. (2019). BERT4Rec introduces bidirectional self-attention and sequence masking techniques to sequential recommendation that underpin modern sequence augmentation and transformation strategies.
Paper: Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding, Jiaxi Tang et al. (2018). This work establishes core principles of modeling sequential transition patterns and skip behaviors in user interaction histories for sequential recommendation.
Paper: Collaborative filtering with temporal dynamics, Yehuda Koren (2009). This seminal paper provides foundational insights into modeling temporal dynamics and user preference drift over time in recommender systems.
Paper: Session-based Recommendations with Recurrent Neural Networks, Balázs Hidasi et al. (2016). This paper establishes sequence-based neural modeling for user interaction sessions that serves as the benchmark paradigm for modern sequential recommenders.

Paper: Plug-In Diffusion Model for Sequential Recommendation, Haokai Ma et al. (2024). PDRec extends time-interval-aware sequential recommendation by employing a diffusion plugin framework to model dynamic user preferences across the entire item space.

Cover for Uniform Sequence Better: Time Interval Aware Data Augmentation for Sequential Recommendation

Abstract

Sequential recommendation is an important task to predict the next-item to access based on a sequence of interacted items. Most existing works learn user preference as the transition pattern from the previous item to the next one, ignoring the time interval between these two items. However, we observe that the time interval in a sequence may vary significantly different, and thus result in the ineffectiveness of user modeling due to the issue of preference drift. In fact, we conducted an empirical study to validate this observation, and found that a sequence with uniformly distributed time interval (denoted as uniform sequence) is more beneficial for performance improvement than that with greatly varying time interval. Therefore, we propose to augment sequence data from the perspective of time interval, which is not studied in the literature. Specifically, we design five operators (Ti-Crop, Ti-Reorder, Ti-Mask, Ti-Substitute, Ti-Insert) to transform the original non-uniform sequence to uniform sequence with the consideration of variance of time intervals. Then, we devise a control strategy to execute data augmentation on item sequences in different lengths. Finally, we implement these improvements on a state-of-the-art model CoSeRec and validate our approach on four real datasets. The experimental results show that our approach reaches significantly better performance than the other 9 competing methods. Our implementation is available: https://github.com/KingGugu/TiCoSeRec.

Topics

data augmentation personalized recommendation recommendation

Keywords

sequential recommendation time interval aware data augmentation data augmentation preference drift temporal patterns user-item interactions recurrent neural networks

Introduction
Related Work
Sequential Recommendation Models
Data Augmentation for Recommendation
Assumption Validation
Ours: Data Augmentation by Time Intervals
Experimentation
Conclusions and Future Work
Acknowledgements
References

Knowls

Knowl 1 — Empirical Superiority of Uniform Time-Interval Sequences in Sequential Recommendation
empirical result
Sequential recommendation models are typically trained on item interaction sequences ordered chronologically, $S_u = (v_1, \dots, v_j, \dots, v_N)$ , where user $u$ interacted with item $v_j$ at step $j$ . Associated with $S_u$ is a time-interval sequence $T_u = (t_1, \dots, t_{N-1})$ , where $t_j$ is the time elapsed between interaction $v_j$ and interaction $v_{j+1}$ . A sequence is defined as uniform if the standard deviation $\text{std}(T_u)$ of its time intervals is small, and non-uniform if $\text{std}(T_u)$ is large.

To evaluate the impact of time-interval uniformity, sequential interaction datasets are partitioned into uniform subsets (denoted $\mathcal{U}$ , containing sequences with smaller interval variance) and non-uniform subsets (denoted $\mathcal{N}$ , containing sequences with larger interval variance) using two distinct balance strategies:
- Strategy S (Equal sequence count): Subset $\mathcal{U}$ contains the top 50% sequences ranked by ascending interval standard deviation; subset $\mathcal{N}$ contains the remaining 50%.
- Strategy I (Equal interaction count): Sequences ranked by ascending interval standard deviation are added to $\mathcal{U}$ until 50% of all user-item interactions are included; the remainder forms $\mathcal{N}$ .
Comparing three sequential recommenders (SASRec, TiSASRec, and CoSeRec) trained separately on $\mathcal{U}$ , $\mathcal{N}$ , and a 50% randomly sampled baseline across four benchmark datasets (Amazon Beauty, Amazon Sports, Amazon Home, Yelp) at NDCG@10 and Hit@10 yields the following results:
1. In 54 out of 64 total comparison pairs, models trained on uniform sequences $\mathcal{U}$ achieve relative performance gains greater than 15% compared to models trained on non-uniform sequences $\mathcal{N}$ , with 33 cases exceeding 20% improvement and peak improvement reaching 32.41%.
2. The minimum observed improvement of $\mathcal{U}$ over $\mathcal{N}$ is 8.22%.
3. The random baseline performance consistently falls strictly between that of $\mathcal{U}$ and $\mathcal{N}$ .
This validates that sequences with uniform time intervals mitigate preference drift and provide substantially higher signal for learning user transition patterns than sequences with irregular, widely varying time intervals.
Knowl 2 — Time-Interval-Aware Insertion Operator (Ti-Insert)
model/method
The Ti-Insert (TI) data augmentation operator reduces the standard deviation of time intervals in a user interaction sequence $S_u = (v_1, \dots, v_N)$ with time intervals $T_u = (t_1, \dots, t_{N-1})$ .

Unlike traditional random insertion, which risks inserting items into already temporally dense regions and inflating interval variance, Ti-Insert targets the largest gaps in the sequence. For an insertion ratio $\beta \in [0, 1]$ , it computes $k = \lfloor \beta N \rfloor$ insertion positions corresponding to the largest time intervals:

$(p_1, p_2, \dots, p_k) = \text{top-}k\text{-indices}(\text{descend}(T_u))$

where $\text{descend}(T_u)$ sorts the time intervals in descending order. At each selected position $p_i$ between items $v_{p_i}$ and $v_{p_i+1}$ (which originally had interval $d$ ), a correlated candidate item $v_{t_i}$ (an item not previously interacted with by the user, selected via item-item similarity) is inserted at the midpoint timestamp. This replaces the single large interval $d$ with two smaller intervals of length $d/2$ , thereby directly decreasing the standard deviation of $T_u$ and transforming non-uniform sequences into more uniform sequences.
Knowl 3 — Time-Interval-Aware Cropping Operator (Ti-Crop)
model/method
The Ti-Crop (TC) data augmentation operator extracts a contiguous sub-sequence of length $c = \lfloor \eta N \rfloor$ from an original sequence $S_u = (v_1, \dots, v_N)$ , where $\eta \in [0, 1]$ is the crop ratio and $N$ is the sequence length.

Rather than selecting the start index uniformly at random, Ti-Crop chooses the contiguous sub-sequence of length $c$ that minimizes the standard deviation of time intervals:

$p^* = \arg\min_{p \in [1, N-c+1]} \text{std}(\text{sub-sequence}(S_u, c, p))$

where $\text{sub-sequence}(S_u, c, p) = (v_p, v_{p+1}, \dots, v_{p+c-1})$ . The resulting cropped sequence preserves the most temporally uniform segment of user interactions, discarding irregular peripheral interaction bursts or long-gap drifts.
Knowl 4 — Time-Interval-Aware Masking and Substitution Operators (Ti-Mask and Ti-Substitute)
model/method
Ti-Mask (TM) and Ti-Substitute (TS) are data augmentation operators designed to modify items located in temporally dense sub-sequences, preventing disruption to macro user preference shifts that occur over large time intervals.

Given a sequence $S_u = (v_1, \dots, v_N)$ and its time intervals $T_u = (t_1, \dots, t_{N-1})$ , both operators identify $h = \lfloor \mu N \rfloor$ target indices corresponding to the smallest time intervals, where $\mu \in [0, 1]$ is the modification ratio:

$(p_1, p_2, \dots, p_h) = \text{top-}h\text{-indices}(\text{ascend}(T_u))$

where $\text{ascend}(T_u)$ sorts intervals in ascending order.
- Ti-Mask (TM): Deletes items at the target positions $(p_1, \dots, p_h)$ . Deleting an item between two short intervals merges them into a moderate interval, reducing overall interval variance without widening the maximum time intervals.
- Ti-Substitute (TS): Replaces items at the target positions $(p_1, \dots, p_h)$ with contextually correlated fake items (selected via item similarity). Operating on items separated by small time intervals ensures that substitutions do not corrupt the critical preference drift boundaries that occur across large time gaps.
Knowl 5 — Time-Interval-Aware Reordering Operator (Ti-Reorder)
model/method
The Ti-Reorder (TR) operator generates augmented sequences by permuting the order of items within a selected contiguous sub-sequence of length $c = \lfloor \eta N \rfloor$ , where $\eta \in [0, 1]$ is the sub-sequence ratio.

To prevent the destruction of cross-category user preference transitions, TR does not select the sub-sequence randomly. Instead, it locates the contiguous sub-sequence with the minimal time interval standard deviation:

$p^* = \arg\min_{p \in [1, N-c+1]} \text{std}(\text{sub-sequence}(S_u, c, p))$

TR then randomly shuffles the items strictly within $(v_{p^*}, \dots, v_{p^*+c-1})$ while keeping items outside this sub-sequence in their original positions and order. Because interactions within this window occurred with relatively uniform and close timestamps, shuffling them minimizes semantic distortion of the user's overarching preference trajectory.
Knowl 6 — Length-Aware Augmentation Control and the TiCoSeRec Framework
model/method
TiCoSeRec integrates time-interval-aware augmentation operators into a contrastive sequential recommendation framework with a length-dependent operator application strategy.

Given a dataset of sequences, the top $\sigma$ proportion of sequences ranked with the lowest time-interval standard deviations are classified as uniform and left unaugmented. The remaining $(1 - \sigma)$ sequences undergo data augmentation conditioned on a sequence length threshold $K$ :

$S'_u = \begin{cases} \text{aug}(S_u, op), \quad op \in \{\text{TS}, \text{TI}, \text{TM}\}, & N \le K \\ \text{aug}(S_u, op), \quad op \in \{\text{TS}, \text{TI}, \text{TM}, \text{TC}, \text{TR}\}, & N > K \end{cases}$

Short sequences ( $N \le K$ ) are restricted from cropping (TC) and reordering (TR) because removing items or shuffling in short sequences severely damages sequential dependency.

TiCoSeRec adopts SASRec (a multi-layer Transformer encoder) as its backbone. For each sequence $S_u$ , augmented views $S'_u$ are generated, and a contrastive self-supervised loss maximizes the mutual information/agreement between the representation of the original sequence and its augmented counterparts. The model is jointly optimized using the contrastive objective alongside the next-item prediction cross-entropy loss.

Knowl 7 — Sequential Recommendation Performance Across Benchmark Datasets

data/table

The recommendation performance of TiCoSeRec was evaluated against nine baselines across four public datasets (Amazon Beauty, Sports, Home, and Yelp) under a full-item ranking, leave-one-out evaluation protocol using Hit Ratio@K (H@K) and Normalized Discounted Cumulative Gain@K (N@K) for $K \in \{10, 20\}$ .

Dataset	Metric	BPR	LightGCN	BERT4Rec	TiSASRec	SASRec	LightSANs	CL4SRec	CoSeRec	DuoRec	Ours (TiCoSeRec)
Beauty	H@10	0.0438	0.0540	0.0526	0.0552	0.0562	0.0567	0.0569	0.0675	0.0686	0.0737
	H@20	0.0626	0.0803	0.0825	0.0845	0.0817	0.0857	0.0885	0.1015	0.1022	0.1079
	N@10	0.0229	0.0292	0.0263	0.0294	0.0305	0.0290	0.0277	0.0381	0.0359	0.0421
	N@20	0.0276	0.0358	0.0338	0.0372	0.0369	0.0363	0.0356	0.0467	0.0470	0.0499
Sports	H@10	0.0291	0.0378	0.0332	0.0297	0.0401	0.0437	0.0414	0.0421	0.0392	0.0507
	H@20	0.0460	0.0578	0.0538	0.0469	0.0565	0.0668	0.0637	0.0629	0.0630	0.0767
	N@10	0.0151	0.0200	0.0164	0.0158	0.0181	0.0209	0.0215	0.0241	0.0195	0.0282
	N@20	0.0194	0.0251	0.0216	0.0201	0.0224	0.0245	0.0271	0.0292	0.0256	0.0345
Home	H@10	0.0064	0.0096	0.0165	0.0129	0.0191	0.0123	0.0184	0.0232	0.0251	0.0267
	H@20	0.0104	0.0155	0.0224	0.0200	0.0263	0.0187	0.0285	0.0336	0.0347	0.0391
	N@10	0.0031	0.0051	0.0089	0.0066	0.0121	0.0068	0.0092	0.0136	0.0124	0.0153
	N@20	0.0043	0.0065	0.0103	0.0085	0.0131	0.0085	0.0118	0.0165	0.0152	0.0184
Yelp	H@10	0.0478	0.0542	0.0508	0.0594	0.0572	0.0624	0.0581	0.0613	0.0627	0.0658
	H@20	0.0683	0.0760	0.0786	0.0941	0.0908	0.0963	0.0896	0.0965	0.0976	0.1032
	N@10	0.0302	0.0328	0.0279	0.0323	0.0311	0.0339	0.0312	0.0322	0.0345	0.0371
	N@20	0.0354	0.0383	0.0352	0.0417	0.0396	0.0426	0.0390	0.0425	0.0432	0.0459

TiCoSeRec consistently outperforms all non-sequential baselines (BPR, LightGCN), sequential baselines without augmentation (BERT4Rec, TiSASRec, SASRec, LightSANs), and sequential models with contrastive augmentation (CL4SRec, CoSeRec, DuoRec). Relative improvements over the second-best performing method range from 4.94% to 18.15% across all datasets and metrics.

Knowl 8 — Ablation and Parameter Sensitivity of Time-Interval Augmentation Operators
empirical result
Ablation studies on individual time-interval operators and sensitivity analyses on the uniform sequence threshold ratio $\sigma$ yield the following insights:
1. Operator Replacement Ablation: Replacing any single time-aware operator with its standard random counterpart (e.g., $\text{TI} \rightarrow \text{Insert}$ , $\text{TC} \rightarrow \text{Crop}$ , $\text{TM} \rightarrow \text{Mask}$ , $\text{TS} \rightarrow \text{Substitute}$ , $\text{TR} \rightarrow \text{Reorder}$ ) results in a decline in HR@10 across datasets. Among all single-operator ablations, replacing Ti-Substitute with random Substitute ( $\text{TS} \rightarrow \text{S}$ ) causes the largest performance drop, indicating that preserving critical preference transitions by substituting only temporally dense items is the most crucial operator. Replacing Ti-Reorder with random Reorder ( $\text{TR} \rightarrow \text{R}$ ) causes the smallest drop.
2. Sensitivity to Uniform Ratio $\sigma$ : Parameter $\sigma$ controls the proportion of sequences treated as uniform (unaugmented). Sweeping $\sigma \in [0.0, 0.5]$ reveals that setting $\sigma = 0.2$ or $0.3$ achieves the peak HR@10 improvement (up to 3--4% improvement over augmenting all sequences at $\sigma = 0$ ). Setting $\sigma > 0.4$ leads to sharp performance degradation because too many non-uniform sequences remain unregularized.

Coverage note — None was omitted. All key contributions—empirical validation of time-interval uniformity, the five time-aware augmentation operators, the length-aware augmentation policy, baseline comparisons, and ablation/sensitivity analyses—are fully represented.

References

1.Du, N.; Wang, Y.; He, N.; Sun, J.; and Song, L. 2015. Time-sensitive recommendation from recurrent user activities. NIPS, 28.
2.Fan, X.; Liu, Z.; Lian, J.; Zhao, W. X.; Xie, X.; and Wen, J.-R. 2021. Lighter and better: low-rank decomposed self-attention networks for next-item recommendation. In SIGIR, 1733–1737.
3.He, R.; and McAuley, J. 2016. Fusing similarity models with markov chains for sparse sequential recommendation. In ICDM, 191–200. IEEE.
4.He, X.; Deng, K.; Wang, X.; Li, Y.; Zhang, Y.; and Wang, M. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. In SIGIR, 639–648.
5.Hidasi, B.; Karatzoglou, A.; Baltrunas, L.; and Tikk, D. 2015. Session-based recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939.
6.Jiang, J.; Luo, Y.; Kim, J. B.; Zhang, K.; and Kim, S. 2021. Sequential Recommendation with Bidirectional Chronological Augmentation of Transformer. arXiv preprint arXiv:2112.06460.
7.Kang, W.-C.; and McAuley, J. 2018. Self-attentive sequential recommendation. In ICDM, 197–206. IEEE.
8.Kingma, D. P.; and Ba, J. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.
9.Krichene, W.; and Rendle, S. 2020. On sampled metrics for item recommendation. Communications of the ACM, 65(7): 75–83.
10.Kurashima, T.; Althoff, T.; and Leskovec, J. 2018. Modeling interdependent and periodic real-world action sequences. In WWW, 803–812.
11.Li, J.; Wang, Y.; and McAuley, J. 2020. Time interval aware self-attention for sequential recommendation. In WSDM, 322–330.
12.Li, Y.; Luo, Y.; Zhang, Z.; Sadiq, S.; and Cui, P. 2019. Context-aware attention-based data augmentation for POI recommendation. In ICDEW, 177–184. IEEE.
13.Liu, Q.; Wu, S.; Wang, D.; Li, Z.; and Wang, L. 2016. Context-aware sequential recommendation. In ICDM, 1053–1058. IEEE.
14.Liu, Z.; Chen, Y.; Li, J.; Yu, P. S.; McAuley, J.; and Xiong, C. 2021. Contrastive self-supervised sequential recommendation with robust augmentation. arXiv preprint arXiv:2108.06479.
15.Ma, J.; Zhou, C.; Yang, H.; Cui, P.; Wang, X.; and Zhu, W. 2020. Disentangled self-supervision in sequential recommenders. In KDD, 483–491.
16.McAuley, J.; Pandey, R.; and Leskovec, J. 2015. Inferring networks of substitutable and complementary products. In KDD, 785–794.
17.Qiu, R.; Huang, Z.; Yin, H.; and Wang, Z. 2022. Contrastive learning for representation degeneration problem in sequential recommendation. In WSDM, 813–823.
18.Qu, S.; Yuan, F.; Guo, G.; Zhang, L.; and Wei, W. 2022. CmnRec: Sequential Recommendations with Chunk-accelerated Memory Network. IEEE Transactions on Knowledge and Data Engineering.
19.Rendle, S.; Freudenthaler, C.; Gantner, Z.; and Schmidt-Thieme, L. 2012. BPR: Bayesian personalized ranking from implicit feedback. arXiv preprint arXiv:1205.2618.
20.Rendle, S.; Freudenthaler, C.; and Schmidt-Thieme, L. 2010. Factorizing personalized markov chains for next-basket recommendation. In WWW, 811–820.
21.Sun, F.; Liu, J.; Wu, J.; Pei, C.; Lin, X.; Ou, W.; and Jiang, P. 2019. BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer. In CIKM, 1441–1450.
22.Sun, Y.; Guo, G.; Chen, X.; Zhang, P.; and Wang, X. 2020. Exploiting review embedding and user attention for item recommendation. Knowledge and Information Systems, 62(8): 3015–3038.
23.Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, Ł.; and Polosukhin, I. 2017. Attention is all you need. NIPS, 30.
24.Wang, Z.; Zhang, J.; Xu, H.; Chen, X.; Zhang, Y.; Zhao, W. X.; and Wen, J.-R. 2021. Counterfactual data-augmented sequential recommendation. In SIGIR, 347–356.
25.Xie, X.; Sun, F.; Liu, Z.; Wu, S.; Gao, J.; Ding, B.; and Cui, B. 2020. Contrastive learning for sequential recommendation. arXiv preprint arXiv:2010.14395.
26.Xu, D.; Ruan, C.; Korpeoglu, E.; Kumar, S.; and Achan, K. 2019. Self-attention with functional time representation learning. NIPS, 32.
27.Yao, T.; Yi, X.; Cheng, D. Z.; Yu, F.; Chen, T.; Menon, A.; Hong, L.; Chi, E. H.; Tjoa, S.; Kang, J.; et al. 2021. Self-supervised learning for large-scale item recommendations. In CIKM, 4321–4330.
28.Ye, W.; Wang, S.; Chen, X.; Wang, X.; Qin, Z.; and Yin, D. 2020. Time matters: Sequential recommendation with complex temporal information. In SIGIR, 1459–1468.
29.Yuan, F.; He, X.; Jiang, H.; Guo, G.; Xiong, J.; Xu, Z.; and Xiong, Y. 2020. Future data helps training: Modeling future contexts for session-based recommendation. In WWW, 303–313.
30.Zhao, W. X.; Hou, Y.; Pan, X.; Yang, C.; Zhang, Z.; Lin, Z.; Zhang, J.; Bian, S.; Tang, J.; Sun, W.; et al. 2022. RecBole 2.0: Towards a More Up-to-Date Recommendation Library. arXiv preprint arXiv:2206.07351.
31.Zhao, W. X.; Mu, S.; Hou, Y.; Lin, Z.; Chen, Y.; Pan, X.; Li, K.; Lu, Y.; Wang, H.; Tian, C.; et al. 2021. Recbole: Towards a unified, comprehensive and efficient framework for recommendation algorithms. In CIKM, 4653–4664.
32.Zhou, K.; Wang, H.; Zhao, W. X.; Zhu, Y.; Wang, S.; Zhang, F.; Wang, Z.; and Wen, J.-R. 2020. S3-rec: Self-supervised learning for sequential recommendation with mutual information maximization. In CIKM, 1893–1902.

Citation

MLA

Dang, Y., et al. “Uniform Sequence Better: Time Interval Aware Data Augmentation for Sequential Recommendation”. arXiv, 2022, http://arxiv.org/abs/2212.08262v2.

APA

Dang, Y., Yang, E., Guo, G., Jiang, L., Wang, X., Xu, X., Sun, Q., & Liu, H. (2022). Uniform Sequence Better: Time Interval Aware Data Augmentation for Sequential Recommendation. arXiv. http://arxiv.org/abs/2212.08262v2

Chicago

Dang, Y., E. Yang, G. Guo, et al. 2022. “Uniform Sequence Better: Time Interval Aware Data Augmentation for Sequential Recommendation”. arXiv. http://arxiv.org/abs/2212.08262v2.

Harvard

Dang, Y. et al. (2022) “Uniform Sequence Better: Time Interval Aware Data Augmentation for Sequential Recommendation”, arXiv [Preprint]. Available at: http://arxiv.org/abs/2212.08262v2.

Vancouver

1. Dang Y, Yang E, Guo G, Jiang L, Wang X, Xu X, Sun Q, Liu H (2022) Uniform Sequence Better: Time Interval Aware Data Augmentation for Sequential Recommendation. arXiv

BibTeX

@article{dang2022uniform,
  title = {Uniform Sequence Better: Time Interval Aware Data Augmentation for Sequential Recommendation},
  author = {Dang, Yizhou and Yang, Enneng and Guo, Guibing and Jiang, Linying and Wang, Xingwei and Xu, Xiaoxiao and Sun, Qinghui and Liu, Hong},
  year = {2022},
  journal = {arXiv},
  url = {http://arxiv.org/abs/2212.08262v2},
  eprint = {2212.08262}
}

Metadata:arXiv

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

Uniform Sequence Better: Time Interval Aware Data Augmentation for Sequential Recommendation

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Empirical Superiority of Uniform Time-Interval Sequences in Sequential Recommendation

Knowl 2 — Time-Interval-Aware Insertion Operator (Ti-Insert)

Knowl 3 — Time-Interval-Aware Cropping Operator (Ti-Crop)

Knowl 4 — Time-Interval-Aware Masking and Substitution Operators (Ti-Mask and Ti-Substitute)

Knowl 5 — Time-Interval-Aware Reordering Operator (Ti-Reorder)

Knowl 6 — Length-Aware Augmentation Control and the TiCoSeRec Framework

Knowl 7 — Sequential Recommendation Performance Across Benchmark Datasets

Knowl 8 — Ablation and Parameter Sensitivity of Time-Interval Augmentation Operators

References

Citation

BibTeX

Access the Paper