Temporal Interest-Driven Multimodal Personalized Content Generation

Tian Miao

REVIEW 4 major objections 5 minor 18 references

Temporal Interest-Driven Multimodal Personalized Content Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read One fused interest vector replaces static preferences in content generation

desk verdict A 22-page architecture proposal with no experiments; the central claim of outperforming static-interest methods is unsupported. read the letter →

arxiv 2509.04330 v1 pith:ETP6T34V submitted 2025-09-04 cs.IR

Tian Miao This is my paper

classification cs.IR

keywords temporalinterestmodelingmultimodalfusionpersonalizedcontentgenerationdynamicuservariationalautoencoderlearnablemodalattentionevolutionrecommendation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

TIMGen is a proposed framework for personalized content generation that replaces the static user-preference vector with a time-evolving interest state. It models a user's history as a sequence of multimodal interactions—text, image, video, audio—each tagged with behavior and temporal features, and passes that sequence through a Transformer to obtain an interest representation at every time step. In parallel, a learnable attention mechanism weights the four modalities at each time step, producing a modal-level interest vector that is concatenated with the temporal one. The fused vector conditions a variational autoencoder that generates content, trained with both rating and category labels. The paper's claim is that jointly learning temporal dependencies and modal preferences in this way yields more adaptive and personalized generation than static approaches; the paper does not include experiments to test this.

What carries the argument

The central object is the fused interest vector z_t^final = [z_t; z_t^multi]: a concatenation of the Transformer's temporal interest state and a modality-weighted interest state. The modal weights alpha_m are produced by a learnable attention function over single-modality representations, which is the mechanism that lets the model adaptively decide how much text, image, video, or audio matters at each time. This vector is mapped through a VAE latent space with reparameterized sampling, and the decoder is trained with reconstruction, KL, rating-label, and category-label losses. The interest state at time t is what carries the argument: without it the generation step has no handle on dynamic p

What would settle it

Run TIMGen against a static-preference baseline and a fixed-modal-weight baseline on the same multimodal temporal dataset, measuring generation relevance and personalization; if the static baseline matches or beats TIMGen, the central claim fails. A simpler probe: check whether the learned modal weights change significantly for a user whose behavior shifts from text-heavy to video-heavy consumption; if the weights stay flat, the dynamic-adaptation claim collapses.

Watch

Extended reading notes

Core claim

On its own terms, TIMGen's discovery is that personalization for generation should be formulated as a state sequence, not a single vector: the user's interest at time t is produced by Transformer self-attention over all past interactions, and is fused with a parallel multimodal interest vector whose per-modality weights are learned rather than fixed. This fused representation is what conditions the generator. The intended consequence is that generated content can track long-term preferences while responding to recent, context-specific shifts, and can automatically emphasize the modality that best expresses the user's current interest—for instance, text for movie reviews versus images for e-c

Load-bearing premise

The input data is trustworthy: the hybrid dataset built from public sets and platform APIs, augmented with semi-supervised and feature-aligned labels, is complete and balanced enough for the four pretrained encoders to yield faithful interest states; the paper itself concedes the annotation is unbalanced and cannot fully capture all information and modalities.

Editorial extensions

If this is right

If correct, generation systems in e-commerce and advertising could respond to interest shifts within a session instead of relying on a fixed profile.
Long-term preferences remain in the picture through the Transformer's attention over history, while the latest state drives immediate output, covering both stability and recency.
Learnable modal weights let the same architecture serve domains where different modalities dominate (text in movie reviews, images in shopping) without task-specific tuning.
Dual supervision with rating and category labels pushes generated output to be both appealing and semantically on-target.
The architecture extends to additional modalities or label types without redesigning the fusion or generation layers.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A direct comparison of TIMGen against a static-vector baseline and a fixed-weight multimodal baseline on the same balanced dataset would be the decisive test; the paper's omission of experiments leaves this open.
The learnable modal weights could double as an interpretability tool: inspecting alpha_m over time may reveal which modality a user relies on at different life moments, an implication the paper does not explore.
The paper's own caveat that recent signals are diluted by long history suggests a hierarchical or gated recency mechanism as a natural next extension rather than a pure Transformer stack.
Because the dataset is acknowledged as unbalanced, a testable extension is to verify whether semi-supervised feature alignment actually preserves temporal ordering, since misaligned timestamps would break the time-indexed state.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Dear colleague,

The paper is a system description with zero evaluation. The abstract says TIMGen "overcomes" static-interest methods and jointly models temporal interest evolution, but nothing in the 22 pages demonstrates that. If you referee it, you're reading a survey of MISSRec and PMG plus a module-by-module catalog of standard components.

What's actually good: the paper identifies a real gap — personalized multimodal generation is moving from static preference vectors to dynamic, time-aware interest modeling, and the cited literature does flag temporal modeling as a bottleneck. The proposed architecture is a sensible synthesis: Transformer self-attention, attention-weighted modal fusion, VAE decoding, and pretrained encoders. As a conceptual sketch, it hangs together.

The problems are large. There is no evaluation: no dataset with results, no baseline comparison, no metric, no error bars, no code. The load-bearing claim that dynamic modeling beats static preferences is asserted, not shown. Section 6 itself concedes that recent interactions get diluted by long histories, that modal fusion can't capture high-order interactions, and that VAE is weaker than GAN/diffusion for visual quality. Those concessions undercut the abstract's advertising. Several equations are corrupted as printed (Sec 5.2.3, 5.3.4, 5.3.5), Fig 1 is referenced but missing, and in-text citations don't match the reference list. None of these are minor typos; they make the method non-reproducible.

The architecture's components are all imported from the cited literature. The dynamic fusion idea appears in MISSRec; the Transformer, VAE, and pretrained encoders are off-the-shelf. The only new elements are hand-specified weighting parameters for the score labels, which are not validated either.

I agree with the reader's take: without experiments, the paper cannot support its claims. It's a position statement plus a design sketch, closer to a blog post than a research paper.

Who would value this? Someone entering the subfield might use the literature review as a shortcut, and the limitations section is an honest list of open problems. But a serious referee would have nothing concrete to check.

Recommendation: desk-reject. If the authors want this considered, they need to run experiments, report numbers against static-interest baselines, fix the equations, and provide code and data. Until then, it doesn't advance the field.

Referee Report

4 major / 5 minor

Summary. The paper proposes TIMGen, a framework for temporal interest-driven multimodal personalized content generation. It combines a Transformer-based temporal interest modeling layer, a learnable multimodal attention fusion layer, and a VAE-based generation layer, supervised by both personalized scoring labels and category labels. The manuscript reviews two prior systems (MISSRec and PMG), describes a hybrid dataset construction approach using public datasets plus API/crawler collection, and presents the TIMGen architecture in detail. The central claim is that TIMGen overcomes the limitations of static-preference-based personalized content generation by capturing dynamic user interests and adaptively weighting modalities. The paper contains no experimental evaluation: no datasets, baselines, metrics, error bars, or code are reported.

Significance. If the claimed capabilities were demonstrated, TIMGen would be a relevant contribution to personalized multimodal generation, addressing a recognized gap in dynamic interest modeling and adaptive modal weighting. The architectural components—Transformer sequence modeling, attention-based fusion, and VAE generation—are standard but plausibly combinable. However, the paper's central claim that TIMGen 'overcomes' static-interest methods is entirely unsupported empirically. The manuscript itself concedes (Sec. 6) that recent interaction signals are diluted, high-order modal interactions are not captured, and VAE generation is inferior to GAN/diffusion approaches. As submitted, the work is a descriptive proposal with no validation, so its significance remains conditional on future experiments that are not present.

major comments (4)

[Abstract and Sec. 5-6] There is no experimental evaluation anywhere in the manuscript. No datasets are evaluated, no baselines are compared, no metrics are reported, and no code is provided. The abstract's assertion that TIMGen 'overcomes the shortcomings of personalized content recommendation methods based on static preferences' is therefore not a demonstrated result but an unverified claim. This is load-bearing: without quantitative evidence or even a qualitative comparison protocol, the central contribution cannot be assessed.
[Sec. 6, limitations (3)-(5)] The paper's own stated limitations directly undermine its headline contributions. Limitation (3) says recent interaction signals are easily diluted by long histories, which challenges the claim of modeling 'strong temporal dependencies' for immediate interests. Limitation (4) says the fusion mechanism cannot capture high-order modal interactions, narrowing the claimed 'multimodal interaction mechanism.' Limitation (5) says the VAE generator is inferior to GANs and diffusion for visual quality, which weakens the 'personalized content generation' promise. These limitations are not addressed by experiments, so the manuscript does not establish that the proposed design achieves its stated goals.
[Sec. 5.2.1, dataset construction] The dataset construction section is too underspecified to support the method. Public datasets (MovieLens-1M, Amazon Review, MM-Rec) are listed but no concrete integration protocol is given for combining them with IMDb/Twitter/YouTube API and crawler data. The phrase 'supplemented using feature alignment and semi-supervised methods' is vague: no alignment objective, no semi-supervised algorithm, and no data statistics (size, modality distribution, time span) are provided. The paper later admits the data annotation is unbalanced (Sec. 6), but no mitigation or discussion of how this affects interest modeling is offered.
[Sec. 5.2.3 and Sec. 5.3.5, equations] Several key equations are corrupted as printed, preventing a precise understanding of the method. In Sec. 5.2.3, the e-commerce score label equation is typeset with missing binary operators between terms: '() () () ()score t click t cart t purchase t comment ty I i Ii I i I iαβγ δ=⋅ +⋅ +⋅ +⋅' is not mathematically well-formed. In Sec. 5.3.5, the reparameterization equation and the VAE objective contain garbled symbols (e.g., '', ' Iµσ= + ' and '(| ) [log ( | )] ( ( | ) ( ))final t final VAE tqz p x KL q z pφ θφ= −    '), making the latent-sampling procedure and the final loss function ambiguous.

minor comments (5)

[Sec. 1.1] The Introduction contains near-duplicate paragraphs (pages 1-2), including repeated sentences about digital-age data and e-commerce/advertising examples. This should be consolidated.
[Sec. 1.2] Typo: 'the overall preferences of the whole world' should likely be 'the whole user' or 'the overall preferences of the user'; also 'interaction' is used inconsistently.
[Sec. 5.2.3] The notation for category labels is inconsistent: 'classe' is used without definition, and the embedding dimension 'cd' is not specified. Also, the score-label weights α, β, γ, δ and λ1-λ4 are stated as adjustable but no guidance or learning procedure is given.
[Sec. 5.3.5] The text says 'Compared to GAN-based methods, the VAE has a more ordered distribution in the latent space' but this is a broad assertion without citation or formal justification.
[General] The paper includes many Chinese punctuation marks (e.g., '（' and '）') and nonstandard spacing; a thorough copyedit is needed.

Circularity Check

1 steps flagged · score 4.0 of 10

The personalized-scoring supervision is defined from the same action indicators used as input features, making that loss self-referential; the main temporal/multimodal claim is unsupported but not circular.

self definitional [Sec 5.2.3 (score label definition) with Sec 5.3.2 (input behavior features) and Sec 5.3.6 (score loss)]
"It can be modeled through multi-dimensional behavioral signals, including click, add-to-cart, purchase, and comment. ... y_t^score = α·I_click(i_t)+β·I_cart(i_t)+γ·I_purchase(i_t)+δ·I_comment(i_t). ... User behavior characteristics mainly describe the user's actions in a certain context, such as clicks, browsing, purchases, likes, and comments. ... Interaction features e_action include five categories: Click, View, Purchase, Like, Comment. After one-hot encoding, it is mapped to a d_ad dimensional vector space through a trainable embedding matrix."

The score label y_t^score is not an independent observation: it is defined as a hand-weighted sum of the same binary action indicators (click/cart/purchase/comment) that Sec 5.3.2 one-hot encodes as input behavior features h_t. The output loss L_score = ||ŷ_score − y_score||^2 then trains the generator to match a function of the input action features using those same input action features as conditioning. Thus the 'personalized scoring' supervision reduces by construction to reproducing a transformation of the input, making the score-based personalization signal self-referential. This is a genuine circular step, although it is auxiliary to the paper's main architectural proposal rather than the whole claim.

full rationale

No self-citations are load-bearing, no uniqueness theorem is imported, and no fitted parameter is renamed as a prediction. The paper is a proposal/position piece with no experiments, baselines, or fitted results, so the usual fitted-value circularity does not arise. The one concrete circular step is in the supervision design: Sec 5.2.3 constructs the personalized scoring label y_t^score as a weighted sum of click/cart/purchase/comment indicators, while Sec 5.3.2 feeds those same interaction types (click/view/purchase/like/comment) as one-hot input features h_t; Sec 5.3.6 then uses L_score to enforce agreement with y_t^score. The target is therefore a deterministic function of the input features, so the model is being asked to predict (a function of) its own inputs. This makes the score-supervision part of the pipeline circular. However, the paper's headline claim—that temporal interest modeling and adaptive multimodal weights improve personalized generation over static methods—is not itself derived from this circular loss; it is simply unvalidated by any experiment. The manuscript's own Sec 6 limitations further concede that the data is unbalanced, recent signals are diluted, high-order modal interactions are not captured, and VAE generation is visually inferior, all of which are evidence problems rather than circularity. Overall, the circularity is real but partial: one auxiliary 'prediction' reduces by construction, while the central proposal remains independent (and unverified).

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The claim that TIMGen captures dynamic interests and outperforms static-interest generation rests entirely on assumed modeling premises: that Transformer attention over concatenated features yields interest states, that the hybrid scraped dataset supports four pretrained encoders, that a VAE latent space supports content generation, and that hand-built labels are valid supervision. No parameter is fitted because no experiment is run; the only explicit free parameters are the hand-set label weights in Sec 5.2.3. No new theoretical entity is introduced.

free parameters (2)

E-commerce score label weights {alpha, beta, gamma, delta} = not specified; 'can be set according to the business scenario'
Sec 5.2.3 defines y_score = alpha*I(click) + beta*I(cart) + gamma*I(purchase) + delta*I(comment); the weights are chosen by hand per scenario and are never fit or validated.
Video score label weights {lambda_1, lambda_2, lambda_3, lambda_4} = not specified; 'can be adjusted according to the application objectives'
Sec 5.2.3 defines the video-scenario score label as a weighted sum of watch ratio, like, comment, and share indicators; weights are hand-set.

assumptions (4)

domain assumption Transformer self-attention over the interaction history produces a valid interest state at each time step.
Sec 5.3.3 asserts z_t = Transformer(x_1,...,x_t) captures 'long-term relationships in the evolution process'; no evidence links attention outputs to actual user behavior.
domain assumption The concatenated vector of behavior, temporal, and modality features is a sufficient representation of each interaction.
Sec 5.3.2 defines x_t = h_t + t_t + x_t^MM by concatenation across one-hot actions, sine-coded time, and frozen encoder vectors, with no empirical alignment check.
domain assumption A VAE latent space can represent the interest-content correspondence such that sampling yields content consistent with user preferences.
Sec 5.3.5 uses a VAE with reparameterization; Sec 6 concedes VAE quality is below GANs and diffusion, weakening the generation premise.
domain assumption The rating and category labels are valid supervision for personalization.
Sec 5.2.3 builds y_score from the same action indicators used as input features in Sec 5.3.2, and y_class is a genre/style/category one-hot; label reliability is never validated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Temporal Interest-Driven Multimodal Personalized Content Generation." pith.science (2026). https://pith.science/paper/ETP6T34V

@misc{pith2026250904330,
  author       = {Pith},
  title        = {Pith review of: Temporal Interest-Driven Multimodal Personalized Content Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ETP6T34V}},
  note         = {Machine review of arXiv:2509.04330}
}

read the original abstract

With the dynamic evolution of user interests and the increasing multimodal demands in internet applications, personalized content generation strategies based on static interest preferences struggle to meet practical application requirements. The proposed TIMGen (Temporal Interest-driven Multimodal Generation) model addresses this challenge by modeling the long-term temporal evolution of users' interests and capturing dynamic interest representations with strong temporal dependencies. This model also supports the fusion of multimodal features, such as text, images, video, and audio, and delivers customized content based on multimodal preferences. TIMGen jointly learns temporal dependencies and modal preferences to obtain a unified interest representation, which it then generates to meet users' personalized content needs. TIMGen overcomes the shortcomings of personalized content recommendation methods based on static preferences, enabling flexible and dynamic modeling of users' multimodal interests, better understanding and capturing their interests and preferences. It can be extended to a variety of practical application scenarios, including e-commerce, advertising, online education, and precision medicine, providing insights for future research.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 17 canonical work pages

[1]

Multimodal Recommendation Systems in the LLM Era: A Survey of Feature Representation and Fusion Methods,

M. Varun and B. Das, "Multimodal Recommendation Systems in the LLM Era: A Survey of Feature Representation and Fusion Methods," in 2024 4th International Conference on Advanced Enterprise Information System (AEIS) , 2024: IEEE, pp. 89-95

work page 2024
[2]

Personalized Recommendation Systems Based on Multi-Source Heterogeneous Data,

P. Chen, "Personalized Recommendation Systems Based on Multi-Source Heterogeneous Data," in 2025 5th International Conference on Consumer Electronics and Computer Engineering (ICCECE) , 2025: IEEE, pp. 752-757

work page 2025
[3]

Modality-Aware Diffusion Augmentation with Consistent Subspace Disentanglement for Session-based Recommendation,

J. Su et al. , "Modality-Aware Diffusion Augmentation with Consistent Subspace Disentanglement for Session-based Recommendation," in Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2 , 2025, pp. 2668- 2679

work page 2025
[4]

Point-of-Interest Recommendations and Large Language Models: A Powerful Combination,

T. Wang and C. Wang, "Point-of-Interest Recommendations and Large Language Models: A Powerful Combination," IEEE Intelligent Systems, vol. 40, no. 3, pp. 5-9, 2025

work page 2025
[5]

A Framework for Adaptive Learning and Context- Aware Knowledge Enhancement: An Art Education System Based on DeepSeek,

Z. Hong, W. Bao, and Y. Qingqing, "A Framework for Adaptive Learning and Context- Aware Knowledge Enhancement: An Art Education System Based on DeepSeek," in 2025 4th International Symposium on Computer Applications and Information Technology (ISCAIT) , 2025: IEEE, pp. 97-101

work page 2025
[6]

Where to Go Next: Modeling Long- and Short-Term User Preferences for Point-of-Interest Recommendation,

K. Sun, T. Qian, T. Chen, Y. Liang, QVH Nguyen, and H. Yin, "Where to Go Next: Modeling Long- and Short-Term User Preferences for Point-of-Interest Recommendation," in National Conference on Artificial Intelligence , 2020

work page 2020
[7]

Personalized long-and short-term preference learning for next POI recommendation,

Y. Wu, K. Li, G. Zhao, and X. Qian, "Personalized long-and short-term preference learning for next POI recommendation," IEEE Transactions on Knowledge and Data Engineering, vol. 34, no. 4, pp. 1944-1957, 2020

work page 1944
[8]

M5: Multi-Modal Multi-Interest Multi-Scenario Matching for Over-the-Top Recommendation,

XGP Zhao, C. Xu, L. Chen., "M5: Multi-Modal Multi-Interest Multi-Scenario Matching for Over-the-Top Recommendation," Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2023

work page 2023

Show all 18 references

[9]

Multimodal recommender systems: A survey,

Q. Liu et al. , "Multimodal recommender systems: A survey," ACM Computing Surveys, vol. 57, no. 2, pp. 1-17, 2024

2024
[10]

Quantization as a Foundation for Deployable High Performance Diffusion Models within the Landscape of Large Scale Generative AI,

M. Sørensen et al. , "Quantization as a Foundation for Deployable High Performance Diffusion Models within the Landscape of Large Scale Generative AI," Authorea Preprints, 2025

2025
[11]

Leveraging multimodal large language model for multimodal sequential recommendation,

Z. Wang, B. Liu, W. Huang, T. Hao, H. Zhou, and Y. Guo, "Leveraging multimodal large language model for multimodal sequential recommendation," Scientific Reports, vol. 15, no. 1, p. 28960, 2025

2025
[12]

Multimodal fusion framework based on knowledge graph for personalized recommendation,

J. Wang et al. , "Multimodal fusion framework based on knowledge graph for personalized recommendation," Expert Systems with Applications, vol. 268, p. 126308, 2025

2025
[13]

Generate, Not Recommend

transfers modal knowledge from a multimodal pre-trained model and introduces external semantics, achieving interest-aware long-term sequence modeling based on multimodal signals. This solves the cold start problem while enhancing the model's ability to generalize multimodal re...
[14]

MISSRec: Pre-training and Transferring Multi-modal Interest-aware Sequence Representation for Recommendation,

J. Wang et al. , "MISSRec: Pre-training and Transferring Multi-modal Interest-aware Sequence Representation for Recommendation," presented at the Proceedings of the 31st ACM International Conference on Multimedia, 2023

2023
[15]

PMG: Personalized Multimodal Generation with Large Language Models,

X. Shen, R. Zhang, X. Zhao, J. Zhu, and X. Xiao, "PMG: Personalized Multimodal Generation with Large Language Models," presented at the Proceedings of the ACM Web Conference 2024, 2024

2024
[16]

Generate, Not Recommend: Personalized Multimodal Content Generation,

J. Liu, Z. Dou, N. Hu, and C. Xiong, "Generate, Not Recommend: Personalized Multimodal Content Generation," arXiv preprint arXiv:2506.01704, 2025

2025 arXiv
[17]

Towards unified multi-modal personalization: Large vision-language models for generative recommendation and beyond,

T. Wei et al. , "Towards unified multi-modal personalization: Large vision-language models for generative recommendation and beyond," arXiv preprint arXiv:2403.10667, 2024. Page 22 of 22

2024 arXiv
[18]

Personalized generation in large model era: A survey,

Y. Xu et al. , "Personalized generation in large model era: A survey," arXiv preprint arXiv:2503.02614, 2025

2025 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Multimodal Recommendation Systems in the LLM Era: A Survey of Feature Representation and Fusion Methods,

M. Varun and B. Das, "Multimodal Recommendation Systems in the LLM Era: A Survey of Feature Representation and Fusion Methods," in 2024 4th International Conference on Advanced Enterprise Information System (AEIS) , 2024: IEEE, pp. 89-95

work page 2024

[2] [2]

Personalized Recommendation Systems Based on Multi-Source Heterogeneous Data,

P. Chen, "Personalized Recommendation Systems Based on Multi-Source Heterogeneous Data," in 2025 5th International Conference on Consumer Electronics and Computer Engineering (ICCECE) , 2025: IEEE, pp. 752-757

work page 2025

[3] [3]

Modality-Aware Diffusion Augmentation with Consistent Subspace Disentanglement for Session-based Recommendation,

J. Su et al. , "Modality-Aware Diffusion Augmentation with Consistent Subspace Disentanglement for Session-based Recommendation," in Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2 , 2025, pp. 2668- 2679

work page 2025

[4] [4]

Point-of-Interest Recommendations and Large Language Models: A Powerful Combination,

T. Wang and C. Wang, "Point-of-Interest Recommendations and Large Language Models: A Powerful Combination," IEEE Intelligent Systems, vol. 40, no. 3, pp. 5-9, 2025

work page 2025

[5] [5]

A Framework for Adaptive Learning and Context- Aware Knowledge Enhancement: An Art Education System Based on DeepSeek,

Z. Hong, W. Bao, and Y. Qingqing, "A Framework for Adaptive Learning and Context- Aware Knowledge Enhancement: An Art Education System Based on DeepSeek," in 2025 4th International Symposium on Computer Applications and Information Technology (ISCAIT) , 2025: IEEE, pp. 97-101

work page 2025

[6] [6]

Where to Go Next: Modeling Long- and Short-Term User Preferences for Point-of-Interest Recommendation,

K. Sun, T. Qian, T. Chen, Y. Liang, QVH Nguyen, and H. Yin, "Where to Go Next: Modeling Long- and Short-Term User Preferences for Point-of-Interest Recommendation," in National Conference on Artificial Intelligence , 2020

work page 2020

[7] [7]

Personalized long-and short-term preference learning for next POI recommendation,

Y. Wu, K. Li, G. Zhao, and X. Qian, "Personalized long-and short-term preference learning for next POI recommendation," IEEE Transactions on Knowledge and Data Engineering, vol. 34, no. 4, pp. 1944-1957, 2020

work page 1944

[8] [8]

M5: Multi-Modal Multi-Interest Multi-Scenario Matching for Over-the-Top Recommendation,

XGP Zhao, C. Xu, L. Chen., "M5: Multi-Modal Multi-Interest Multi-Scenario Matching for Over-the-Top Recommendation," Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2023

work page 2023