Pith. sign in

REVIEW 4 major objections 5 minor 5 cited by

Research on E-Commerce Long-Tail Product Recommendation Mechanism Based on Large-Scale Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that fusing LLM-generated product text embeddings with attention-weighted user behavior and collaborative and generative scores improves long-tail e-commerce recommendation, reporting gains of 12% recall, 9% hit rate…

desk verdict The paper's headline gains (+12%/+9%/+15%) never appear in the body, which reports only a +4.9% recall gain; the central empirical claim is unverifiable. read the letter →

arxiv 2506.06336 v1 pith:NOBUKI63 submitted 2025-05-31 cs.IR

Qingyi Lu , Haotian Lyu , Jiayun Zheng , Yang Wang , Li Zhang , Chengrui Zhou This is my paper
classification cs.IR
keywords large-scalelanguagemodellong-tailproductsrecommendationsystemsemanticrepresentationattentionmechanismscorefusione-commerce
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

The paper's proposal is that long-tail product recommendation can be improved by letting a large-scale language model read product text and by fusing three scoring signals: semantic similarity, collaborative filtering, and LLM-generated candidate probability. It presents a three-part mechanism: an LLM-based 'semantic visor' that embeds product titles, descriptions, and review summaries; an attention-based user intent encoder with a learnable time-decay weight; and a weighted fusion ranking model. On a private three-month e-commerce dataset with 50,000 users and 200,000 products, the method is reported to beat collaborative filtering, BPR-MF, shallow text embedding, and BERT4Rec, with Recall@1000 of 0.732, a 12% overall recall gain, 9% higher hit rate, and 15% higher user coverage. It also reports an average latency of 182 ms, which is under the 200 ms threshold for online e-commerce. If these results hold, the practical consequence is that text-aware LLM representations can substitute for missing interaction data in the long tail without sacrificing real-time service.

What carries the argument

The semantic visor is the pooling step that converts product text into one embedding per item, $e_i = \mathrm{Pool}(\mathrm{LLM}_\theta(\mathrm{text}_i))$, where pooling can be average or CLS-token pooling. The attention-based user intent encoder then maps a user's behavior sequence to item embeddings and applies self-attention with a learnable time-decay parameter to form the intent vector $h_u$. The hybrid ranking model combines three scores, $s_{\mathrm{sem}} = \cos(h_u, e_j)$, a collaborative filtering score $s_{\mathrm{cf}}$, and a generative LLM log-probability $s_{\mathrm{llm}}$, into $S(u,j) = \lambda_1 s_{\mathrm{sem}} + \lambda_2 s_{\mathrm{cf}} + \lambda_3 s_{\mathrm{llm}}$. This machinery is what lets otherwise interaction-sparse long-tail items be represented and ranked through text semantics rather than popularity.

What would settle it

Re-run the same five models on a publicly available e-commerce dataset using the same top-10% head split and check whether the LLM hybrid still beats the BERT sequence baseline at Recall@1000, hit rate, and tail coverage by the reported margins; additionally, re-running with head cutoffs of 5% and 20% would reveal whether the gains are tied to the arbitrary definition of long-tail.

Watch

Extended reading notes

Core claim

The central claim is that the sparsity and cold-start failures of traditional long-tail recommenders can be addressed by using an LLM as a semantic backbone rather than as a pure chatbot. The paper's model encodes each product as $e_i = \mathrm{Pool}(\mathrm{LLM}_\theta(\mathrm{text}_i))$, forms a user intent embedding $h_u$ by self-attention over the item embeddings in the user's behavior sequence with a learnable time-decay term, and computes a final score $S(u,j) = \lambda_1 \cos(h_u,e_j) + \lambda_2 \mathrm{CF}(u,j) + \lambda_3 \log P_{\mathrm{LLM}}(j \mid \mathrm{history}_u)$. It reports that on its dataset this fusion achieves 0.732 Recall@1000 (4.9% above BERT4Rec), NDCG@100 of 0.487, NDCG@10 of 0.312, and an average latency of 182 ms, and it interprets these numbers as evidence that LLM-based semantic understanding, collaborative signals, and generative candidate diversity are complementary for surfacing long-tail products.

Load-bearing premise

The reported gains all rest on defining 'long-tail' as the 90% of products outside the top 10% in sales on one private three-month dataset, and the paper does not test whether the improvements survive other head/tail cutoffs or other datasets.

Editorial extensions

If this is right

  • A platform adopting this architecture could keep its recommendation service responsive (under 200 ms) while exposing customers to products outside the top 10% of sales volume, because the LLM embedding step happens offline and the online stage is a lightweight fusion of precomputed scores.
  • Text-aware embeddings give cold-start items a representation before they accumulate user interactions, so the method is a candidate solution for new product launches and niche inventory.
  • Because the final score is a weighted sum, the system can degrade gracefully: if the LLM candidate generator misses a query, the collaborative and semantic signals still produce a recommendation, and vice versa.
  • The reported 15% improvement in user coverage implies that the method does not merely re-rank the same popular items for the same users, but systematically reaches users who saw fewer recommendations under head-focused baselines.

Reading between the lines

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

  • A direct test the paper leaves open is whether the gains come from the LLM embeddings or from the score-fusion layer; running the same fusion with static non-contextual text embeddings would isolate the contribution of the semantic visor.
  • The reported recall gain is at candidate size 1000, while NDCG is reported for 10 and 100; a reader should not assume recall improvements persist at the short list sizes a real storefront displays.
  • If the mechanism is as general as claimed, it should transfer to other text-rich domains such as job matching or scholarly article recommendation, where the same combination of semantic content and sparse interaction data appears.
  • The paper's own note that future work should tune fusion weights adaptively suggests the fixed initial weights are a known limitation, and category-specific weighting is a natural next experiment.
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.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes a hybrid long-tail product recommendation mechanism that combines a pre-trained LLM ('semantic visor') for product text embeddings, an attention-based user intent encoder, and a weighted fusion of semantic similarity, collaborative filtering, and LLM-generated candidate scores. Experiments are reported on a private e-commerce dataset described as 1 million interactions, 50,000 users, and 200,000 products. The abstract claims improvements of +12% recall, +9% hit rate, and +15% user coverage over baselines, but the body reports only Recall@1000 = 0.732, NDCG@100 = 0.487, NDCG@10 = 0.312, and an average latency of 182 ms; the headline improvements are not presented in any table or figure.

Significance. If validated, the architecture is a plausible, practically oriented combination of LLM semantic representations with behavior-sequence modeling and multi-source score fusion, and a sub-200 ms latency is in a deployable range. The paper's strength is the integration of several established ideas into one pipeline; the authors do not provide code, data, or machine-checked proofs. The central empirical claim, however, is not verifiable from the submitted manuscript: the figures containing the Recall and NDCG results are missing, and the only concrete number in the body (Recall@1000 = 0.732, '4.9% more accurate than BERT4Rec') is inconsistent with the abstract's '+12%' recall. The contribution therefore cannot currently be assessed on its merits.

major comments (4)
  1. [Abstract; §3.2] The abstract's headline numbers (+12% recall, +9% hit rate, +15% user coverage) are not reported anywhere in the body. The only concrete result in §3.2 is a recall of 0.732 at 1000, described as '4.9% more accurate than BERT4Rec'; that value is about 4.9% higher than a baseline near 0.698, not 12% higher. No hit-rate or user-coverage result appears in any table. Because the contribution is empirical, this mismatch makes the central claim unverifiable.
  2. [§3.2] The text states that Recall results appear in Figure 2 and NDCG results appear in Figure 3, and that latency appears in Table 1; however, Figures 1, 2, and 3 are referenced but not actually present in the manuscript, and only Table 1 (latency) is present. The section heading 'Model Performance Comparison and Ablation Analysis' promises an ablation, but no ablation results are reported anywhere. This is not a cosmetic issue: the evidence for the main quantitative claims is missing from the submitted version.
  3. [§3.1] The long-tail partition is defined as all products outside the top 10% by sales volume, with no justification or sensitivity analysis for this cutoff. All reported improvements in recall, coverage, and diversity are computed against this arbitrary partition of an unnamed private dataset, so the central empirical premise is untested. The authors should either justify the cutoff from domain considerations or report sensitivity to the threshold.
  4. [§3.2; Eq. (6)] No error bars, confidence intervals, or significance tests are reported for any metric, and the number of repeated runs/seeds is not stated. The claimed gains are small (4.9% recall, 6.5% NDCG) and could be within run-to-run variation. Additionally, the fusion weights λ1, λ2, λ3 in Eq. (6) are tuned on the validation set and the LLM is fine-tuned on the same platform's data, yet the individual contributions of the components and of the tuning procedure are not separated; the promised ablation is absent.
minor comments (5)
  1. [§2.1, Eq. (2)] The self-attention formula appears malformed (unbalanced parentheses and undefined symbols such as α&), and the 'learnable time decay parameter βt' is described in the text but its role in the equation is unclear; please rewrite the formula with all variables defined.
  2. [§2.1] The paper refers to 'multimodal textual content' (titles, descriptions, reviews); since all inputs are textual, this should be described as multi-field or multi-source text, not multimodal.
  3. [§3.1] The dataset is described only as being from 'a large e-commerce platform'; for reproducibility, please provide the domain, language, splitting method, and whether the 20% hold-out is temporal or random.
  4. [§4.2] Claims about auto-scaling, microservice deployment, and A/B testing are speculative and unsupported by any experiment; please clearly mark them as future work or remove them from the discussion.
  5. [References] The reference list contains many entries that appear unrelated to the paper's topic (e.g., [1], [24], [36], [39]-[43]); please replace these with relevant prior work on long-tail recommendation, LLM-based recommenders, and hybrid ranking.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the proposed fusion is not defined by the target metrics and evaluation uses a held-out set; the reported empirical gaps are verification issues, not circularity.

full rationale

The derivation chain is a standard ML pipeline: Eq. (1) pools LLM text embeddings, Eq. (2) computes an attention-based user intent vector, Eqs. (3)-(5) define semantic, collaborative-filtering, and LLM candidate scores, and Eq. (6) fuses them with weights tuned on a validation set. None of these quantities is defined in terms of Recall@K, Hit Rate@K, NDCG@K, diversity, or tail coverage, and the metrics are said to be evaluated on a 20% hold-out set, so the reported improvements are not forced by construction. The only clearly identifiable overlapping-author citation (ref. [21], by Yang, Lu, Wang, Zheng, and Xiang, cited for multi-instruction SFT data construction) is peripheral and not load-bearing. The paper does have a serious verification problem that is separate from circularity: Section 3.2 gives only one concrete number (Recall@1000 = 0.732, '4.9% more accurate than BERT4Rec'), while the abstract claims +12% recall, +9% hit rate, and +15% user coverage, and Figures 2-3 and any hit-rate/coverage table are absent from the submitted text. Those omissions prevent checking the empirical claim, but they do not make the derivation equivalent to its inputs.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central comparison depends on an arbitrary private dataset partition and on several unverified modeling assumptions about the complementarity of LLM embeddings, attention-based user encoding, and linear score fusion. The fusion weights are tuned on validation data, and the LLM is fine-tuned on the same platform's logs, so the reported gains are not independent of the fitting procedure. No new entities are introduced beyond named components of the architecture, which are re-labelings of standard building blocks.

free parameters (5)
  • fusion weights lambda_1, lambda_2, lambda_3 = initialized 0.4:0.4:0.2, then grid-searched; final values not reported
    Section 2.1 Formula 6 and Section 3.1: these weights are tuned on the validation set and directly determine the final recommendation score.
  • time decay parameter beta_t = not reported
    Section 2.1 Formula 2: a learnable parameter associated with each historical item, intended to reflect recency, but no values or analysis are given.
  • attention projection W and bias b = not reported
    Section 2.1 Formula 2: trainable linear transformation parameters in the attention mechanism that computes the user intent embedding.
  • LLM parameters theta = not reported
    Section 2.2: the LLM is fine-tuned with multi-instruction supervised fine-tuning and alignment; no model size, dataset size, or checkpoints are disclosed.
  • head/long-tail cutoff (top 10% sales volume) = 10%
    Section 3.1: an arbitrary partition that defines the long-tail set and all coverage, diversity, and recall claims depend on this cutoff.
assumptions (5)
  • domain assumption Product text embeddings produced by a pre-trained LLM are semantically meaningful enough to represent long-tail item relevance.
    Section 2.1 Formula 1 uses pooled LLM output as the item representation; no evidence is given that these embeddings are better than cheaper text models for long-tail products.
  • domain assumption A self-attention weighted sum of historical item embeddings captures a user's latent long-tail intent.
    Section 2.1 Formula 2; the attention parameters are fit to data and no analysis supports the long-tail-specific interpretation.
  • domain assumption A linear weighted sum of semantic, collaborative-filtering, and LLM-generation scores is sufficient to rank long-tail items.
    Section 2.1 Formulas 3 to 6; assumes the three signals are complementary and that fixed fusion weights generalize across items and users.
  • domain assumption Offline CTR and conversion metrics produce reliable partial-order labels for LLM ranking alignment.
    Section 2.2 Formula 7; the weak supervision labels are taken as ground truth with no validation of their stability or noise.
  • ad hoc to paper Products in the bottom 90% by sales volume form a valid long-tail set, and the 10% cutoff is not sensitive.
    Section 3.1 defines head versus long-tail by an arbitrary sales-volume cutoff; all long-tail metrics depend on this partition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Research on E-Commerce Long-Tail Product Recommendation Mechanism Based on Large-Scale Language Models." pith.science (2026). https://pith.science/paper/NOBUKI63

@misc{pith2026250606336,
  author       = {Pith},
  title        = {Pith review of: Research on E-Commerce Long-Tail Product Recommendation Mechanism Based on Large-Scale Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NOBUKI63}},
  note         = {Machine review of arXiv:2506.06336}
}
read the original abstract

As e-commerce platforms expand their product catalogs, accurately recommending long-tail items becomes increasingly important for enhancing both user experience and platform revenue. A key challenge is the long-tail problem, where extreme data sparsity and cold-start issues limit the performance of traditional recommendation methods. To address this, we propose a novel long-tail product recommendation mechanism that integrates product text descriptions and user behavior sequences using a large-scale language model (LLM). First, we introduce a semantic visor, which leverages a pre-trained LLM to convert multimodal textual content such as product titles, descriptions, and user reviews into meaningful embeddings. These embeddings help represent item-level semantics effectively. We then employ an attention-based user intent encoder that captures users' latent interests, especially toward long-tail items, by modeling collaborative behavior patterns. These components feed into a hybrid ranking model that fuses semantic similarity scores, collaborative filtering outputs, and LLM-generated recommendation candidates. Extensive experiments on a real-world e-commerce dataset show that our method outperforms baseline models in recall (+12%), hit rate (+9%), and user coverage (+15%). These improvements lead to better exposure and purchase rates for long-tail products. Our work highlights the potential of LLMs in interpreting product content and user intent, offering a promising direction for future e-commerce recommendation systems.

Figures

Figures reproduced from arXiv: 2506.06336 by the authors.

Figure 1
Figure 1. Training and Alignment Process of the Large-Scale Language Model for Recommendation Multi-Instruction Supervised Fine-Tuning: We use D to obtain the query and then, as a first step, apply rejection sampling to it in order to delete query-rewrite pairs that are considered incorrect and therefore result in the refined subset D7>&. After that, we introduce the model to task examples that are from other areas (Q&A, summ… view at source ↗
Figure 2
Figure 2. Recall Performance Comparison Across Models [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Ranking Quality (NDCG) Comparison Across Models [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Multimodal Foundation Model-Driven User Interest Modeling and Behavior Analysis on Short Video Platforms

    cs.IR 2025-09 reject novelty 3.0 of 10

    A standard attention-fusion plus Transformer sequence model is applied to short-video recommendation, with claimed gains over weak baselines and no reproducible artifacts.

  2. Meta-Learning for Cold-Start Personalization in Prompt-Tuned LLMs

    cs.LG 2025-07 reject novelty 3.0 of 10

    A meta-learned prompt-tuning method for cold-start LLM recommendations reports better Hit@10 and nDCG@10 on MovieLens-1M, but with no code, no error bars, and no shown results for Amazon or Recbole.

  3. Research on Low-Latency Inference and Training Efficiency Optimization for Graph Neural Network and Large Language Model-Based Recommendation Systems

    cs.LG 2025-06 reject novelty 3.0 of 10

    A hybrid GNN-LLM recommender with FPGA, DeepSpeed, and LoRA reportedly reaches NDCG@10 of 0.75 at 40-60ms latency while cutting training time by 66%, but the supporting artifacts are absent.

  4. LLM-Augmented Symptom Analysis for Cardiovascular Disease Risk Prediction: A Clinical NLP

    cs.CL 2025-07 reject novelty 2.0 of 10

    A small synthetic study reports that Bio_ClinicalBERT embeddings with Random Forest classify CVD risk in about 20 hand-written symptom texts, but the claims of MIMIC-III and CARDIO-NLP evaluation are unsupported.

  5. Research on Model Parallelism and Data Parallelism Optimization Methods in Large Language Model-Based Recommendation Systems

    cs.DC 2025-06 reject novelty 2.0 of 10

    A hybrid model-plus-data parallel scheme is reported to boost training throughput and GPU utilization for LLM-based recommenders, but the supporting experiments are not reproducible from the paper.

Reference graph

Works this paper leans on

39 extracted references · 18 canonical work pages · cited by 5 Pith papers

  1. [1]

    Automated Parking Trajectory Generation Using Deep Reinforcement Learning[J]

    Zhang Z, Luo Y, Chen Y, et al. Automated Parking Trajectory Generation Using Deep Reinforcement Learning[J]. arXiv preprint arXiv:2504.21071,

    arXiv

  2. [2]

    Rag-rlrc-laysum at biolaysumm: Integrating retrieval-augmented generation and readability control for layman summarization of biomedical texts[J]

    Ji Y, Li Z, Meng R, et al. Rag-rlrc-laysum at biolaysumm: Integrating retrieval-augmented generation and readability control for layman summarization of biomedical texts[J]. arXiv preprint arXiv:2405.13179,

    arXiv

  3. [3]

    An ensemble approach to stock price prediction using deep learning and time series models

    Sui, Mujie, et al. "An ensemble approach to stock price prediction using deep learning and time series models." (2024)

    work page 2024

  4. [4]

    Optimized path planning for logistics robots using ant colony algorithm under multiple constraints[J]

    Zhao H, Ma Z, Liu L, et al. Optimized path planning for logistics robots using ant colony algorithm under multiple constraints[J]. arXiv preprint arXiv:2504.05339,

    arXiv

  5. [5]

    Assertion detection in clinical natural language processing using large language models[C]//2024 IEEE 12th International Conference on Healthcare Informatics (ICHI)

    Ji Y, Yu Z, Wang Y. Assertion detection in clinical natural language processing using large language models[C]//2024 IEEE 12th International Conference on Healthcare Informatics (ICHI). IEEE, 2024: 242-247

    work page 2024

  6. [6]

    Dral: Deep reinforcement adaptive learning for multi-uavs navigation in unknown indoor environment[J]

    Mo K, Chu L, Zhang X, et al. Dral: Deep reinforcement adaptive learning for multi-uavs navigation in unknown indoor environment[J]. arXiv preprint arXiv:2409.03930,

    arXiv

  7. [7]

    Generative Prompt Controlled Diffusion for weakly supervised semantic segmentation[J]

    Wu W, Dai T, Chen Z, et al. Generative Prompt Controlled Diffusion for weakly supervised semantic segmentation[J]. Neurocomputing, 2025, 638: 130103

    work page 2025

  8. [8]

    text understanding and candidate generation

    π(y#∣x)>π(y$∣x)>⋯>π(y?∣x) (8) Through this step, not only in the case of the offline situation where the items that are better should be the higher ones but also during the process of actual online generation, the process is sufficiently reliable, which therefore may continuously complement the retrieval and hit rate for long-tail products [26-27]. This f...

    work page 2022

Show all 39 references
  1. [9]

    A Generative Adversarial Network-Based Investor Sentiment Indicator: Superior Predictability for the Stock Market[J]

    Qiu S, Wang Y, Ke Z, et al. A Generative Adversarial Network-Based Investor Sentiment Indicator: Superior Predictability for the Stock Market[J]. Mathematics, 2025, 13(9):

    2025

  2. [10]

    Contextual bandits for unbounded context distributions[J]

    Zhao P, Wu J, Liu Z, et al. Contextual bandits for unbounded context distributions[J]. arXiv preprint arXiv:2408.09655,

    arXiv

  3. [12]

    Integrated segmentation and subspace clustering for RSS-based localization under blind calibration[C]//GLOBECOM 2022-2022 IEEE Global Communications Conference

    Xing Z, Chen J, Tang Y. Integrated segmentation and subspace clustering for RSS-based localization under blind calibration[C]//GLOBECOM 2022-2022 IEEE Global Communications Conference. IEEE, 2022: 5360-5365

    2022

  4. [13]

    Prompt Categories Cluster for Weakly Supervised Semantic Segmentation[J]

    Wu W, Qiu X, Song S, et al. Prompt Categories Cluster for Weakly Supervised Semantic Segmentation[J]. arXiv preprint arXiv:2412.13823,

    arXiv

  5. [14]

    Score: Story coherence and retrieval enhancement for ai narratives[J]

    Yi Q, He Y, Wang J, et al. Score: Story coherence and retrieval enhancement for ai narratives[J]. arXiv preprint arXiv:2503.23512,

  6. [15]

    Xin, Y., Yang, J., Luo, S., Zhou, H., Du, J., Liu, X., ... & Du, Y. (2024). Parameter-efficient fine-tuning for pre-trained vision models: A survey. arXiv preprint arXiv:2402.02242

    2024

  7. [16]

    Xin, Y., Luo, S., Liu, X., Zhou, H., Cheng, X., Lee, C. E., ... & Liu, X. (2024). V-petl bench: A unified visual parameter-efficient transfer learning benchmark. Advances in Neural Information Processing Systems, 37, 80522-80535

    2024

  8. [17]

    (2025, May)

    Wang, Y., Zhao, T., & Wang, X. (2025, May). Fine-grained Heartbeat Waveform Monitoring with RFID: A Latent Diffusion Model. In Proceedings of the 3rd International Workshop on Human-Centered Sensing, Modeling, and Intelligent Systems (pp. 86-91)

    2025

  9. [18]

    Zhong, J., & Wang, Y. (2025). Enhancing Thyroid Disease Prediction Using Machine Learning: A Comparative Study of Ensemble Models and Class Balancing Techniques

    2025

  10. [19]

    Wang, Y., Zhong, J., & Kumar, R. (2025). A Systematic Review of Machine Learning Applications in Infectious Disease Prediction, Diagnosis, and Outbreak Forecasting

    2025

  11. [20]

    Wang, L., Cheng, Y., Xiang, A., Zhang, J., & Yang, H. (2024). Application of natural language processing in financial risk detection. arXiv preprint arXiv:2406.09765

    2024 arXiv

  12. [21]

    Yang, H., Lu, Q., Wang, Y., Liu, S., Zheng, J., & Xiang, A. (2025). User Behavior Analysis in Privacy Protection with Large Language Models: A Study on Privacy Preferences with Limited Data

    2025

  13. [22]

    Multidimensional precipitation index prediction based on CNN-LSTM hybrid framework[J]

    Wang Y, Jia P, Shu Z, et al. Multidimensional precipitation index prediction based on CNN-LSTM hybrid framework[J]. arXiv preprint arXiv:2504.20442,

    arXiv

  14. [23]

    Analysis of collective response reveals that covid-19-related activities start from the end of 2019 in mainland china[J]

    Liu J, Huang T, Xiong H, et al. Analysis of collective response reveals that covid-19-related activities start from the end of 2019 in mainland china[J]. medRxiv, 2020: 2020.10. 14.20202531

    2019

  15. [24]

    Research on improved u-net based remote sensing image segmentation algorithm

    Yang, Qiming, et al. "Research on improved u-net based remote sensing image segmentation algorithm." 2024 6th International Conference on Internet of Things, Automation and Artificial Intelligence (IoTAAI). IEEE,

    2024

  16. [25]

    Deep Learning for Weather Forecasting: A CNN-LSTM Hybrid Model for Predicting Historical Temperature Data

    Gong, Yuhao, et al. "Deep Learning for Weather Forecasting: A CNN-LSTM Hybrid Model for Predicting Historical Temperature Data." arXiv preprint arXiv:2410.14963 (2024)

    2024 arXiv

  17. [26]

    Accurate Prediction of Temperature Indicators in Eastern China Using a Multi-Scale CNN-LSTM-Attention model[J]

    Shen J, Wu W, Xu Q. Accurate Prediction of Temperature Indicators in Eastern China Using a Multi-Scale CNN-LSTM-Attention model[J]. arXiv preprint arXiv:2412.07997,

    arXiv

  18. [27]

    Deep learning-based anomaly detection and log analysis for computer networks[J]

    Wang S, Jiang R, Wang Z, et al. Deep learning-based anomaly detection and log analysis for computer networks[J]. arXiv preprint arXiv:2407.05639,

    arXiv

  19. [28]

    Gao Z, Tian Y, Lin S C, et al. A ct image classification network framework for lung tumors based on pre-trained mobilenetv2 model and transfer learning, and its application and market analysis in the medical field[J]. arXiv preprint arXiv:2501.04996,

    arXiv

  20. [30]

    Xiang, A., Zhang, J., Yang, Q., Wang, L., & Cheng, Y. (2024). Research on splicing image detection algorithms based on natural image statistical characteristics. arXiv preprint arXiv:2404.16296

    2024 arXiv

  21. [31]

    (2024, August)

    Xiang, A., Huang, B., Guo, X., Yang, H., & Zheng, T. (2024, August). A neural matrix decomposition recommender system model based on the multimodal large language model. In Proceedings of the 2024 7th International Conference on Machine Learning and Machine Intelligence (MLMI)...

    2024

  22. [32]

    Attention mechanism and context modeling system for text mining machine translation

    Bo, Shi, et al. "Attention mechanism and context modeling system for text mining machine translation." 2024 6th International Conference on Data-driven Optimization of Complex Systems (DOCS). IEEE,

    2024

  23. [33]

    Deep learning solutions for pneumonia detection: Performance comparison of custom and transfer learning models

    Zhong, Yihao, et al. "Deep learning solutions for pneumonia detection: Performance comparison of custom and transfer learning models." International Conference on Automation and Intelligent Technology (ICAIT 2024). Vol. 13401. SPIE, 2024.1

    2024

  24. [34]

    Applications of explainable ai in natural language processing

    Xu, Qiming, et al. "Applications of explainable ai in natural language processing." Global Academic Frontiers 2.3 (2024): 51-64

    2024

  25. [35]

    Exploiting Diffusion Prior for Out-of-Distribution Detection

    Zhu, Armando, et al. "Exploiting Diffusion Prior for Out-of-Distribution Detection." arXiv preprint arXiv:2406.11105 (2024)

    2024 arXiv

  26. [36]

    A Mixed-Heuristic Quantum-Inspired Simplified Swarm Optimization Algorithm for scheduling of real-time tasks in the multiprocessor system

    Su, Pei-Chiang, et al. "A Mixed-Heuristic Quantum-Inspired Simplified Swarm Optimization Algorithm for scheduling of real-time tasks in the multiprocessor system." Applied Soft Computing 131 (2022): 109807

    2022

  27. [39]

    Real-Time Prediction for Athletes' Psychological States Using BERT-XGBoost: Enhancing Human-Computer Interaction

    Duan, Chenming, et al. "Real-Time Prediction for Athletes' Psychological States Using BERT-XGBoost: Enhancing Human-Computer Interaction." arXiv preprint arXiv:2412.05816 (2024)

    2024 arXiv

  28. [40]

    Zhang T, Zhang B, Zhao F, et al. COVID-19 localization and recognition on chest radiographs based on Yolov5 and EfficientNet[C]//2022 7th International Conference on Intelligent Computing and Signal Processing (ICSP). IEEE, 2022: 1827-1830

    2022

  29. [41]

    Research on Effectiveness Evaluation and Optimization of Baseball Teaching Method Based on Machine Learning[J]

    Sun S, Yuan J, Yang Y. Research on Effectiveness Evaluation and Optimization of Baseball Teaching Method Based on Machine Learning[J]. arXiv preprint arXiv:2411.15721,

    arXiv

  30. [42]

    Avocado Price Prediction Using a Hybrid Deep Learning Model: TCN-MLP-Attention Architecture[J]

    Zhang L, Liang R. Avocado Price Prediction Using a Hybrid Deep Learning Model: TCN-MLP-Attention Architecture[J]. arXiv preprint arXiv:2505.09907,

    arXiv

  31. [43]

    CCi-YOLOv8n: Enhanced Fire Detection with CARAFE and Context-Guided Modules[J]

    Lv K. CCi-YOLOv8n: Enhanced Fire Detection with CARAFE and Context-Guided Modules[J]. arXiv preprint arXiv:2411.11011,

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.