Pith. sign in

REVIEW 4 major objections 5 minor 56 references

AutoCluster, AutoTopicModeling, AutoTrendAnalysis: A Complete AutoML Pipeline for Predicting Emerging Trends

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read An AutoML pipeline can automatically cluster, label, and forecast emerging topic trends from dated text, achieving a best RMSE of 7.099.

desk verdict A coherent end-to-end AutoML trend pipeline whose central accuracy claim rests on a single cherry-picked RMSE with no baseline and an unvalidated synthetic meta-learning knowledge base. read the letter →

arxiv 2607.22641 v1 pith:56FTYAPS submitted 2026-06-23 cs.DL cs.AIcs.LG

Ahmed Abolfadl , Marwa Mahmoud Abla , Mervat Abu-Elkheir , Maggie Mashaly This is my paper
classification cs.DLcs.AIcs.LG
keywords AutoMLtrendpredictiontopicmodelingmeta-learningtimeseriesforecastingclusteringtextminingsignalclassification
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 tries to show that trend prediction from text—usually a manual, expertise-heavy process—can be fully automated. It claims the complete pipeline, from preprocessing to forecast, produces accurate predictions, citing a best root-mean-square error of 7.099 on annual topic-frequency forecasts. The author's intended contribution is a democratized system: anyone with a text-plus-date dataset can obtain clusters, topic labels, and strong/weak/noise signals with minimal machine-learning knowledge.

What carries the argument

The mechanism that carries the argument is a meta-learning knowledge base built from 1,000 synthetic two-dimensional datasets, each containing 100 random points. Nine clustering algorithms are run on these, optimized for silhouette score, and at inference the three most similar entries are retrieved by cosine similarity and evaluated on the target data. The winning clusterer then feeds a successive-halving procedure that compares four topic models on coherence and four forecasters on RMSE, with a penalty-adjusted coherence metric to control topic-word-length inflation.

What would settle it

On the same NIPS data, compute the RMSE of a naive forecaster that predicts next year's topic count to equal the current year's count (or the historical mean). If the naive RMSE equals or beats 7.099, the central claim of high predictive accuracy is unsupported. Alternatively, run AutoCluster with random algorithm selection on the same datasets; if the silhouette scores match the meta-learning selection, the knowledge base is not contributing.

Watch

Extended reading notes

Core claim

The central claim is that a modular AutoML stack—embedding and dimension reduction, meta-learning-based clustering, per-cluster topic model selection, and time-series forecasting—can turn dated text into a ranked list of emerging topics. On the NIPS corpus the best trial reported an RMSE of 7.099, which the paper interprets as high predictive accuracy. The paper also finds that no single topic model wins everywhere: within one dataset different clusters favor different algorithms, supporting the need for per-cluster selection.

Load-bearing premise

The load-bearing premise is that clustering performance on 1,000 synthetic datasets, each a 100×2 matrix of random numeric values, transfers to real text-derived two-dimensional embeddings; the paper reports no test of this transfer, and if it fails, the clustering stage—and the topic models and forecasts built on it—are effectively arbitrary.

Editorial extensions

If this is right

  • The system reduces the user's job to supplying text and a date field; clustering, topic selection, hyperparameter tuning, and forecasting happen automatically.
  • Forecasts are converted into three signal categories, so a stakeholder can immediately see which topics are rising, which are early signals, and which are noise.
  • Because topic model choice is made per cluster, the pipeline adapts to heterogeneous sub-communities within one dataset.
  • The reported RMSE of 7.099 on annual NIPS forecasts suggests the approach can track topic prominence at least on research corpora.
  • Multi-horizon forecasting (weekly, monthly, yearly) shows that a topic's signal can change with the aggregation window, so time resolution matters.

Reading between the lines

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

  • The synthetic knowledge base may be the weakest link: if clustering behavior on random 2D point clouds doesn't transfer to real text embeddings, the cluster assignments—and every downstream topic and forecast—would be effectively arbitrary, so the reported RMSE could be contingent on an untested assumption.
  • A fair evaluation would compare RMSE 7.099 against a naive baseline (e.g., forecasting last year's count) and against a fixed default pipeline (e.g., one clustering algorithm, one topic model, one forecaster) to isolate the value added by automation.
  • The percentile-based signal thresholds (50th and 10th) are a design choice; a testable extension would calibrate them against external emerging-trend lists to check whether the strong/weak/noise labels correspond to real shifts.
  • The modular design invites ablation: replacing the meta-learning recommendation with random algorithm selection would directly test whether the knowledge base contributes predictive accuracy.
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 paper proposes an end-to-end AutoML pipeline for trend detection from timestamped text. The pipeline comprises preprocessing and SPECTER embedding, PCA/UMAP reduction, a meta-learning-based clustering selector (AutoCluster), successive-halving topic modeling (AutoTopicModeling) comparing LDA, NMF, LSA, and BERTopic, LLM-based topic labeling via the DeepSeek API, and a forecasting stage (AutoTrendAnalysis) that evaluates Prophet, ARIMA, STL, and LSTM and labels topics as strong/weak/noise based on forecast percentiles. Experiments are conducted on three paper corpora (arXiv CS, NIPS, Software Engineering). The headline result is a best-trial RMSE of 7.099 for the NIPS dataset, which the abstract interprets as indicating high predictive accuracy.

Significance. If the system performed as claimed, it would be a useful integration of AutoML, topic modeling, and forecasting, and the modular architecture is a reasonable starting point. The authors use appropriate off-the-shelf components and are transparent about general limitations in Section V, which is commendable. However, the evaluation does not currently substantiate the central accuracy claim: there is no baseline, no repeated-split or multi-run variability, no validation of the meta-learning knowledge base, and the signal classification is an operational relabeling of the forecast rather than an independently validated detection of emerging trends. At present the paper is a system description with an unverified headline number; substantial additional evidence is required before the contribution can be assessed.

major comments (4)
  1. [Abstract and §IV-F] The central claim—'the proposed system's best trial achieves a final RMSE of 7.099, indicating high predictive accuracy'—is not interpretable without a baseline or error bars. RMSE is scale-dependent; in the NIPS series (Table V), topic counts range up to 69, so RMSE ≈ 7 is roughly 10% of the scale and not obviously 'high.' The 'best trial' wording suggests selection across Optuna trials or models, but no mean, standard deviation, repeated splits, or random seeds are reported. The paper needs at least a persistence or historical-mean baseline and a distribution over multiple splits/trials to support the accuracy claim.
  2. [§III-D, Fig. 2] AutoCluster's meta-learning knowledge base is built from 1,000 synthetic datasets, each a 100×2 matrix 'populated with random numerical values,' and no experiment tests whether clustering performance on such random matrices transfers to real text-derived 2D UMAP embeddings. The observation that Agglomerative clustering is selected on all three real datasets (Table I) is consistent with the meta-learner being uninformative. Without an ablation—for example, comparing the meta-learner's choice against running all nine clustering algorithms on the real data, or evaluating the downstream topic and trend outputs—the clustering stage is not grounded. This is load-bearing because clustering is the foundation for the topic and trend analyses.
  3. [§III-F.2] In the successive-halving option, model selection is performed on the 'latest 25%' and then the 'latest 50%' of the full time series. If the final 20% is later used as the test set for RMSE, then the test segment has been used during model selection. The text is ambiguous ('retrained on the full dataset in an 80-20 train-test split'), but as written it risks test-set contamination. The manuscript should state explicitly that selection is performed only on a training prefix, and the reported RMSEs should be evaluated strictly on held-out data. Otherwise the reported numbers are not honest out-of-sample results.
  4. [§III-F.4] The strong/weak/noise classification uses thresholds on the system's own forecasted values: 'Values above the 50th percentile' are strong, 'between the 10th and 50th percentile with positive trends' are weak, and the rest are noise. This makes the signal taxonomy a relabeling of the forecast rather than an independently validated detection of emerging trends. The paper does not validate these labels against external evidence, future ground truth, or the BERTrend taxonomy cited in Section II. If this is intended as an operational definition, it should be stated as such and the 'emerging trends' language should be softened; otherwise a validation experiment is needed.
minor comments (5)
  1. [§III-D] The construction of the synthetic knowledge base is underspecified: no distribution family, parameter ranges, or rationale for the 100×2 size are given. This makes the transferability concern harder to assess and should be clarified.
  2. [Table III] The reported forecast dates are inconsistent: the '1 Week' row gives 06/02/2025 and the '1 Month' row gives 01/02/2025, which cannot both be the same forecast origin. The date convention (MM/DD/YYYY vs DD/MM/YYYY) should be unified and the alignment with weekly/monthly/yearly steps checked.
  3. [Table V] The table lists strong signals and noise signals, but no weak-signal category, while the text describes a three-way classification. The absence of weak signals should be explained or the table formatting clarified.
  4. [§III-E] The penalty-adjusted coherence formula is described only in words: '5% is deducted per word beyond 10 (up to 50%).' An explicit equation or a concrete example would remove ambiguity about how the penalty is computed and applied.
  5. [General] No code, data, or random-seed documentation is provided. Given the dependence on external APIs and stochastic optimization, sharing code, seeds, and API/model versions would materially improve reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

Emerging-trend classification is a thresholded relabeling of the model's own forecast; RMSE itself is genuine out-of-sample.

  1. self definitional [Section III-F, subsection 4 (Signal Classification)]
    "Strong Signals: Values above the 50th percentile, indicating likely sustained or emerging prominence. Weak Signals: Values between the 10th and 50th percentile with positive trends, representing potential early indicators. Noise: Values below the 10th percentile or non-increasing trends, suggesting limited future relevance."

    The categories 'strong', 'weak', and 'noise' are defined directly as thresholds (50th and 10th percentiles) and trend signs applied to the model's own forecasted topic frequencies. Therefore the advertised identification of 'emerging trends' is a quantile relabeling of the forecast output, not an independent detection validated against future reality. The abstract's claim that classification 'enabl[es] the identification of emerging trends' reduces by construction to sorting the forecast values. The forecast itself is independent, but the trend-identification result is definitionally tied to the forecast's distribution.

full rationale

The forecasting module (AutoTrendAnalysis) trains models on an 80/20 split and reports out-of-sample RMSE (e.g., 7.099 for NIPS), so the numerical prediction of topic counts is not circular. AutoCluster's meta-learning over synthetic 2D data and AutoTopicModeling's coherence-based selection are methodologically weak but not circular. The one self-definitional step is the signal classification: strong/weak/noise labels are defined by percentiles of the forecasted values themselves, making 'emerging trend detection' a relabeling of the forecast rather than an independent discovery. The RMSE accuracy claim lacks baselines and variance, but that is a correctness threat, not circularity. Overall, the core forecast is independent, while the headline trend-identification contribution partially reduces to a thresholding operation, warranting a score of 6.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claims rest on three unsupported pillars: transferability of the random synthetic knowledge base to semantic text embeddings, the exclusivity of coherence as the topic-quality signal, and the adequacy of a single 80/20 split for accuracy claims. None are validated, so the ledger captures those hand-set constants and assumptions.

free parameters (6)
  • Signal classification percentile thresholds = strong: >50th pct; weak: 10th–50th; noise: <10th
    Chosen by hand without sensitivity analysis; these thresholds define the strong/weak/noise taxonomy that constitutes the paper's 'emerging trend' output.
  • Coherence penalty rate = 5% per word beyond 10, cap 50%
    Ad hoc penalty to normalize coherence inflation from longer topic word lists; affects which topic models are selected in AutoTopicModeling.
  • Cluster count validity bounds = 3–15 for knowledge base; penalty outside 4–15 in final optimization
    Hand-set bounds used to filter valid clustering configurations and to penalize invalid cluster counts; no justification or ablation.
  • Successive halving data fractions = 25% then 50%
    Hand-chosen fractions for both topic-model and forecaster selection; no evidence that these approximate full-data rankings.
  • Train/test split ratio = 80/20
    Single split used for all RMSE reports; no cross-validation, no repeated seeds, which inflates variance and selection bias.
  • Meta-feature set = unspecified (six meta-features)
    The six meta-features are never named or defined; they are the input to the knowledge-base matching, so the reported clustering recommendation cannot be inspected or replicated.
assumptions (5)
  • ad hoc to paper Random 100×2 matrices are representative of real text-derived 2D embeddings for clustering algorithm selection
    Section III-D builds the meta-learning knowledge base from 1,000 synthetic random matrices; no evidence supports transfer to UMAP/SPECTER text embeddings.
  • domain assumption Coherence score is a valid proxy for topic quality and predictive usefulness
    Section III-E selects topic models by coherence only; no human evaluation or downstream task validation.
  • domain assumption SPECTER embeddings are appropriate for all three corpora
    Section III-C chooses SPECTER due to scientific nature; no comparison to other embeddings, and NIPS/SE corpora span different eras and styles.
  • domain assumption Topic frequency series are predictable from their own history by one of the four models
    Section III-F treats each topic count series independently and assumes no external covariates are needed.
  • domain assumption An 80/20 split gives a reliable estimate of next-step forecasting error
    Section III-F2 uses one 80/20 split for model selection and RMSE, yet reports one-step-ahead forecasts, creating a mismatch between evaluation and reported predictions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AutoCluster, AutoTopicModeling, AutoTrendAnalysis: A Complete AutoML Pipeline for Predicting Emerging Trends." pith.science (2026). https://pith.science/paper/56FTYAPS

@misc{pith2026260722641,
  author       = {Pith},
  title        = {Pith review of: AutoCluster, AutoTopicModeling, AutoTrendAnalysis: A Complete AutoML Pipeline for Predicting Emerging Trends},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/56FTYAPS}},
  note         = {Machine review of arXiv:2607.22641}
}
read the original abstract

Predicting emerging trends is vital for businesses, researchers, and policymakers; yet traditional approaches often lack scalability and adaptability. This paper presents a trend prediction framework based on Automated Machine Learning (AutoML), designed to extract insights from textual datasets with temporal attributes. The system ingests subject-specific textual entries accompanied by a date field. The pipeline begins with preprocessing and embedding, followed by AutoClustering, which uses meta-learning to select the optimal clustering algorithm. AutoTopicModeling then applies successive halving to identify the best topic modeling method: Latent Dirichlet Allocation (LDA), Latent Semantic Analysis (LSA), BERTopic, or Non-negative Matrix Factorization (NMF) based on the coherence score for each cluster. For trend forecasting, AutoTrendAnalysis evaluates multiple models: Facebook Prophet, AutoRegressive Integrated Moving Average (ARIMA), Seasonal-Trend decomposition using Loess (STL), and Long Short-Term Memory (LSTM) selecting the most accurate based on Root Mean Square Error (RMSE), either through successive halving or exhaustive comparison. Topics are classified as strong signals, weak signals, or noise based on forecasting outcomes, enabling the identification of emerging trends. By automating clustering, topic modeling, and time series forecasting, this research enhances trend prediction accuracy while reducing manual effort. The proposed system offers a scalable and user-friendly solution suitable for real-time applications and stakeholders with limited machine learning expertise. Experimental results demonstrate that the proposed system's best trial achieves a final RMSE of 7.099, indicating high predictive accuracy.

Figures

Figures reproduced from arXiv: 2607.22641 by the authors.

Figure 1
Figure 1. Methodology Diagram D. AutoCluster To generate high-quality clusters for topic modeling, our approach relies on 2D reduced embeddings and dataset meta￾features that are independent of the dataset’s dimensions. As shown in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. AutoTopicModeling Diagram The top model is then passed to Optuna for Bayesian hyper￾parameter optimization over 10 trials, maximizing coherence score. The final output includes the top 8 topic words per cluster, appended to the original dataset under a new topic column. To generate interpretable labels, we apply the DeepSeek API to each list of topic words, using the following prompt: "Analyze these topic words gene… view at source ↗
Figure 4
Figure 4. AutoTrendAnalysis Diagram dataset, reinforcing the importance of per-cluster model selec￾tion in topic modeling pipelines. Topic labels are displayed in Tables III, IV and V. D. Experiment 3: AutoTrendAnalysis Results In this experiment, we apply AutoTrendAnalysis to forecast the evolution of research topics in three datasets. The goal is to predict the number of documents published per topic in future time steps. T… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 2 canonical work pages

  1. [1]

    and Klein, A

    Feurer, M. and Klein, A. and Eggensperger, K. and Springenberg, J. T. and Blum, M. and Hutter, F. , title =. Proceedings of the 19th International Conference on Artificial Intelligence and Statistics (AISTATS) , year =

  2. [2]

    Query-Based Topic Modeling and Trend Analysis in Scientific Literature , booktitle =

    Tarek, Ahmed and Mahmoud, Marwa and Afifi, Basma and Mashaly, Maggie and Abuelkheir, Mervat , year =. Query-Based Topic Modeling and Trend Analysis in Scientific Literature , booktitle =

  3. [3]

    and Zhao, Z

    He, J. and Zhao, Z. and Li, C. and Zhao, Z. , title =. Knowledge and Information Systems , volume =

  4. [4]

    and Springenberg, J

    Domhan, T. and Springenberg, J. T. and Hutter, F. , title =. International Conference on Machine Learning (ICML) , year =

  5. [5]

    and Picault, J

    Boutaleb, A. and Picault, J. and Grosjean, G. , title =. arXiv preprint arXiv:2411.05930 , year =

    arXiv

  6. [6]

    and Wang, L

    Tang, J. and Wang, L. and Sun, J. , title =. IEEE Transactions on Knowledge and Data Engineering , volume =

  7. [7]

    and Zhang, X

    Wang, H. and Zhang, X. , title =. Knowledge and Information Systems , volume =

  8. [8]

    and Manca, G

    Rodrigues, M. and Manca, G. and Picault, J. , title =. Proceedings of the International Conference on Web Intelligence , year =

Show all 56 references
  1. [9]

    and Wunsch, D

    Xu, R. and Wunsch, D. , title =. IEEE Transactions on Neural Networks , volume =

  2. [10]

    and Granell, C

    Anglada, L. and Granell, C. and Garcia, R. , title =. Journal of Computational Science , volume =

  3. [11]

    and Canu, S

    Lacoste, A. and Canu, S. and Bedin, L. , title =. Journal of Machine Learning Research , volume =

  4. [12]

    Blei, D. M. and Ng, A. Y. and Lafferty, J. D. , title =. Journal of Machine Learning Research , volume =

  5. [13]

    and Casanova, M

    Wolf, T. and Casanova, M. , title =. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =

  6. [14]

    and Hsieh, J

    Yeh, C. and Hsieh, J. and Cheng, Y. , title =. Expert Systems with Applications , volume =

  7. [15]

    and Costa, E

    Silva, D. and Costa, E. , title =. Journal of Information Science , volume =

  8. [16]

    Box, G. E. P. and Jenkins, G. M. and Reinsel, G. C. , title =

  9. [17]

    and Schmidhuber, J

    Hochreiter, S. and Schmidhuber, J. , title =. Neural Computation , volume =

  10. [18]

    Taylor, S. J. and Letham, B. , title =. The American Statistician , volume =

  11. [19]

    Sezer, O. B. and Gudelek, M. U. and Ozbayoglu, A. M. , title =. Computers , volume =

  12. [20]

    Proceedings of the 23rd European Conference on Artificial Intelligence,

    Gijsbers, Sander and Leijnen, Stefan and Vanschoren, Joaquin , title =. Proceedings of the 23rd European Conference on Artificial Intelligence,. 2022 , doi =

    2022

  13. [21]

    IEEE Transactions on Knowledge and Data Engineering , year =

    Huang, Zixuan and Zhang, Jia and Li, Zongyuan and Chen, Huajun and Cheng, Xueqi , title =. IEEE Transactions on Knowledge and Data Engineering , year =. doi:10.1109/TKDE.2022.3228488 , url =

    2022

  14. [22]

    Sirajus Salekin and Akter, Nasrin , title =

    Karmaker, Gourab and Hossain, Md. Sirajus Salekin and Akter, Nasrin , title =. 2021 5th International Conference on Informatics and Computational Intelligence for Data Science (. 2021 , doi =

    2021

  15. [23]

    International Conference on Artificial Intelligence and Soft Computing , pages =

    Kaur, Harleen and Singh, Gurjot and Wozniak, Michal , title =. International Conference on Artificial Intelligence and Soft Computing , pages =. 2022 , series =. doi:10.1007/978-3-031-09027-7\_18 , url =

    2022 doi

  16. [24]

    Veerappa, Rakesh and van der Aa, Han A. H. , title =. Business Process Management Forum , pages =. 2021 , series =. doi:10.1007/978-3-030-85243-1\_20 , url =

    2021 doi

  17. [25]

    Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =

    Zhao, Junjie and Zhang, Chao and Lyu, Chenghao and Wang, Changdong , title =. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =. 2022 , publisher =. doi:10.1145/3534678.3539423 , url =

    2022

  18. [26]

    Proceedings of the 2022 International Conference on Management of Data , pages =

    Niemann, Lukas and Rausch, Tobias and Grabner, Helmut , title =. Proceedings of the 2022 International Conference on Management of Data , pages =. 2022 , publisher =. doi:10.1145/3514221.3517251 , url =

    2022

  19. [27]

    Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining , pages =

    Zhao, Junjie and Zhang, Chao and Wang, Changdong , title =. Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining , pages =. 2023 , publisher =. doi:10.1145/3539597.3570429 , url =

    2023

  20. [28]

    IEEE Transactions on Knowledge and Data Engineering , year =

    Zhou, Yue and Chen, Huajun and Zhang, Ning and Xie, Guotong and Cheng, Xueqi , title =. IEEE Transactions on Knowledge and Data Engineering , year =. doi:10.1109/TKDE.2021.3053144 , url =

    2021

  21. [29]

    Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =

    Wang, Jing and Zhang, Chao and Zhang, Junchi and Shen, Huawei and Cheng, Xueqi , title =. Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =. 2021 , publisher =. doi:10.1145/3447548.3467169 , url =

    2021

  22. [30]

    Thirty-sixth Conference on Neural Information Processing Systems , year =

    Bengio, Samuel and Jomaa, Nesrine and Caccia, Meher and Chartier, Samuel and Gregor, Karol and Chapados, Nicolas , title =. Thirty-sixth Conference on Neural Information Processing Systems , year =

  23. [31]

    , title =

    Liu, Jia and Shao, Yuan and Yu, Philip S. , title =. IEEE Transactions on Knowledge and Data Engineering , year =. doi:10.1109/TKDE.2019.2956774 , url =

    2019

  24. [32]

    Journal of Machine Learning Research , volume=

    AutoML: A Survey of the State-of-the-Art , author=. Journal of Machine Learning Research , volume=

  25. [33]

    IEEE Transactions on Artificial Intelligence , volume=

    AutoML to Date and Beyond: Challenges and Opportunities , author=. IEEE Transactions on Artificial Intelligence , volume=

  26. [34]

    Journal of Computational Science , volume=

    A Comparison of AutoML Tools for Machine Learning, Deep Learning, and XGBoost , author=. Journal of Computational Science , volume=

  27. [35]

    Pattern Recognition Letters , volume=

    AutoML: State of the Art with a Focus on Anomaly Detection, Challenges, and Research Directions , author=. Pattern Recognition Letters , volume=

  28. [36]

    IEEE Access , volume=

    AutoML: A Systematic Review on Automated Machine Learning with Neural Architecture Search , author=. IEEE Access , volume=

  29. [37]

    Journal of the American Society for Information Science , volume=

    Indexing by Latent Semantic Analysis , author=. Journal of the American Society for Information Science , volume=

  30. [38]

    Nature , volume=

    Learning the Parts of Objects by Non-negative Matrix Factorization , author=. Nature , volume=

  31. [39]

    arXiv preprint arXiv:2203.05794 , year=

    BERTopic: Neural Topic Modeling with Class-based TF-IDF , author=. arXiv preprint arXiv:2203.05794 , year=

    arXiv

  32. [40]

    Mining Text Data , author=

  33. [41]

    International Symposium on Intelligent Systems Technologies and Applications , pages=

    Full-Text or Abstract? Examining Topic Coherence Scores Using Latent Dirichlet Allocation , author=. International Symposium on Intelligent Systems Technologies and Applications , pages=. 2017 , publisher=

    2017

  34. [42]

    Communication Methods and Measures , volume=

    Quantitative Analysis of Large Amounts of Text Using Topic Modeling , author=. Communication Methods and Measures , volume=

  35. [43]

    Journal of Official Statistics , volume=

    STL: A Seasonal-Trend Decomposition Procedure Based on Loess , author=. Journal of Official Statistics , volume=

  36. [44]

    2018 , note=

    Forecasting: Principles and Practice , author=. 2018 , note=

    2018

  37. [45]

    Philosophical Transactions of the Royal Society A , volume=

    Time-Series Forecasting with Deep Learning: A Survey , author=. Philosophical Transactions of the Royal Society A , volume=

  38. [46]

    Advances in Neural Information Processing Systems , volume=

    Algorithms for Hyper-Parameter Optimization , author=. Advances in Neural Information Processing Systems , volume=

  39. [47]

    Advances in Neural Information Processing Systems , volume=

    Practical Bayesian Optimization of Machine Learning Algorithms , author=. Advances in Neural Information Processing Systems , volume=

  40. [48]

    Advances in Neural Information Processing Systems , volume=

    Efficient and Robust Automated Machine Learning , author=. Advances in Neural Information Processing Systems , volume=

  41. [49]

    Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=

    Optuna: A Next-generation Hyperparameter Optimization Framework , author=. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=. 2019 , organization=

    2019

  42. [50]

    International Conference on Learning Representations (ICLR) , year=

    Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization , author=. International Conference on Learning Representations (ICLR) , year=

  43. [51]

    Proceedings of the Tenth International Workshop on Multimedia Data Mining , pages=

    Emerging topic detection on twitter based on temporal and social terms evaluation , author=. Proceedings of the Tenth International Workshop on Multimedia Data Mining , pages=. 2010 , organization=

    2010

  44. [52]

    Nature Human Behaviour , volume=

    The growing alignment and changing semantics of American political discourse , author=. Nature Human Behaviour , volume=. 2021 , publisher=

    2021

  45. [53]

    IEEE Transactions on Knowledge and Data Engineering , year=

    Forecasting Scientific Trends Using Topic Embedding and Citation Networks , author=. IEEE Transactions on Knowledge and Data Engineering , year=

  46. [54]

    Proceedings of the International Conference on Weblogs and Social Media (ICWSM) , year=

    Event detection and tracking in social streams using community-based graphs , author=. Proceedings of the International Conference on Weblogs and Social Media (ICWSM) , year=

  47. [55]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    SepTrends: Learning from history to forecast the future in multivariate event sequences , author=. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    2020

  48. [56]

    ACM SIGKDD Explorations Newsletter , volume=

    Meta-learning for selecting a clustering algorithm using multiple data characterizations , author=. ACM SIGKDD Explorations Newsletter , volume=. 2011 , publisher=

    2011

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.