Two-Stage Feature Generation with Transformer and Reinforcement Learning

Jun Gao; Kunpeng Liu; Wanfu Gao; Yuhao Tang; Zebin He; Zengyao Man

REVIEW 4 major objections 5 minor 33 references

Two-Stage Feature Generation with Transformer and Reinforcement Learning

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 TSFG: a two-stage framework where a Transformer encoder-decoder generates arithmetic features and PPO tunes them against the task's own metric, beating existing feature engineering baselines.

desk verdict A plausible two-stage feature generation idea that is undercut by an unspecified gradient path in pretraining and single-run experiments, but the combination is worth a serious referee. read the letter →

arxiv 2505.21978 v1 pith:BUW4XEGB submitted 2025-05-28 cs.LG

Wanfu Gao , Zengyao Man , Zebin He , Yuhao Tang , Jun Gao , Kunpeng Liu This is my paper

classification cs.LG

keywords automatedfeaturegenerationengineeringTransformerencoder-decoderProximalPolicyOptimizationreinforcementlearningfine-tuningtabulardatatransformationsequence

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 is trying to establish that automated feature generation for tabular data can be made reliable and adaptable by splitting it into two stages: a Transformer-based encoder-decoder first learns to propose arithmetic feature transformations, and then Proximal Policy Optimization fine-tunes those proposals against a downstream task's own performance metric. The benefit of this design, if it works, is that a user no longer needs domain expertise to construct new features; the framework can produce explicit new columns such as $V_1+V_2$ or $-V_3 \times V_1$ that improve a model's predictions. Across 13 classification and regression datasets the paper reports that the generated feature sets beat several existing automated feature engineering methods on accuracy, F1-score, precision, 1-RAE, and R2. This would turn feature engineering from a largely manual bottleneck into an automated, task-specific step.

What carries the argument

The central machinery is the tokenized feature-transformation language combined with the encoder-decoder architecture. Each token is a single operation-feature pair, such as +V1 or *V2, so a decoded segment between EOS tokens corresponds to one new arithmetic feature; this makes the search space discrete and the generated features directly interpretable. The Transformer's self-attention provides the encoder with a representation of inter-feature dependencies, while the decoder generates the transformation sequence autoregressively with temperature-scaled sampling for exploration. The same network then becomes the PPO policy in the fine-tuning stage, with the reward defined as the downstream metric difference between the augmented and original dataset. Pretraining gives a stable initialization, and PPO's clipped objective keeps the subsequent policy updates from destabilizing it.

What would settle it

Run TSFG's released code on any reported dataset and check how the augmented dataset is passed to the pre-trained MLP once the number of features changes; if the MLP's input dimension stays fixed, the described prediction loss cannot be computed on the generated features, and the pretraining loop is not implemented as written. A second check is to replace the PPO reward with random values while keeping pretraining identical; if downstream performance does not drop, the fine-tuning stage is not the source of the reported gains.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a single Transformer encoder-decoder can serve both as a feature-transformation generator and as the policy network for reinforcement learning, and that this duality is what lets the framework first learn a useful feature-generation strategy and then refine it for a specific task. The encoder compresses the raw feature set into a latent representation; the decoder emits a sequence of tokens, each binding an arithmetic operator to an original feature, with EOS marking the end of one new feature and STOP marking the end of the sequence. In the first stage the prediction loss of a pre-trained MLP on the augmented dataset is used to update the generator; in the second stage PPO uses the improvement in a downstream metric as the reward. The paper concludes that this staged training reduces feature redundancy and inefficient exploration, and that the resulting feature sets improve downstream model performance across different classifiers and regressors.

Load-bearing premise

The load-bearing assumption is that the pretraining step can transmit a useful learning signal through the whole pipeline: the prediction loss of a pre-trained MLP must be able to update the generator even though the generator outputs are discrete sampled tokens and the new arithmetic features are built by non-differentiable operators, and that same MLP must be able to score a dataset whose feature set has been augmented beyond its original input dimension. If either part of that assumption fails, the pretraining stage cannot work as described.

Editorial extensions

If this is right

Applied to a tabular dataset, TSFG outputs explicit new features like $V_1+V_2$ and $-V_3 \times V_1$, so a data scientist gets an enriched dataset whose new columns are interpretable arithmetic combinations rather than opaque embeddings.
Because the reward is a downstream performance metric, the same trained generator can be re-tuned for a new task without re-running an exhaustive feature search.
The paper reports competitive runtime against deep learning baselines, since the pretrained generator converges faster than reinforcement learning from scratch.
The framework is downstream-model-agnostic in the paper's experiments: gains appear with Random Forest, XGBoost, SVM, and CatBoost.
Automated feature generation of this kind removes the manual step of guessing which ratios, products, or differences of raw columns will help a model.

Reading between the lines

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

The same two-stage recipe likely transfers to non-tabular data or to feature selection by swapping the operation vocabulary and the reward function, since nothing in the design is specific to tables.
If the discrete-sampling gradient problem is real, the pretraining gain may actually come from the temperature-scaled exploration that widens the search, not from the prediction loss; an experiment that keeps exploration fixed while randomizing the pretraining target would separate these effects.
The fixed pre-trained MLP that scores the augmented dataset is a scalability bottleneck: once feature generation changes the feature count, the scoring head either must be re-trained or replaced by a metric that does not require a fixed input dimension.
The paper's 10 PPO iterations and small-to-medium UCI/OpenML datasets leave open whether the reward signal stays stable on larger benchmarks where evaluating each generated feature set is more expensive.

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

Here's my take on arXiv:2505.21978. The paper proposes TSFG, a two-stage feature generation method: first pretrain a Transformer encoder-decoder to output feature transformation sequences (like "+V1 +V2 EOS -V3 *V1 STOP"), then fine-tune that same model with PPO using the downstream metric as reward. The combination of supervised pretraining plus PPO fine-tuning is genuinely new in the feature-generation literature; GRFG and DIFER don't do this. The experiments cover 13 UCI/Kaggle/OpenML datasets and compare against OpenFE, DIFER, DFS, GRFG, with ablations and runtime. If the numbers are trustworthy, this would be a solid contribution to automated feature engineering.

But the method section has a gap that's hard to ignore. Section 3.3 says the decoder samples discrete transformation tokens, those operations build a new dataset, a pre-trained MLP scores it, and "the parameters of the encoder-decoder model are adjusted through backpropagation." No gradient estimator is described — no Gumbel-Softmax, straight-through, REINFORCE, or any relaxation. Eq. (4) defines a temperature-softmax distribution, but the paper never says whether sampling is replaced by a differentiable approximation. That's not a minor omission; the pretraining stage is exactly where the paper claims to gain stability and efficiency. The second issue is dimensional: the MLP is pretrained on the original feature set, but the augmented dataset has more columns. The paper never explains how the MLP accepts the new features. It would need to be retrained, expanded, or otherwise adapted; none of that is stated. Both gaps are fixable, and the linked code might clarify them, but the manuscript alone is not implementable from the text.

The experiments also need more care. Table 1 reports single values with no standard deviations, no number of runs, no significance tests, and no specification of the downstream model used for those headline numbers. That's a real weakness for a paper whose main claim is "significantly improves predictive performance." The ablation and runtime sections are useful but secondary.

On balance, I think the idea is worth engaging with. The two-stage recipe makes sense, the baselines are the right ones, and the gap in the method description is an omission rather than an internal contradiction. I'd send this to peer review — not desk reject — but with the clear expectation that the authors must specify the gradient estimator, explain the MLP adaptation, and provide multi-run statistics. The code could also be checked. If those pieces land, the paper would be a decent contribution to AutoML feature engineering.

Who's the audience? People working on automated feature generation and tabular AutoML. Not a field-changer, but a legitimate method paper if revised. I'd probably not cite it until the method is pinned down, but I'd put it on the reading list for my group's AutoML discussion.

Referee Report

4 major / 5 minor

Summary. The paper proposes TSFG, a two-stage automated feature generation framework for tabular data. In the first stage, a Transformer-based encoder-decoder is pre-trained to emit sequences of feature-transformation tokens (e.g., '+V1,+V2,EOS,-V3,*V1,STOP'); the resulting new dataset is evaluated by a pre-trained MLP, and the MLP prediction loss is backpropagated into the encoder-decoder. In the second stage, the same encoder-decoder is treated as a policy network and fine-tuned with PPO, with the reward defined as the downstream metric improvement. Experiments compare TSFG with five baselines on 13 classification and regression datasets, and include ablation, runtime, downstream-model, and feature-importance analyses.

Significance. If the proposed method worked as described, it would be a relevant contribution to automated feature engineering, combining a sequence-generation pretraining stage with RL-based fine-tuning and offering a concrete efficiency argument. The paper provides a code link and evaluates on 13 datasets, which is a useful empirical starting point. However, the manuscript does not specify the gradient path through discrete transformation tokens, does not explain how the fixed-input MLP evaluates augmented feature sets, and reports experimental results without variance or significance information. These gaps directly affect the central claim that TSFG 'outperforms existing state-of-the-art methods,' so the paper's significance cannot be assessed from the manuscript as written.

major comments (4)

[§3.3 and Algorithm 1 (lines 5–7)] The pretraining loss in Eq. (5) is computed from predictions of a pre-trained MLP on data produced by discretely sampled transformation tokens, but the manuscript never states how gradients of L_classification or L_regression pass through the discrete token selection or through the arithmetic feature-building operations. Temperature scaling in Eq. (4) only reshapes a softmax distribution over actions; it is not a differentiable relaxation of sampling. Algorithm 1 lines 5–7 simply call for backpropagation without specifying a Gumbel-Softmax, straight-through estimator, REINFORCE-type estimator, or any other mechanism. This is load-bearing because the pretraining stage is credited with the stability and efficiency that distinguish TSFG from GRFG and DIFER; without a concrete gradient estimator, the stage is not implementable as described.
[§3.3 and Algorithm 1 (line 12)] The manuscript says the generated transformation operations are applied to Dtrain to create Dnew_train, and that 'the dataset Dnew_train is then fed into a pre-trained MLP.' The pre-trained MLP's input dimension is fixed by the original feature set, while the generated sequence such as '+V1,+V2,EOS,-V3,*V1,STOP' produces two new features and thereby changes the number of columns. No input-size adaptation, retraining, or feature-subset mechanism is described for the MLP. As written, the evaluation of the generated features by this MLP is undefined, so the pretraining objective cannot be computed.
[§4.2, Table 1] Table 1 reports only single point estimates for each method and metric, with no standard deviations, no number of runs, and no significance tests. Many of the reported improvements are small (e.g., australian F1-score 0.8832 for TSFG versus 0.8757 for GRFG), and without repeated trials or error bars the abstract's claim that TSFG 'significantly improves' predictive performance and 'outperforms existing state-of-the-art methods' is not statistically substantiated. The authors should report mean and variance over multiple runs and, ideally, pairwise significance tests or at least confidence intervals.
[§4.3, Figure 3] The ablation study is the only direct evidence that pretraining and PPO fine-tuning each contribute to the final result, but Figure 3 is presented as unlabeled curves for four datasets with no numeric values, no axis labels, and no error bars. The claim that pre-training and PPO fine-tuning 'contribute significantly' cannot be checked from the figure. At minimum, the authors should provide a table with final metric values for 'Base,' 'TSFG+,' 'TSFG#,', and 'TSFG' on the ablated datasets.

minor comments (5)

[Eq. (4)] The notation P(a,z) = exp(a/T) / sum_i exp(ai/T) uses the action a itself in the exponential, but in a softmax the exponential should be taken over logits or scores associated with actions. Please clarify whether a denotes a score and how the denominator is indexed.
[Eq. (5)] The displayed cross-entropy loss is the binary form, while several classification datasets in Table 2 have more than two classes. If the framework is intended for multiclass tasks, a multiclass softmax cross-entropy formula should be used or the binary form should be explicitly restricted to binary datasets.
[Table 1] In the f5 row, the OpenFE precision value is listed as 0.888 while all other OpenFE values for that dataset are between 0.77 and 0.81; this appears to be a typographical error and should be checked against the original results.
[Figures 1 and 4] Figure 1 contains the misspelling 'Obecit' in the dataset illustration, and Figure 4's caption says 'hepapits' instead of 'hepatitis.' Please correct these typos.
[§4.4, Table 3] The runtime table reports times in seconds but does not state the hardware, software versions, or whether all methods were run under the same environment; adding this information would make the efficiency comparison reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the framework optimizes the same metric it reports, but this is a standard training objective with external test-set evaluation, not a definitional reduction.

full rationale

The paper's derivation chain is a constructive ML pipeline, not a logical proof that reduces to its inputs. Eq. (7) defines the PPO reward as r = Metric(Dnew) - Metric(D), and Table 1 reports F1/accuracy/1-RAE/R2 improvements; this is an optimization objective, not a fitted parameter renamed as a prediction. The pretraining losses (Eqs. 5-6) are supervised training objectives evaluated by a pre-trained MLP; they do not presuppose the final performance claim. The only author-overlapping citations ([Liu et al., 2019], [Wang et al., 2022], [Zhang and Gao, 2021], [Zhang et al., 2021], [Li et al., 2024]) appear in related-work or baseline descriptions, and none is invoked as a load-bearing justification, uniqueness theorem, or ansatz source. The identified implementability gaps (backpropagation through discrete tokens, fixed-input MLP with augmented features) are correctness/rigor concerns, not circularity, because no equation reduces to another by construction. Thus the central empirical claim is externally checkable against the reported benchmarks.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on two unstated implementation assumptions: differentiability through discrete feature generation, and MLP compatibility with changed input dimension. These are load-bearing because the pretraining stage, which the paper credits for stability and speed, cannot be executed as described without them.

free parameters (5)

Maximum transformation sequence length = 100
Chosen by hand; caps the number and order of generated features and therefore the search space.
PPO iterations = 10
Chosen by hand; controls how much reward-based fine-tuning occurs, influencing final performance.
Temperature range T = 0.1 to 1
Chosen by hand; controls exploration and exploitation balance in pretraining sampling.
Entropy coefficient = 1e-4
Chosen by hand; PPO exploration regularizer.
Transformer dimensions = embedding 128, hidden 128, 8 heads
Chosen by hand; sets model capacity.

assumptions (3)

domain assumption The dataset can be tokenized as a sequence of discrete feature tokens with operations, preserving sufficient information for the Transformer encoder.
Section 3.2 and Eq. (2)-(3) assume the raw features can be embedded as tokens; no preprocessing details for continuous values are given.
ad hoc to paper Gradients can be backpropagated from the downstream MLP loss through discretely sampled feature transformations to the encoder-decoder parameters.
Section 3.3 (Eq. 5-6) says parameters are updated by backpropagation, but the sampling step is discrete; no Gumbel-softmax, REINFORCE, or other estimator is described.
ad hoc to paper The pre-trained MLP can accept datasets with augmented feature sets.
The paper does not state how the MLP input dimension is adjusted when new features are appended or substituted, as required in Section 3.3.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Two-Stage Feature Generation with Transformer and Reinforcement Learning." pith.science (2026). https://pith.science/paper/BUW4XEGB

@misc{pith2026250521978,
  author       = {Pith},
  title        = {Pith review of: Two-Stage Feature Generation with Transformer and Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BUW4XEGB}},
  note         = {Machine review of arXiv:2505.21978}
}

read the original abstract

Feature generation is a critical step in machine learning, aiming to enhance model performance by capturing complex relationships within the data and generating meaningful new features. Traditional feature generation methods heavily rely on domain expertise and manual intervention, making the process labor-intensive and challenging to adapt to different scenarios. Although automated feature generation techniques address these issues to some extent, they often face challenges such as feature redundancy, inefficiency in feature space exploration, and limited adaptability to diverse datasets and tasks. To address these problems, we propose a Two-Stage Feature Generation (TSFG) framework, which integrates a Transformer-based encoder-decoder architecture with Proximal Policy Optimization (PPO). The encoder-decoder model in TSFG leverages the Transformer's self-attention mechanism to efficiently represent and transform features, capturing complex dependencies within the data. PPO further enhances TSFG by dynamically adjusting the feature generation strategy based on task-specific feedback, optimizing the process for improved performance and adaptability. TSFG dynamically generates high-quality feature sets, significantly improving the predictive performance of machine learning models. Experimental results demonstrate that TSFG outperforms existing state-of-the-art methods in terms of feature quality and adaptability.

Figures

Figures reproduced from arXiv: 2505.21978 by the authors.

**Figure 2.** Overview of TSFG. In the first phase, the encoder-decoder model is optimized by calculating the loss. In the second phase, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Results of ablation studies on different datasets. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Feature importance analysis on different datasets. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 26 canonical work pages

[1]

Neural machine translation by jointly learning to align and translate

[Bahdanau, 2014] Dzmitry Bahdanau. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473,

arXiv 2014
[8]

On calibration of modern neural net- works

[Guo et al., 2017] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural net- works. In International conference on machine learning , pages 1321–1330. PMLR,

work page 2017
[15]

Kingma and Jimmy Ba

[Kingma and Ba, 2015] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, 3rd International Con- ference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Pro- ceedings,

work page 2015
[16]

Hyperopt-sklearn: Automatic hyperpa- rameter configuration for scikit-learn

[Komer et al., 2014] Brent Komer, James Bergstra, and Chris Eliasmith. Hyperopt-sklearn: Automatic hyperpa- rameter configuration for scikit-learn. In Scipy, pages 32– 37,

work page 2014
[17]

Deep learning

[LeCun et al., 2015] Yann LeCun, Yoshua Bengio, and Ge- offrey Hinton. Deep learning. nature, 521(7553):436–444,

work page 2015
[18]

Multi-label feature selection with high-sparse personalized and low-redundancy shared common features.Information Processing & Management, 61(3):103633,

[Li et al., 2024] Yonghao Li, Liang Hu, and Wanfu Gao. Multi-label feature selection with high-sparse personalized and low-redundancy shared common features.Information Processing & Management, 61(3):103633,

2024
[19]

Libsvm dataset download

[Lin, 2024] Chih-Jen Lin. Libsvm dataset download. https: //www.csie.ntu.edu.tw/∼cjlin/libsvmtools/datasets/,

work page 2024
[20]

[Liu et al., 2019] Kunpeng Liu, Yanjie Fu, Pengfei Wang, Le Wu, Rui Bo, and Xiaolin Li

Accessed: 2024-05-01. [Liu et al., 2019] Kunpeng Liu, Yanjie Fu, Pengfei Wang, Le Wu, Rui Bo, and Xiaolin Li. Automating feature sub- space exploration via multi-agent reinforcement learning. In Ankur Teredesai, Vipin Kumar, Ying Li, R ´omer Ros- ales, Evimaria Terzi, and George Karypis, editors, Pro- ceedings of the 25th ACM SIGKDD International Confer- ...

work page 2024

Show all 33 references

[22]

Body mass index (bmi)

[Obese, 1998] HJOR Obese. Body mass index (bmi). Obes Res, 6(2):51S–209S,

1998
[23]

[Public, 2024b] Public

Accessed: 2024-05-01. [Public, 2024b] Public. UCI Dataset Download. https:// archive.ics.uci.edu/,

2024
[24]

[Raffel et al., 2020] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu

Accessed: 2024-05-01. [Raffel et al., 2020] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the lim- its of transfer learning with a unified text-to-text trans- former. Journal of machine le...

2024
[25]

Proximal policy optimization algorithms

[Schulman et al., 2017] John Schulman, Filip Wolski, Pra- fulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. CoRR, abs/1707.06347,

2017 arXiv
[26]

Sequence to sequence learn- ing with neural networks.arXiv preprint arXiv:1409.3215,

[Sutskever, 2014] I Sutskever. Sequence to sequence learn- ing with neural networks.arXiv preprint arXiv:1409.3215,

2014 arXiv
[27]

Gomez, Lukasz Kaiser, and Illia Polosukhin

[Vaswani et al., 2017] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V . N. Vish- wan...

2017
[28]

Generative partial multi-view clustering with adaptive fusion and cy- cle consistency

[Wang et al., 2021] Qianqian Wang, Zhengming Ding, Zhiqiang Tao, Quanxue Gao, and Yun Fu. Generative partial multi-view clustering with adaptive fusion and cy- cle consistency. IEEE Transactions on Image Processing, 30:1771–1783,

2021
[29]

Group-wise reinforcement feature generation for optimal and explainable representa- tion space reconstruction

[Wang et al., 2022] Dongjie Wang, Yanjie Fu, Kunpeng Liu, Xiaolin Li, and Yan Solihin. Group-wise reinforcement feature generation for optimal and explainable representa- tion space reconstruction. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Dat...

2022
[31]

A conditional-weight joint relevance metric for feature relevancy term

[Zhang et al., 2021] Ping Zhang, Wanfu Gao, Juncheng Hu, and Yonghao Li. A conditional-weight joint relevance metric for feature relevancy term. Engineering Applica- tions of Artificial Intelligence, 106:104481,

2021
[32]

Openfe: automated feature generation with expert-level performance

[Zhang et al., 2023] Tianping Zhang, Zheyu Aqa Zhang, Zhiyuan Fan, Haoyan Luo, Fengyuan Liu, Qian Liu, Wei Cao, and Li Jian. Openfe: automated feature generation with expert-level performance. In International Confer- ence on Machine Learning , pages 41880–41901. PMLR,

2023
[33]

Feature engineering for machine learning: principles and techniques for data scientists

[Zheng and Casari, 2018] Alice Zheng and Amanda Casari. Feature engineering for machine learning: principles and techniques for data scientists . ” O’Reilly Media, Inc.”,

2018
[34]

Difer: differentiable automated feature engineering

[Zhu et al., 2022] Guanghui Zhu, Zhuoer Xu, Chunfeng Yuan, and Yihua Huang. Difer: differentiable automated feature engineering. In International Conference on Auto- mated Machine Learning, pages 17–1. PMLR, 2022

2022
[2012]

Feature engineering for machine learning and data analytics

[Dong and Liu, 2018] Guozhu Dong and Huan Liu. Feature engineering for machine learning and data analytics. CRC press,

2018
[2013]

Generating sequences with re- current neural networks

[Graves, 2013b] Alex Graves. Generating sequences with re- current neural networks. CoRR, abs/1308.0850,

arXiv
[2014]

A survey on deep reinforcement learning for data processing and ana- lytics

[Cai et al., 2023] Qingpeng Cai, Can Cui, Yiyuan Xiong, Wei Wang, Zhongle Xie, and Meihui Zhang. A survey on deep reinforcement learning for data processing and ana- lytics. IEEE Trans. Knowl. Data Eng., 35(5):4446–4465,

2023
[2015]

Explorekit: Automatic feature generation and selection

[Katz et al., 2016] Gilad Katz, Eui Chul Richard Shin, and Dawn Song. Explorekit: Automatic feature generation and selection. In 2016 IEEE 16th International Conference on Data Mining (ICDM), pages 979–984. IEEE,

2016
[2016]

[Khurana et al., 2018] Udayan Khurana, Horst Samulowitz, and Deepak S. Turaga. Feature engineering for predic- tive modeling using reinforcement learning. In Proceed- ings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applicati...

2018
[2017]

Automl: A survey of the state-of-the-art

[He et al., 2021] Xin He, Kaiyong Zhao, and Xiaowen Chu. Automl: A survey of the state-of-the-art. Knowledge- based systems, 212:106622,

2021
[2018]

Strength- ening learning algorithms by feature discovery

[Dor and Reich, 2012] Ofer Dor and Yoram Reich. Strength- ening learning algorithms by feature discovery. Informa- tion Sciences, 189:176–190,

2012
[2019]

Autocross: Automatic feature cross- ing for tabular data in real-world applications

[Luo et al., 2019] Yuanfei Luo, Mengshuo Wang, Hao Zhou, Quanming Yao, Wei-Wei Tu, Yuqiang Chen, Wenyuan Dai, and Qiang Yang. Autocross: Automatic feature cross- ing for tabular data in real-world applications. InProceed- ings of the 25th ACM SIGKDD International Conference on...

2019
[2020]

Kaggle dataset download

[Howard, 2024] Jeremy Howard. Kaggle dataset download. https://www.kaggle.com/datasets,

2024
[2021]

The autofeat python library for automated feature engineering and selection

[Horn et al., 2020] Franziska Horn, Robert Pack, and Michael Rieger. The autofeat python library for automated feature engineering and selection. In Machine Learning and Knowledge Discovery in Databases: International Workshops of ECML PKDD 2019, W ¨urzburg, Germany, September...

2020
[2022]

Feature relevance term variation for multi-label feature selection

[Zhang and Gao, 2021] Ping Zhang and Wanfu Gao. Feature relevance term variation for multi-label feature selection. Applied Intelligence, 51:5095–5110,

2021
[2023]

A few useful things to know about machine learning

[Domingos, 2012] Pedro Domingos. A few useful things to know about machine learning. Communications of the ACM, 55(10):78–87,

2012
[2024]

[Kanter and Veeramachaneni, 2015] James Max Kanter and Kalyan Veeramachaneni

Accessed: 2024- 05-01. [Kanter and Veeramachaneni, 2015] James Max Kanter and Kalyan Veeramachaneni. Deep feature synthesis: Towards automating data science endeavors. In 2015 IEEE interna- tional conference on data science and advanced analytics (DSAA), pages 1–10. IEEE,

2024

Pith tools

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

[1] [1]

Neural machine translation by jointly learning to align and translate

[Bahdanau, 2014] Dzmitry Bahdanau. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473,

arXiv 2014

[2] [8]

On calibration of modern neural net- works

[Guo et al., 2017] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural net- works. In International conference on machine learning , pages 1321–1330. PMLR,

work page 2017

[3] [15]

Kingma and Jimmy Ba

[Kingma and Ba, 2015] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, 3rd International Con- ference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Pro- ceedings,

work page 2015

[4] [16]

Hyperopt-sklearn: Automatic hyperpa- rameter configuration for scikit-learn

[Komer et al., 2014] Brent Komer, James Bergstra, and Chris Eliasmith. Hyperopt-sklearn: Automatic hyperpa- rameter configuration for scikit-learn. In Scipy, pages 32– 37,

work page 2014

[5] [17]

Deep learning

[LeCun et al., 2015] Yann LeCun, Yoshua Bengio, and Ge- offrey Hinton. Deep learning. nature, 521(7553):436–444,

work page 2015

[6] [18]

Multi-label feature selection with high-sparse personalized and low-redundancy shared common features.Information Processing & Management, 61(3):103633,

[Li et al., 2024] Yonghao Li, Liang Hu, and Wanfu Gao. Multi-label feature selection with high-sparse personalized and low-redundancy shared common features.Information Processing & Management, 61(3):103633,

2024

[7] [19]

Libsvm dataset download

[Lin, 2024] Chih-Jen Lin. Libsvm dataset download. https: //www.csie.ntu.edu.tw/∼cjlin/libsvmtools/datasets/,

work page 2024

[8] [20]

[Liu et al., 2019] Kunpeng Liu, Yanjie Fu, Pengfei Wang, Le Wu, Rui Bo, and Xiaolin Li

Accessed: 2024-05-01. [Liu et al., 2019] Kunpeng Liu, Yanjie Fu, Pengfei Wang, Le Wu, Rui Bo, and Xiaolin Li. Automating feature sub- space exploration via multi-agent reinforcement learning. In Ankur Teredesai, Vipin Kumar, Ying Li, R ´omer Ros- ales, Evimaria Terzi, and George Karypis, editors, Pro- ceedings of the 25th ACM SIGKDD International Confer- ...

work page 2024