FROM TABULAR DATA TO EXPLAINABLE CLASSIFICATION

Using alphaML.

A practical guide to the alphaML 2.0 research workbench.

Application 0.2.0 · Updated 19 October 2026

BringNumeric features and a separate labels table.

BuildA model distinguishing two chosen classes.

ReviewScores, explanations and new predictions.

Default server settings are shown below. Check the application and privacy notice for the settings of your server.

Your first analysis

alphaML is an AutoML application for binary classification: predicting one of two labels from numeric features. It was developed for research, including pharmacogenomic expression data. CLETE stands for Clear, Legible, Explainable, Transparent and Elucidative; explanation plots help inspect a model, but do not establish that it is scientifically valid.

Sign inRead the privacy notice.
UploadFeatures and labels.
ConfigureClasses, selection and model.
ReviewJob results and limitations.
PredictApply the saved model.

A small, reproducible first run

Download the synthetic example files. Open Train → Upload new…, give the data set a name, and upload example_labeled_data.csv as Labeled data and example_data_labels.csv as Labels. Leave optional files empty and select Upload.
Set Class column = response, Positive class = resistant, and Negative class = sensitive.
Set all four Feature selection methods to None. Select Single algorithm → Ridge → Predefined. Set Normalization = None, Sampling = None, Test size = 0.2, and Random seed = 12. Leave explanation options off for this first run.
Select Start training. On Jobs, wait for success and inspect the scores, confusion matrix and files. The 120-row example produces 96 training and 24 test rows with this setup.
Open Predict, choose that finished training job, and upload example_new_samples.csv. Keep Same as the model and leave class-name overrides blank. Select Start prediction, then download the prediction spreadsheet.

This recipe checks the file-to-model workflow; its settings and scores are not a recommendation for your study. To explore AutoML with the same data, try only LASSO and RandomForest, tune 2 algorithms with 2 trials each and 3 CV folds, and keep explanations off initially.

Load example data in the app is a separate, shared pharmacogenomic example downloaded from figshare. It needs internet access the first time. The synthetic downloads in this handbook work without that external download.

What alphaML can and cannot do

Scope of this web application
It can	It does not currently provide
Classify numeric tabular samples into two selected classes.	Regression, multiclass/multilabel prediction, survival models or time-series forecasting workflows.
Apply up to four feature-selection methods, in sequence.	Native categorical/text/image predictors, automatic batch correction or a complete assay preprocessing pipeline.
Compare algorithms with AutoML, tune candidates and try a probability-averaging ensemble.	A guarantee of the globally best model, calibrated probabilities, or an unbiased estimate from an exhaustive nested validation pipeline.
Provide test metrics, selected features, optional importance/SHAP/LIME plots and repeated-CV ROC plots.	Causal effects, proof of a biological mechanism, clinical validation or regulatory approval.
Save model bundles and predict new samples in background jobs.	Automatic retraining, drift monitoring, a threshold editor, or a browser control to refit the final model on all labeled rows.

Options available through the Python API are not necessarily exposed in the web interface. For example, the combinatorial IterativeFeatSel method is not offered in the browser.

Prepare input files

Training requires two separate tables. Use UTF-8 CSV, or tab-separated .tsv/.txt, with a header row. The first column contains sample names and is read as the row index, never as a predictor. Samples go down rows and features go across columns. Export Excel sheets as CSV first; .xlsx is a result format, not an accepted input format.

Files accepted by Train → Upload new…
Upload field	Required?	Contents
Labeled data	Yes	First column: sample code. Remaining columns: numeric features such as gene measurements. Do not include the target among these features.
Labels	Yes	First column: matching sample code. Remaining columns: possible outcomes, such as a drug response. Select one outcome and two class values in the app.
Data for feature selection	No	A numeric sample × feature table for unsupervised selection. Without it, those methods use the labeled table. Supervised methods still use the labeled data and its classes.
Unlabeled data	No	Numeric features for additional samples. Predictions are generated after training; this table does not supply validation labels.
Suggested features	No	A single column of feature names with a header. Select Suggested as a feature-selection method to use the list. This file has no sample-index column.

Labeled data

sample,GENE01,GENE02,GENE03
S001,3.12,4.81,2.67
S002,4.76,2.15,3.09

Labels

sample,response
S001,resistant
S002,sensitive

This miniature preview illustrates layout only. Use the full downloadable examples for a runnable first analysis.

Before uploading

Use unique, non-empty sample codes and feature names. Prefer simple sample codes without commas, tabs or embedded newlines. Keep sample codes as text and preserve their spelling and leading zeros in every file.
Check that the sample codes in the labels match the feature table. Review the app’s matched-label count before training. Identical row order alone is a fragile way to link separate files.
All predictor columns must represent numeric quantities. Encode categorical variables deliberately outside alphaML if appropriate; it does not infer categories or one-hot encode them. Use a decimal point, no unit suffixes, percent signs or thousands separators, and no infinite values.
Check missing values explicitly. The numeric file reader replaces missing values and nonnumeric text with zero; it does not learn medians or report a missingness audit. A misspelled measurement can therefore become an apparently valid zero.
Keep units, preprocessing, assay platform and feature naming compatible between training and prediction. alphaML does not perform RNA-seq count normalization, batch correction or cross-platform harmonization for you.
Remove target-derived variables, post-outcome information and identifiers encoding the answer. Only upload anonymised data allowed by the server’s privacy notice.

Feature overlap affects the model. During training, the application intersects feature columns across labeled data, optional feature-selection data and optional unlabeled data. Features missing from another supplied matrix are dropped before selection. Check the log and selected_features.csv; supplying an incomplete optional table can change the model.

Download synthetic examples

These files contain no real people or patient data. The two training tables share 120 sample codes and 16 numeric features; the new-samples table contains 16 independent samples, reordered feature columns and no labels.

Labeled data CSV ↓120 samples × 16 features Labels CSV ↓response: resistant / sensitive New samples CSV ↓16 samples · same 16 features Suggested features CSV ↓Optional four-feature list

Use the first-run recipe with all features. To try the feature list, upload the suggested-features CSV in its optional field and choose Suggested for Method 1, with the other methods None. Uploading the file by itself does not activate that method.

The values are synthetic log2-scale expression-like measurements. If exploring HVF, its data-transform setting is log2(x+1) for this example; that setting describes the input scale rather than requesting an additional log transform.

Labels and normalization

Choose the meaning of positive and negative

Select one Class column, then two different class values. Positive is encoded as 1, negative as 0. Matching trims spaces and ignores class-label case. Other values, including unknown labels, are excluded from this run; choosing two classes from a larger label column still produces a binary model. The web app offers label columns containing between 2 and 50 distinct non-missing values.

Samples are aligned by their sample names. If there are no matching sample names and both files have the same number of rows, alphaML falls back to row-order matching and records this in the log. If only some names match, it uses those matches and drops unmatched samples. Verify the log, retained sample count and class counts; a row-order fallback can silently assign the wrong labels if files are ordered differently.

Normalization acts across features within each sample

Normalization choices, applied after feature selection
Choice	What happens	When to think carefully
Min-max per sample (`min_max`, default)	Each row is rescaled between its own minimum and maximum.	Absolute level differences between samples are removed. This may be unsuitable when a sample’s overall level carries information.
Standardize per sample (`standardization`)	Each row is centered at its own mean and divided by its population standard deviation.	This is not column-wise feature standardization learned on training samples.
None	Keeps numeric input values after the file reader’s zero filling.	You must provide a scientifically appropriate and consistent scale yourself.

A constant row becomes all zeros under min-max or standardization. With only one selected feature, both methods turn every sample into zero, so use None for a one-feature analysis. Predicting one sample or a batch uses the same row-wise rule and the saved feature order.

Feature selection

Up to four methods run in order, each on the features retained by the preceding method. None keeps the current set. Additional steps do not necessarily improve performance, and the order can change the result. Most computed selectors log a fallback to the current set if they return no features; an empty Suggested match can instead make the run fail. Inspect the output rather than assuming the requested reduction happened.

Methods offered in the browser
Method	Purpose and key settings
HVF	Selects highly variable features using dispersion within mean bins. Choose the input’s actual `log2(x+1)`, `log10(x+1)`, `ln(x+1)` or non-log scale. It reverses that transform to calculate dispersion; it does not log-transform the model input. Top features ≥2 uses a ranked cut-off; 0 uses the dispersion/mean cut-offs. Ties can retain more than the requested number.
MedianMAD	Keeps features with median absolute deviation above the median across features. No count setting; constant or tied-dispersion data may retain no new subset.
SelByPCA	Ranks original features by aggregated loadings from components covering about 80% of variance. Set top features. The classifier receives selected original features, not PCA component scores.
RandomProjection	Ranks original features by their association with a seeded random projection. Set top features; this is not a classifier operating directly on projection coordinates.
SelByClustering	Clusters feature profiles and chooses representative features. Set a number of clusters compatible with the available feature count.
SelByNMF	Selects features using non-negative matrix factorization. Requires non-negative input and a feasible number of components.
SelectByRF	A supervised random-forest selector. Set an importance threshold such as `1.5*mean` and an optional maximum feature count; 0 means no count limit.
RecursiveFeatElim	Supervised recursive elimination with a linear SVM and internal CV. Set minimum retained features. Removing features one at a time can be expensive.
SeqFeatSel	Supervised sequential selection with XGBoost and internal CV. Set the retained feature count; use only after a substantial reduction for wide data.
RemoveHighCorrFeat	Greedily removes a feature if its absolute correlation with an earlier retained feature is at least 0.9. Column order affects which feature survives.
Suggested	Keeps names from the supplied one-column list that exist in the current data. Missing names are logged. A list fixed independently of the study labels can help define a prespecified panel.
None	Keeps all current features. Useful for a known panel and for testing the input workflow.

Feature selection is not nested within evaluation. Selection currently runs on the labeled data before the train/test split and before the model-selection CV folds. In particular, supervised selection can use labels from rows later called “test.” Test and CV scores can therefore be optimistic. See validation limits before using these scores in a manuscript or to support a decision.

Models and settings

AutoML: screen, tune, optionally combine

Screen: cross-validate every selected algorithm with predefined settings on the training portion.
Tune: take the best Algorithms to tune candidates and run Optuna searches, sharing the tuning time budget.
Combine: if enabled and at least two algorithms are in the shortlist, try a soft-voting ensemble that averages class probabilities. Select the best candidate by the CV objective; the ensemble is not guaranteed to win.

The 14 available algorithms are AdaBoost, CatBoost, ElasticNet, ExtraTrees, GradientBoosting, HistGradientBoosting, LASSO, LightGBM, MLPC, RandomForest, Ridge, SVC, NuSVC and XGBoost. LASSO, Ridge and ElasticNet here are regularized binary classifiers, not regression tasks. MLPC is a neural-network classifier. AutoML only compares algorithms you selected.

Ranking choices are Cohen’s kappa × MCC (implemented as the square root of the product after clipping negative values to zero), ROC AUC, Balanced accuracy and F1. Higher scores are better. The leaderboard’s objective, rather than the rounded displayed score alone, determines rank; Control fitting changes the objective using the train/validation gap. It does not prove that overfitting has been prevented.

Single algorithm

Select one algorithm and choose Optuna, Bayes, Grid or Predefined hyperparameters. Predefined fits the bundled settings without a hyperparameter search. Grid uses the bundled parameter grid, so its number of fits is not capped by the trials field. The AutoML ranking-metric selector and time budget apply to AutoML, not to a single-algorithm search.

Starting defaults, before you change the form
Setting	Default	Practical meaning
Mode / candidates	AutoML / all 14	A broad screen can be expensive. Reduce the candidate list for a workflow check.
Algorithms to tune	3	The shortlist shares the tuning time budget.
Tuning time budget	10 minutes	Applies to tuning, not the whole job. 0 removes this budget. A running trial may finish after the nominal limit.
Trials / CV folds	50 / 5	Trials are per tuned algorithm. AutoML uses stratified, shuffled folds within the training portion; supply enough samples of each class.
Test size / seed	0.2 / 12	About 20% of eligible rows become the test portion. The outer split is random and not stratified; it is not a group-aware or time-aware split.
Selection / normalization	HVF / min-max per sample	These expression-oriented defaults may be unsuitable for general tabular measurements. Confirm the input’s log scale before HVF.
Sampling / control fitting	None / off	Oversample uses SMOTE. Undersample uses Tomek links then random undersampling. Some estimators also use built-in class balancing.
Ensemble / explanations / ROC	Ensemble on; other outputs off	Additional fits and explanations can greatly increase runtime.

The final saved model is fitted on the training portion, possibly resampled. It is not automatically refitted on all labeled rows. A displayed training count after sampling may differ from the number of original training samples.

Follow a job

After Start training or Start prediction, open Jobs and select the item. Watch its status, stage, settings and log. Jobs can be queued, running, succeeded, failed, cancelled or interrupted. With accounts enabled, the job continues if you close the browser or log out; sign in again to inspect it.

Cancel stops a queued or running job. A cancelled, interrupted or failed run is not a completed model and is not available as a finished training job in Predict. Read the last log lines, fix the setup and submit a new job. Some result files may already exist from a partial run.

The default server runs 1 job at a time. With accounts on, the default quota is 3 queued/running jobs per user, configurable by the administrator. Queued time is separate from model runtime. A server restart interrupts running work; save valuable results before maintenance.

Use Delete only when you intend to remove that job and its files. Download the bundle and results first. Deleting a training job removes the model used for future browser predictions. A data set cannot be deleted while a queued/running job uses it.

Read scores and the leaderboard

Separate the AutoML leaderboard (CV comparisons used to choose a model) from the Train / Test scores of the final fit. A large Train–Test gap or a large Train–CV gap suggests instability or overfitting. Close scores do not establish validity; small samples, leakage and repeated tuning can also hide problems.

Key metrics and how to interpret them
Metric	Meaning
Accuracy	Fraction of correct labels. A majority-class classifier can score highly on an imbalanced data set.
Sensitivity / specificity	Fraction of true positive / true negative samples correctly identified. Their interpretation depends on the positive class you chose.
Precision / NPV	Fraction of positive / negative predictions that are correct in this evaluation sample. These depend on class prevalence.
F1 / MCC / Cohen’s kappa	Complementary classification summaries. F1 balances positive-class precision and recall; MCC and kappa account for both classes or chance agreement. No single metric replaces class counts and the confusion matrix.
AUC (probability)	ROC AUC using positive-class probabilities. This is the more informative AUC row for ranking predictions. Chance ranking is about 0.5; AUC does not assess calibration or a chosen decision threshold.
AUC	A legacy row calculated from hard class predictions, equal to balanced accuracy when both classes are present. Do not mistake it for probability-based ROC AUC. The headline Test AUC uses the probability row when available.
Average Precision Score	Summarizes positive-class precision/recall using probabilities; its baseline depends on positive-class prevalence.
Brier Score Loss	Mean squared probability error. Lower is better; a good value alone does not establish calibration in a new population.
NegLog2-RMSL	A custom composite of test losses and train/test differences; higher is generally better. It is not a log-likelihood, probability, standard calibration score or universal acceptance threshold.
Test loss column	Usually 1 minus the metric, but Brier loss and negative likelihood ratio are kept as raw losses. Do not assume every row is scaled the same way.

The confusion matrix shows counts of true versus predicted classes; check the axes and class names. Undefined metrics, for example when there is only one test class or no predicted positives, may appear blank. They are not evidence of a perfect result.

The leaderboard includes screened/tuned candidates and an optional ensemble. Score is the chosen CV metric, SD is variability across those folds (not a confidence interval), and Train-CV gap is the training minus validation score. Candidates reuse the same CV folds, so the winning CV score is itself selected and can be optimistic.

Understand explanations

Feature importance: exports built-in tree importance or signed coefficients where available, plus the mean change in test accuracy from permuting a feature five times. Unsupported built-in importance is blank. Negative permutation importance can occur. Correlated predictors can share or mask importance; values from different algorithms need not have the same meaning.
SHAP: a PDF of summary plots and per-sample waterfalls for the positive class. The explanation scale depends on the explainer/model: inspect its axes rather than assuming every value is a probability change. A fallback explainer may be slower; unsupported individual plots can be skipped and logged.
LIME: a PDF with local approximations for individual samples, generally showing the top 10 features. It describes a local surrogate, not the full classifier. Perturbed samples and the reference data influence the explanation.
ROC curve: an optional repeated-CV summary for the chosen model recipe, using all eligible labeled data after selection. It is a different analysis from the single test split. The current implementation uses repeated random K-folds, not stratified repeats, and does not repeat feature selection and model search within each fold.

SHAP and LIME use the samples being explained as reference/background data. Explanations can change when you submit a different prediction batch even if an individual sample’s model prediction is unchanged. They do not identify causes or independently validate biomarkers.

Estimate the cost before enabling ROC. With the default 5 folds and 50 trials, the ROC option requests 250 extra fits of the selected recipe; an ensemble fits each member. SHAP and LIME add work for each explained sample and can create large PDFs. Start without them, then enable only what your analysis needs.

Predict new samples

Open Predict and choose a succeeded training job from your account. Train a model first if none is available.
Upload a CSV/TSV/TXT sample × numeric-feature table, with sample names in the first column. A target column and labels file are not needed.
Provide every selected model feature, with the exact same column names and compatible measurements. Column order may differ: the saved bundle reorders them. Extra columns are ignored; missing required feature columns cause an error. Missing cells within a present column are zero-filled by the file reader.
Keep Normalization = Same as the model and leave class fields blank for the usual workflow. Overriding normalization changes the input scale. Class-name overrides relabel outputs; they do not retrain, reverse or change the model’s definition of positive.
Select optional SHAP/LIME outputs if required, then Start prediction. Read the completed job and download the prediction spreadsheet.

The full *_test_prediction.xlsx contains one row per new sample, a predicted class name and pred_proba, the model’s estimated probability of the training positive class. It is not automatically calibrated or a confidence interval. The class label follows the estimator’s decision rule; for some models it need not equal a simple 0.5 cut-off on the exported probability. There is no browser setting for choosing another threshold.

The Jobs prediction preview shows at most 500 rows. Download the spreadsheet for the complete output. Predictions alone do not measure accuracy; assess them against independently obtained outcomes outside this prediction form.

Model-bundle uploads are disabled by default when accounts are enabled. Use your own finished training job. On a server that permits bundle uploads, only trusted alphaML 2.0 *_bundle.pkl files are supported; loading a pickle can execute code. Models from alphaML 0.x require retraining.

Save results and models

Open a completed job to preview tables and plots, download individual files or select Download all. The ZIP includes the job’s result files and log; it does not include the original uploaded input files or the complete submitted job settings. Keep your original files and record the job’s Settings panel separately.

Common result files (the prefix contains target/model settings)
File	Use
`*_bundle.pkl`	Trained estimator, selected feature order, normalization, class names, algorithm, parameters and alphaML version. Required for reuse outside the current job. Preserve a compatible Python/software environment.
`selected_features.csv`, `_Feat_M.csv`	Final feature names and intermediate selection outputs, where applicable.
`*_scores.csv`, `summary.json`	Final train/test scores and a machine-readable summary. The summary is not a full copy of all input settings.
`*_test_prediction.xlsx`	In a training job: test rows only, including encoded true/predicted labels and class names. In a prediction job: all submitted new samples and predicted class names. `pred_proba` means positive-class probability.
`*_unlabeled_prediction.xlsx`	Predictions for an optional unlabeled table supplied during training.
`_normalized_.csv`	Selected numeric matrices after the chosen row-wise normalization. They can contain sensitive measurements; treat downloads according to your data policy.
`*_AutoML_leaderboard.csv` / `.png`	Candidate ranking and a visual overview for AutoML runs.
`*_parameters.json`, confusion matrix, training curve	Chosen parameters and diagnostics. A training curve is available for supported single-model fits; it is not a learning curve across different study sample sizes.
Importance, SHAP/LIME PDFs, repeated-CV files	Appear when requested and successfully generated. Check the log for skipped plots.

For a reproducible analysis, retain the exact inputs and sample alignment, class definitions, four selection methods and their settings, normalization, test size, seed, sampling, candidate list, ranking metric, CV folds, search/trial/time settings, output options, package versions, bundle, selected features, scores and log. A fixed seed helps repeat a run but does not ensure identical results across platforms, library versions or time-limited searches.

Scientific and practical limits

Use alphaML for research exploration. A successful job or a high score does not establish clinical usefulness, generalization, causal relevance or deployment readiness.

What the internal evaluation does not protect against

Selection before splitting: feature selection happens before the test split. Supervised methods can see future test labels; unsupervised methods can see future test feature distributions. Selection is not fitted anew within the model CV folds. An externally prespecified panel or None avoids label-driven feature selection within alphaML, but other biases still need attention.
Sampling before model CV: optional SMOTE/undersampling is applied to the training portion before its model-selection folds are made. It is not repeated separately inside each fold. This can affect validation estimates, particularly with synthetic samples. Repeated ROC sampling is likewise done before its folds.
Random split limitations: the outer test split is not stratified, grouped or chronological. Small or imbalanced data can leave too few samples of a class in a split or fold. Repeated samples from one person, family, site or batch can leak information between splits.
Reuse of evaluation: choosing among many algorithms on the same CV folds, repeatedly changing settings after viewing test results, or selecting a threshold from those results can make reported performance optimistic. Repeated ROC reuses an already selected recipe and feature set; it is not nested independent validation.
Distribution shift: new assay platforms, units, batches, populations or prevalence can change performance and probability interpretation. The app does not automatically flag out-of-distribution samples or monitor a deployed model.
Uncertainty and explanations: the interface does not provide a full model-selection uncertainty analysis, calibrated prediction intervals, causal effects, fairness assessment or prospective clinical validation.

For a manuscript or substantive application, use a validation design suited to the study: keep an external cohort untouched by feature/model/threshold selection, and use an independently implemented nested, group-aware or time-aware pipeline where needed. The web interface does not configure those designs for you. Download enough information to report the actual analysis and its limitations.

Files and computation

The default per-file upload limit is 2048 MB, sent in chunks of 8 MB. This is an upload limit, not a guarantee that a file of that size can be parsed or modeled in memory. Wide data, grids, sequential/RFE selection, SHAP/LIME and repeated CV may need substantial RAM and CPU. Available resources determine feasibility.

The AutoML time budget covers tuning only. Screening, selection, an in-progress trial, final fitting and optional outputs can continue beyond it. Reduce feature count, candidates, trials or optional outputs when work becomes too expensive. A zero-minute budget means no tuning time limit, not zero work.

Accounts and privacy

With accounts enabled, register using your email and a password of at least 8 characters, agree to the privacy terms and sign in. A sign-up code may be required, or registration may be restricted by the administrator.

Your own data sets, jobs and downloads are visible to your account. The built-in example is shared and read-only. Account isolation does not mean that files stay in your browser: uploads and results are processed and stored on the server, and the operator can access server storage.

What is stored: your email, a salted password hash, privacy-agreement version/time, session/reset records, uploaded files, job settings, logs and results. A necessary login cookie keeps you signed in; the app’s notice states it has no advertising/tracking cookies.
Retention: with accounts enabled, data sets and jobs are normally removed 14 days after upload/submission. Cleanup runs periodically and protects active inputs while in use. Check each item’s deletion date and download results before it expires. Account records remain until you delete the account.
Upload policy: follow the linked privacy notice and supply only anonymised data. Do not include names, dates of birth, linkable patient identifiers or identifiable health/genetic data. An arbitrary sample code alone does not establish that a data set is anonymous.
Password changes: use the account menu. For a forgotten password, choose Forgot your password? on the login screen. If email reset is configured, the requested link is single-use and valid for 60 minutes; the response does not reveal whether an email has an account. Check spam, or contact the administrator if mail is unavailable.
Deletion: individual Delete controls remove data sets or jobs. The account menu’s Delete account and data requires your password and removes the account and its associated data. Download anything you need first.
Questions and rights: use the operator/contact in the current privacy notice. This handbook explains the application and does not replace that notice.

If accounts are disabled on a local server, anyone who can open that server can see its data. Check the current server configuration; do not assume a local-looking URL means all processing happens on your own device.

Troubleshooting

Common problems and next steps
What you see	What to check
No class column / no usable labels	Use a separate labels file with sample names first and an outcome column containing 2–50 distinct values. Check delimiter, header and missing labels.
Zero/few matched labels or unexpected sample count	Compare sample codes exactly, inspect row-order fallback/dropped-label messages, and confirm the two chosen classes. Fix alignment outside the app before continuing.
Both classes must be present / CV or SMOTE error	Check retained class counts after matching and filtering. Reduce folds or turn sampling off. SMOTE’s default neighborhood needs more than five minority training rows. A passing run still does not establish adequate study size.
HVF fails or selects an unexpected subset	Check the input log scale, finite numeric values, count/dispersion settings and log fallback. Try None or another appropriate method for a workflow check.
NMF, PCA, clustering, RFE or sequential selection fails	Use feasible counts for the current sample/feature dimensions; NMF requires non-negative data. Reduce the input first and check each preceding selection step.
Constant/uninformative predictions	Inspect labels, zero-filled values, constant columns, selection and normalization. A single feature with row-wise min-max/standardization becomes zero for every sample. Refit with a justified setup.
Prediction reports missing model features	Use the saved selected-feature list; fix spelling/case and supply all required columns. Do not fabricate zero columns for measurements that were never obtained.
Job seems slow or the time budget passed	Check the stage/log. The budget is not a whole-job timeout. ROC costs folds × trials extra fits, and explanations can be expensive. Cancel and reduce the setup if appropriate.
Upload fails / file too large	Use CSV/TSV/TXT within the current per-file limit, keep the page open while chunks transfer, and retry after checking the connection. Unused uploads are cleaned up after about a day.
No model available in Predict / missing result files	Choose a succeeded training job from your account. Failed/partial runs do not qualify. Check retention/deletion and whether optional outputs were requested.
Cannot load the built-in example	Its first load needs the external figshare download. Use the handbook’s small synthetic files or ask the administrator to check connectivity.
Login, consent or reset problem	Check the email, current privacy agreement and sign-up policy. Request a fresh reset link if expired/used; avoid repeated attempts during a rate-limit lockout. Contact the operator if email is unavailable.

When asking for help, include the job type, alphaML version, selected methods/model, dimensions, class counts and relevant error text. Do not email passwords, reset links, raw sensitive data or identifiable sample names.

Glossary and support

Feature: A numeric input variable, such as an expression measurement.
Target / class column: The label to predict. One target and two chosen class values are used per run.
Training / test portion: Rows used to fit/select a model versus rows reserved by the outer random split. The preprocessing limitations above still apply.
Cross-validation (CV): Repeated fitting on subsets of training data to compare candidates. It is not a substitute for an untouched external cohort.
Hyperparameter / trial: A model setting chosen before fitting / one candidate configuration evaluated during a search.
Ensemble: A predictor averaging probabilities from several fitted classifiers.
Model bundle: The saved estimator plus feature order, normalization and class metadata needed to predict consistently.
Leakage: Information entering model development that would not have been available for a genuinely new sample.

For support or privacy questions: use the contact in the application’s privacy notice.