Source: https://github.com/aipoch/medical-research-skills

LightGBM Analysis

Use this skill to build a LightGBM model on tabular data and export feature importance ranking results as both a table and a figure.

Use This Skill When

• You need a command-line LightGBM workflow written in R.
• You need classification or regression on structured tabular data.
• You need ranked feature importance outputs for reporting or interpretation.
• You need standardized outputs under table/, figure/, and data/.

Primary Command

Rscript scripts/main.R \
  --data_file <input_file> \
  --target_var <target_column> \
  --output_dir <output_dir>

Prerequisites

• Rscript is available in the shell.
• Required R packages: optparse, data.table, lightgbm.
• Install basic dependencies with Rscript -e 'install.packages(c("optparse", "data.table"), repos="https://cloud.r-project.org")'.
• Install the R lightgbm package from the LightGBM project because it is usually not available from CRAN.

Core Arguments

| Argument | Required | Description | |----------|----------|-------------| | --data_file | Yes | Input data file in CSV format or tab-delimited TXT/TSV format | | --target_var | Yes | Target column used for modeling | | --output_dir | No | Output directory, default ./LightGBM_Results | | --fail_if_output_exists | No | Stop instead of overwriting when output_dir already contains files | | --task_type | No | auto, regression, binary, or multiclass. Default auto | | --feature_cols | No | Comma-separated feature columns. Default uses all columns except target and dropped columns | | --drop_cols | No | Comma-separated columns to exclude before modeling | | --importance_type | No | gain or split. Default gain | | --top_n | No | Number of features to show in the importance plot. Default 20 | | --output_format | No | csv or txt table export. Default csv |

Modeling Arguments

| Argument | Default | Description | |----------|---------|-------------| | --metric | auto | Evaluation metric matched to task type | | --test_size | 0.2 | Test-set proportion | | --valid_size | 0.2 | Validation proportion taken from the training partition | | --nrounds | 500 | Maximum boosting rounds | | --learning_rate | 0.05 | Shrinkage rate | | --num_leaves | 31 | Maximum leaf count per tree | | --max_depth | -1 | Maximum tree depth, -1 means no explicit limit | | --min_data_in_leaf | 5 | Minimum samples per leaf | | --feature_fraction | 0.8 | Column sampling ratio | | --bagging_fraction | 0.8 | Row sampling ratio | | --bagging_freq | 1 | Bagging frequency | | --lambda_l1 | 0 | L1 regularization | | --lambda_l2 | 0 | L2 regularization | | --early_stopping_rounds | 50 | Early stopping patience | | --seed | 42 | Random seed |

Input Requirements

• The input file must include the target column.
• Prefer .csv or .tsv inputs. .txt files must be tab-delimited.
• The skill expects at least 20 rows after removing missing target values.
• Features may be numeric, integer, logical, character, or factor-like text.
• Character features are label-encoded internally for LightGBM.
• Missing target values are removed before modeling.
• Missing feature values are left for LightGBM to handle.
• If task_type=auto, the script infers regression or classification from the target values.

Bundled test data examples:

V1,fustat,CAMK2N2,GGT6,GPR161,RAB26,RIBC2
TCGA-C5-A1M5,1,2.248291938,5.274690305,2.825215762,3.121114894,5.35318565
TCGA-EA-A5O9,0,3.346176843,5.404368414,2.604616977,0.629473197,4.429314674
TCGA-C5-A3HL,0,3.363100974,5.363314779,4.124799581,4.127228806,4.916596068

Minimal Workflow

1. Confirm the input file exists and the target column name is correct.
2. Remove identifier or sensitive columns such as id, sample_id, patient_id, accession numbers, or the bundled sample identifier column V1 before training.
3. Set --drop_cols and optionally --feature_cols so the model only sees intended predictors.
4. If you need overwrite protection, add --fail_if_output_exists or choose a fresh --output_dir.
5. Run scripts/main.R.
6. Check table/ for the importance table, model metrics, and remediation guidance.
7. Check figure/ for the feature importance ranking plot and data/ for the run summary.

Avoid ambiguous text exports. If a .txt file is parsed as one column, re-export it as tab-delimited text or CSV before rerunning.

For quick validation in small audit environments, prefer the bundled dt_sample3.txt smoke test shown below with reduced --nrounds and --early_stopping_rounds. The full binary example on dt_sample1.csv is still useful as a complete workflow example, but it can exceed short runtime budgets.

If you omit --data_file or --target_var, the script exits with SKILL_MISSING_INPUT.

Outputs

Expected output structure:

<output_dir>/
├── table/
├── figure/
└── data/

Primary result files:

• table/lightgbm_feature_importance.<output_format>
• table/lightgbm_model_metrics.<output_format>
• table/lightgbm_remediation.<output_format>
• figure/lightgbm_feature_importance_<importance_type>.pdf
• data/lightgbm_run_summary.txt
• data/lightgbm_categorical_levels.txt when categorical or character predictors were encoded

Feature importance table fields include:

• feature
• gain
• split
• cover
• importance_type
• importance_value
• rank
• gain_share
• split_share

Model metrics include:

• task_type
• metric_primary
• best_iteration
• train_rows
• valid_rows
• test_rows
• prediction_collapse_flag
• model_quality_flag
• interpretation_status
• primary_issue
• model_quality_issues
• rerun_hint
• model_quality_note
• task-specific evaluation metrics such as rmse, mae, accuracy, auc, or logloss

Remediation table fields include:

• task_type
• model_quality_flag
• interpretation_status
• issue_code
• issue_detail
• recommended_action
• suggested_rerun_change

Run summary file includes the task type, best iteration, primary quality fields, top features, and artifact paths for the completed run.

Overwrite Behavior

• Rerunning into an existing output_dir replaces prior result files with the new metrics, importance table, remediation table, figure, and session metadata.
• Set --fail_if_output_exists when you want the run to stop instead of replacing prior artifacts.
• If you need an audit trail, prefer a timestamped or per-run output_dir.
• The script now warns when output_dir already contains files.

Success And Failure Contract

Success:

• Console output should end with LightGBM analysis completed successfully.
• table/lightgbm_model_metrics.<output_format> and table/lightgbm_feature_importance.<output_format> should exist.
• table/lightgbm_remediation.<output_format> and data/lightgbm_run_summary.txt should exist.
• figure/lightgbm_feature_importance_<importance_type>.pdf should exist.
• The importance table should contain at least one non-zero gain or split value.

Failure or caution:

• If parsing fails, expect a SKILL_* message instead of a raw stack trace.
• If best_iteration <= 1, predictions collapse to one class, recall is 0, f1 is NA, or the selected importance values are mostly zero, do not treat the ranking as reliable.
• Review model_quality_flag and model_quality_note in table/lightgbm_model_metrics.csv before interpreting the exported ranking.
• Use interpretation_status to decide whether the run is report-ready: eligible means interpretation-ready, eligible_with_caveats means the ranking may still be usable with caveats, and caution_only means diagnostic-only.
• Review table/lightgbm_remediation.csv and rerun_hint for the exact failure mode and recommended rerun changes.
• Recheck delimiter choice, identifier leakage, and --min_data_in_leaf before trusting the outputs.

Caution Remediation

• best_iteration<=1: lower --min_data_in_leaf and verify that the selected predictors have usable signal.
• single_predicted_class: review class balance and feature selection before using the ranking downstream.
• recall=0 or no_positive_predictions: revisit --feature_cols and the target balance before treating the run as report-ready.
• <importance_type>_importance_sparse: compare against the alternate importance type and review whether the retained predictors have enough signal.

Agent Response Contract

When this skill completes, the agent should report:

• resolved task_type
• best_iteration
• primary evaluation metrics from table/lightgbm_model_metrics.<output_format>
• top ranked features from table/lightgbm_feature_importance.<output_format>
• model_quality_flag and interpretation_status
• artifact paths for the metrics table, importance table, remediation table, figure, and run summary file

If model_quality_flag is not ok, the agent must explicitly say the run is diagnostic-only or caveat-limited and include the recommended rerun changes from rerun_hint or table/lightgbm_remediation.<output_format>.

Feature Importance Guidance

• Use gain when you care about overall contribution to loss reduction.
• Use split when you care about how often a feature is used in tree splits.
• Prefer gain for most ranking summaries and reports.
• Low importance does not imply no business value, especially under correlated features.

Read These Files When Needed

| Need | File | |------|------| | LightGBM method details and importance interpretation | references/algorithm.md | | CLI examples | references/cli-guide.md | | Error diagnosis | references/troubleshooting.md | | Main entry point | scripts/main.R | | Sample test data | tests/data/ |

Quick Examples

Fast smoke test with dt_sample3.txt:

Rscript scripts/main.R \
  --data_file tests/data/dt_sample3.txt \
  --target_var Group \
  --drop_cols V1 \
  --task_type binary \
  --nrounds 80 \
  --early_stopping_rounds 20 \
  --top_n 15 \
  --output_dir tests/output_smoke_txt

Audit-friendly binary preset for short runtime budgets:

Rscript scripts/main.R \
  --data_file tests/data/dt_sample1.csv \
  --target_var fustat \
  --drop_cols V1 \
  --task_type binary \
  --nrounds 120 \
  --early_stopping_rounds 20 \
  --output_dir tests/output_binary_fast

Full binary workflow example with dt_sample1.csv:

Rscript scripts/main.R \
  --data_file tests/data/dt_sample1.csv \
  --target_var fustat \
  --drop_cols V1 \
  --task_type binary \
  --output_dir tests/output_binary

Split-based importance export example with dt_sample2.csv:

Use this to verify split-based ranking output. Review model_quality_flag and interpretation_status before treating the bundled example as report-ready because this path can remain diagnostic-only on small test splits.

Rscript scripts/main.R \
  --data_file tests/data/dt_sample2.csv \
  --target_var fustat \
  --feature_cols CAMK2N2,GGT6,GPR161,RAB26,RIBC2 \
  --drop_cols V1 \
  --task_type binary \
  --importance_type split \
  --output_dir tests/output_binary_split

Audit-friendly regression preset with dt_sample1.csv and RIBC2 as the target:

Rscript scripts/main.R \
  --data_file tests/data/dt_sample1.csv \
  --target_var RIBC2 \
  --drop_cols V1 \
  --task_type regression \
  --nrounds 120 \
  --early_stopping_rounds 20 \
  --output_dir tests/output_regression_fast

Full regression workflow with dt_sample1.csv and RIBC2 as the target:

Rscript scripts/main.R \
  --data_file tests/data/dt_sample1.csv \
  --target_var RIBC2 \
  --drop_cols V1 \
  --task_type regression \
  --output_dir tests/output_regression

Tab-delimited TXT input with automatic binary target encoding from Group:

Rscript scripts/main.R \
  --data_file tests/data/dt_sample3.txt \
  --target_var Group \
  --drop_cols V1 \
  --task_type binary \
  --top_n 15 \
  --output_dir tests/output_group_txt

Validation

Rscript scripts/main.R --help

Use the smoke test under ## Quick Examples for a fast validation pass. After a successful run, verify that these files exist under the selected output_dir:

• table/lightgbm_feature_importance.csv
• table/lightgbm_model_metrics.csv
• table/lightgbm_remediation.csv
• figure/lightgbm_feature_importance_<importance_type>.pdf
• data/lightgbm_run_summary.txt
• data/lightgbm_categorical_levels.txt if categorical or character predictors were encoded

When Not To Use

• The input file is an unstructured note, JSON blob, or free-text report.
• The text file delimiter is unknown and you cannot inspect or re-export it.
• The table still contains sample IDs, patient IDs, accession numbers, or similar identifiers that should not be model features.
• The input still contains direct identifiers or sensitive fields that you have not reviewed and removed from modeling.

Common Errors

• SKILL_FILE_NOT_FOUND: Input file path is wrong or inaccessible.
• SKILL_MISSING_COLUMNS: The target or requested feature columns are missing.
• SKILL_INVALID_DATA: Data types, target encoding, or row count are unsuitable for LightGBM.
• SKILL_DEGENERATE_MODEL: Training finished but the exported importance table is all zero and should not be interpreted.
• SKILL_INVALID_PARAMETER: An argument value is invalid.
• SKILL_DEPENDENCY_MISSING: Required package such as lightgbm is unavailable.
• SKILL_TRAINING_FAILED: LightGBM training failed.

Before sharing exported artifacts, verify that identifier-like columns such as V1, sample IDs, or patient IDs were excluded from modeling and from any published tables. If model_quality_flag is not ok, treat the run as a diagnostic result rather than an interpretable ranking.

If the issue is not obvious, read references/troubleshooting.md.