---

name: single-task-modeling description: "Operate DeepCTR-Torch single-output binary classification and regression model workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Single-task modeling

Use this sub-skill when the user needs a single-output DeepCTR-Torch model for binary CTR-style classification or scalar regression.

Owns

• Model families: WDL, DeepFM, xDeepFM, AFM, AFN, AutoInt, DCN, DCNMix, FiBiNET, IFM, DIFM, MLR, NFM, ONN, PNN, CCPM.
• Choosing a single-task model family and constructor shape.
• compile, fit, predict, evaluate, metrics, CPU/GPU device, callbacks, checkpointing, and save/load for single-task flows.
• Binary-to-regression workflow conversion.
• A self-contained DeepFM binary smoke helper.

Route away

• Feature-column construction, feature names, dense/sparse/sequence input arrays, and data validation: use ../feature-column-inputs/SKILL.md.
• DIN/DIEN sequence-interest models: use ../sequence-and-interest-models/SKILL.md.
• SharedBottom, ESMM, MMOE, and PLE multi-task models: use ../multitask-modeling/SKILL.md.

Operating sequence

1. Confirm the task is binary or regression and that the target is a single column.
2. Obtain valid feature columns and model_input dictionaries from the feature-column skill.
3. Select a model from the catalog; watch constructor exceptions for PNN and MLR.
4. Compile with one of the supported strings:
   • optimizers: sgd, adam, adagrad, rmsprop
   • losses: binary_crossentropy, mse, mae
   • metrics: binary_crossentropy, logloss, auc, mse, accuracy, acc
5. Train with fit, generate predictions with predict, and use evaluate or external sklearn metrics for final reporting.
6. Add EarlyStopping or ModelCheckpoint only after the monitored metric is present in logs.

References

• Model catalog
• Training and prediction workflows
• API reference
• Troubleshooting

Bundled helper

• scripts/deepfm_binary_smoke.py: inline tiny DeepFM binary example with --help, default one-epoch CPU run, optional CUDA device, prediction shape check, and final LogLoss/AUC reporting.