autoseqmodels

Automated weekly sequence-model workflow for customer transaction prediction. Provides an end-to-end pipeline from raw transaction tables to trained LSTM / Transformer models, with column-type inference, encoding strategy proposal, per-customer sequence construction, training, tuning (Optuna), and holdout evaluation.

Installation

pip install autoseqmodels

Or from a local clone:

pip install -e .

Expected input format

build_transaction_panel does not infer column roles — auto-detection silently degrades if the inputs don't match these expectations:

• tx_df (raw transactions, one row per purchase) must contain a customer-identifier column (the name you pass as merge_on, e.g. "Id") and exactly one date-typed column (either datetime64 dtype, or a string column whose name contains "date" and parses as a date). Any other columns are ignored at panel-building time but flow through to the panel and are typed in step 3.
• cov_df (optional covariate calendar) must contain the same merge_on column, exactly one date-typed column (the per-week observation date), and one row per (customer, week). Merge-key dtypes are coerced automatically across the two tables, so an int64 Id on one side and a string Id on the other is fine.
• The downstream build_transaction_sequences step expects the panel to carry a transaction_count column (added by build_transaction_panel) plus the customer and date columns; everything else is treated as a feature.

Workflow

Steps 1-7 prepare the data and are identical for both training paths. Step 8 has two variants — pick one:

• 8a trains a single model with hyperparameters you set yourself.
• 8b runs an Optuna search, then refits with the best trial.

Steps 1-7 — data preparation (shared)

from autoseqmodels import (
    loader, inspection, encoders, sequence_builder,
)

# 1. Load data (CSV / Excel / RData)
df = loader.load_table("transactions.csv")

# 2. Aggregate raw transactions to a (customer, week) panel
#    Two modes:
#      a) With a covariate calendar — pass tx_df + cov_df:
panel = loader.build_transaction_panel(tx_df, cov_df, merge_on="Id")
#      b) Transactions only (no covariates) — omit cov_df:
panel = loader.build_transaction_panel(tx_df, merge_on="Id")
#    The transactions-only mode returns one row per (customer, purchase-week);
#    sequence_builder fills zero-transaction weeks itself in step 7.

# 3. Detect column types (user-editable)
detected = inspection.infer_column_types(panel)
panel    = inspection.cast_columns_by_detected_type(panel, detected)

# 4. Resolve entity / date / target + covariate plan
structure, plan = inspection.analyze_structure(panel, detected)

# 5. Propose encoding strategy (user-editable)
strategy = encoders.propose_encodings(panel, detected, plan)

# 6. Fit encoders on training rows only
enc_df, spec = encoders.apply_encodings(panel, strategy, ...)
plan         = encoders.expand_plan(plan, spec)

# 7. Build per-customer sequences
seqs = sequence_builder.build_transaction_sequences(enc_df, ...)

Step 8a — train with fixed hyperparameters

Use this when you already know the hyperparameters you want or just need a quick baseline. No Optuna involved.

from autoseqmodels import training, sequence_lstm

ds                = training.SequenceDataset(seqs)
train_ds, val_ds  = training.make_train_val_split(ds, val_fraction=0.1, seed=42)

model = sequence_lstm.SequenceLSTM.from_sequences(
    seqs, hidden_size=128, n_layers=1, dropout=0.1,
)
history = training.train_model(model, train_ds, val_ds, lr=1e-3, epochs=30)
preds   = sequence_lstm.predict_holdout(model, seqs)

Step 8b — train with Optuna hyperparameter search

tune_lstm runs the search; train_tuned_lstm is the post-tuning refit and requires the resulting study plus the same train/val subsets. The study persists to SQLite when storage is given, so re-running the same study_name resumes from where it left off.

from autoseqmodels import training, sequence_lstm

ds                = training.SequenceDataset(seqs)
train_ds, val_ds  = training.make_train_val_split(ds, val_fraction=0.1, seed=42)

study = sequence_lstm.tune_lstm(
    seqs,
    train_ds,
    val_ds,
    n_trials   = 30,
    max_epochs = 100,
    storage    = "sqlite:///optuna_lstm.db",
    study_name = "lstm_v1",
)

model, history = sequence_lstm.train_tuned_lstm(seqs, study, train_ds, val_ds)
preds          = sequence_lstm.predict_holdout(model, seqs)

The Transformer variant mirrors both step-8 paths under sequence_transformer: build with SequenceTransformer.from_sequences

• training.train_model (8a), or tune_transformer → train_tuned_transformer → predict_holdout_transformer (8b).

Minimal runnable example

End-to-end on the bundled Electronic.csv (Id, Date, Price, 829 customers, transactions from 1999-01-01 to 2004-11-30). Uses the transactions-only mode of build_transaction_panel so no covariate file is needed; the same script works with a covariate calendar by passing it as the second positional argument.

from autoseqmodels import (
    loader, inspection, encoders, sequence_builder,
    training, sequence_lstm,
)

# 1-2. Load and aggregate to a (customer, week) panel
tx_df = loader.load_table("Datasets/Electronic.csv")
panel = loader.build_transaction_panel(tx_df, merge_on="Id")

# 3. Detect column types and cast
detected = inspection.infer_column_types(panel)
panel    = inspection.cast_columns_by_detected_type(panel, detected)

# 4. Resolve entity / date / target + covariate plan
structure, plan = inspection.analyze_structure(
    panel, detected, target_col="transaction_count"
)

# 5-6. Encoding strategy + fit on training rows
strategy        = encoders.propose_encodings(panel, detected, plan)
train_mask      = panel["Date"] <= "2003-12-31"
enc_df, spec    = encoders.apply_encodings(panel, strategy, train_mask=train_mask)
plan            = encoders.expand_plan(plan, spec)

# 7. Build per-customer sequences (3-year calibration, 11-month holdout)
seqs = sequence_builder.build_transaction_sequences(
    enc_df,
    account_col     = "Id",
    date_col        = "Date",
    training_start  = "2001-01-01",
    training_end    = "2003-12-31",
    holdout_start   = "2004-01-01",
    holdout_end     = "2004-11-30",
    transaction_col = "transaction_count",
    plan            = plan,
    seasonality     = ["woy"],
)

# 8. Tune with Optuna, refit on the best trial, then roll out the holdout
ds                = training.SequenceDataset(seqs)
train_ds, val_ds  = training.make_train_val_split(ds, val_fraction=0.1, seed=42)

study = sequence_lstm.tune_lstm(
    seqs, train_ds, val_ds,
    n_trials   = 10,
    max_epochs = 30,
    storage    = "sqlite:///optuna_lstm.db",
    study_name = "electronic_demo",
)
model, history = sequence_lstm.train_tuned_lstm(seqs, study, train_ds, val_ds)
preds          = sequence_lstm.predict_holdout(model, seqs)

To verify the install in a few minutes, drop n_trials and max_epochs to small values (e.g. 3 and 10). Re-running the same script with the same study_name resumes the existing Optuna study from the SQLite file.

Modules

loader — input I/O and panel aggregation

• load_table(path, ...) — read .csv, .xlsx/.xls, .RData/.rda. For R files, r_covariates_object_name (and optional r_base_object_name) returns a (base_df, covariates_df) tuple in one call.
• build_transaction_panel(tx_df, cov_df=None, merge_on=...) — aggregate raw transactions into a weekly (customer, week) panel. With cov_df, left-joins counts onto a covariate calendar (missing weeks → 0). Without cov_df, returns one row per (customer, purchase-week); the sequence builder fills zero weeks itself. Auto-detects the date column and reconciles merge-key dtypes across the two tables.

inspection — column typing & structural analysis

• infer_column_types(df, config=TypeDetectionConfig()) — Auto-Prep-style detection: id / bool / date / time / category / string, plus a statistical role (identifier / discrete / continuous / binary / categorical / temporal / text). Output is a typed summary DataFrame (detected) the user can edit before casting.
• cast_columns_by_detected_type(df, detected) — apply the inferred dtypes column by column.
• propose_structure(detected) → DataStructure — pick entity_col, date_col (highest-cardinality date), and surface alternatives as entity_candidates / date_candidates.
• classify_time_variance(df, entity_col, detected) — label every column as invariant / variant per entity, and pick the main transaction date.
• plan_covariates(df, detected, structure) → CovariatePlan — classify every column as static_cols, time_varying_cols, or skip_cols, with per-column encoding hints.
• analyze_structure(df, detected, unknown_future_cols=None, target_col=None) — one-shot wrapper that runs propose_structure + plan_covariates and removes columns whose future values aren't known at prediction time from time_varying_cols.

encoders — turning columns into model-ready numerics

• propose_encodings(df, detected, plan, ...) — pick a strategy per column: EMBED (id or high-cardinality category), ONE-HOT (low-cardinality), SCALE (numeric, Standard or MinMax), or PASSTHROUGH (binary). Embedding dimension follows the Guo & Berkhahn rule. Strategy is user-editable.
• apply_encodings(df, strategy, training_mask=...) — fit encoders on training rows only, transform the whole panel, return (enc_df, spec) where spec: EncodingSpec carries the embedding vocabularies, scaler parameters, one-hot category orders, and the total input_width.
• expand_plan(plan, spec) — rewrite the CovariatePlan so one-hot expansion columns replace their source column.
• prepare_encodings(...) / encoding_report(spec) — convenience wrapper and a human-readable summary table.

sequence_builder — per-customer weekly sequences

• build_transaction_sequences(df, *, account_col, date_col, training_start, training_end, holdout_start, holdout_end, transaction_col="transaction_count", plan=None, covariate_cols=None, time_varying_cols=None, seasonality=None, clip_transactions=None) — constructs its own universal weekly calendar from (training_start, holdout_end), merges each customer's transactions onto it (zero-filling missing weeks), and emits a dict with samples / targets (calibration shifted by one), calibration (full training tensor), holdout (rollout tensor), account_ids, n_features, feature_names, total_transactions. Optional seasonality features: woy / moy / dow / year (sin/cos pairs).

models.training — dataset, split, train loop, evaluation

• SequenceDataset — wraps the dict from build_transaction_sequences for use with DataLoader.
• make_train_val_split(dataset, val_fraction=..., seed=...) — customer-level random split (no leakage between train and val).
• train_model(model, train_ds, val_ds, ...) — generic training loop with early stopping, returns the per-epoch loss history.
• evaluate_predictions(...) — MAE / RMSE / aggregate-count metrics on the holdout rollout.
• plot_training_and_predictions(...) — quick-look loss curves and predicted-vs-actual transaction plots.

models.sequence_lstm — LSTM model + Optuna tuning

• SequenceLSTM(nn.Module) — embedding layers (one per id / categorical), multi-layer LSTM, regression head; consumes the feature_names order produced by build_transaction_sequences.
• predict_holdout(model, seqs) — autoregressive holdout rollout, returns predicted weekly counts per customer.
• tune_lstm(seqs, train_ds, val_ds, *, n_trials, max_epochs, storage=None, study_name=None, fixed_params=None) — Optuna search over hidden_size, n_layers, dropout, learning rate, batch size, … Studies persist to SQLite (sqlite:///optuna_lstm.db) when storage is given, so runs are resumable. Any param can be pinned via fixed_params.
• train_tuned_lstm(seqs, study, train_ds, val_ds) — refit with the best trial's hyperparameters, returning (model, history).

models.sequence_transformer — Transformer mirror

• SequenceTransformer(nn.Module) — same input contract as SequenceLSTM but with positional encoding and a TransformerEncoder stack; n_heads is auto-picked to divide d_model.
• predict_holdout_transformer, tune_transformer, train_tuned_transformer — direct counterparts of the LSTM helpers, with their own SQLite study (optuna_transformer.db).

Requirements

Python ≥ 3.10. See pyproject.toml for the full dependency list.

License

MIT — see LICENSE.