Bảng ML Pipeline

TÓM TẮT

12 bảng hỗ trợ toàn bộ ML lifecycle: training → prediction → backtest → self-improvement → audit.

forecast_results — Kết quả dự báo

CREATE TABLE forecast_results (
    id                  bigserial PRIMARY KEY,
    series_key          text NOT NULL,
    client_id           text NOT NULL,
    product_id          bigint,
    forecast_date       date NOT NULL,
    predicted_quantity  numeric NOT NULL,
    confidence_lower    numeric,        -- P10
    confidence_upper    numeric,        -- P90
    event_tier          int,
    model_version       text,
    created_at          timestamptz DEFAULT now()
);

training_jobs — Tracking Training

CREATE TABLE training_jobs (
    id              bigserial PRIMARY KEY,
    job_id          uuid UNIQUE NOT NULL,
    client_id       text NOT NULL,
    status          text NOT NULL DEFAULT 'pending',
    started_at      timestamptz,
    completed_at    timestamptz,
    metrics         jsonb,          -- {wmape, bias, fnr, mae}
    model_version   text,
    error_message   text
);

backtest_results — Walk-Forward Validation

CREATE TABLE backtest_results (
    id              bigserial PRIMARY KEY,
    result_id       uuid UNIQUE NOT NULL,
    cycle_id        uuid,                    -- FK → improvement_cycles
    client_id       text NOT NULL,
    forecast_level  text NOT NULL DEFAULT 'order',  -- 'order' | 'sku'
    fold_index      int,
    train_start     date,
    train_end       date,
    test_start      date,
    test_end        date,
    metrics         jsonb NOT NULL,          -- {wmape, bias, fnr, mae, event_wmape...}
    model_version   text
);

improvement_cycles — AutoML State Machine

CREATE TABLE improvement_cycles (
    id              bigserial PRIMARY KEY,
    cycle_id        uuid UNIQUE NOT NULL,
    client_id       text NOT NULL,
    state           text NOT NULL DEFAULT 'PROFILING',
    -- States: PROFILING → BACKTEST → ANALYZE → SUGGEST → APPLY → RETRAIN → VERIFY → COMPLETE
    max_iterations  int DEFAULT 5,
    current_iteration int DEFAULT 0,
    convergence_threshold numeric DEFAULT 0.005,
    mape_before     numeric,
    mape_after      numeric,
    config_snapshot jsonb,
    error_message   text
);

llm_suggestions — Gợi ý AI

CREATE TABLE llm_suggestions (
    id              bigserial PRIMARY KEY,
    suggestion_id   uuid UNIQUE NOT NULL,
    cycle_id        uuid,
    client_id       text NOT NULL,
    category        text NOT NULL,
    -- Categories: hyperparameter, feature, event, calendar, cross_audit, other
    summary         text NOT NULL,
    reasoning       text,
    parameter_diff  jsonb,           -- Before/after config diff
    before_metrics  jsonb,
    after_metrics   jsonb,
    status          text NOT NULL DEFAULT 'pending'
    -- Status: pending, applied, rejected, rollback
);

Các bảng phụ trợ

Bảng	Mục đích
model_metadata	Registry model: version, metrics, feature_count, training_samples
sku_metadata	Phân loại SKU: sku_type (dead/intermittent/lumpy/smooth/seasonal), demand_frequency
discovered_events	Sự kiện auto-discovered bởi LLM từ residual analysis
llm_providers	Cấu hình LLM API: provider_type, model_name, API key
model_comparisons	A/B test giữa các model/LLM providers
data_profiles	EDA snapshots: summary_stats, demand_patterns, temporal_patterns
anomaly_log	Cross-audit anomalies giữa order-level và SKU-level forecasts

Tài liệu liên quan

• Schema tổng quan
• AutoML — Cách improvement_cycles được sử dụng
• Bảo mật & RLS