Train Model Page (16/11/2017)

[FireFlyTy]

CONTROLS:

• Model Selection (Dynamic Thresholds (default), Prophet):
• Model Train section
• Model Tuning section

---

DYNAMIC THESHOLDS

Model Train section

• Train Mode (Simple (default), Expert):
  • Simple Mode
    • Sensitivity (Low, Medium (default), High)
  • Expert Mode
    • Thresholds Aggregation Level [0.05, 0.9] (0.6 - default)
    • Local Trend Basis [0.05, 0.9] (0.3 - default)
    • Neighbours similarity [0.01, 0.99] (0.1 - default)
• Train button

Model Tuning section

• Low-type Anomalies (if Low or Both anomalies type)
  • Correction [-1, 1] (0 - default)
  • Scale [0.25, 4] (1 - default)
• High-type Anomalies (if High or Both anomalies type)
  • Correction [-1, 1] (0 - default)
  • Scale [0.25, 4] (1 - default)
• Auto-Tune button
• Precent of anomaly cases for auto turning [0.001:0.05] (0.01 - default)

ACTIONS

Train execution - caused by Train button:

1. Train Model:
   • Simple Mode:
     MODEL <- dynamicThreshold.train( ts.agg = list( data.agg = <aggregated time series>, ts_type=<aggregation type - days, hours, etc>, ts_val=<aggregation step>, ts_func = <aggregation function>), train.params = list(mode = 'simple', sensitivity = <Sensitivity>), type_th = 'Both')
   • Expert Mode:
     MODEL <- dynamicThreshold.train( ts.agg = list( data.agg = <aggregated time series - train set>, ts_type=<aggregation type - days, hours, etc>, ts_val=<aggregation step>, ts_func = <aggregation function>), train.params = list(mode = 'expert', params = list(agg_th = <Thresholds Aggregation Level >, local_trend = <Local Trend Basis>, similar = <Neighbours similarity>)), type_th ='Both')
2. Apply Model:
   RES = dynamicThreshold.apply( ts.agg = list( data.agg = <aggregated time series - train set>, ts_type=<aggregation type - days, hours, etc>, ts_val=<aggregation step>, ts_func = <aggregation function>), model = MODEL, type_th = 'Both', correction =list("Low" = c(coef = 0, scale = 1), "High" = c(coef = 0, scale = 1)))
3. Show Results with interpretation:
   • plot_time_series(RES, treshhold_type = 'both')
   • ANAYSIS.STAT = anomalies.stat(ad_results = RES, data, ts_type=<aggregation type - days, hours, etc>, ts_val=<aggregation step>)
   • anomalies.detail(ANAYSIS.STAT$ad_res[index,], data, ts_func = <aggregation function>)

Model Manual Tuning - caused by Correction and Scale sliders:

1. RES = dynamicThreshold.apply( ts.agg = list( data.agg = <aggregated time series - train set>, ts_type=<aggregation type - days, hours, etc>, ts_val=<aggregation step>, ts_func = <aggregation function>), model = MODEL, type_th = 'Both', correction =list("Low" = c(coef = <Low-type Anomalies: Correction>, scale = <Low-type Anomalies: Scale>), "High" = c(coef = < High-type Anomalies: Correction >, scale = <High-type Anomalies: Scale>)))
2. Show Results with interpretation

Model Auto Tuning - caused by Auto-Tune button:

1. Calculate correcting coefficients:
   COEF = dynamicThreshold.autoturn( ts.agg = list( data.agg = <aggregated time series - train set>, ts_type=<aggregation type - days, hours, etc>, ts_val=<aggregation step>, ts_func = <aggregation function>), model = MODEL, type_th = 'Both', p_anomalies = <Precent of anomaly cases for auto turning>)
2. Set values for corresponded sliders:
   • <Low-type Anomalies: Correction> = COEF$Low[1]
   • <Low-type Anomalies: Scale> = COEF$Low[2]
   • <High-type Anomalies: Correction> = COEF$High[1]
   • <High-type Anomalies: Scale> = COEF$High[2]
3. Apply Model:
   RES = dynamicThreshold.apply( ts.agg = list( data.agg = <aggregated time series - train set>, ts_type=<aggregation type - days, hours, etc>, ts_val=<aggregation step>, ts_func = <aggregation function>), model_ = MODEL, type_th = 'Both', correction =list("Low" = c(coef = < Low-type Anomalies: Correction >, scale = < Low-type Anomalies: Scale >), "High" = c(coef = <High-type Anomalies: Correction>, scale = <High-type Anomalies: Scale>)))
4. Show Results with interpretation

---

PROPHET
Model Train section

• Train Mode (Simple (default), Expert):
  • Simple Mode
  • Expert Mode
    • yearly seasonality ('auto', TRUE, FALSE)
    • weekly seasonality ('auto', TRUE, FALSE)
    • daily seasonality ('auto', TRUE, FALSE)
    • uncertainty intervals [0.6, 0.99] (0.9 - default)
• Train button

Model Tuning section

• Low-type Anomalies (if Low or Both anomalies type)
  • Scale [-100, 100] (0 - default)
• High-type Anomalies (if High or Both anomalies type)
  • Scale [-100, 100] (0 - default)

ACTIONS
Train execution - caused by Train button:

1. Train Model:
   • Simple Mode:
     MODEL <- model_prophet_train_test(< aggregated time series - train >, < aggregated time series - test>, method = 'train')
   • Expert Mode:
     MODEL = model_prophet_train_test( < aggregated time series - train >, < aggregated time series - test>, yearly=< yearly seasonality >, weekly=< weekly seasonality >, daily=< daily seasonality >, interval_width=< uncertainty intervals > method = 'train')
2. Apply Model:
   • RES <- model_prophet_new_interval(MODEL$data_train, method='Both')
3. Tuning Model:
   • RES_tune <- model_prophet_new_interval(RES, percent_up=< Scale UP>, percent_low=< Scale LOW>, method='Both')
4. Show Results with interpretation:
   • plot_time_series(RES, treshhold_type = 'both') or
     plot_time_series(RES_tune, treshhold_type ='both')
   • ANAYSIS.STAT = anomalies.stat(ad_results = RES, data, ts_type=< aggregation type - days, hours, etc >, ts_val=< aggregation step >)
   • anomalies.detail(aANAYSIS.STAT$ad_res[index,], data, ts_func = < aggregation function >)

Model Manual Tuning - caused by Scale sliders:

1. RES <- model_prophet_new_interval( MODEL$data_train, percent_up= < Low-type Anomalies Scale >, percent_low= < High-type Anomalies Scale >, method='Both')
2. Show Results with interpretation