Anomaly detection-Platform For AI(PAI)-阿里云帮助中心

The Anomaly Detection component identifies data points or patterns in a dataset that significantly deviate from normal behavior. It supports two detection methods: Box Plot for continuous (numeric) features and Attribute Value Frequency (AVF) for categorical (enumeration) features, making it suitable for detecting errors, frauds, and outliers across a wide range of datasets.

Choose a detection method

Select the method that matches your column types before configuring the component.

Method	Feature type	Use when
Box Plot	Continuous (numeric)	Columns contain numeric measurements, such as prices, rates, or counts
Attribute Value Frequency (AVF)	Categorical (enumeration)	Columns contain discrete categories, such as status codes, labels, or identifiers

Each method applies only to columns that match its feature type.

Configure the component

Method 1: Configure on the pipeline page (recommended)

In Machine Learning Designer, add the Anomaly Detection component to your pipeline, then set the following parameters in the component panel.

Parameter	Description
Feature Columns	The columns to include in anomaly detection. Select all columns that belong to the same feature type.
Anomaly Detection Method	The detection method. Select Box Plot for continuous features or Attribute Value Frequency (AVF) for categorical features.

Method 2: Use PAI commands

Run the Anomaly Detection component through Platform for AI (PAI) commands. Use the SQL Script component to call PAI commands. For more information, see SQL Script (Scenario 4: Execute PAI commands within the SQL script component).

PAI -name fe_detect_runner -project algo_public
     -DselectedCols="emp_var_rate,cons_price_rate,cons_conf_idx,euribor3m,nr_employed" \
     -Dlifecycle="28"
     -DdetectStrategy="boxPlot"
     -DmodelTable="pai_temp_2458_23565_2"
     -DinputTable="pai_bank_data"
     -DoutputTable="pai_temp_2458_23565_1";

Parameters

Parameter	Required	Type	Default	Valid values	Description
`inputTable`	Yes	String	—	—	The name of the input table.
`inputTablePartitions`	No	String	All partitions	—	The partitions in the input table. If not specified, all partitions are used. See Partition format for syntax.
`selectedCols`	Yes	String	—	—	The feature columns to analyze. Accepts any data type.
`detectStrategy`	Yes	String	—	`boxPlot`, `avf`	The detection method. Use `boxPlot` for continuous features or `avf` for categorical features.
`outputTable`	Yes	String	—	—	The name of the output table. Contains data with anomalous features.
`modelTable`	Yes	String	—	—	The name of the table used to store the anomaly detection model.
`lifecycle`	No	Integer	7	—	The lifecycle of the output table.
`coreNum`	No	Integer	—	1–9999	The number of cores. Must be set together with `memSizePerCore`.
`memSizePerCore`	No	Integer (MB)	—	2048–65536	The memory allocated per core. Must be set together with `coreNum`.

Partition format

Use the inputTablePartitions parameter to filter input data by partition.

Scenario	Format	Example
Single partition	`partition_name=value`	`ds=20240101`
Multiple partitions	`name1=value1,name2=value2`	`ds=20240101,region=cn`
Multi-level partitions	`name1=value1/name2=value2`	`year=2024/month=01`

Separate multiple partitions with commas (,).

Output

The outputTable contains data with anomalous features from the input table. After the component runs, query the output table to review the flagged records and determine which data points deviate from normal patterns.