Platform For AI:Label propagation classification

How it works

The algorithm builds a graph where each node is a data point and each edge represents the similarity between two nodes. Labels spread through the graph according to edge weights: a label moves easily across strongly connected nodes (high similarity) but has difficulty crossing weakly connected regions (low similarity). This means nodes in a dense cluster converge to the same label, while sparse connections act as natural barriers between groups.

The propagation follows these steps:

1. Initialize: labeled nodes start with their known labels; unlabeled nodes start with no assignment.

2. Propagate: in each iteration, every node updates its label distribution based on the weighted labels of its neighbors.

3. Anchor: labeled nodes are pinned — their labels never change during propagation, ensuring they remain the source of truth.

4. Converge: iterations continue until label distributions stabilize (change falls below the convergence coefficient) or the maximum number of iterations is reached.

5. Classify: after convergence, each node is assigned the category with the highest probability in its label distribution.

Configure the component

PAI -name LabelPropagationClassification
    -project algo_public
    -DinputEdgeTableName=LabelPropagationClassification_func_test_edge
    -DfromVertexCol=flow_out_id
    -DtoVertexCol=flow_in_id
    -DinputVertexTableName=LabelPropagationClassification_func_test_node
    -DvertexCol=node
    -DvertexLabelCol=label
    -DoutputTableName=LabelPropagationClassification_func_test_result
    -DhasEdgeWeight=true
    -DedgeWeightCol=edge_weight
    -DhasVertexWeight=true
    -DvertexWeightCol=label_weight
    -Dalpha=0.8
    -Depsilon=0.000001;

Example

This example creates a small graph with two labeled nodes (a with label X, d with label Y) and two unlabeled nodes (b and c), then runs label propagation to classify b and c.

1. Add the SQL Script component. Deselect Use Script Mode and Whether the system adds a create table statement. Enter the following SQL statements to create the edge and vertex tables.

   Data structure

   drop table if exists LabelPropagationClassification_func_test_edge;
   create table LabelPropagationClassification_func_test_edge as
   select * from
   (
       select 'a' as flow_out_id, 'b' as flow_in_id, 0.2 as edge_weight
       union all
       select 'a' as flow_out_id, 'c' as flow_in_id, 0.8 as edge_weight
       union all
       select 'b' as flow_out_id, 'c' as flow_in_id, 1.0 as edge_weight
       union all
       select 'd' as flow_out_id, 'b' as flow_in_id, 1.0 as edge_weight
   )tmp
   ;
   drop table if exists LabelPropagationClassification_func_test_node;
   create table LabelPropagationClassification_func_test_node as
   select * from
   (
       select 'a' as node,'X' as label, 1.0 as label_weight
       union all
       select 'd' as node,'Y' as label, 1.0 as label_weight
   )tmp;

2. Add another SQL Script component. Deselect Use Script Mode and Whether the system adds a create table statement. Enter the following PAI command and connect this component to the one from step 1.

   drop table if exists ${o1};
   PAI -name LabelPropagationClassification
       -project algo_public
       -DinputEdgeTableName=LabelPropagationClassification_func_test_edge
       -DfromVertexCol=flow_out_id
       -DtoVertexCol=flow_in_id
       -DinputVertexTableName=LabelPropagationClassification_func_test_node
       -DvertexCol=node
       -DvertexLabelCol=label
       -DoutputTableName=${o1}
       -DhasEdgeWeight=true
       -DedgeWeightCol=edge_weight
       -DhasVertexWeight=true
       -DvertexWeightCol=label_weight
       -Dalpha=0.8
       -Depsilon=0.000001;

3. Click in the upper left corner to run the pipeline.

4. Right-click the SQL Script component from step 2 and choose View Data > SQL Script Output to view the results.

   • Node a (labeled X) and node d (labeled Y) retain their original labels with full confidence (weight = 1.0).

   • Node c is classified as X with approximately 54% probability, meaning it is more likely in X's category.

   • Node b is classified as Y with approximately 83% probability, pulled more strongly toward d through the high-weight edge from d to b (weight 1.0).

   | node | tag | weight              |
   | ---- | --- | ------------------- |
   | a    | X   | 1.0                 |
   | c    | X   | 0.5370370370370371  |
   | c    | Y   | 0.4629629629629629  |
   | b    | X   | 0.16666666666666666 |
   | b    | Y   | 0.8333333333333333  |
   | d    | Y   | 1.0                 |

   The output contains one row per label per node. The weight column shows the probability that a node belongs to that label's category. In this example: