Prize-Collecting Steiner Tree (PCST)

Overview

The Prize-Collecting Steiner Tree (PCST) algorithm finds a subtree that balances two competing goals: collecting node prizes and minimizing edge costs. Unlike the standard Steiner tree which must connect all specified terminals, PCST decides which nodes are worth including based on the trade-off between their prize value and the cost of reaching them.

Applications:

• Network design: Deciding which customers to connect when wiring is expensive.
• Feature selection: Selecting a connected subgraph of high-value nodes in biological networks.
• Revenue optimization: Balancing service coverage cost against customer value.

Concepts

Prize-Collecting Steiner Tree

Each node has a prize (value gained by including it) and each edge has a cost (price paid to use it). The algorithm finds a connected subtree that maximizes total prize - λ × total cost.

The λ parameter controls the trade-off:

• High lambda: Penalizes edge costs more → smaller tree, only high-prize nodes included.
• Low lambda: Penalizes costs less → larger tree, more nodes included.

The algorithm works by:

1. Build a minimum spanning tree of the graph, connecting all nodes.
2. Examine each leaf node (node with only one edge in the tree). If the leaf's prize is less than λ × edge cost, remove it — the cost of reaching it outweighs its value.
3. Repeat step 2 — removing a node may create new leaf nodes that can also be pruned.
4. Stop when every remaining leaf node is worth keeping (prize ≥ λ × edge cost).

Considerations

• The algorithm treats all edges as undirected.

Example Graph

GQL
INSERT (A:default {_id: "A", prize: 10}), (B:default {_id: "B", prize: 5}),
       (C:default {_id: "C", prize: 8}), (D:default {_id: "D", prize: 3}),
       (E:default {_id: "E", prize: 12}), (F:default {_id: "F", prize: 6}),
       (G:default {_id: "G", prize: 4}), (H:default {_id: "H", prize: 2}),
       (A)-[:default {distance: 1}]->(B), (A)-[:default {distance: 2.4}]->(C),
       (A)-[:default {distance: 1.2}]->(D), (A)-[:default {distance: 0.7}]->(E),
       (A)-[:default {distance: 2.2}]->(F), (A)-[:default {distance: 1.6}]->(G),
       (A)-[:default {distance: 0.4}]->(H), (B)-[:default {distance: 1.3}]->(C),
       (C)-[:default {distance: 1}]->(D), (D)-[:default {distance: 1.65}]->(H),
       (E)-[:default {distance: 1.27}]->(F), (E)-[:default {distance: 0.9}]->(G),
       (F)-[:default {distance: 0.45}]->(G)

Parameters

Run Mode

Returns:

GQL
CALL algo.pcst({
  prizeProperty: "prize",
  weightProperty: "distance",
  lambda: 1
}) YIELD nodeId, inTree, totalCost, totalPrize

Stream Mode

Returns the same columns as run mode, streamed for memory efficiency.

GQL
CALL algo.pcst.stream({
  prizeProperty: "prize",
  weightProperty: "distance",
  lambda: 1
}) YIELD nodeId, inTree
FILTER inTree = true
RETURN nodeId

Stats Mode

Returns:

GQL
CALL algo.pcst.stats({
  prizeProperty: "prize",
  weightProperty: "distance",
  lambda: 1
}) YIELD nodeCount, nodesInTree, totalCost, totalPrize