Using Neo4j for Heterogeneous nodes link prediction

alanwtmec · March 28, 2024, 6:00am

Dear neo4j developer

I have a question, I want to use Neo4j for Heterogeneous nodes link prediction.
There are 2 types of node labels, and I need to predict the connections between these 2 types of nodes.
Do you have any actual code examples for this?"

I tested using the example from Training the pipeline - Neo4j Graph Data Science.
Then I used my own data (code as follows:), and made many changes (e.g. sourceNodeLabel: 'Person', targetNodeLabel: 'Product', targetRelationshipType: 'BUY').
However, the results seemed unreasonable, with all probabilities being 0.4999.

Could you please advise how I should modify the code? Where did I go wrong? Are there any other examples I can refer to? Thank you.

code: (below)

CREATE
  (alice:Person {name: 'Alice', age: 38, gender:1}),
  (michael:Person {name: 'Michael', age: 67, gender:0}),
  (karin:Person {name: 'Karin', age: 30, gender:1}),
  (chris:Person {name: 'Chris', age: 52, gender:0}),
  (will:Person {name: 'Will', age: 18, gender:0}),
  (mark:Person {name: 'Mark', age: 20, gender:0}),
 
  (milk:Product {name: 'milk', price:80}),
  (apple:Product {name: 'apple', price:25}),
  (orange:Product {name: 'orange', price:35}),
  (piapple:Product {name: 'piapple', price:65}),
  (watermelon:Product {name: ' watermelon', price:45}),
  (cantaloupe:Product {name: 'cantaloupe', price:55}),
  (notebook:Product {name: 'notebook', price:25000}),
  (tv:Product {name: 'TV', price:19999}),
  (aircon:Product {name: 'aircondition', price:29000}),
  (shoe_air:Product {name: 'shoe_air', price:5000}),
  (runshoes:Product {name: 'runshoes', price:1999}),
  (sportshoes:Product {name: 'sportshoes', price:3900}),
  (goodshoes:Product {name: 'goodshoes', price:7900}),
  (underwear:Product {name: 'Underwear', price:390}),
  (pants:Product {name: 'Pants', price:499}),
  (jen:Product {name: 'Jen', price:600}),

  (fruit:fruit {name: 'fruit'}),
  (ele:ele {name: 'ele'}),
  (shoe:shoe {name: 'shoe'}),
  (clothe:clothe {name: 'clothe'}),
  
  (milk)-[:BELONG]->(fruit),
  (apple)-[:BELONG]->(fruit),
  (orange)-[:BELONG]->(fruit),
  (piapple)-[:BELONG]->(fruit),
  (watermelon)-[:BELONG]->(fruit),
  (cantaloupe)-[:BELONG]->(fruit),
  (notebook)-[:BELONG]->(ele),
  (tv)-[:BELONG]->(ele),
  (aircon)-[:BELONG]->(ele),
  (shoe_air)-[:BELONG]->(shoe),
  (runshoes)-[:BELONG]->(shoe),
  (sportshoes)-[:BELONG]->(shoe),
  (goodshoes)-[:BELONG]->(shoe),
  (underwear)-[:BELONG]->(clothe),
  (pants)-[:BELONG]->(clothe),
  (jen)-[:BELONG]->(clothe),    

  (alice)-[:BUY]->(milk),
  (alice)-[:BUY]->(apple),
  (alice)-[:BUY]->(orange),
  (alice)-[:BUY]->(piapple), 
  (michael)-[:BUY]->(apple),
  (michael)-[:BUY]->(orange),
  (michael)-[:BUY]->(notebook),
  (michael)-[:BUY]->(shoe_air),
  (karin)-[:BUY]->(milk),
  (karin)-[:BUY]->(pants),
  (karin)-[:BUY]->(jen),
  (karin)-[:BUY]->(underwear),
  (chris)-[:BUY]->(sportshoes),
  (chris)-[:BUY]->(runshoes),
  (chris)-[:BUY]->(goodshoes),  
  (will)-[:BUY]->(aircon),
  (will)-[:BUY]->(tv),
  (will)-[:BUY]->(notebook),
  (mark)-[:BUY]->(watermelon),
  (mark)-[:BUY]->(cantaloupe),
  
  (alice)-[:FRIEND]->(michael),
  (michael)-[:FRIEND]->(will),
  (michael)-[:FRIEND]->(mark),
  (alice)-[:FRIEND]->(karin),
  (will)-[:FRIEND]->(mark),
  (mark)-[:FRIEND]->(chris),
  (chris)-[:FRIEND]->(alice),
  (will)-[:FRIEND]->(karin)



CALL gds.graph.project(
  'fullGraph',
  {
    Person: 
    {
      properties: 
        {
        age: {defaultValue: 1},
        price: {defaultValue: 1}
        }    
    },
    Product: 
    {
      properties: 
        {
        age: {defaultValue: 1},
        price: {defaultValue: 1}
        }
    }
  },
  {
    BUY: {
      orientation: 'UNDIRECTED'
    },
    FRIEND: {},
    BELONG: {}
  }
)

CALL gds.beta.pipeline.linkPrediction.create('pipe-with-context')

CALL gds.beta.pipeline.linkPrediction.addNodeProperty('pipe-with-context', 'fastRP', {
  mutateProperty: 'embedding',
  embeddingDimension: 256,
  randomSeed: 42,
  contextRelationshipTypes: ['FRIEND', 'BELONG']
})

CALL gds.beta.pipeline.linkPrediction.addFeature('pipe-with-context', 'hadamard', {
  nodeProperties: ['embedding', 'age', 'price']
})

CALL gds.beta.pipeline.linkPrediction.configureSplit('pipe-with-context', {
  testFraction: 0.25,
  trainFraction: 0.6,
  validationFolds: 3
})

CALL gds.alpha.pipeline.linkPrediction.addMLP('pipe-with-context',
{hiddenLayerSizes: [4, 2], penalty: 1, patience: 2})

CALL gds.beta.pipeline.linkPrediction.train('fullGraph', {
  pipeline: 'pipe-with-context',
  modelName: 'lp-pipeline-model-filtered',
  metrics: ['AUCPR', 'OUT_OF_BAG_ERROR'],
  sourceNodeLabel: 'Person',
  targetNodeLabel: 'Product',
  targetRelationshipType: 'BUY',
  randomSeed: 12
}) YIELD modelInfo, modelSelectionStats
RETURN
  modelInfo.bestParameters AS winningModel,
  modelInfo.metrics.AUCPR.train.avg AS avgTrainScore,
  modelInfo.metrics.AUCPR.outerTrain AS outerTrainScore,
  modelInfo.metrics.AUCPR.test AS testScore,
  [cand IN modelSelectionStats.modelCandidates | cand.metrics.AUCPR.validation.avg] AS validationScores
  
CALL gds.beta.pipeline.linkPrediction.predict.stream('fullGraph', {
  modelName: 'lp-pipeline-model-filtered',
  topN: 50,
  threshold: 0
})
 YIELD node1, node2, probability
 RETURN gds.util.asNode(node1).name AS Person, gds.util.asNode(node2).name AS Product, probability
 ORDER BY Product

Result : (below)

╒═══════════╤══════════════╤══════════════════╕
│Person     │Product       │probability       │
╞═══════════╪══════════════╪══════════════════╡
│"Alice"    │" watermelon" │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Alice"    │"cantaloupe"  │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Alice"    │"TV"          │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Alice"    │"shoe_air"    │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Alice"    │"sportshoes"  │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Alice"    │"runshoes"    │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Alice"    │"aircondition"│0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Alice"    │"notebook"    │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Chris"    │"milk"        │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Chris"    │"cantaloupe"  │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Chris"    │"TV"          │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Chris"    │"shoe_air"    │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Chris"    │"aircondition"│0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Chris"    │"notebook"    │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Chris"    │" watermelon" │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Chris"    │"piapple"     │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Chris"    │"orange"      │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Chris"    │"apple"       │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Karin"    │"piapple"     │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Karin"    │"cantaloupe"  │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Karin"    │"TV"          │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Karin"    │"shoe_air"    │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Karin"    │"sportshoes"  │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Karin"    │"runshoes"    │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Karin"    │"aircondition"│0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Karin"    │"notebook"    │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Karin"    │" watermelon" │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Karin"    │"orange"      │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Karin"    │"apple"       │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Mark"     │"milk"        │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Mark"     │"orange"      │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Mark"     │"TV"          │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Mark"     │"shoe_air"    │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Mark"     │"runshoes"    │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Mark"     │"sportshoes"  │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Mark"     │"aircondition"│0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Mark"     │"notebook"    │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Mark"     │"piapple"     │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Mark"     │"apple"       │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Pants"    │"Alice"       │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Pants"    │"Michael"     │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Pants"    │"Will"        │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Pants"    │"Mark"        │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Pants"    │"Chris"       │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Underwear"│"Alice"       │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Underwear"│"Michael"     │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Underwear"│"Mark"        │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Underwear"│"Will"        │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Underwear"│"Chris"       │0.4999999999999998│
├───────────┼──────────────┼──────────────────┤
│"Will"     │"milk"        │0.4999999999999998│
└───────────┴──────────────┴──────────────────┘

Neo4j Desktop-1.5.7
Version 5.18.0Preformatted text

adam_schill-col · April 2, 2024, 9:14am

Hi @alanwtmec,

The output does indeed look a bit fishy. Unfortunately there are no other examples I'm aware of that does this.

What is the output of the call to gds.beta.pipeline.linkPrediction.train? Do the scores seem resonable at that point in the pipeline?

It looks like you're leaving out some node labels from your projection: fruit, ele, shoe, clothe. Should these perhaps share a node label instead? That should then also be included in the projection?

Apart from that the code looks ok to me. But it's very noisy to train a model on such a small graph, so hard to predict what you're going to get. To see if you can get interesting scores, here are some things to experiment with:

Use a different training method, like logistic regression or random forest
Try different parameters for the MLP
Try using fewer features for the pipeline, maybe only 'embedding'
Experiment with different default values for age and price

Hope this is helpful,
Adam

alanwtmec · April 3, 2024, 2:36am

Hi, Adam
Thank you for your reply.
Below is output of gds.beta.pipeline.linkPrediction.train.

I have tried another dataset( the Hetionet dataset), but the answer is still all 0.49999.
Is there something I haven't noticed?
Thank you!

adam_schill-col · April 3, 2024, 7:42am

Hi again @alanwtmec,

So the train output looks reasonable, that's good.

It's concerning that you get strange scores also on the other dataset.

Is there something I haven't noticed?

I cannot see anything more than what I mentioned in my last post. Did you try any of the things I suggested as experiments? I wonder if there's something fishy going on with the MLP implementation.

Thanks,
Adam

Topic		Replies	Views
Link prediction predicting wrong links Neo4j Graph Platform migrated	8	258	July 18, 2022
Does Neo4j GraphSage work for Heterogeneous graphs? Graph Algorithms/Graph Data Science	5	665	October 18, 2021
Finding associations between nodes from different types General	0	230	November 8, 2021
Online Meetup: Link Prediction with Neo4j Community Content & Blogs	15	2362	May 25, 2021
Link Prediction with Neo4j Part 1: An Introduction Neo4j Developer Blog Archive	3	2346	November 27, 2019

Using Neo4j for Heterogeneous nodes link prediction

Related topics