Machine Learning Procedures for Neo4j (WIP)

ToDo:

• Prediction for DL4j RNN

• Implementation with h2o (?)

Description:

My specific version of Machine Learning Stored Procedures for the Neo4j Database.

List of Implementations:

• 'nd4j': Linear Regression with Nd4j

• 'dl4j': Recurrent Neural Networks with Dl4j (WIP)

List of Stored Procedures:

• ml.create

• ml.add

• ml.train

• ml.predict

• ml.remove

• ml.info

• ml.models

New in 0.1.0:

New procedure ml.models to list the existing models in the database

Setup:

• Build the application with Maven

• Location of ZIP file: /target

• Put the ZIP file in the plugins folder of your Neo4j installation directory

• Restart Neo4j

HowTo:

To list all available Stored Procedures in Neo4j Browser execute:

CALL dbms.procedures

This procedures have the prefix "ml.".

How to call the procedures:

CALL ml.create("model", {types}, {params}, {extraAttributes}, timePeriod, "implementation")

MATCH data CALL ml.add("model", {data.features}, data.label) YIELD result RETURN result

CALL ml.train("model")

CALL ml.predict("model", {features})

CALL ml.remove("model")

CALL ml.info("model")

Parameters explained:

• "model": String with the name of the model

• {"types"}: Map of attribute names and their respective data types

• {params}: Map with hyperparameters for special Machine Learning implementation

• {extraAttributes}: Map of attributes with constant values the model shall not be trained with but displayed with the result of the prediction. Possible data types are "numeric" and "class".

• timePeriod: Boolean indicating whether a time period shall be predicted or not

• "implementation": String with the name of the implementation you want to use

• {data.features}: Map that assigns the features of the matched data to the types that were defined in the create call

• data.label: column with label values of the matched data

• {features}: Map with features for prediction (for timePeriod == false the feature names again must match the types defined in the create call), there are several ways to define this parameter:

  • for several predictions at once: {feature1: [f1_value1, f1_value2, …​, f1_valueN], feature2:[f2_value1, f2_value2, …​, f2_valueN], …​}

  • for only one prediction: {feature1: f1_value, feature2: f2_value, …​} or {feature1: [f1_value], feature2: [f2_value], …​} (List is optional)

  • for timePeriod == true: {start: startDate, end: endDate} (must exactly look like this, with the dates in the format like '20170516' / yyyyMMdd)

Examples (only for Linear Regression Implementation with Nd4j):

Predict the user counts for the dates 15.05.2017 and 16.05.2017 ('test' is an additional class feature always set to dummy value "true" and has no effect on the prediction. It is only included for the purpose of the example.):

CALL ml.create('user', {date: "numeric", test: "class"}, {alpha: 0.1, iter: 300, theta: [0.0, 0.0, 0.0]}, null, false, 'nd4j')

MATCH (n:User {stage:"cpd-ece-prod"}) WHERE toInteger(n.date) < 20170515 CALL ml.add('user', {date: n.date, test: "true"}, n.count) YIELD result RETURN result

CALL ml.train('user')

CALL ml.predict('user', {date: [20170515, 20170516], test: ["true", "true"]})

CALL ml.remove('user')

If you want to have such a dummy value outputted with the prediction result it is recommended to use the {extraAttributes} parameter in the Create Procedure instead:

CALL ml.create('user', {date: "numeric"}, {alpha: 0.1, iter: 300, theta: [0.0, 0.0]}, {test: 'true'}, false, 'nd4j')

MATCH (n:User {stage:"cpd-ece-prod"}) WHERE toInteger(n.date) < 20170515 CALL ml.add('user', {date: n.date}, n.count) YIELD result RETURN result

CALL ml.train('user')

CALL ml.predict('user', {date: [20170515, 20170516]})

CALL ml.remove('user')

Predict user counts for the time period from 15.05.2017 to 22.05.2017:

CALL ml.create('user', {date: "numeric"}, {alpha: 0.1, iter: 300, theta: [0.0, 0.0]}, null, true, 'nd4j')

MATCH (n:User {stage:"cpd-ece-prod"}) WHERE toInteger(n.date) < 20170515 CALL ml.add('user', {date: n.date}, n.count) YIELD result RETURN result

CALL ml.train('user')

CALL ml.predict('user', {start: 20170515, end: 20170522})

CALL ml.remove('user')

It is also possible to call all the procedures sequentially in one call by yielding the results:

CALL ml.create('user', {date: "numeric"}, {alpha: 0.1, iter: 300, theta: [0.0, 0.0]}, null, true, 'nd4j')
YIELD result AS createresult
MATCH (n:User {stage:"cpd-ece-prod"}) WHERE toInteger(n.date) < 20170515
CALL ml.add('user', {date: n.date}, n.count)
YIELD result
WITH collect(distinct result) AS addresult, createresult
CALL ml.info('user')
YIELD result
WITH collect(distinct result) AS inforesult, addresult, createresult
CALL ml.train('user')
YIELD result AS trainresult
CALL ml.predict('user', {start: 20170515, end: 20170522})
YIELD result
WITH collect(distinct result) AS predictresult, trainresult, inforesult, addresult, createresult
CALL ml.remove('user')
YIELD result AS removeresult
RETURN  createresult, addresult, inforesult, trainresult, predictresult, removeresult

Additional test data and examples can be found at "src/test/resources".