This is the implementation of the paper "Exception-Enriched Rule Learning from Knowledge Graphs" Link.

Requirements:

maven 2
Java 8
For development, IntelliJ (recommended)

Prepare Project:

Run scripts:

sh scripts/local_libs.sh

sh scripts/download_large_data.sh

Import the project as maven project to intellij if needed

Installation:

To install the project run scripts

mvn compile

mvn package -Dmaven.test.skip=true

mvn install -Dmaven.test.skip=true

Running

To mine rules with or without exceptions use mine_riles.sh use the following options

usage: mine_rules.sh
 -cPM,--cautious-materialization <arg>    Use partial materialization cautiously, here the minimum support for exceptions
 -de                                     Decode the output
 -dPM,--Debug_materialization <file>     debug Materialization file
 -en                                     Encode the input
 -ex                                     Mine the output
 -exMinSup <EXCEPTION_MIN_SUPP_RATIO>    Exception Minimum support for the rule
 -expOnly                                Output rules with exceptions only
 -exRank,--exception_ranking <order>     Exception ranking method(LIFT|PNCONF|SUPP|CONF|PNCONV|PNJACC)
 -f1,--first_filter                      first filter based on size (4 body atoms at most, 1 head and Max conf)
 -f2,--second_filter                     Second filter based on type hierarchy
 -i,--input <file>                       Input file inform of RDF or Integer transactions
 -m,--mapping_file <file>                Mapping RDF to Integer
 -maxConf <MAX_CONF_RATIO>               Maximum Confidence for the rule (default=1.0)
 -minConf <MIN_CONF_RATIO>               Minimum Confidence for the rule (default=0.001)
 -minS <MIN_SUPP_RATIO>                  Minimum support for the rule(default=0.0001)
 -o,--output <file>                      Input file inform of RDF or Integer transactions
 -oDLV,--output_DLV                      Export rules as PrASP
 -oDLV_CONFLICT,--export_DLVConflict     Export rules to count conflict to file
 -oPrASP,--output_PrASP                  Export rules as PrASP
 -pm,--materialization                   Use partial materialization
 -PMo,--materialization_order            Materialize with order. Only useful with Materialization
 -s,--sorting <order>                    Output sorting(CONF|HEAD|BODY|LIFT|HEAD_CONF|HEAD_LIFT|NEW_LIFT|CONV)
 -stats,--export_statistics              Export statistics to file
 -w,--weighted_transactions              Count transactions with weights. Only useful with Materialization

Output sorting Methods

HEAD: Sort according to the head predicates (useful for grouping)

BODY: According to the rules body

CONF: Association Rules Confidence (Original horn rule confidence is used)

LIFT: Association Rules Lift measurement (Original horn rule confidence is used)

HEAD_CONF: Sort according to head then confidence.

HEAD_LIFT: Sort according to head then lift.

NEW_LIFT: Revised Rules Lift

CONV: [Conviction measurement] (www3.di.uminho.pt/~pja/ps/conviction.pdf).

Exception ranking Methods

SUPP: Used in the naive approach. Only consider increase in support. LIFT: Only consider increase in lift CONF: Only consider increase in confidence

PNCONF: Used in partial materialization. Considers the increase of average confidence of positive and negative predictions.

PNCONV: Used in partial materialization. Considers the increase of average conviction of positive and negative predictions.

PNJACC: Used in partial materialization. Considers the increase of average Jaccard Coefficient of positive and negative predictions.

Running Experiments:

To Run YAGO experiments

to Run IMDB experiments

Note: fix the directories inside the scripts to point to facts_to_mine.tsv file

Other Important Scripts:

To convert the KB from RDF to different formats

rdf2int.sh <required conversion [SPMF|DLV_SAFE|PrASP]> <input file path> <output prefix>

Ex: sh assemble/bin/rdf2int.sh DLV_SAFE /GW/D5data-5/gadelrab/imdb/facts_to_mine_imdb.tsv /GW/D5data-5/gadelrab/imdb/in/facts_to_mine_imdb

SPMF : outputs transactional KB in numbers 1,2,3 for projected predicates DLV_SAFE : outputs unary Encoding in format p1234t(s1234o) for projected predicates. PrASP : outputs in PrASP format without encoding for example isMarriedToScientist(X). (Good fro viewing but causes problems with PrASP)

A mapping will be generated in case of encoding

P.S: Other scripts to be added

References

This is an implementation of the paper:

Exception-enriched Rule Learning from Knowledge Graphs Mohamed Gad-Elrab, Daria Stepanova, Jacopo Urbani, Gerhard Weikum In 15th International Semantic Web Conference (ISWC 2016),234-251, Springer 2016.

gadelrab/Exception_Enriched_Rules

About

Resources

Stars

Watchers

Forks

Releases

Languages