GPFL算法的整理

算法的论文是《Towards Learning Instantiated Logical Rules from Knowledge Graphs》
https://arxiv.org/pdf/2003.06071.pdf
算法的源码来源于
https://github.com/irokin/GPFL

按target读取数据

                Set<Pair> trainPairs = IO.readPair(graph, trainFile, target);
                Settings.TARGET_FUNCTIONAL = IO.isTargetFunctional(trainPairs);
                Set<Pair> validPairs = IO.readPair(graph, validFile, target);
                Set<Pair> testPairs = IO.readPair(graph, testFile, target);

拓展path，生成Closed Rule(CAR),利用instantiatedRule(oar)生成Head Anchored Rule(HAR)和Both Anchored Rule (BAR)

1.针对每一条三元组（head，relation，tail），用head或者tail拓展path。
2.head与path组成一条规则的雏形。同时，用字母标注生成规则的雏形，生成template。
3.如果head与path[0]并且tail与path[path.size()-1]相同，则该template可以生成ClosedRule。
4.如果不相同，则是instantiatedRule,实例化生成HAR和BAR

用head或者tail拓展path

        @Override
        public void run() {
            Random rand = new Random();
            try(Transaction tx = graph.beginTx()) {
                while(consumer.isAlive()) {
                    Pair pair = trainPairs.get(rand.nextInt(trainPairs.size()));
                    addVisitedPair(pair);
                    Traverser traverser = GraphOps.buildStandardTraverser(graph, pair, Settings.RANDOM_WALKERS);
                    List<Path> paths=new ArrayList<>();
                    for (Path path : traverser) {
                        paths.add(path);
                        Rule rule = Context.createTemplate(path, pair);
                        while (consumer.isAlive()) {
                            if (ruleQueue.offer(rule, 100, TimeUnit.MILLISECONDS))
                                break;
                        }
                        if (!consumer.isAlive())
                            break;
                    }
                }
                tx.success();
            } catch (InterruptedException e) {
                e.printStackTrace();
                System.exit(-1);
            }
        }

head与path组成一条规则的雏形。同时，用字母标注生成规则的雏形，生成template。

    public static Rule createTemplate(Path path, Pair pair) {
        List<Atom> bodyAtoms = buildBodyAtoms(path);
        Atom head = new Atom(pair);
        return new Template(head, bodyAtoms);
    }

标注template是否是Closed,如果是Closed则这个template就是ClosedRule

    Rule(Atom head, List<Atom> bodyAtoms) {
        this.head = head;
        this.bodyAtoms = new ArrayList<>(bodyAtoms);
        Atom lastAtom = bodyAtoms.get(bodyAtoms.size() - 1);
        closed = head.getObjectId() == lastAtom.getObjectId();
        Atom firstAtom = bodyAtoms.get( 0 );
        fromSubject = head.getSubjectId() == firstAtom.getSubjectId();
    }

对非Closed的Rule生成HAR和BAR

                /*
                 * 对path=1的instantied rule拓展，生成HAR或者BAR
                 * 定义规则的读写类
                 * tempFiltercontents只写入ABS:instantiated开头的，后接HAR或者BAR，对应的是各种类型的文件夹
                 * ruleFilecontens只写入HAR或者BAR，对应只有一个rules.txt文件夹
                 *
                 */
                if(Settings.ESSENTIAL_TIME != -1 && Settings.INS_DEPTH != 0)
                    EssentialRuleGenerator.generateEssentialRules(trainPairs, validPairs
                            , context, graph, ruletempFile, ruleFile);
                /*
                 * 对path>1的instantied rule拓展，生成HAR或者BAR
                 * ！！！！！！！！！！！！！！！！！！！！！！！！！！！
                 * generalization没把closed的规则写进去，specialization把closed的规则也写进去了
                 * tempFiltercontents两种方式写入，一类是ABS:closed的，另一类是ABS:instantiated开头的，后接HAR或者BAR，对应的是各种类型的文件夹
                 * ruleFilecontens写入HAR或者BAR以及Closed的，对应只有一个rules.txt文件夹
                 */
                specialization(context, trainPairs, validPairs, ruletempFile);

用生成的规则CAR,HAR,BAR,生成答案集合

                /*
                 * ！！！！！！！！！！！query中会生成newTriple
                 * 生成predictionMap （答案集合），key是query，value是所有的answer值
                 */
                ruleApplication(context, ruletempFile);