KB_modeling

Now Support

models: TransE/TransD/TransR/DistmMul/ProjSoftmax
margin based; sampled sigmoid/softmax. i prefer second.
data_loader on WN, FB, Wikidata
metrics tensorboard show on (MR, MRR, Hits@N)*(Raw, Filter) [(all, 1-n, n-1, n-m)]

Usage

修改config里的数据，模型，训练等参数
python main.py train
观察log的tensorboard上述指标
python main.py test

data description

├── dev2id.txt  "\t".join([h,t,r]), h,t,r分别是头结点，尾结点，关系的唯一id 
├── entity2id.txt   结点id映射名字，训练非必须，可以看效果
├── relation2id.txt 关系id映射名字
├── test2id.txt 
└── train2id.txt

目前有3种数据 WN, FB, Wikidata WN，FB在本文件夹里， Wikidata在ftp上 wiki_data/Wikidata_3_kg.zip

model description

TransR/TransD/TransE/DistMul 使用了不同的score函数，f(h,r,t)
其他部分: 负采样+Max_margin损失函数(对应main.py MarginBasedModel)
另外一种选择：使用N个2分类或者1个N分类(对应main.py BasedLogitsModel), [代码需要验证 ]

TODO

add Model:IRN, ConvE, ConvKB, IRN
Retest
using features
trainfrom into application
joint model text and graph
link graph to text for text-auto-extarct and reasoning

Objects/applications

vectorize of relations, nodes, also the arithmic of relation(more than vector)
相似点/边查询
异常发现；图补齐；
推理 ** dense vectorize 是信息量减少的过程，做分类是可以的，做一些具体的推理不科学，比如"***的爸爸是谁"，向量化不靠谱 ** 这个问题以后再仔细探索。

referneces

Nguyen, Dat Quoc. “An Overview of Embedding Models of Entities and Relationships for Knowledge Base Completion.” ArXiv:1703.08098 [Cs], March 23, 2017. http://arxiv.org/abs/1703.08098.

Results

FB15K transE MR_tail=105(d=512,L2,margin=1) transR MR_tail=104(d=50,L1,margin=1,batch_size=50) distmul MR_tail=153(Do more)

hxsnow10 / kb_modeling