Statistical Type Inference (both fully automatic and semi supervised) for RDF datasets in the N3 or N-Quads format.
\cite: StaTIX — Statistical Type Inference on Linked Data
@inproceedings{Stx18,
title={StaTIX — Statistical Type Inference on Linked Data},
author={Artem Lutov and Soheil Roshankish and Mourad Khayati and Philippe Cudré-Mauroux},
booktitle={IEEE BigData 2018 - 4th Special Session on Intelligent Data Mining},
doi={10.1109/BigData.2018.8622285}
year={2018}
}The StaTIX Java client application was initially developed in the scope of the Master Project of Soheil Roshankish.
\authors: (c) Artem Lutov [email protected], Soheil Roshankish
\organization: eXascale Infolab
\license: Apache License, Version 2.0
StaTIX performs statistical type inference for the RDF datasets in fully automatic fashion with possibility to use semi supervised mode. In the semi supervised mode, either a) the sample of the processing dataset prelabeled with the type properties should be provided, or b) another dataset should be specified with the present type properties and, desirably, similar structure to the processing dataset. The input RDF dataset(s) should be specified in the N3/N4 (N-Quads) formats: <subject> <property> <object> ..
Types (rdf objects, which correspond rdf:type predicate value) are inferred as clusters of the RDF triple subjects in the scope of the whole input dataset with automatic scale identification for each cluster. The scale for all clusters can be manually forced in case specific macro or micro level clustering is required.
For the semi supervised mode, similarity between the RDF subjects is evaluated with respect to the #type properties using TF-IDF based similarity weighting for the weighted cosin similarity.
Currently StaTIX infers semantic types based only on the statistical analysis of the input dataset. The accuracy can be improved further with additional semantic analyses leveraging both logical inference and embedding techniques to better grasp the differences and relationships between various instances.
The output results are clusters in the .cnl format (space separated list of members). Each cluster correspond to the type and has members represented by the subject ids. Subject ids are generated sequentially starting from 0 for all unique subjects in the input dataset.
Evaluation of the StaTIX type inference (see details in TInfES)
StaTIX is executed as ./run.sh -f -m -r m -o <outp_cls> <inp_dataset> and compared with SDA and SDType.
Accuracy (the higher, the better):
Execution time (the lower, the better):
StaTIX uses DAOC clustering library and Apache Commons CLI arguments parser. Both libraries are included into the repository and located in the /lib dir.
Java 1.8 is required to compile (JDK) or execute (JRE) StaTIX, on Linux Ubuntu Java can be install executing: $ sudo apt-get install openjdk-8-jdk-headless. On other platforms Java can be downloaded from the OpenJDK or Oracle sites.
DAOC is a C++ clustering library with SWIG-generated Java interface. The provided native library (libdaoc.so) is built on Ubuntu 16.04 x64 (and might also work in the Ubuntu console of Windows 10 x64). DAOC should be rebuilt from the sources to run StaTIX on other platforms.
./run.sh -h
Usage: ./run.sh [OPTIONS...] <inputfile.rdf>
Statistical type inference in fully automatic and semi supervised modes
Options:
-b,--brief-hints <arg> Brief hits, possible values:
'--' - interactive hints
'<filename.ipl>' - read from the file
having the format for each line:
<indicativity> <property>
where indicativity E [0, 1]; 0 - the
property has no any impact on the entity
type, 1 - the property fully specifies
the entity type, line comments starting
with '#' are allowed.
'-[<nopts=INF>]' - automatic generation
of the hints to the <inpfile_marks.ipl>,
where <marks> is the range of marks (>=
1) on supervision, which defines the
indicativity precision eps=0.5/(marks +
1): eps=0.167 for 2 marks
-c,--cut-ratio <arg> Cut the graph links (similarity matrix)
iteratively on the graph construction
before the construction is completed
discarding instance (node) links lighter
than cut-ratio * avg_ndlinks_weight,
cut-ratio E [0, 1), recommended value if
applied: ~0.25, 0 means skip the
cutting. Reduces the memory consumption
and speedups the clustering but affects
the accuracy
-e,--extract-groundtruth <arg> Extract ground-truth (ids of the
subjects per each type) to the specified
file in the .cnl format, optionally with
subjects and type labels
-f,--filter Filter out from the resulting clusters
all subjects that do not have the
'#type' property in the input dataset,
used for the type inference evaluation
-g,--groundtruth-sample <arg> The ground-truth sample (subset of the
input dataset or another similar dataset
with the specified type properties)
-h,--help Show usage
-j,--jaccard-similarity Use (weighted) Jaccard instead of the
Cosine similarity
-l,--cl-label <arg> Output map of the cluster labels (names)
(<inpfile>.clb in the label per line
format, default: disabled, requires: -e
-m,--multi-level Output type inference for multiple
scales (representative clusters from all
hierarchy levels) besides the macro
scale (top level, root)
-n,--id-name <arg> Output map of the instance id names
(labels) to the <inpfile>.idm in tab
separated format: <id> <subject_name>,
default: disabled. Note: all instances
are mapped including non-typed ones
-o,--output <arg> Output file, default: <inpfile>.cnl
-p,--network <arg> Produce .rcg input network file for the
clustering without the type inference
itself and respecting the "cut-ratio",
"filter", "weigh-instance" and
"jaccard-similarity" options
-r,--reduce <arg> Reduce graph links (similarity matrix)
on the graph clustering (after the graph
is constructed) by non-significant
relations to reduce memory consumption
and speedup the clustering (recommended
for large datasets). Options X[Y]; X: a
- accurate, m - mean, s - severe; Y: o -
use optimization function for the links
reduction (default), w - reduce links by
their raw weight. Examples: -r m, -r mw.
Note: all non-zero unique items (half of
the symmetric matrix) are supplied for
the graph construction, which is the
memory consumption bottleneck
-s,--scale <arg> Scale (resolution, gamma parameter of
the clustering), -1 is automatic scale
inference for each cluster, >=0 is the
forced static scale (<=1 for the macro
clustering); default: -1
-u,--unique-triples Unique triples only are present in the
ground-truth dataset (natty, clean data
without duplicates), so there is no need
of the possible duplicates
identification and omission
-v,--version Show version number
-w,--weigh-instance Weight RDF instances (subjects, consider
the self-relation) or use only the
weighted relations between the instances
To infer types without the ground-truth available with the implicit output to the inpDataset.cnl: ./run.sh inpDataset.rdf.
To infer types with available ground-truth for the sampled reduced dataset or using another typed dataset with similar structure, performing output to the results.cnl: ./run.sh -g gtSample.rdf -o results.cnl inpDataset.rdf.
To infer types on multiple resolution levels (besides the whole dataset scope): ./run.sh -a inpDataset.rdf.
To produce the input network in the .rcg format for the clustering without the type inference itself: $ ./run.sh -f -p networks/gendr-f.rcg datasets/biomedical/gendr.rdf.
./build.sh [-p] [-c] [<outdir>]
-p,--pack - build the tarball besides the executables
-c,--classes - retain classes after the build, useful for the frequent
modification and recompilation of some files.
Compilation or the single file (.java to .class):
$ javac -cp lib/\*:src -d classes/ src/info/exascale/statix/main.java
The compilation requires JDK and verified on OpenJDK 8/9 x64.
The build yields statix.jar with all requirements in the output directory (. by default) and optionally packs all these files to the tarball statix.tar.gz.
Compilation generates statix.tar.gz tarball with all requirements ready for the distribution. Also the tarball can be generated from the executables using the pack.sh script.
Note: Please, star this project if you use it.
- TInfES - Type inference evaluation scripts and accessory apps used for the StaTIX benchmarking.
- xmeasures - Extrinsic clustering measures evaluation for the multi-resolution clustering with overlaps (covers): F1_gm for overlapping multi-resolution clusterings with possible unequal node base and standard NMI for non-overlapping clustering on a single resolution.
- GenConvNMI - Overlapping NMI evaluation that is (unlike
onmi) compatible with the original NMI and suitable for both overlapping and multi resolution (hierarchical) clusterings. - OvpNMI - Another method of the NMI evaluation for the overlapping clusters (communities) that is not compatible with the standard NMI value unlike GenConvNMI, but it is much faster than GenConvNMI.