tnreason 0.0

A package for reasoning with tensors

tensor-reasoning

Represent Knowledge Graphs as tensors to perform logic calculus and regression.

Usage

This module can be used for different tasks, as demonstrated in the examples folder.

Generation of Random Knowledge Graphs

Given a set of logical formulas a Markov Logic Network can be created using the Basis Calculus. One can then sample from the model to generate random data, in this case interpreted as a Random Knowledge Graph.

An example can be found in examples/generation/generate_accounting_kg.py.

Learning of logical formulas

Given a Knowledge Graph and positive and negative examples (each a pair of individuals), one can learn a logical formula true on the positive and false on the negative examples. To this end optimization via Alternating Least Squares has been implemented. Examples can be found in examples/learning/.

Packages

Logic

Coordinate Calculus: CoordinateCalculus main class for coordinate-based calculus of logical formulas.

Basis Calculus: BasisCalculus main class for basis-vector-based calculus of logical formulas.

Expression Calculus: Evaluation of expressions given dictionaries of CoordinateCalculus/BasisCalculus objects.

Optimization

generalized_als.py Performs the Alternating Least Squares to solve tensor regression problems.

Learning

expression_learning.py Optimizes formulas using Coordinate Calculus and the Alternating Least Squares.

mln_learning.py Learns Markov Logic Networks based on data.

Models

markov_logic_network.py Creates a Markov Logic Network using Basis Calculus based on pgmpy.models.MarkovNetwork.

Representation

On KG represented in turtle files:

ttl_to_csv.py Transform turtle file into a DataFrame containing facts.

factdf_to_cores.py Transforms the fact DataFrame into CoordinateCalculus Cores in the variable-based representation.

pairdf_to_cores.py Uses the pair DataFrame to initialize the targetCore.

sampledf_to_cores.py Transform sample DataFrame into CoordinateCalculus Cores in the atom-based representation.