Knowledge Graphs lecture mini-project
We have a classic machine learning approach with positive (included) and negative (excluded) classes. That means we have to generate feature vectors for each instance. We split the problem into 4 subproblems by using independent models for the 4 types of instances: Bonds, Atoms, Structs and Compounds. This makes it easier to generate feature vectors. We apply simple feature engineering using the ontology. These features consists often of binary "is-of-type"/"is-of-subtype" statements or simple occurrence counts. This means we extract the following features for our 4 subproblems:
Bonds: bond type, 2x(atom type and subatom type) for the 2 atoms connected via this bond
Atoms: atom type, subatom type
Structs: struct type, substruct type
Compounds: counting all occurrances of - bonds (bond type), - atoms (atom type, subatom type), - structs (struct type, substruct type) and adding data properties, which are of type owl:DatatypeProperty in the ontology.
Generating these feature vectors we can train our 4 models. We decided to use Decision Trees as they were used previously on these type of chemical data.
Clone the repository https://github.com/alexanderwerning/chemMAP.git or download it. Install the libary chemMAP with pip install chemMAP.
To predict the remaining individuals based on learning problems, inside the root project directory chemMAP, execute python3 -m chemMAP.predict_remaining[ <path/to/learning-problems.ttl>]. The squared bracket part is optional. Without a learning problem file, this function will use the grading data, provided in chemMAP/data. You might need to exchange python3 for your python installation.
The result is stored as predictions.ttl in the package root directory.