nodeclass 0.0.3

Node classification package

Welcome to Node Classification

• Written by Miguel Romero

• Last update: 18/08/21

Node classification

This package aims to provide different approaches to the node classification problem (also known as attribute prediction) using machine learning techniques. There are two approaches available: flat node classification (fnc) and hierarchical classification (hc). Both approaches are based on a gradient boosting decision tree algorithm called XGBoost, in addition the approaches are equipped with an over-sampling technique call SMOTE.

Flat classification

Flat node classification aims to valuate whether the structural (topological) properties of a network are useful for predicting node attributes of nodes (i.e., node classification), without considering the (possible) relationships between the classes of the node attribute to be predicted, i.e., the classes are predicted independently.

Hierarchical classification

Hierarchical node classification considers the hierarchical organization of the classes of a node attribute to be predicted. Using a top-down approach a binary classifier is trained per class according to the hierarchy, which is represented as a DAG.

Installation

The node classification package can be install using pip, the requirements will be automatically installed:

python3 -m pip install nodeclass

The source code and examples can be found in this GitHub repository.

Example

Flat classification

This example illustrates how the node classification package can be used to check whether the structural properties of the gene co-expression network improve the performance of the prediction of gene functions for rice (Oryza sativa Japonica). In this example, a gene co-expression network gathered from ATTED II is used.

How to run the example?

The complete source code of the example can be found in the GitHub repository. First, the xgbfnc package need to be imported:

from nodeclass.models.xgbfnc import XGBfnc
from nodeclass.tools import data

After creating adjacency matrix adj for the network, the structural properties are computed using the module data of the package:

df, strc_cols = data.compute_strc_prop(adj)

This method returns a DataFrame with the structural properties of the network and a list of the names of these properties (i.e., column names). After adding the additional features of the network to the DataFrame, the XGBhc module is used to instantiate the XGBhc class:

test = XGBhc()
test.load_data(df, strc_cols, y, term, output_path='output')
ans, pred, params = test.structural_test()

The data of the network is loaded using the load_data method. And the structural test is execute using the structural_test method. The test returns a boolean value which indicates whether the structural properties help to improve the prediction performance, the prediction for the model including the structural properties and its best parameters.

To run the example execute the following commands:

cd test/flat_classification
python3 test_small.py

Hierarchical classification

This example illustrates how the hierarchical classification package can be used to predict gene functions considering the hierachical structure of gene functions (as determined by Gene Ontology) based on the gene co-expression network. This example uses the data for rice (Oryza sativa Japonica),the gene co-expression network (GCN) was gathered from ATTED II.

How to run the example?

The complete source code of the example can be found in the GitHub repository. First, the xgbhc package need to be imported:

from nodeclass.models.xgbhc import XGBhc
from nodeclass.tools import data

The adjacency matrix for the GCN and the gene functions (from ancestral relations of biological processes), and the matrix of associations between genes and functions are created using the packaga data as follows:

gcn, go_by_go, gene_by_go, G, T = data.create_matrices(data_ppi, data_isa, data_term_def, data_gene_term, OUTPUT_PATH, True)

The tree representation of the hierarchy is generated from the adjacency matrix of the classes by removing the isolated classes, filtering the classes according to the number of nodes associated (if required) and finding the sub-hierarchies remaining. Then a minimum spanning tree (MST) algorithm is applied to each sub-hierarchy to get the its tree representation (the order and ancestors of the classes will be calculated):

roots, subh_go_list = data.generate_hierarchy(gcn, go_by_go, gene_by_go, data_term_def, G, T, OUTPUT_PATH, filter=[5,300], trace=True)
root, subh_go = roots[13], subh_go_list[13]
subh_adj = data.hierarchy_to_tree(gcn, go_by_go, gene_by_go, T, subh_go, OUTPUT_PATH)

Additionally, the structural properties of the sub-graph of the GCN, corresponding to the set of nodes associated to the classes in the sub-hierarchy, are computed using the module data:

data.compute_strc_prop(subh_adj, path=OUTPUT_PATH)

Finally, the XGBhc class is instantiated, the data of the sub-hierarchy is loaded and the prediction is done as follows:

model = XGBhc()
model.load_data(data, root, hierarchy, ancestors, DATA_PATH, OUTPUT_PATH)
model.train_hierarchy()

The results of the prediction are saved on the OUTPUT_PATH, including the roc and precision-recall curve, the confusion matrix and a csv file with some performance metrics (such as the auc roc, average precision, recall, precision and F1, true positive and true negative rate and the execution time).

To run the example execute the following commands:

cd test/hierarchical_classification
python3 test_data.py
python3 test.py

Documentation

Documentation of the package can be found here.