multilabel-knn 0.0.2

A lightweight toolbox for multilabel classification algorithms based on the k-nearest neighbors

multilabel_knn

multilabel_knn is a lightweight toolbox for the multilabel classifications based on the k-nearest neighbor algorithms [Doc].

The following algorithms are implemented:

• k-nearest neighbor classifier
• multilabel k-nearest neighbor classifier
• Binomial multilabel k-nearest neighbor classifier
• Binomial multilabel graph neighbor classifer

Usage

k-nearest neighbor algorithm (Predict a single label per sample)

import multilabel_knn as mlk
model = mlk.kNN(k=10, metric = "cosine") #k: number of neighbors, metric: distance metric {"euclidean", "cosine"}
model.fit(X, Y) # X :2d feature vectors. Y: label matrix, where Y[i,k] = 1 if i has label k.
Y_pred = model.predict(X_test) # Y_pred[i,k] = 1 is i is predicted to have label k.

mutilabel kNN (Can predict multiple labels per sample) [1]

import multilabel_knn as mlk
model = mlk.multilabel_kNN(k=10, metric = "cosine")
model.fit(X, Y)
Y_pred = model.predict(X_test) 

Binomial mutilabel kNN (Can predict multiple labels per sample)

import multilabel_knn as mlk
model = mlk.binomial_multilabel_kNN(k=10, metric = "cosine")
model.fit(X, Y) 
Y_pred = model.predict(X_test) 

Binomial multilabel kNN is a mobidifed version of multilabel kNN. It can perform well for data with a large number of samples and labels. See the docstring for details.

Binomial mutilabel graph (Take a graph as input. Can predict multiple labels per node)

import multilabel_knn as mlk
model = mlk.binomial_multilabel_graph()
model.fit(A, Y) # A is the adjacency matrix of the graph for training. A[i,j] =1 if node i has a link to node j. 
Y_pred = model.predict(B) # B is the adjacency matrix of the biparite network, where B[i,j] =1 if node i has a link to node j in the training graph.

Install

Requirements: Python 3.7 or later

pip install multilabel_knn

multilabel_knn uses faiss library, which has two versions, faiss-cpu and faiss-gpu. As the name stands, faiss-gpu can leverage GPUs, thureby faster if you have GPUs. multilabel_knn uses faiss-cpu by default to avoid unnecessary GPU-related troubles. But, if you have gpus compatible with the faiss-gpu, you can benefit the gpu accelarations by installing faiss-gpu by

with conda:

conda install -c conda-forge faiss-gpu

or with pip:

pip install faiss-gpu

Don't forget to pass gpu_id to the init argument to enable GPU

Maintenance

Code Linting:

conda install -y -c conda-forge pre-commit
pre-commit install

Docsctring: sphinx format

Test:

python -m unittest tests/simple_test.py

Reference

[1] Zhang, Min-Ling, and Zhi-Hua Zhou. 2007. “ML-KNN: A Lazy Learning Approach to Multi-Label Learning.” Pattern Recognition 40 (7): 2038–48.