baseline_1.py

# -*- coding: utf-8 -*-
# @Author: Ashutosh Tiwari
# @Email: checkashu@gmail.com
# @Date:   2023-02-13 11:09:50


import sys
from models import fast_knn_cpu
import numpy as np
from sklearn.decomposition import PCA
from baseline.debias_graph import debias_wrapper
from baseline.we_utils import get_direction, doPCA
from os.path import join as j

BASE = "../final_/polbook/polbook_one/"
NODE2VEC = False
DATASET = "polbook"

if "snakemake" in sys.modules:
    print("snakemake" in sys.modules)

    from utils import graph_utils, snakemake_utils
    BASE = snakemake.params["DATA_ROOT"]
    
    DATASET = snakemake.params["DATASET"]
    NODE2VEC = snakemake_utils.get_string_boolean(snakemake.params["NODE2VEC"])

from utils import graph_utils, snakemake_utils

METHOD_NAME = 'node2vec' if NODE2VEC else 'deepwalk'
print("BASE: ", BASE, "NODE2VEC: ", NODE2VEC, "DATASET: ", DATASET)

def get_embs(dataset, node2vec=NODE2VEC):
    y = snakemake_utils.get_dataset(dataset).get_grouped_col().numpy()
    deepwalk = np.load(j(BASE, "{}_{}.npy".format(dataset, METHOD_NAME)))
    
    
    centroids = graph_utils.get_centroid_per_group(deepwalk, y)
    # definitional words, these are supposed to be represent the group,
    # in this case lets take these to be the nodes closest to centroid of group
    # in this case are the centroids of the groups
    # definitional = graph_utils.get_n_nearest_neighbors_for_nodes(
    #     nodes=centroids, 
    #     embs=deepwalk,
    #     k=1,
    #     metric='cosine'
    # )
    
    N, dim = deepwalk.shape
    K = np.unique(y).shape[0]
    
    gender_specific_nodes = graph_utils.get_n_nearest_neighbors_for_nodes(
        nodes=centroids, 
        embs=deepwalk,
        k=int (.2 * N) // K,
        metric='cosine'
    )
    # equalize = graph_utils.get_farthest_pairs(deepwalk, y, same_class=False, per_class_count=int((.2 * N) / K))
    # equalize all nodes
    equalize = np.arange(N)
    print("number of gender specific pairs: ", gender_specific_nodes.shape)
    # direction = get_direction(deepwalk, y, "PCA")
    direction = doPCA(gender_specific_nodes, deepwalk, num_components=1).components_[0]
    
    return debias_wrapper(emb=deepwalk, equalize=equalize, direction=direction,)
            

embs = get_embs(dataset=DATASET, node2vec=NODE2VEC)

np.save(j(BASE, "{}_baseline_man_woman+{}_embs.npy".format(DATASET, METHOD_NAME)), embs)