data_utils.py

import torch
import dgl
import numpy as np
import scipy.sparse as ssp

# This is the train-test split method most of the recommender system papers running on MovieLens
# takes.  It essentially follows the intuition of "training on the past and predict the future".
# One can also change the threshold to make validation and test set take larger proportions.
def train_test_split_by_time(df, timestamp, item):
    df = df.copy()
    df['train_mask'] = np.ones((len(df),), dtype=np.bool)
    df['val_mask'] = np.zeros((len(df),), dtype=np.bool)
    df['test_mask'] = np.zeros((len(df),), dtype=np.bool)
    df = df.sort_values([item, timestamp])
    for track_id in df[item].unique():
        idx = (df[item] == track_id).to_numpy().nonzero()[0]
        idx = df.index[idx]
        if len(idx) > 1:
            df.loc[idx[-1], 'train_mask'] = False
            df.loc[idx[-1], 'test_mask'] = True
        if len(idx) > 2:
            df.loc[idx[-2], 'train_mask'] = False
            df.loc[idx[-2], 'val_mask'] = True
    df = df.sort_index()
    return df['train_mask'].to_numpy().nonzero()[0], \
           df['val_mask'].to_numpy().nonzero()[0], \
           df['test_mask'].to_numpy().nonzero()[0]

def build_train_graph(g, train_indices, utype, itype, etype, etype_rev):
    train_g = g.edge_subgraph(
        {etype: train_indices, etype_rev: train_indices},
        preserve_nodes=True)
    # remove the induced node IDs - should be assigned by model instead
    del train_g.nodes[utype].data[dgl.NID]
    del train_g.nodes[itype].data[dgl.NID]

    # copy features
    for ntype in g.ntypes:
        for col, data in g.nodes[ntype].data.items():
            train_g.nodes[ntype].data[col] = data
    for etype in g.etypes:
        for col, data in g.edges[etype].data.items():
            train_g.edges[etype].data[col] = data[train_g.edges[etype].data[dgl.EID]]

    return train_g

def build_val_test_matrix(g, val_indices, test_indices, utype, itype, etype):
    n_users = g.number_of_nodes(utype)
    n_items = g.number_of_nodes(itype)
    val_src, val_dst = g.find_edges(val_indices, etype=etype)
    test_src, test_dst = g.find_edges(test_indices, etype=etype)
    val_src = val_src.numpy()
    val_dst = val_dst.numpy()
    test_src = test_src.numpy()
    test_dst = test_dst.numpy()
    val_matrix = ssp.coo_matrix((np.ones_like(val_src), (val_src, val_dst)), (n_users, n_items))
    test_matrix = ssp.coo_matrix((np.ones_like(test_src), (test_src, test_dst)), (n_users, n_items))

    return val_matrix, test_matrix

def linear_normalize(values):
    return (values - values.min(0, keepdims=True)) / \
        (values.max(0, keepdims=True) - values.min(0, keepdims=True))