utils.py

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import random

from tqdm import tqdm
import pandas as pd
from torch_geometric.datasets import AmazonBook, MovieLens100K, MovieLens1M
from torch_geometric.nn import GCNConv, LGConv
from torch_geometric.utils import degree
from torch_geometric.nn.conv import MessagePassing
from torch_geometric.data import HeteroData, Data
import  torch_geometric.transforms as T

def predict(model, device, data, num_users, num_items, user_id, train_edge_label_index, k=5):    
    with torch.no_grad():
        
        ## ML100k
        interaction_dataframe = pd.read_csv('./u1.base', delim_whitespace=True, header=None)
        meta_dataframe = pd.read_csv('./u.item', sep='|', encoding='latin-1', header=None)
        interaction_dataframe = interaction_dataframe[[0, 1]]
        interaction_dataframe.columns = ['reviewerID', 'asin']

        meta_dataframe = meta_dataframe[[0, 1]]
        meta_dataframe.columns = ['asin', 'title']
        

        out = model.get_embedding(data.edge_index)
        user_emb, item_emb = out[:num_users], out[num_users:]
        logits = user_emb @ item_emb.t()
        logits = torch.nn.Sigmoid()(logits)
        logits[train_edge_label_index[0], train_edge_label_index[1]-num_users] = float('-inf')

        # Create unique users to find the index of it in embedding table
        unique_users = interaction_dataframe['reviewerID'].unique().tolist()
        unique_items = interaction_dataframe['asin'].unique().tolist()
        
        random_row = random.randint(0, len(interaction_dataframe))
        user_to_rec = interaction_dataframe.iloc[random_row]['reviewerID']
        user_to_rec = user_id
        #user_to_rec = 923
        user_rates = logits[unique_users.index(user_to_rec)]
        
        # print(f"ID of user we want to recommend to: {user_to_rec}")
        
        ground_truth_asins = interaction_dataframe[interaction_dataframe['reviewerID'] == user_to_rec]['asin'].to_list()
        
        ground_truth_items = meta_dataframe[meta_dataframe['asin'].isin(ground_truth_asins)].head(5)
        
        # print(f"Ground truth items: {ground_truth_items['title'].values.tolist()}")
        
        _, top_ratings = torch.topk(user_rates, k)
        
        recommended_items = []
        for index in top_ratings:
            asin_of_item = unique_items[index]
            recommended_item = meta_dataframe[meta_dataframe['asin'] == asin_of_item]['title'].values
            recommended_items.append(recommended_item.item())

        return ground_truth_items, recommended_items