User-Item Recommender

Hint

Author: Yifan Feng (丰一帆)
Proof: Xinwei Zhang

In the following examples, two typical collaborative filtering neural networks are used to perform item recommender on the User-Item bipartite graph.

Models

NGCF (dhg.models.GCN), Neural Graph Collaborative Filtering paper (SIGIR 2019).
LightGCN (dhg.models.LightGCN), LightGCN: Lightweight Graph Convolutional Networks paper (SIGIR 2020).

Dataset

The Gowalla dataset (dhg.data.Gowalla) is collected for user-item recommendation task. Locations are viewed as items. The full dataset can be found in this website.

Results

Model	NDCG@20 on Testing	Recall@20 on Testing
NGCF	0.1307	0.1547
LightGCN	0.1550	0.1830

NGCF on Gowalla

Import Libraries

import time
from copy import deepcopy

import torch
import torch.nn as nn
from torch.utils.data import DataLoader

from dhg import BiGraph
from dhg.data import Gowalla
from dhg.models import NGCF
from dhg.nn import BPRLoss, EmbeddingRegularization
from dhg.metrics import UserItemRecommenderEvaluator as Evaluator
from dhg.random import set_seed
from dhg.utils import UserItemDataset, adj_list_to_edge_list

Define Functions

class BPR_Reg(nn.Module):
    def __init__(self, weight_decay):
        super().__init__()
        self.reg = EmbeddingRegularization(p=2, weight_decay=weight_decay)
        self.bpr = BPRLoss(activation="softplus")

    def forward(self, emb_users, emb_items, users, pos_items, neg_items, raw_emb_users, raw_emb_items):
        cur_u = emb_users[users]
        cur_pos_i, cur_neg_i = emb_items[pos_items], emb_items[neg_items]
        pos_scores, neg_scores = (cur_u * cur_pos_i).sum(dim=1), (cur_u * cur_neg_i).sum(dim=1)
        loss_bpr = self.bpr(pos_scores, neg_scores)
        raw_u = raw_emb_users[users]
        raw_pos_i, raw_neg_i = raw_emb_items[pos_items], raw_emb_items[neg_items]
        loss_l2 = self.reg(raw_u, raw_pos_i, raw_neg_i)
        loss = loss_bpr + loss_l2

        return loss


def train(net, data_loader, optimizer, criterion, epoch):
    net.train()

    loss_mean, st = 0, time.time()
    for users, pos_items, neg_items in data_loader:
        users, pos_items, neg_items = users.to(device), pos_items.to(device), neg_items.to(device)
        optimizer.zero_grad()
        emb_users, emb_items = net(ui_bigraph)
        loss = criterion(
            emb_users, emb_items, users, pos_items, neg_items, net.u_embedding.weight, net.i_embedding.weight,
        )
        loss.backward()
        optimizer.step()
        loss_mean += loss.item() * users.shape[0]
    loss_mean /= len(data_loader.dataset)
    print(f"Epoch: {epoch}, Time: {time.time()-st:.5f}s, Loss: {loss_mean:.5f}")


@torch.no_grad()
def validate(net, data_loader):
    net.eval()

    for users, train_mask, true_rating in data_loader:
        users, train_mask, true_rating = (
            users.to(device),
            train_mask.to(device),
            true_rating.to(device),
        )

        emb_users, emb_items = net(ui_bigraph)
        pred_rating = emb_users[users] @ emb_items.t()
        pred_rating += train_mask
        evaluator.validate_add_batch(true_rating, pred_rating)
    return evaluator.validate_epoch_res()


@torch.no_grad()
def test(net, data_loader):
    net.eval()

    for users, train_mask, true_rating in data_loader:
        users, train_mask, true_rating = (
            users.to(device),
            train_mask.to(device),
            true_rating.to(device),
        )
        emb_users, emb_items = net(ui_bigraph)
        pred_rating = emb_users[users] @ emb_items.t()
        pred_rating += train_mask
        evaluator.test_add_batch(true_rating, pred_rating)
    return evaluator.test_epoch_res()

Main

Note

More details about the metric Evaluator can be found in the Building Evaluator section.

if __name__ == "__main__":
    dim_emb = 64
    lr = 0.001
    num_workers = 0
    batch_sz = 2048
    val_freq = 20
    epoch_max = 1000
    weight_decay = 1e-4
    set_seed(2022)
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    evaluator = Evaluator([{"ndcg": {"k": 20}}, {"recall": {"k": 20}}])

    data = Gowalla()
    num_u, num_i = data["num_users"], data["num_items"]
    train_adj_list = data["train_adj_list"]
    test_adj_list = data["test_adj_list"]
    ui_bigraph = BiGraph.from_adj_list(num_u, num_i, train_adj_list)
    ui_bigraph = ui_bigraph.to(device)
    train_edge_list = adj_list_to_edge_list(train_adj_list)
    test_edge_list = adj_list_to_edge_list(test_adj_list)
    train_dataset = UserItemDataset(num_u, num_i, train_edge_list)
    test_dataset = UserItemDataset(num_u, num_i, test_edge_list, train_user_item_list=train_edge_list, phase="test")
    train_loader = DataLoader(train_dataset, batch_size=batch_sz, shuffle=True, num_workers=num_workers)
    test_loader = DataLoader(test_dataset, batch_size=batch_sz, shuffle=False, num_workers=num_workers)

    net = NGCF(num_u, num_i, dim_emb)
    net = net.to(device)
    criterion = BPR_Reg(weight_decay)
    optimizer = torch.optim.Adam(net.parameters(), lr=lr)

    best_state, best_val, best_epoch = None, 0, -1
    for epoch in range(epoch_max):
        train(net, train_loader, optimizer, criterion, epoch)
        if epoch % val_freq == 0:
            val_res = validate(net, test_loader)
            print(f"Validation: NDCG@20 -> {val_res}")
            if val_res > best_val:
                best_epoch = epoch
                best_val = val_res
                best_state = deepcopy(net.state_dict())
    print("train finished")
    print(f"best val: {best_val}")
    print(f"best epoch: {best_epoch}")
    print("testing...")
    net.load_state_dict(best_state)
    test_res = test(net, test_loader)
    print(f"test res: {test_res}")

Outputs

Epoch: 0, Time: 17.58267s, Loss: 0.57975
Validation: NDCG@20 -> 0.059597898178090525
Epoch: 1, Time: 17.35420s, Loss: 0.53948
Epoch: 2, Time: 16.67790s, Loss: 0.49900
Epoch: 3, Time: 16.58108s, Loss: 0.45402
Epoch: 4, Time: 16.49324s, Loss: 0.41055
Epoch: 5, Time: 16.47659s, Loss: 0.37169
Epoch: 6, Time: 16.50014s, Loss: 0.33842
Epoch: 7, Time: 16.53070s, Loss: 0.31066
Epoch: 8, Time: 16.50077s, Loss: 0.28642
Epoch: 9, Time: 16.47511s, Loss: 0.26709
Epoch: 10, Time: 16.48582s, Loss: 0.25038
Epoch: 11, Time: 16.48268s, Loss: 0.23593
Epoch: 12, Time: 16.55598s, Loss: 0.22323
Epoch: 13, Time: 16.59955s, Loss: 0.21264
Epoch: 14, Time: 16.57648s, Loss: 0.20171
Epoch: 15, Time: 16.61875s, Loss: 0.19399
Epoch: 16, Time: 16.60113s, Loss: 0.18529
Epoch: 17, Time: 16.67147s, Loss: 0.17826
Epoch: 18, Time: 16.39420s, Loss: 0.17196
Epoch: 19, Time: 16.43819s, Loss: 0.16577
Epoch: 20, Time: 16.39095s, Loss: 0.16056
Validation: NDCG@20 -> 0.0796889192830519
Epoch: 21, Time: 16.38881s, Loss: 0.15489
Epoch: 22, Time: 16.38985s, Loss: 0.15104
Epoch: 23, Time: 16.34736s, Loss: 0.14618
Epoch: 24, Time: 16.45004s, Loss: 0.14248
Epoch: 25, Time: 16.47847s, Loss: 0.13837
Epoch: 26, Time: 16.37476s, Loss: 0.13460
Epoch: 27, Time: 16.39726s, Loss: 0.13152
Epoch: 28, Time: 16.46616s, Loss: 0.12831
Epoch: 29, Time: 16.42933s, Loss: 0.12524
Epoch: 30, Time: 16.42688s, Loss: 0.12196
Epoch: 31, Time: 16.39388s, Loss: 0.11980
Epoch: 32, Time: 16.45209s, Loss: 0.11667
Epoch: 33, Time: 16.38611s, Loss: 0.11485
Epoch: 34, Time: 16.38676s, Loss: 0.11186
Epoch: 35, Time: 16.43171s, Loss: 0.11052
Epoch: 36, Time: 16.42350s, Loss: 0.10853
Epoch: 37, Time: 16.46971s, Loss: 0.10643
Epoch: 38, Time: 16.41361s, Loss: 0.10481
Epoch: 39, Time: 16.40113s, Loss: 0.10274
Epoch: 40, Time: 16.45297s, Loss: 0.10065
Validation: NDCG@20 -> 0.09484630939006403
Epoch: 41, Time: 16.20096s, Loss: 0.09951
Epoch: 42, Time: 16.33159s, Loss: 0.09786
Epoch: 43, Time: 16.41295s, Loss: 0.09629
Epoch: 44, Time: 16.29521s, Loss: 0.09473
Epoch: 45, Time: 16.31462s, Loss: 0.09310
Epoch: 46, Time: 16.30070s, Loss: 0.09155
Epoch: 47, Time: 16.42125s, Loss: 0.09050
Epoch: 48, Time: 16.34268s, Loss: 0.08982
Epoch: 49, Time: 16.38188s, Loss: 0.08804
Epoch: 50, Time: 16.35001s, Loss: 0.08735
Epoch: 51, Time: 16.30478s, Loss: 0.08568
Epoch: 52, Time: 16.26564s, Loss: 0.08473
Epoch: 53, Time: 16.38207s, Loss: 0.08340
Epoch: 54, Time: 16.34952s, Loss: 0.08275
Epoch: 55, Time: 16.41525s, Loss: 0.08166
Epoch: 56, Time: 16.34333s, Loss: 0.08030
Epoch: 57, Time: 16.43872s, Loss: 0.07994
Epoch: 58, Time: 16.37634s, Loss: 0.07892
Epoch: 59, Time: 16.37193s, Loss: 0.07846
Epoch: 60, Time: 16.36561s, Loss: 0.07732
Validation: NDCG@20 -> 0.10073506573468528
Epoch: 61, Time: 16.31512s, Loss: 0.07683
Epoch: 62, Time: 16.48562s, Loss: 0.07560
Epoch: 63, Time: 16.38161s, Loss: 0.07542
Epoch: 64, Time: 16.38181s, Loss: 0.07415
Epoch: 65, Time: 16.37734s, Loss: 0.07392
Epoch: 66, Time: 16.35093s, Loss: 0.07365
Epoch: 67, Time: 16.42241s, Loss: 0.07198
Epoch: 68, Time: 16.39753s, Loss: 0.07206
Epoch: 69, Time: 16.43910s, Loss: 0.07088
Epoch: 70, Time: 16.40806s, Loss: 0.07004
Epoch: 71, Time: 16.38006s, Loss: 0.07041
Epoch: 72, Time: 16.42882s, Loss: 0.06922
Epoch: 73, Time: 16.41414s, Loss: 0.06855
Epoch: 74, Time: 16.34444s, Loss: 0.06793
Epoch: 75, Time: 16.40675s, Loss: 0.06769
Epoch: 76, Time: 16.41324s, Loss: 0.06697
Epoch: 77, Time: 16.38147s, Loss: 0.06661
Epoch: 78, Time: 16.42382s, Loss: 0.06648
Epoch: 79, Time: 16.41072s, Loss: 0.06594
Epoch: 80, Time: 16.38907s, Loss: 0.06481
Validation: NDCG@20 -> 0.10532317576637099
Epoch: 81, Time: 16.42970s, Loss: 0.06468
Epoch: 82, Time: 16.45658s, Loss: 0.06442
Epoch: 83, Time: 16.38556s, Loss: 0.06388
Epoch: 84, Time: 16.32818s, Loss: 0.06370
Epoch: 85, Time: 16.36058s, Loss: 0.06294
Epoch: 86, Time: 16.34388s, Loss: 0.06260
Epoch: 87, Time: 16.33080s, Loss: 0.06234
Epoch: 88, Time: 16.36727s, Loss: 0.06197
Epoch: 89, Time: 16.32790s, Loss: 0.06154
Epoch: 90, Time: 16.43729s, Loss: 0.06101
Epoch: 91, Time: 16.38772s, Loss: 0.06070
Epoch: 92, Time: 16.42943s, Loss: 0.06037
Epoch: 93, Time: 16.36849s, Loss: 0.06043
Epoch: 94, Time: 16.39440s, Loss: 0.05969
Epoch: 95, Time: 16.33486s, Loss: 0.05954
Epoch: 96, Time: 16.34549s, Loss: 0.05876
Epoch: 97, Time: 16.37610s, Loss: 0.05866
Epoch: 98, Time: 16.39110s, Loss: 0.05857
Epoch: 99, Time: 16.38359s, Loss: 0.05788
Epoch: 100, Time: 16.42878s, Loss: 0.05773
Validation: NDCG@20 -> 0.10774315184649631
Epoch: 101, Time: 16.37178s, Loss: 0.05742
Epoch: 102, Time: 16.50821s, Loss: 0.05743
Epoch: 103, Time: 16.38737s, Loss: 0.05706
Epoch: 104, Time: 16.38123s, Loss: 0.05672
Epoch: 105, Time: 16.38323s, Loss: 0.05625
Epoch: 106, Time: 16.39332s, Loss: 0.05609
Epoch: 107, Time: 16.38817s, Loss: 0.05554
Epoch: 108, Time: 16.39039s, Loss: 0.05561
Epoch: 109, Time: 16.40110s, Loss: 0.05534
Epoch: 110, Time: 16.42629s, Loss: 0.05496
Epoch: 111, Time: 16.40456s, Loss: 0.05436
Epoch: 112, Time: 16.42960s, Loss: 0.05448
Epoch: 113, Time: 16.41036s, Loss: 0.05448
Epoch: 114, Time: 16.38433s, Loss: 0.05405
Epoch: 115, Time: 16.38922s, Loss: 0.05338
Epoch: 116, Time: 16.37122s, Loss: 0.05375
Epoch: 117, Time: 16.39454s, Loss: 0.05359
Epoch: 118, Time: 16.37232s, Loss: 0.05301
Epoch: 119, Time: 16.38497s, Loss: 0.05317
Epoch: 120, Time: 16.44990s, Loss: 0.05326
Validation: NDCG@20 -> 0.11050138281284864
Epoch: 121, Time: 16.42819s, Loss: 0.05270
Epoch: 122, Time: 16.43767s, Loss: 0.05240
Epoch: 123, Time: 16.33994s, Loss: 0.05205
Epoch: 124, Time: 16.37961s, Loss: 0.05193
Epoch: 125, Time: 16.40023s, Loss: 0.05187
Epoch: 126, Time: 16.44434s, Loss: 0.05143
Epoch: 127, Time: 16.44631s, Loss: 0.05155
Epoch: 128, Time: 16.42970s, Loss: 0.05141
Epoch: 129, Time: 16.43539s, Loss: 0.05119
Epoch: 130, Time: 16.41379s, Loss: 0.05097
Epoch: 131, Time: 16.43115s, Loss: 0.05080
Epoch: 132, Time: 16.41100s, Loss: 0.05077
Epoch: 133, Time: 16.42312s, Loss: 0.05043
Epoch: 134, Time: 16.39068s, Loss: 0.05028
Epoch: 135, Time: 16.37832s, Loss: 0.05016
Epoch: 136, Time: 16.39196s, Loss: 0.04994
Epoch: 137, Time: 16.38732s, Loss: 0.04976
Epoch: 138, Time: 16.41807s, Loss: 0.04935
Epoch: 139, Time: 16.37651s, Loss: 0.04916
Epoch: 140, Time: 16.39615s, Loss: 0.04923
Validation: NDCG@20 -> 0.11280099123452347
Epoch: 141, Time: 16.41225s, Loss: 0.04903
Epoch: 142, Time: 16.46800s, Loss: 0.04892
Epoch: 143, Time: 16.39678s, Loss: 0.04835
Epoch: 144, Time: 16.38563s, Loss: 0.04838
Epoch: 145, Time: 16.37892s, Loss: 0.04874
Epoch: 146, Time: 16.46196s, Loss: 0.04824
Epoch: 147, Time: 16.39248s, Loss: 0.04801
Epoch: 148, Time: 16.37935s, Loss: 0.04801
Epoch: 149, Time: 16.44855s, Loss: 0.04773
Epoch: 150, Time: 16.94777s, Loss: 0.04736
Epoch: 151, Time: 17.25382s, Loss: 0.04770
Epoch: 152, Time: 17.55223s, Loss: 0.04734
Epoch: 153, Time: 17.03791s, Loss: 0.04729
Epoch: 154, Time: 17.59021s, Loss: 0.04759
Epoch: 155, Time: 17.50267s, Loss: 0.04705
Epoch: 156, Time: 17.43284s, Loss: 0.04690
Epoch: 157, Time: 16.67660s, Loss: 0.04659
Epoch: 158, Time: 17.15853s, Loss: 0.04668
Epoch: 159, Time: 16.93252s, Loss: 0.04653
Epoch: 160, Time: 16.66944s, Loss: 0.04636
Validation: NDCG@20 -> 0.11396838930066855
Epoch: 161, Time: 16.75059s, Loss: 0.04627
Epoch: 162, Time: 16.80186s, Loss: 0.04613
Epoch: 163, Time: 16.75320s, Loss: 0.04616
Epoch: 164, Time: 16.79349s, Loss: 0.04604
Epoch: 165, Time: 16.82817s, Loss: 0.04579
Epoch: 166, Time: 16.78084s, Loss: 0.04599
Epoch: 167, Time: 16.83057s, Loss: 0.04553
Epoch: 168, Time: 16.83778s, Loss: 0.04554
Epoch: 169, Time: 16.83636s, Loss: 0.04548
Epoch: 170, Time: 16.76483s, Loss: 0.04547
Epoch: 171, Time: 16.85442s, Loss: 0.04487
Epoch: 172, Time: 16.83118s, Loss: 0.04475
Epoch: 173, Time: 16.80676s, Loss: 0.04518
Epoch: 174, Time: 16.82507s, Loss: 0.04470
Epoch: 175, Time: 16.87042s, Loss: 0.04485
Epoch: 176, Time: 17.00146s, Loss: 0.04471
Epoch: 177, Time: 17.02007s, Loss: 0.04455
Epoch: 178, Time: 16.63682s, Loss: 0.04445
Epoch: 179, Time: 17.08953s, Loss: 0.04450
Epoch: 180, Time: 16.89926s, Loss: 0.04419
Validation: NDCG@20 -> 0.11516925413130324

LightGCN on Gowalla