README.md

PMGT

Implementation of "Pre-training Graph Transformer with Multimodal Side Information for Recommendation" (MM'21)

Experiments

Datasets	Metrics	Top-N Recommendation				CTR Prediction
		GMF	MLP	NeuMF	NeuMF-PMGT	DCN	DCN-PMGT
VG
	N@10	0.1426	0.0972	0.1621	0.1810	-	-
	N@20	0.1602	0.1209	0.1815	0.2067	-	-
	R@10	0.2057	0.1724	0.2365	0.2748	-	-
	R@20	0.2687	0.2592	0.3060	0.3661	-	-
	AUC	-	-	-	-	0.8178	0.8667
TG
	N@10	0.1730	0.1163	0.1995	0.2192	-	-
	N@20	0.1837	0.1369	0.2189	0.2384	-	-
	R@10	0.2104	0.1828	0.2733	0.2889	-	-
	R@20	0.2497	0.2589	0.3445	0.3590	-	-
	AUC	-	-	-	-	0.8387	0.8486

Stats

Datasets	Data for Downstream tasks			Item Graph		Multimodal Feat.
	# Users	# Items	# Interact.	# Nodes	# Edges	# Visual Feat.	# Textual Feat.
VG	27,988	6,551	98,278	7,252	88,606	502	7,252
TG	134,697	10,337	378,138	10,834	38,252	1,279	10,834

Log

[2022.01.16]

• Add TG dataset & experiment

[2022.01.15]

• Fixed bug for Validation & test set diff
• Re-experiment for NeuMF-end & NeuMF-PMGT
• Implement DCN (pmgt/dcn/models.py)
• Implement DCN training (pmgt/dcn/trainer.py)

[2022.01.14]

• Experiment for NeuMF-PMGT

[2022.01.13]

• Implement PMGT pre-training

[2022.01.12]

• Implement PMGT training (pmgt/pmgt/trainer.py)

[2022.01.11]

• Implement Graph structure reconstruction loss (pmgt/pmgt/models.py)
• Implement Node feature reconstruction loss

[2022.01.10]

• Implement MCNSampling (pmgt/pmgt/datasets.py)

[2022.01.09]

• Implement PMGT (pmgt/pmgt/modeling_pmgt.py)

[2022.01.08]

• HPO for NCF (VG dataset)

[2022.01.07]

• Implment NCF Training
• Experiment on VG dataset over NCF Baseline (except NeuMF-pre)

[2022.01.06]

• Report stats for VG dataset
• Implement NCF Dataset
• Implement NDCG@k
• Implement Recall@k

[2022.01.05]

• Implement Item Graph
• Implement NCF

[2022.01.04]

• Download Images (notebooks/PMGT.ipynb)
• Extract Visual (Inception-v4) & Textual (BERT) features

[2022.01.03]

• Download Amazon Review (notebooks/PMGT.ipynb)