Awesome

PMGT

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

Experiments

<table> <tr> <td rowspan="2">Datasets</td> <td rowspan="2">Metrics</td> <td colspan="4">Top-N Recommendation</td> <td colspan="2">CTR Prediction</td> </tr> <tr> <td>GMF</td> <td>MLP</td> <td>NeuMF</td> <td>NeuMF-PMGT</td> <td>DCN</td> <td>DCN-PMGT</td> </tr> <tr> <td rowspan="7">VG</td> </tr> <tr> <td>N@10</td> <td>0.1426</td> <td>0.0972</td> <td>0.1621</td> <td>0.1810</td> <td>-</td> <td>-</td> </tr> <tr> <td>N@20</td> <td>0.1602</td> <td>0.1209</td> <td>0.1815</td> <td>0.2067</td> <td>-</td> <td>-</td> </tr> <tr> <td>R@10</td> <td>0.2057</td> <td>0.1724</td> <td>0.2365</td> <td>0.2748</td> <td>-</td> <td>-</td> </tr> <tr> <td>R@20</td> <td>0.2687</td> <td>0.2592</td> <td>0.3060</td> <td>0.3661</td> <td>-</td> <td>-</td> </tr> <tr> <td>AUC</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> <td>0.8178</td> <td>0.8667</td> </tr> <tr> <td></td> <td></td> <td></td> <td></td> <td></td> <td></td> <td></td> </tr> <tr> <td rowspan="6">TG</td> </tr> <tr> <td>N@10</td> <td>0.1730</td> <td>0.1163</td> <td>0.1995</td> <td>0.2192</td> <td>-</td> <td>-</td> </tr> <tr> <td>N@20</td> <td>0.1837</td> <td>0.1369</td> <td>0.2189</td> <td>0.2384</td> <td>-</td> <td>-</td> </tr> <tr> <td>R@10</td> <td>0.2104</td> <td>0.1828</td> <td>0.2733</td> <td>0.2889</td> <td>-</td> <td>-</td> </tr> <tr> <td>R@20</td> <td>0.2497</td> <td>0.2589</td> <td>0.3445</td> <td>0.3590</td> <td>-</td> <td>-</td> </tr> <tr> <td>AUC</td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> <td>0.8387</td> <td>0.8486</td> </tr> </table>

Stats

<table> <tr> <td rowspan="2" style="text-align:center">Datasets</td> <td colspan="3" style="text-align:center">Data for Downstream tasks</td> <td colspan="2" style="text-align:center">Item Graph</td> <td colspan="2" style="text-align:center">Multimodal Feat.</td> </tr> <tr> <td style="text-align:center"># Users</td> <td style="text-align:center" ># Items</td> <td style="text-align:center"># Interact.</td> <td style="text-align:center"># Nodes</td> <td style="text-align:center"># Edges</td> <td style="text-align:center"># Visual Feat.</td> <td style="text-align:center"># Textual Feat.</td> </tr> <tr> <td style="text-align:center">VG</td> <td style="text-align:right">27,988</td> <td style="text-align:right">6,551</td> <td style="text-align:right">98,278</td> <td style="text-align:right">7,252</td> <td style="text-align:right">88,606</td> <td style="text-align:right">502</td> <td style="text-align:right">7,252</td> </tr> <tr> <td style="text-align:center">TG</td> <td style="text-align:right">134,697</td> <td style="text-align:right">10,337</td> <td style="text-align:right">378,138</td> <td style="text-align:right">10,834</td> <td style="text-align:right">38,252</td> <td style="text-align:right">1,279</td> <td style="text-align:right">10,834</td> </tr> </table>

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)