Learning Neural Set Functions Under the Optimal Subset Oracle

EquiVSet

Learning Neural Set Functions Under the Optimal Subset Oracle

Zijing Ou1,2,3, Tingyang Xu1, Qinliang Su2, Yingzhen Li3, Peilin Zhao1, Yatao Bian1
1Tencent AI Lab, China
2Sun Yat-sen University, China
3Imperial College London, United Kingdom

(NeurIPS 2022 Oral Presentation)

Project Description

Learning neural set functions becomes increasingly more important in many applications like product recommendation and compound selection in AI-aided drug discovery. The majority of existing works study methodologies of set function learning under the function value oracle, which, however, requires expensive supervision signals. This renders it impractical for applications with only weak supervisions under the Optimal Subset (OS) oracle, the study of which is surprisingly overlooked. In this work, we present a principled yet practical maximum likelihood learning framework, termed as EquiVSet, that simultaneously meets the following desiderata of learning set functions under the OS oracle: i) permutation invariance of the set mass function being modeled; ii) permission of varying ground set; iii) minimum prior; and iv) scalability. The main components of our framework involve: an energy-based treatment of the set mass function, DeepSet-style architectures to handle permutation invariance, mean-field variational inference, and its amortized variants. Thanks to the elegant combination of these advanced architectures, empirical studies on three real-world applications (including Amazon product recommendation, set anomaly detection, and compound selection for virtual screening) demonstrate that EquiVSet outperforms the baselines by a large margin.

Paper

Preprint: Arxiv

Slides

The oral presentation can be downloaded here.

Code and Document

Code is released at: https://github.com/SubsetSelection/EquiVSet/tree/main

Experiments

contents
Exp 1: Product Recommendation
Exp 2: Set Anomaly Detection
Exp 3: Compound Selection in AI-aided Drug Discovery


Experiment 1: Product Recommendation

Table 1: Product recommendation results in the MJC metric on the Amazon dataset with different categories.
Categories Random PGM DeepSet DiffMF (ours) EquiVSetind (ours) EquiVSetcopula (ours)
Toys 0.0832 0.4414 ± 0.0036 0.4287 ± 0.0047 0.6147 ± 0.0102 0.6491 ± 0.0152 0.6762 ± 0.0221
Furniture 0.0651 0.1746 ± 0.0069 0.1758 ± 0.0072 0.1744 ± 0.0121 0.1775 ± 0.0108 0.1724 ± 0.0091
Gear 0.0771 0.4712 ± 0.0037 0.3806 ± 0.0019 0.5622± 0.0171 0.6103 ± 0.0193 0.6973 ± 0.0119
Carseats 0.0659 0.2330 ± 0.0115 0.2121 ± 0.0096 0.2229 ± 0.0104 0.2141 ± 0.0073 0.2149 ± 0.0123
Bath 0.0763 0.5638 ± 0.0077 0.4241 ± 0.0058 0.6901 ± 0.0061 0.6457 ± 0.0200 0.7567 ± 0.0095
Health 0.0758 0.4493 ± 0.0024 0.4481 ± 0.0041 0.5650± 0.0092 0.6315 ± 0.0153 0.7003 ± 0.0159
Diaper 0.0839 0.5802± 0.0092 0.4572 ± 0.0050 0.7011 ± 0.0112 0.7344 ± 0.0199 0.8275 ± 0.0136
Bedding 0.0791 0.4799 ± 0.0061 0.4824 ± 0.0081 0.6408± 0.0093 0.6287 ± 0.0195 0.7688 ± 0.0121
Safety 0.0648 0.2495 ± 0.0060 0.2211 ± 0.0044 0.2007± 0.0527 0.2250 ± 0.0287 0.2524 ± 0.0285
Feeding 0.0925 0.5596 ± 0.0081 0.4295 ± 0.0021 0.7496 ± 0.0114 0.6955 ± 0.0063 0.8101 ± 0.0074
Apparel 0.0918 0.5333 ± 0.0050 0.5074 ± 0.0036 0.6708± 0.0225 0.6465 ± 0.0150 0.7521 ±
0.0114
Media 0.0944 0.4406 ± 0.0092 0.4241 ± 0.0105 0.5145 ± 0.0105 0.5506 ± 0.0072 0.5694 ± 0.0105

Experiment 2: Set Anomaly Detection

Figure 1: A sampled data for the Double MNIST dataset, which consists of |S*| images with the same digit (red box, 83 in this case) and 20 − |S*|images with different digits.

Double MNIST

Figure 2: Sampled data points for the CelebA dataset. Each row is a sample, consisting of |S*| anomalies (red box) and 8 −|S*| normal images. In each sample, a normal image has two attributes (rightmost column) while anomalies do not have neither of them.

CelebA

Table 2: Set anomaly detection results in the MJC metric.
Method Double MNIST CelebA
Random 0.0816 0.2187
PGM 0.3031 ± 0.0118 0.4812 ± 0.0064
DeepSet 0.1108 ± 0.0031 0.3915 ± 0.0133
DiffMF (ours) 0.6064 ± 0.0133 0.5455 ± 0.0079
EquiVSetind (ours) 0.4054 ± 0.0122 0.5310 ± 0.0123
EquiVSetcopula (ours) 0.5878 ± 0.0068 0.5549 ± 0.0053

Experiment 3: Compound Selection in AI-aided Drug Discovery

Table 3: Compound selection results in the MJC metric.
Method PDBBind BindingDB
Random 0.0725 0.0267
PGM 0.3499 ± 0.0087 0.1760 ± 0.0055
DeepSet 0.3189 ± 0.0034 0.1615 ± 0.0074
DiffMF (ours) 0.3534 ± 0.0143 0.1894 ± 0.0021
EquiVSetind (ours) 0.3553 ± 0.0049 0.1904 ± 0.0034
EquiVSetcopula (ours) 0.3536 ± 0.0083 0.1875 ± 0.0032