An off-center technique: Learning a feature transformation to improve the performance of clustering and classification

Yan, Dasen; Zhou, Xinlei; Wang, Xizhao

doi:10.1016/j.ins.2019.06.068

Cited by 9 publications

(9 citation statements)

References 19 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…As the objective function E(W) quantifies a loss, the latter must be minimized. For more information, see the main reference of Dasen et al [11] Given two ligands p and q whose initial representations are denoted x p ! and x q !…”

Section: Wmlmentioning

confidence: 99%

“…x train i Tc train k cluster train k (11) In order to predict the membership of a new ligand in an already constructed cluster, this ligand is assigned to the cluster which minimizes the similarity between its centroid and this point [37] as defined in Equation 12.…”

Section: Predictive Clusteringmentioning

confidence: 99%

“…The first one is called Weight-Matrix Learning (named WML) and is uses unsupervised learning metrics, identified as an off-center technique. [11] The underlying idea is to obtain a chemical representation in a new feature space so that ligands with similarity above a given threshold are closer to each other and ligands with similarity below that threshold are farther apart. Based on WML, we propose a supervised way to transform the data.…”

Section: Introductionmentioning

confidence: 99%

“…This second feature transformation method is called SWML for Supervised Weight-Matrix. [11,12] The main idea is to use the prior knowledge on the ligands' label (active vs inactive) by optimizing a given objective function, in our case categorical cross entropy. The learning process leads to a learned matrix such that ligands from the same class are close to each other while those from different classes are moved away from each other.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

New Pharmacophore Fingerprints and Weight‐matrix Learning for Virtual Screening. Application to Bcr‐Abl Data

Rehioui

Cuissart

Ouali

et al. 2022

Molecular Informatics

View full text Add to dashboard Cite

In this work, we propose to analyze the potential of a new type of pharmacophoric descriptors coupled to a novel feature transformation technique, called Weight‐Matrix Learning (WML, based on a feed‐forward neural network). The application concerns virtual screening on a tyrosine kinase named BCR‐ABL. First, the compounds were described using three different families of descriptors: our new pharmacophoric descriptors, and two circular fingerprints, ECFP4 and FCFP4. Afterwards, each of these original molecular representations were transformed using either an unsupervised WML method or a supervised one. Finally, using these transformed representations, K‐Means clustering algorithm was applied to automatically partition the molecules. Combining our pharmacophoric descriptors with supervised Weight‐Matrix Learning (SWMLR) leads to clearly superior results in terms of several quality measures.

show abstract

Section: Wmlmentioning

confidence: 99%

Section: Predictive Clusteringmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

New Pharmacophore Fingerprints and Weight‐matrix Learning for Virtual Screening. Application to Bcr‐Abl Data

Rehioui

Cuissart

Ouali

et al. 2022

Molecular Informatics

View full text Add to dashboard Cite

show abstract

“…It is worth noting that the Min-Max property is similar to an off-center technique proposed in metric learning [40]. The off-center technique is to achieve a better representation in feature space based on such an idea that, given the center being 0.5, the similarity after transformation is require to tend towards zero (the minimum) or one (the maximum) if the similarity before the transformation is less than or more than 0.5, respectively.…”

Section: Min-max Propertymentioning

confidence: 99%

Adversarial Learning with Cost-Sensitive Classes

Shen,

Chen,

Wang

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

It is necessary to improve the performance of some special classes or to particularly protect them from attacks in adversarial learning. This paper proposes a framework combining cost-sensitive classification and adversarial learning together to train a model that can distinguish between protected and unprotected classes, such that the protected classes are less vulnerable to adversarial examples. We find in this framework an interesting phenomenon during the training of deep neural networks, called Min-Max property, that is, the absolute values of most parameters in the convolutional layer approach zero while the absolute values of a few parameters are significantly larger becoming bigger. Based on this Min-Max property which is formulated and analyzed in a view of random distribution, we further build a new defense model against adversarial examples for adversarial robustness improvement. An advantage of the built model is that it does no longer need adversarial training, and thus, has a higher computational efficiency than most existing models of needing adversarial training. It is experimentally confirmed that, regarding the average accuracy of all classes, our model is almost as same as the existing models when an attack does not occur and is better than the existing models when an attack occurs. Specifically, regarding the accuracy of protected classes, the proposed model is much better than the existing models when an attack occurs.

show abstract