Classifier chains

About: Classifier chains is a research topic. Over the lifetime, 170 publications have been published within this topic receiving 20989 citations.

A Hierarchical Matcher using Local Classifier Chains

Abstract: This paper focuses on improving the performance of current convolutional neural networks in visual recognition without changing the network architecture. A hierarchical matcher is proposed that builds chains of local binary neural networks after one global neural network over all the class labels, named as Local Classifier Chains based Convolutional Neural Network (LCC-CNN). The signature of each sample as two components: global component based on the global network; local component based on local binary networks. The local networks are built based on label pairs created by a similarity matrix and confusion matrix. During matching, each sample travels through one global network and a chain of local networks to obtain its final matching to avoid error propagation. The proposed matcher has been evaluated with image recognition, character recognition and face recognition datasets. The experimental results indicate that the proposed matcher achieves better performance when compared with methods using only a global deep network. Compared with the UR2D system, the accuracy is improved significantly by 1% and 0.17% on the UHDB31 dataset and the IJB-A dataset, respectively.

Risk neutrality in learning classifier systems

Abstract: Both economics and biology have come to agree that successful behavior in a stochastic environment responds to the variance of potential outcomes. Unfortunately, when biological and economic paradigms are mated together in a learning classifier system (LCS), decision-making agents called classifiers typically simply ignore risk. Since a fundamental problem of learning is risk management, LCS have not always performed as well as theoretically predicted. This paper develops a novel model of risk-neutral reinforcement learning in a traditional Bucket Brigade credit-allocation market under the pressure of a Genetic Algorithm. I demonstrate the applicability of the basic model to the classical LCS design and reexamine two basic issues where traditional LCS performance fails to meet expectations: default hierarchies and long chains of coupled classifiers. Risk-neutrality and noisy probabilistic auctions create dynamic instability in both areas, while identical preferences result in market failure in default hierarchies and exponential attenuation of price signals down classifier chains. Despite the limitations of simple risk-neutral classifiers, I show they’re capable of cheap short-run emulation of more rational behaviors. Still, risk-neutral information markets are a dead end. The model suggests a path toward a new type of LCS built on stable, heterogeneous, and risk-averse preferences under efficient auctions and access to more complete markets exploitable by competing risk management strategies. This will require a radical rethinking of the evolutionary and economic algorithms, but ultimately heralds a return to a market-based approach to LCS.

Practical Significance of GA PartCC in Multi-Label Classification

Abstract: Multi-label classification (MLC) can be defined as the objective of learning a classification model which has the capability to infer the accurate labels of new, previously unseen, objects where it is a likely situation that each object of the dataset may rightfully belong to multiple class labels. While single-label classification problems have been thoroughly researched, the same cannot be said for MLC. A gradually increasing number of problems are now being tackled as multi-label, allowing for richer and more accurate knowledge mining in real-world domains, such as medical diagnoses, social media, text classification, etc. Currently, there are two ways of solving MLC problems; Problem Transformation Approach and Algorithm Adaptation Method. Of the two, the former has in its domain Classifier Chains (CC) which is the most effective and popular method of solving MLC problems because of its simplicity in implementation. Unfortunately, CC is not favoured due to 2 drawbacks, [1] ordering of the labels for classification are randomly decided without a fixed logic or algorithm to it which results in varying accuracy, [2] all the labels, even those which may be redundant for a particular dataset are put into the chain despite the probability that some may be carrying irrelevant details. Through the research conducted for the purpose of this study, both challenges are tackled along with others detailed further on simultaneously using Genetic Algorithms (GA) over a Partial CC (PartCC) model, which is a modification over CC. A toxic comments dataset is used since its classification is a multi-label text classification problem with a highly imbalanced dataset. This paper aims to create a prototype model that is capable of detecting various types of toxicity like neutral, toxic, severe toxic, threats, obscenity, insults and identity hate. With the explosion of social media in the modern world and the resulting increasing phenomenon of social media hatred and bullying, there is a need for an advanced prototype model to predict the toxicity of each class of comments.

Learning Label Structures with Neural Networks for Multi-label Classification

Abstract: Multi-label classification (MLC) is the task of predicting a set of labels for a given input instance. A key challenge in MLC is how to capture underlying structures in label spaces. Due to the computational cost of learning from all possible label combinations, it is crucial to take into account scalability as well as predictive performance when we deal with large scale MLC problems. Another problem that arises when building MLC systems is which evaluation measures need to be used for performance comparison. Unlike traditional multi-class classification, several evaluation measures are often used together in MLC because each measure prefers a different MLC system. In other words, we need to understand the properties of MLC evaluation measures and build a system which performs well in terms of those evaluation measures in which we are particularly interested. In this thesis, we develop neural network architectures that efficiently and effectively utilize underlying label structures in large-scale MLC problems. In the literature, neural networks (NNs) that learn from pairwise relationships between labels have been used, but they do not scale well on large-scale label spaces. Thus, we propose a comparably simple NN architecture that uses a loss function which ignores label dependencies. We demonstrate that simpler NNs using cross-entropy per label works better than more complex NNs, particularly in terms of rank loss, an evaluation measure that takes into account the number of incorrectly ranked label pairs. Another commonly considered evaluation measure is subset 0/1 loss. Classifier chains (CCs) have shown state-of-the-art performance in terms of that measure because the joint probability of labels is optimized explicitly. CCs essentially convert the problem of learning the joint probability into a sequential prediction problem. Then, the task is to predict a sequence of binary values for labels. Contrary to the aforementioned NN architecture which ignores label structures, we study recurrent neural networks (RNNs) so as to make use of sequential structures on label chains. The proposed RNNs are advantageous over CC approaches when dealing with a large number of labels due to parameter sharing effects in RNNs and their abilities to learn from long sequences. Our experimental results also confirm that their superior performance on very large label spaces. In addition to NNs that learn from label sequences, we present two novel NN-based methods that learn a joint space of instances and labels efficiently while exploiting label structures. The proposed joint space learning methods project both instances and labels into a lower dimensional space in a way that minimizes the distance between an instance and its relevant labels in that space. While the goal of both joint space learning methods is same, they use different additional information on label spaces during training: One approach makes use of hierarchical structures of labels and can be useful when such label structures are given by human experts. The other uses latent label spaces learned from textual label descriptions so that we can apply it to more general MLC problems where no explicit label structures are available. Notwithstanding the difference between the two approaches, both approaches allow us to make predictions with respect to labels that have not been seen during training.

Multi-dimensional Classification Automation with Human Interface Based on Metromaps

Abstract: There are several existing multidimensional classification methods which attempt to build meaningful multidimensional structures from simple Binary Relevance (BR) classifiers. One of the recent methods is the Classifier Chains (CC) method which applies several binary classes in sequence. While the method offers a major reduction in complexity it is not clear how to define the order of binary classes in the chain. This paper presents a novel method called Metro Map Classifier (MMC) in which binary classes are connected by a metro map and can be applied in any order. Results show that MMCs perform between 10% and 50% better depending on the type of content. The ultimate target of this research is automation when selecting a small subset of content from Big Data.