scholarly journals Like trainer, like bot? Inheritance of bias in algorithmic content moderation

2017 ◽  
Author(s):  
Reuben Binns ◽  
Michael Veale ◽  
Max Van Kleek ◽  
Nigel Shadbolt
Keyword(s):  
Machine Learning ◽  
Training Data ◽  
The Internet ◽  
Men And Women ◽  
Machine Learning Classifiers ◽  
Ethical Choices ◽  
Test Sets ◽  
Networked Publics ◽  
Important Objective

The internet has become a central medium through which 'networked publics' express their opinions and engage in debate. Offensive comments and personal attacks can inhibit participation in these spaces. Automated content moderation aims to overcome this problem using machine learning classifiers trained on large corpora of texts manually annotated for offence. While such systems could help encourage more civil debate, they must navigate inherently normatively contestable boundaries, and are subject to the idiosyncratic norms of the human raters who provide the training data. An important objective for platforms implementing such measures might be to ensure that they are not unduly biased towards or against particular norms of offence. This paper provides some exploratory methods by which the normative biases of algorithmic content moderation systems can be measured, by way of a case study using an existing dataset of comments labelled for offence. We train classifiers on comments labelled by different demographic subsets (men and women) to understand how differences in conceptions of offence between these groups might affect the performance of the resulting models on various test sets. We conclude by discussing some of the ethical choices facing the implementers of algorithmic moderation systems, given various desired levels of diversity of viewpoints amongst discussion participants.

scholarly journals Using Machine Learning Methods to Develop a Short Tree-Based Adaptive Classification Test: Case Study With a High-Dimensional Item Pool and Imbalanced Data

2020 ◽  
Vol 44 (7-8) ◽  
pp. 499-514
Author(s):  
Yi Zheng ◽  
Hyunjung Cheon ◽  
Charles M. Katz
Keyword(s):  
Machine Learning ◽  
Classification Accuracy ◽  
Imbalanced Data ◽  
Training Data ◽  
Adaptive Tests ◽  
Promising Alternative ◽  
Adaptive Classification ◽  
Short Tree ◽  
Classification Test

This study explores advanced techniques in machine learning to develop a short tree-based adaptive classification test based on an existing lengthy instrument. A case study was carried out for an assessment of risk for juvenile delinquency. Two unique facts of this case are (a) the items in the original instrument measure a large number of distinctive constructs; (b) the target outcomes are of low prevalence, which renders imbalanced training data. Due to the high dimensionality of the items, traditional item response theory (IRT)-based adaptive testing approaches may not work well, whereas decision trees, which are developed in the machine learning discipline, present as a promising alternative solution for adaptive tests. A cross-validation study was carried out to compare eight tree-based adaptive test constructions with five benchmark methods using data from a sample of 3,975 subjects. The findings reveal that the best-performing tree-based adaptive tests yielded better classification accuracy than the benchmark method IRT scoring with optimal cutpoints, and yielded comparable or better classification accuracy than the best benchmark method, random forest with balanced sampling. The competitive classification accuracy of the tree-based adaptive tests also come with an over 30-fold reduction in the length of the instrument, only administering between 3 to 6 items to any individual. This study suggests that tree-based adaptive tests have an enormous potential when used to shorten instruments that measure a large variety of constructs.

Debiasing Algorithms for Protein Ligand Binding Data do not Improve Generalisation

2019 ◽  
Author(s):  
Vikram Sundar ◽  
Lucy Colwell
Keyword(s):  
Machine Learning ◽  
Ligand Binding ◽  
Training Data ◽  
Chemical Data ◽  
Learning Models ◽  
Binding Data ◽  
Test Sets ◽  
Machine Learning Models

The structured nature of chemical data means machine learning models trained to predict protein-ligand binding risk overfitting the data, impairing their ability to generalise and make accurate predictions for novel candidate ligands. To address this limitation, data debiasing algorithms systematically partition the data to reduce bias. When models are trained using debiased data splits, the reward for simply memorising the training data is reduced, suggesting that the ability of the model to make accurate predictions for novel candidate ligands will improve. To test this hypothesis, we use distance-based data splits to measure how well a model can generalise. We first confirm that models perform better for randomly split held-out sets than for distant held-out sets. We then debias the data and find, surprisingly, that debiasing typically reduces the ability of models to make accurate predictions for distant held-out test sets. These results suggest that debiasing reduces the information available to a model, impairing its ability to generalise.

Debiasing Algorithms for Protein Ligand Binding Data do not Improve Generalisation

2019 ◽  
Author(s):  
Vikram Sundar ◽  
Lucy Colwell
Keyword(s):  
Machine Learning ◽  
Ligand Binding ◽  
Training Data ◽  
Chemical Data ◽  
Learning Models ◽  
Binding Data ◽  
Test Sets ◽  
Machine Learning Models

The structured nature of chemical data means machine learning models trained to predict protein-ligand binding risk overfitting the data, impairing their ability to generalise and make accurate predictions for novel candidate ligands. To address this limitation, data debiasing algorithms systematically partition the data to reduce bias. When models are trained using debiased data splits, the reward for simply memorising the training data is reduced, suggesting that the ability of the model to make accurate predictions for novel candidate ligands will improve. To test this hypothesis, we use distance-based data splits to measure how well a model can generalise. We first confirm that models perform better for randomly split held-out sets than for distant held-out sets. We then debias the data and find, surprisingly, that debiasing typically reduces the ability of models to make accurate predictions for distant held-out test sets. These results suggest that debiasing reduces the information available to a model, impairing its ability to generalise.

Improving Auto-Detection of Phishing Websites using Fresh-Phish Framework

2020 ◽  
pp. 326-340
Author(s):  
Hossein Shirazi ◽  
Kyle Haefner ◽  
Indrakshi Ray
Keyword(s):  
Machine Learning ◽  
Social Engineering ◽  
The Internet ◽  
Privacy And Security ◽  
Machine Learning Classifiers ◽  
Learning Classifiers ◽  
Phishing Attacks ◽  
Website Features ◽  
Learning Data

Denizens of the Internet are under a barrage of phishing attacks of increasing frequency and sophistication. Emails accompanied by authentic looking websites are ensnaring users who, unwittingly, hand over their credentials compromising both their privacy and security. Methods such as the blacklisting of these phishing websites become untenable and cannot keep pace with the explosion of fake sites. Detection of nefarious websites must become automated and be able to adapt to this ever-evolving form of social engineering. There is an improved framework that was previously implemented called “Fresh-Phish”, for creating current machine-learning data for phishing websites. The improved framework uses a total of 28 different website features that query using python, then a large labeled dataset is built and analyze over several machine learning classifiers against this dataset to determine which is the most accurate. This modified framework improves the accuracy of modeling those features by using integer rather than binary values where possible. This article analyzes not just the accuracy of the technique, but also how long it takes to train the model.

scholarly journals Multicenter validation of a machine learning algorithm for 48 hour all-cause mortality prediction

2018 ◽  
Author(s):  
Hamid Mohamadlou ◽  
Saarang Panchavati ◽  
Jacob Calvert ◽  
Anna Lynn-Palevsky ◽  
Christopher Barton ◽  
...  
Keyword(s):  
Machine Learning ◽  
Operating Characteristic ◽  
Learning Algorithm ◽  
Clinical Care ◽  
Characteristic Curve ◽  
Mortality Prediction ◽  
Health Centers ◽  
Training Data ◽  
All Cause Mortality ◽  
Test Sets

AbstractPurposeThis study evaluates a machine-learning-based mortality prediction tool.Materials and MethodsWe conducted a retrospective study with data drawn from three academic health centers. Inpatients of at least 18 years of age and with at least one observation of each vital sign were included. Predictions were made at 12, 24, and 48 hours before death. Models fit to training data from each institution were evaluated on hold-out test data from the same institution and data from the remaining institutions. Predictions were compared to those of qSOFA and MEWS using area under the receiver operating characteristic curve (AUROC).ResultsFor training and testing on data from a single institution, machine learning predictions averaged AUROCs of 0.97, 0.96, and 0.95 across institutional test sets for 12-, 24-, and 48-hour predictions, respectively. When trained and tested on data from different hospitals, the algorithm achieved AUROC up to 0.95, 0.93, and 0.91, for 12-, 24-, and 48-hour predictions, respectively. MEWS and qSOFA had average 48-hour AUROCs of 0.86 and 0.82, respectively.ConclusionThis algorithm may help identify patients in need of increased levels of clinical care.

Authorship Attribution for Online Social Media

2018 ◽  
pp. 141-165 ◽  
Author(s):  
Ritu Banga ◽  
Akanksha Bhardwaj ◽  
Sheng-Lung Peng ◽  
Gulshan Shrivastava
Keyword(s):  
Machine Learning ◽  
Social Media ◽  
Authorship Attribution ◽  
The Internet ◽  
Writing Style ◽  
Machine Learning Classifiers ◽  
Online Social Media ◽  
Comprehensive Knowledge ◽  
Authorship Identification ◽  
Cyber Ethics

This chapter gives a comprehensive knowledge of various machine learning classifiers to achieve authorship attribution (AA) on short texts, specifically tweets. The need for authorship identification is due to the increasing crime on the internet, which breach cyber ethics by raising the level of anonymity. AA of online messages has witnessed interest from many research communities. Many methods such as statistical and computational have been proposed by linguistics and researchers to identify an author from their writing style. Various ways of extracting and selecting features on the basis of dataset have been reviewed. The authors focused on n-grams features as they proved to be very effective in identifying the true author from a given list of known authors. The study has demonstrated that AA is achievable on the basis of selection criteria of features and methods in small texts and also proved the accuracy of analysis changes according to combination of features. The authors found character grams are good features for identifying the author but are not yet able to identify the author independently.

Sign in / Sign up

Export Citation Format

Share Document