Recent work has raised concerns on the risk of spurious correlations and unintended biases in statistical machine learning models that threaten model robustness and fairness. In this paper, we propose a simple and intuitive regularization approach to integrate causal knowledge during model training and build a robust and fair model by emphasizing causal features and de-emphasizing spurious features. Specifically, we first manually identify causal and spurious features with principles inspired from the counterfactual framework of causal inference. Then, we propose a regularization approach to penalize causal and spurious features separately. By adjusting the strength of the penalty for each type of feature, we build a predictive model that relies more on causal features and less on non-causal features. We conduct experiments to evaluate model robustness and fairness on three datasets with multiple metrics. Empirical results show that the new models built with causal awareness significantly improve model robustness with respect to counterfactual texts and model fairness with respect to sensitive attributes.

Studies across many disciplines have shown that lexical choice can affect audience perception. For example, how users describe themselves in a social media profile can affect their perceived socio-economic status. However, we lack general methods for estimating the causal effect of lexical choice on the perception of a specific sentence. While randomized controlled trials may provide good estimates, they do not scale to the potentially millions of comparisons necessary to consider all lexical choices. Instead, in this paper, we first offer two classes of methods to estimate the effect on perception of changing one word to another in a given sentence. The first class of algorithms builds upon quasi-experimental designs to estimate individual treatment effects from observational data. The second class treats treatment effect estimation as a classification problem. We conduct experiments with three data sources (Yelp, Twitter, and Airbnb), finding that the algorithmic estimates align well with those produced by randomized-control trials. Additionally, we find that it is possible to transfer treatment effect classifiers across domains and still maintain high accuracy.

As statistical classifiers become integrated into real-world applications, it is important to consider not only their accuracy but also their robustness to changes in the data distribution. Although identifying and controlling for confounding variables Z - correlated with both the input X of a classifier and its output Y - has been assiduously studied in empirical social science, it is often neglected in text classification. This can be understood by the fact that, if we assume that the impact of confounding variables does not change between the time we fit a model and the time we use it, then prediction accuracy should only be slightly affected. We show in this paper that this assumption often does not hold and that when the influence of a confounding variable changes from training time to prediction time (i.e. under confounding shift), the classifier accuracy can degrade rapidly. We use Pearl's back-door adjustment as a predictive framework to develop a model robust to confounding shift under the condition that Z is observed at training time. Our approach does not make any causal conclusions but by experimenting on 6 datasets, we show that our approach is able to outperform baselines 1) in controlled cases where confounding shift is manually injected between fitting time and prediction time 2) in natural experiments where confounding shift appears either abruptly or gradually 3) in cases where there is one or multiple confounders. Finally, we discuss multiple issues we encountered during this research such as the effect of noise in the observation of Z and the importance of only controlling for confounding variables.

As statistical classifiers become integrated into real-world applications, it is important to consider not only their accuracy but also their robustness to changes in the data distribution. In this paper, we consider the case where there is an unobserved confounding variable $z$ that influences both the features $\mathbf{x}$ and the class variable $y$. When the influence of $z$ changes from training to testing data, we find that the classifier accuracy can degrade rapidly. In our approach, we assume that we can predict the value of $z$ at training time with some error. The prediction for $z$ is then fed to Pearl's back-door adjustment to build our model. Because of the attenuation bias caused by measurement error in $z$, standard approaches to controlling for $z$ are ineffective. In response, we propose a method to properly control for the influence of $z$ by first estimating its relationship with the class variable $y$, then updating predictions for $z$ to match that estimated relationship. By adjusting the influence of $z$, we show that we can build a model that exceeds competing baselines on accuracy as well as on robustness over a range of confounding relationships.

Deep learning algorithms have recently produced state-of-the-art accuracy in many classification tasks, but this success is typically dependent on access to many annotated training examples. For domains without such data, an attractive alternative is to train models with light, or distant supervision. In this paper, we introduce a deep neural network for the Learning from Label Proportion (LLP) setting, in which the training data consist of bags of unlabeled instances with associated label distributions for each bag. We introduce a new regularization layer, Batch Averager, that can be appended to the last layer of any deep neural network to convert it from supervised learning to LLP. This layer can be implemented readily with existing deep learning packages. To further support domains in which the data consist of two conditionally independent feature views (e.g. image and text), we propose a co-training algorithm that iteratively generates pseudo bags and refits the deep LLP model to improve classification accuracy. We demonstrate our models on demographic attribute classification (gender and race/ethnicity), which has many applications in social media analysis, public health, and marketing. We conduct experiments to predict demographics of Twitter users based on their tweets and profile image, without requiring any user-level annotations for training. We find that the deep LLP approach outperforms baselines for both text and image features separately. Additionally, we find that co-training algorithm improves image and text classification by 4% and 8% absolute F1, respectively. Finally, an ensemble of text and image classifiers further improves the absolute F1 measure by 4% on average.

As statistical classifiers become integrated into real-world applications, it is important to consider not only their accuracy but also their robustness to changes in the data distribution. In this paper, we consider the case where there is an unobserved confounding variable z that influences both the features x and the class variable y. When the influence of z changes from training to testing data, we find that the classifier accuracy can degrade rapidly. In our approach, we assume that we can predict the value of z at training time with some error. The prediction for z is then fed to Pearl's back-door adjustment to build our model. Because of the attenuation bias caused by measurement error in z, standard approaches to controlling for z are ineffective. In response, we propose a method to properly control for the influence of z by first estimating its relationship with the class variable y, then updating predictions for z to match that estimated relationship. By adjusting the influence of z, we show that we can build a model that exceeds competing baselines on accuracy as well as on robustness over a range of confounding relationships.

Consumer protection agencies are charged with safeguarding the public from hazardous products, but the thousands of products under their jurisdiction make it challenging to identify and respond to consumer complaints quickly. In this paper, we propose a system to mine Amazon.com reviews to identify products that may pose safety or health hazards. Since labeled data for this task are scarce, our approach combines positive unlabeled learning with domain adaptation to train a classifier from consumer complaints submitted to an online government portal. We find that our approach results in an absolute F1 score improvement of 8% over the best competing baseline. Furthermore, when we apply the classifier to Amazon reviews of known recalled products, we identify safety hazard reports prior to the recall date for 45% of the products. This suggests that the system may be able to provide an early warning system to alert consumers to hazardous products before an official recall is announced.

As text classifiers become increasingly used in real-time applications, it is critical to consider not only their accuracy but also their robustness to changes in the data distribution. In this paper, we consider the case where there is a confounding variable Z that influences both the text features X and the class variable Y. For example, a classifier trained to predict the health status of a user based on their online communications may be confounded by socioeconomic variables. When the influence of Z changes from training to testing data, we find that classifier accuracy can degrade rapidly. Our approach, based on Pearl's back-door adjustment, estimates the underlying effect of a text variable on the class variable while controlling for the confounding variable. Although our goal is prediction, not causal inference, we find that such adjustments are essential to building text classifiers that are robust to confounding variables. On three diverse text classifications tasks, we find that covariate adjustment results in higher accuracy than competing baselines over a range of confounding relationships (e.g., in one setting, accuracy improves from 60% to 81%).

Recent work has demonstrated the value of social media monitoring for health surveillance (e.g., tracking influenza or depression rates). It is an open question whether such data can be used to make causal inferences (e.g., determining which activities lead to increased depression rates). Even in traditional, restricted domains, estimating causal effects from observational data is highly susceptible to confounding bias. In this work, we estimate the effect of exercise on mental health from Twitter, relying on statistical matching methods to reduce confounding bias. We train a text classifier to estimate the volume of a user's tweets expressing anxiety, depression, or anger, then compare two groups: those who exercise regularly (identified by their use of physical activity trackers like Nike+), and a matched control group. We find that those who exercise regularly have significantly fewer tweets expressing depression or anxiety; there is no significant difference in rates of tweets expressing anger. We additionally perform a sensitivity analysis to investigate how the many experimental design choices in such a study impact the final conclusions, including the quality of the classifier and the construction of the control group.

A common bottleneck in deploying supervised learning systems is collecting human-annotated examples. In many domains, annotators form an opinion about the label of an example incrementally -- e.g., each additional word read from a document or each additional minute spent inspecting a video helps inform the annotation. In this paper, we investigate whether we can train learning systems more efficiently by requesting an annotation before inspection is fully complete -- e.g., after reading only 25 words of a document. While doing so may reduce the overall annotation time, it also introduces the risk that the annotator might not be able to provide a label if interrupted too early. We propose an anytime active learning approach that optimizes the annotation time and response rate simultaneously. We conduct user studies on two document classification datasets and develop simulated annotators that mimic the users. Our simulated experiments show that anytime active learning outperforms several baselines on these two datasets. For example, with an annotation budget of one hour, training a classifier by annotating the first 25 words of each document reduces classification error by 17% over annotating the first 100 words of each document.