There is an ongoing debate in the NLP community whether modern language models contain linguistic knowledge, recovered through so-called probes. In this paper, we study whether linguistic knowledge is a necessary condition for the good performance of modern language models, which we call the rediscovery hypothesis. In the first place, we show that language models that are significantly compressed but perform well on their pretraining objectives retain good scores when probed for linguistic structures. This result supports the rediscovery hypothesis and leads to the second contribution of our paper: an information-theoretic framework that relates language modeling objectives with linguistic information. This framework also provides a metric to measure the impact of linguistic information on the word prediction task. We reinforce our analytical results with various experiments, both on synthetic and on real NLP tasks in English.

In many application domains such as medicine, information retrieval, cybersecurity, social media, etc., datasets used for inducing classification models often have an unequal distribution of the instances of each class. This situation, known as imbalanced data classification, causes low predictive performance for the minority class examples. Thus, the prediction model is unreliable although the overall model accuracy can be acceptable. Oversampling and undersampling techniques are well-known strategies to deal with this problem by balancing the number of examples of each class. However, their effectiveness depends on several factors mainly related to data intrinsic characteristics, such as imbalance ratio, dataset size and dimensionality, overlapping between classes or borderline examples. In this work, the impact of these factors is analyzed through a comprehensive comparative study involving 40 datasets from different application areas. The objective is to obtain models for automatic selection of the best resampling strategy for any dataset based on its characteristics. These models allow us to check several factors simultaneously considering a wide range of values since they are induced from very varied datasets that cover a broad spectrum of conditions. This differs from most studies that focus on the individual analysis of the characteristics or cover a small range of values. In addition, the study encompasses both basic and advanced resampling strategies that are evaluated by means of eight different performance metrics, including new measures specifically designed for imbalanced data classification. The general nature of the proposal allows the choice of the most appropriate method regardless of the domain, avoiding the search for special purpose techniques that could be valid for the target data.

Outliers widely occur in big-data applications and may severely affect statistical estimation and inference. In this paper, a framework of outlier-resistant estimation is introduced to robustify an arbitrarily given loss function. It has a close connection to the method of trimming and includes explicit outlyingness parameters for all samples, which in turn facilitates computation, theory, and parameter tuning. To tackle the issues of nonconvexity and nonsmoothness, we develop scalable algorithms with implementation ease and guaranteed fast convergence. In particular, a new technique is proposed to alleviate the requirement on the starting point such that on regular datasets, the number of data resamplings can be substantially reduced. Based on combined statistical and computational treatments, we are able to perform nonasymptotic analysis beyond M-estimation. The obtained resistant estimators, though not necessarily globally or even locally optimal, enjoy minimax rate optimality in both low dimensions and high dimensions. Experiments in regression, classification, and neural networks show excellent performance of the proposed methodology at the occurrence of gross outliers.

We introduce a machine-in-the-loop pipeline that aims to address root causes of unwanted bias in natural language based supervised machine learning tasks in the education domain. Learning from the experiences of students is foundational for education researchers, and academic administrators. 21st-century skills learned from experience are becoming a core part of college and career readiness as well as the hiring process in the new knowledge economy. Minoritized students demonstrate these skills in their daily lives, but documenting, assessing, and validating these skills is a huge problem for educational institutions. As an equity focused online platform, LivedX translates minoritized students' lived experiences into the 21st century skills, issues micro-credentials, and creates personal 21st century skills portfolio. To automate the micro credential mining from the natural language texts received from the students' submitted essays, we employed a bag-of-word model to construct a multi-output classifier. Despite our goal, our model initially exacerbated disparate impact on minoritized students. We used a machine-in-the-loop model development pipeline to address the problem and refine the aforementioned model to ensure fairness in its prediction.

The main focus of information retrieval methods is to provide accurate and efficient results which are cost-effective too. LINGO (Label Induction Grouping Algorithm) is a clustering algorithm that aims to provide search results in form of quality clusters but also has a few limitations. In this paper, our focus is based on achieving results that are more meaningful and improving the overall performance of the algorithm. LINGO works on two main steps; Cluster Label Induction by using Latent Semantic Indexing technique (LSI) and Cluster content discovery by using the Vector Space Model (VSM). As LINGO uses VSM in cluster content discovery, our task is to replace VSM with LSI for cluster content discovery and to analyze the feasibility of using LSI with Okapi BM25. The next task is to compare the results of a modified method with the LINGO original method. The research is applied to five different text-based data sets to get more reliable results for every method. Research results show that LINGO produces 40-50% better results when using LSI for content Discovery. From theoretical evidence using Okapi BM25 for scoring method in LSI (LSI+Okapi BM25) for cluster content discovery instead of VSM, also results in better clusters generation in terms of scalability and performance when compares to both VSM and LSI's Results.

Normalizing flows are diffeomorphic, typically dimension-preserving, models trained using the likelihood of the model. We use the SurVAE framework to construct dimension reducing surjective flows via a new layer, known as the funnel. We demonstrate its efficacy on a variety of datasets, and show it improves upon or matches the performance of existing flows while having a reduced latent space size. The funnel layer can be constructed from a wide range of transformations including restricted convolution and feed forward layers.

Sparsity priors are commonly used in denoising and image reconstruction. For analysis-type priors, a dictionary defines a representation of signals that is likely to be sparse. In most situations, this dictionary is not known, and is to be recovered from pairs of ground-truth signals and measurements, by minimizing the reconstruction error. This defines a hierarchical optimization problem, which can be cast as a bi-level optimization. Yet, this problem is unsolvable, as reconstructions and their derivative wrt the dictionary have no closed-form expression. However, reconstructions can be iteratively computed using the Forward-Backward splitting (FB) algorithm. In this paper, we approximate reconstructions by the output of the aforementioned FB algorithm. Then, we leverage automatic differentiation to evaluate the gradient of this output wrt the dictionary, which we learn with projected gradient descent. Experiments show that our algorithm successfully learns the 1D Total Variation (TV) dictionary from piecewise constant signals. For the same case study, we propose to constrain our search to dictionaries of 0-centered columns, which removes undesired local minima and improves numerical stability.

Patient scheduling is a difficult task as it involves dealing with stochastic factors such as an unknown arrival flow of patients. Scheduling radiotherapy treatments for cancer patients faces a similar problem. Curative patients need to start their treatment within the recommended deadlines, i.e., 14 or 28 days after their admission while reserving treatment capacity for palliative patients who require urgent treatments within 1 to 3 days after their admission. Most cancer centers solve the problem by reserving a fixed number of treatment slots for emergency patients. However, this flat-reservation approach is not ideal and can cause overdue treatments for emergency patients on some days while not fully exploiting treatment capacity on some other days, which also leads to delaying treatment for curative patients. This problem is especially severe in large and crowded hospitals. In this paper, we propose a prediction-based approach for online dynamic radiotherapy scheduling. An offline problem where all future patient arrivals are known in advance is solved to optimality using Integer Programming. A regression model is then trained to recognize the links between patients' arrival patterns and their ideal waiting time. The trained regression model is then embedded in a prediction-based approach that schedules a patient based on their characteristics and the present state of the calendar. The numerical results show that our prediction-based approach efficiently prevents overdue treatments for emergency patients while maintaining a good waiting time compared to other scheduling approaches based on a flat-reservation policy.

A major impediment to successful drug development is the complexity, cost, and scale of clinical trials. The detailed internal structure of clinical trial data can make conventional optimization difficult to achieve. Recent advances in machine learning, specifically graph-structured data analysis, have the potential to enable significant progress in improving the clinical trial design. TrialGraph seeks to apply these methodologies to produce a proof-of-concept framework for developing models which can aid drug development and benefit patients. In this work, we first introduce a curated clinical trial data set compiled from the CT.gov, AACT and TrialTrove databases (n=1191 trials; representing one million patients) and describe the conversion of this data to graph-structured formats. We then detail the mathematical basis and implementation of a selection of graph machine learning algorithms, which typically use standard machine classifiers on graph data embedded in a low-dimensional feature space. We trained these models to predict side effect information for a clinical trial given information on the disease, existing medical conditions, and treatment. The MetaPath2Vec algorithm performed exceptionally well, with standard Logistic Regression, Decision Tree, Random Forest, Support Vector, and Neural Network classifiers exhibiting typical ROC-AUC scores of 0.85, 0.68, 0.86, 0.80, and 0.77, respectively. Remarkably, the best performing classifiers could only produce typical ROC-AUC scores of 0.70 when trained on equivalent array-structured data. Our work demonstrates that graph modelling can significantly improve prediction accuracy on appropriate datasets. Successive versions of the project that refine modelling assumptions and incorporate more data types can produce excellent predictors with real-world applications in drug development.

In the context of time series forecasting, it is a common practice to evaluate multiple methods and choose one of these methods or an ensemble for producing the best forecasts. However, choosing among different ensembles over multiple methods remains a challenging task that undergoes a combinatorial explosion as the number of methods increases. In the context of demand forecasting or revenue forecasting, this challenge is further exacerbated by a large number of time series as well as limited historical data points available due to changing business context. Although deep learning forecasting methods aim to simultaneously forecast large collections of time series, they become challenging to apply in such scenarios due to the limited history available and might not yield desirable results. We propose a framework for forecasting short high-dimensional time series data by combining low-rank temporal matrix factorization and optimal model selection on latent time series using cross-validation. We demonstrate that forecasting the latent factors leads to significant performance gains as compared to directly applying different uni-variate models on time series. Performance has been validated on a truncated version of the M4 monthly dataset which contains time series data from multiple domains showing the general applicability of the method. Moreover, it is amenable to incorporating the analyst view of the future owing to the low number of latent factors which is usually impractical when applying forecasting methods directly to high dimensional datasets.