Gradient Boosted Decision Trees (GBDTs) are dominant machine learning algorithms for modeling discrete or tabular data. Unlike neural networks with millions of trainable parameters, GBDTs optimize loss function in an additive manner and have a single trainable parameter per leaf, which makes it easy to apply high-order optimization of the loss function. In this paper, we introduce high-order optimization for GBDTs based on numerical optimization theory which allows us to construct trees based on high-order derivatives of a given loss function. In the experiments, we show that high-order optimization has faster per-iteration convergence that leads to reduced running time. Our solution can be easily parallelized and run on GPUs with little overhead on the code. Finally, we discuss future potential improvements such as automatic differentiation of arbitrary loss function and combination of GBDTs with neural networks.

Random forests (RFs) [Bre01] are machine learning models with various applications in areas like E-commerce, Finance and Medicine. They consist of multiple decision trees that use different subsets of the available features. Given an input, every tree makes an individual decision and the output of the random forest is obtained by a majority vote. They have low risk of overfitting; support both classification and regression tasks and come equipped with some feature importance measures [Bre01]. However, feature importance measures can be too simplistic as they can represent neither joint effects of features (e.g., multi-drug interactions) nor non-monotonicity (e.g., increasing the weight may be healthy for an underweight person, but not for an overweight person). In recent years, a variety of other explanation methods has been proposed. Modelagnostic feature importance measures like LIME [RSG16], SHAP [LL17] and MAPLE [PMT18] have similar limitations like the feature importance measures defined for random forests.

Power transformers are an important component of a nuclear power plant (NPP). Currently, the NPP operates a lot of power transformers with extended service life, which exceeds the designated 25 years. Due to the extension of the service life, the task of monitoring the technical condition of power transformers becomes urgent. An important method for monitoring power transformers is Chromatographic Analysis of Dissolved Gas. It is based on the principle of controlling the concentration of gases dissolved in transformer oil. The appearance of almost any type of defect in equipment is accompanied by the formation of gases that dissolve in oil, and specific types of defects generate their gases in different quantities. At present, at NPPs, the monitoring systems for transformer equipment use predefined control limits for the concentration of dissolved gases in the oil. This study describes the stages of developing an algorithm to detect defects and faults in transformers automatically using machine learning and data analysis methods. Among machine learning models, we trained Logistic Regression, Decision Trees, Random Forest, Gradient Boosting, Neural Networks. The best of them were then combined into an ensemble (StackingClassifier) showing F1-score of 0.974 on a test sample. To develop mathematical models, we used data on the state of transformers, containing time series with values of gas concentrations (H2, CO, C2H4, C2H2). The datasets were labeled and contained four operating modes: normal mode, partial discharge, low energy discharge, low-temperature overheating.

This is the first in a series of posts that illustrate what our data team is up to, experimenting with, and building'under the hood' at CitizenNet. He has been involved in web-scale machine learning and information retrieval for over 10 years. One of the first posts we published spoke at a high level of the technical problem CitizenNet is trying to solve. In essence, we are trying to predict what combinations of demographic and interest targets will be interested in some piece of content. On the CitizenNet platform, a user would create a project that would define (broadly) the target audience, the pieces of Facebook content they are looking to promote, and other campaign and financial information. Behind the scenes, a robust prediction system builds the targets for the project.

The availability of commercial wearable trackers equipped with features to monitor sleep duration and quality has enabled more useful sleep health monitoring applications and analyses. However, much research has reported the challenge of long-term user retention in sleep monitoring through these modalities. Since modern Internet users own multiple mobile devices, our work explores the possibility of employing ubiquitous mobile devices and passive WiFi sensing techniques to predict sleep duration as the fundamental measure for complementing long-term sleep monitoring initiatives. In this paper, we propose SleepMore, an accurate and easy-to-deploy sleep-tracking approach based on machine learning over the user's WiFi network activity. It first employs a semi-personalized random forest model with an infinitesimal jackknife variance estimation method to classify a user's network activity behavior into sleep and awake states per minute granularity. Through a moving average technique, the system uses these state sequences to estimate the user's nocturnal sleep period and its uncertainty rate. Uncertainty quantification enables SleepMore to overcome the impact of noisy WiFi data that can yield large prediction errors. We validate SleepMore using data from a month-long user study involving 46 college students and draw comparisons with the Oura Ring wearable. Beyond the college campus, we evaluate SleepMore on non-student users of different housing profiles. Our results demonstrate that SleepMore produces statistically indistinguishable sleep statistics from the Oura ring baseline for predictions made within a 5% uncertainty rate. These errors range between 15-28 minutes for determining sleep time and 7-29 minutes for determining wake time, proving statistically significant improvements over prior work. Our in-depth analysis explains the sources of errors.

Machine Learning involves algorithms that learn from patterns of data and then apply it to decision making. Deep Learning, on the other hand, is able to learn through processing data on its own and is quite similar to the human brain where it identifies something, analyse it, and makes a decision.

Traditional evaluation metrics for classification in natural language processing such as accuracy and area under the curve fail to differentiate between models with different predictive behaviors despite their similar performance metrics. We introduce sensitivity score, a metric that scrutinizes models' behaviors at the vocabulary level to provide insights into disparities in their decision-making logic. We assess the sensitivity score on a set of representative words in the test set using two classifiers trained for hospital readmission classification with similar performance statistics. Our experiments compare the decision-making logic of clinicians and classifiers based on rank correlations of sensitivity scores. The results indicate that the language model's sensitivity score aligns better with the professionals than the xgboost classifier on tf-idf embeddings, which suggests that xgboost uses some spurious features. Overall, this metric offers a novel perspective on assessing models' robustness by quantifying their discrepancy with professional opinions. Our code is available on GitHub (https://github.com/nyuolab/Model_Sensitivity).

Hyper-parameters (HPs) are an important part of machine learning (ML) model development and can greatly influence performance. This paper studies their behavior for three algorithms: Extreme Gradient Boosting (XGB), Random Forest (RF), and Feedforward Neural Network (FFNN) with structured data. Our empirical investigation examines the qualitative behavior of model performance as the HPs vary, quantifies the importance of each HP for different ML algorithms, and stability of the performance near the optimal region. Based on the findings, we propose a set of guidelines for efficient HP tuning by reducing the search space.

Threshold Autoregressive (TAR) models have been widely used by statisticians for non-linear time series forecasting during the past few decades, due to their simplicity and mathematical properties. On the other hand, in the forecasting community, general-purpose tree-based regression algorithms (forests, gradient-boosting) have become popular recently due to their ease of use and accuracy. In this paper, we explore the close connections between TAR models and regression trees. These enable us to use the rich methodology from the literature on TAR models to define a hierarchical TAR model as a regression tree that trains globally across series, which we call SETAR-Tree. In contrast to the general-purpose tree-based models that do not primarily focus on forecasting, and calculate averages at the leaf nodes, we introduce a new forecasting-specific tree algorithm that trains global Pooled Regression (PR) models in the leaves allowing the models to learn cross-series information and also uses some time-series-specific splitting and stopping procedures. The depth of the tree is controlled by conducting a statistical linearity test commonly employed in TAR models, as well as measuring the error reduction percentage at each node split. Thus, the proposed tree model requires minimal external hyperparameter tuning and provides competitive results under its default configuration. We also use this tree algorithm to develop a forest where the forecasts provided by a collection of diverse SETAR-Trees are combined during the forecasting process. In our evaluation on eight publicly available datasets, the proposed tree and forest models are able to achieve significantly higher accuracy than a set of state-of-the-art tree-based algorithms and forecasting benchmarks across four evaluation metrics.

Tuberculosis (TB), an infectious bacterial disease, is a significant cause of death, especially in low-income countries, with an estimated ten million new cases reported globally in $2020$. While TB is treatable, non-adherence to the medication regimen is a significant cause of morbidity and mortality. Thus, proactively identifying patients at risk of dropping off their medication regimen enables corrective measures to mitigate adverse outcomes. Using a proxy measure of extreme non-adherence and a dataset of nearly $700,000$ patients from four states in India, we formulate and solve the machine learning (ML) problem of early prediction of non-adherence based on a custom rank-based metric. We train ML models and evaluate against baselines, achieving a $\sim 100\%$ lift over rule-based baselines and $\sim 214\%$ over a random classifier, taking into account country-wide large-scale future deployment. We deal with various issues in the process, including data quality, high-cardinality categorical data, low target prevalence, distribution shift, variation across cohorts, algorithmic fairness, and the need for robustness and explainability. Our findings indicate that risk stratification of non-adherent patients is a viable, deployable-at-scale ML solution. As the official AI partner of India's Central TB Division, we are working on multiple city and state-level pilots with the goal of pan-India deployment.