Justin Sybrandt

Daniel Golovin, Gabor Bartok, Eric Chen et al. · mit

In many software systems, heuristics are used to make decisions - such as cache eviction, task scheduling, and information presentation - that have a significant impact on overall system behavior. While machine learning may outperform these heuristics, replacing existing heuristics in a production system safely and reliably can be prohibitively costly. We present SmartChoices, a novel approach that reduces the cost to deploy production-ready ML solutions for contextual bandits problems. SmartChoices' interface cleanly separates problem formulation from implementation details: engineers describe their use case by defining datatypes for the context, arms, and feedback that are passed to SmartChoices APIs, while SmartChoices manages encoding & logging data and training, evaluating & deploying policies. Our implementation codifies best practices, is efficient enough for use in low-level applications, and provides valuable production features off the shelf via a shared library. Overall, SmartChoices enables non-experts to rapidly deploy production-ready ML solutions by eliminating many sources of technical debt common to ML systems. Engineers have independently used SmartChoices to improve a wide range of software including caches, batch processing workloads, and UI layouts, resulting in better latency, throughput, and click-through rates.

David Marasco, Ilya Tyagin, Justin Sybrandt et al.

This project demonstrates how medical corpus hypothesis generation, a knowledge discovery field of AI, can be used to derive new research angles for landscape and urban planners. The hypothesis generation approach herein consists of a combination of deep learning with topic modeling, a probabilistic approach to natural language analysis that scans aggregated research databases for words that can be grouped together based on their subject matter commonalities; the word groups accordingly form topics that can provide implicit connections between two general research terms. The hypothesis generation system AGATHA was used to identify likely conceptual relationships between emerging infectious diseases (EIDs) and deforestation, with the objective of providing landscape planners guidelines for productive research directions to help them formulate research hypotheses centered on deforestation and EIDs that will contribute to the broader health field that asserts causal roles of landscape-level issues. This research also serves as a partial proof-of-concept for the application of medical database hypothesis generation to medicine-adjacent hypothesis discovery.

Justin Sybrandt, Angelo Carrabba, Alexander Herzog et al.

The potential for automatic hypothesis generation (HG) systems to improve research productivity keeps pace with the growing set of publicly available scientific information. But as data becomes easier to acquire, we must understand the effect different textual data sources have on our resulting hypotheses. Are abstracts enough for HG, or does it need full-text papers? How many papers does an HG system need to make valuable predictions? How sensitive is a general-purpose HG system to hyperparameter values or input quality? What effect does corpus size and document length have on HG results? To answer these questions we train multiple versions of knowledge network-based HG system, Moliere, on varying corpora in order to compare challenges and trade offs in terms of result quality and computational requirements. Moliere generalizes main principles of similar knowledge network-based HG systems and reinforces them with topic modeling components. The corpora include the abstract and full-text versions of PubMed Central, as well as iterative halves of MEDLINE, which allows us to compare the effect document length and count has on the results. We find that, quantitatively, corpora with a higher median document length result in marginally higher quality results, yet require substantially longer to process. However, qualitatively, full-length papers introduce a significant number of intruder terms to the resulting topics, which decreases human interpretability. Additionally, we find that the effect of document length is greater than that of document count, even if both sets contain only paper abstracts. Reproducibility: Our code and data are available at github.com/jsybran/moliere, and bit.ly/2GxghpM respectively.

Ilya Tyagin, Ankit Kulshrestha, Justin Sybrandt et al.

In 2020, the White House released the, "Call to Action to the Tech Community on New Machine Readable COVID-19 Dataset," wherein artificial intelligence experts are asked to collect data and develop text mining techniques that can help the science community answer high-priority scientific questions related to COVID-19. The Allen Institute for AI and collaborators announced the availability of a rapidly growing open dataset of publications, the COVID-19 Open Research Dataset (CORD-19). As the pace of research accelerates, biomedical scientists struggle to stay current. To expedite their investigations, scientists leverage hypothesis generation systems, which can automatically inspect published papers to discover novel implicit connections. We present an automated general purpose hypothesis generation systems AGATHA-C and AGATHA-GP for COVID-19 research. The systems are based on graph-mining and the transformer model. The systems are massively validated using retrospective information rediscovery and proactive analysis involving human-in-the-loop expert analysis. Both systems achieve high-quality predictions across domains (in some domains up to 0.97% ROC AUC) in fast computational time and are released to the broad scientific community to accelerate biomedical research. In addition, by performing the domain expert curated study, we show that the systems are able to discover on-going research findings such as the relationship between COVID-19 and oxytocin hormone.

Farah Alshanik, Amy Apon, Alexander Herzog et al.

Text preprocessing is an essential step in text mining. Removing words that can negatively impact the quality of prediction algorithms or are not informative enough is a crucial storage-saving technique in text indexing and results in improved computational efficiency. Typically, a generic stop word list is applied to a dataset regardless of the domain. However, many common words are different from one domain to another but have no significance within a particular domain. Eliminating domain-specific common words in a corpus reduces the dimensionality of the feature space, and improves the performance of text mining tasks. In this paper, we present a novel mathematical approach for the automatic extraction of domain-specific words called the hyperplane-based approach. This new approach depends on the notion of low dimensional representation of the word in vector space and its distance from hyperplane. The hyperplane-based approach can significantly reduce text dimensionality by eliminating irrelevant features. We compare the hyperplane-based approach with other feature selection methods, namely \c{hi}2 and mutual information. An experimental study is performed on three different datasets and five classification algorithms, and measure the dimensionality reduction and the increase in the classification performance. Results indicate that the hyperplane-based approach can reduce the dimensionality of the corpus by 90% and outperforms mutual information. The computational time to identify the domain-specific words is significantly lower than mutual information.

Fei Ding, Xiaohong Zhang, Justin Sybrandt et al.

Graph representation learning based on graph neural networks (GNNs) can greatly improve the performance of downstream tasks, such as node and graph classification. However, the general GNN models do not aggregate node information in a hierarchical manner, and can miss key higher-order structural features of many graphs. The hierarchical aggregation also enables the graph representations to be explainable. In addition, supervised graph representation learning requires labeled data, which is expensive and error-prone. To address these issues, we present an unsupervised graph representation learning method, Unsupervised Hierarchical Graph Representation (UHGR), which can generate hierarchical representations of graphs. Our method focuses on maximizing mutual information between "local" and high-level "global" representations, which enables us to learn the node embeddings and graph embeddings without any labeled data. To demonstrate the effectiveness of the proposed method, we perform the node and graph classification using the learned node and graph embeddings. The results show that the proposed method achieves comparable results to state-of-the-art supervised methods on several benchmarks. In addition, our visualization of hierarchical representations indicates that our method can capture meaningful and interpretable clusters.

Justin Sybrandt, Ilya Safro

Biomedical research papers use significantly different language and jargon when compared to typical English text, which reduces the utility of pre-trained NLP models in this domain. Meanwhile Medline, a database of biomedical abstracts, introduces nearly a million new documents per-year. Applications that could benefit from understanding this wealth of publicly available information, such as scientific writing assistants, chat-bots, or descriptive hypothesis generation systems, require new domain-centered approaches. A conditional language model, one that learns the probability of words given some a priori criteria, is a fundamental building block in many such applications. We propose a transformer-based conditional language model with a shallow encoder "condition" stack, and a deep "language model" stack of multi-headed attention blocks. The condition stack encodes metadata used to alter the output probability distribution of the language model stack. We sample this distribution in order to generate biomedical abstracts given only a proposed title, an intended publication year, and a set of keywords. Using typical natural language generation metrics, we demonstrate that this proposed approach is more capable of producing non-trivial relevant entities within the abstract body than the 1.5B parameter GPT-2 language model.

Justin Sybrandt, Ilya Tyagin, Michael Shtutman et al.

Medical research is risky and expensive. Drug discovery, as an example, requires that researchers efficiently winnow thousands of potential targets to a small candidate set for more thorough evaluation. However, research groups spend significant time and money to perform the experiments necessary to determine this candidate set long before seeing intermediate results. Hypothesis generation systems address this challenge by mining the wealth of publicly available scientific information to predict plausible research directions. We present AGATHA, a deep-learning hypothesis generation system that can introduce data-driven insights earlier in the discovery process. Through a learned ranking criteria, this system quickly prioritizes plausible term-pairs among entity sets, allowing us to recommend new research directions. We massively validate our system with a temporal holdout wherein we predict connections first introduced after 2015 using data published beforehand. We additionally explore biomedical sub-domains, and demonstrate AGATHA's predictive capacity across the twenty most popular relationship types. This system achieves best-in-class performance on an established benchmark, and demonstrates high recommendation scores across subdomains. Reproducibility: All code, experimental data, and pre-trained models are available online: sybrandt.com/2020/agatha

Justin Sybrandt, Ruslan Shaydulin, Ilya Safro

Problems in scientific computing, such as distributing large sparse matrix operations, have analogous formulations as hypergraph partitioning problems. A hypergraph is a generalization of a traditional graph wherein "hyperedges" may connect any number of nodes. As a result, hypergraph partitioning is an NP-Hard problem to both solve or approximate. State-of-the-art algorithms that solve this problem follow the multilevel paradigm, which begins by iteratively "coarsening" the input hypergraph to smaller problem instances that share key structural features. Once identifying an approximate problem that is small enough to be solved directly, that solution can be interpolated and refined to the original problem. While this strategy represents an excellent trade off between quality and running time, it is sensitive to coarsening strategy. In this work we propose using graph embeddings of the initial hypergraph in order to ensure that coarsened problem instances retrain key structural features. Our approach prioritizes coarsening within self-similar regions within the input graph, and leads to significantly improved solution quality across a range of considered hypergraphs. Reproducibility: All source code, plots and experimental data are available at https://sybrandt.com/2019/partition.

Justin Sybrandt, Ilya Safro

Typical graph embeddings may not capture type-specific bipartite graph features that arise in such areas as recommender systems, data visualization, and drug discovery. Machine learning methods utilized in these applications would be better served with specialized embedding techniques. We propose two embeddings for bipartite graphs that decompose edges into sets of indirect relationships between node neighborhoods. When sampling higher-order relationships, we reinforce similarities through algebraic distance on graphs. We also introduce ensemble embeddings to combine both into a "best of both worlds" embedding. The proposed methods are evaluated on link prediction and recommendation tasks and compared with other state-of-the-art embeddings. While being all highly beneficial in applications, we demonstrate that none of the considered embeddings is clearly superior (in contrast to what is claimed in many papers), and discuss the trade offs present among them. Reproducibility: Our code, data sets, and results are all publicly available online at: http://sybrandt.com/2020/fobe_hobe.

Justin Sybrandt, Michael Shtutman, Ilya Safro

The first step of many research projects is to define and rank a short list of candidates for study. In the modern rapidity of scientific progress, some turn to automated hypothesis generation (HG) systems to aid this process. These systems can identify implicit or overlooked connections within a large scientific corpus, and while their importance grows alongside the pace of science, they lack thorough validation. Without any standard numerical evaluation method, many validate general-purpose HG systems by rediscovering a handful of historical findings, and some wishing to be more thorough may run laboratory experiments based on automatic suggestions. These methods are expensive, time consuming, and cannot scale. Thus, we present a numerical evaluation framework for the purpose of validating HG systems that leverages thousands of validation hypotheses. This method evaluates a HG system by its ability to rank hypotheses by plausibility; a process reminiscent of human candidate selection. Because HG systems do not produce a ranking criteria, specifically those that produce topic models, we additionally present novel metrics to quantify the plausibility of hypotheses given topic model system output. Finally, we demonstrate that our proposed validation method aligns with real-world research goals by deploying our method within Moliere, our recent topic-driven HG system, in order to automatically generate a set of candidate genes related to HIV-associated neurodegenerative disease (HAND). By performing laboratory experiments based on this candidate set, we discover a new connection between HAND and Dead Box RNA Helicase 3 (DDX3). Reproducibility: code, validation data, and results can be found at sybrandt.com/2018/validation.

Ruslan Shaydulin, Justin Sybrandt

Since the Agile Manifesto, many organizations have explored agile development methods to replace traditional waterfall development. Interestingly, waterfall remains the most widely used practice, suggesting that there is something missing from the many "flavors" of agile methodologies. We explore seven of the most common practices to explore this, and evaluate each against a series of criteria centered around product quality and adherence to agile practices. We find that no methodology entirely replaces waterfall and summarize the strengths and weaknesses of each. From this, we conclude that agile methods are, as a whole, unable to cope with the realities of technical debt and large scale systems. Ultimately, no one methodology fits all projects.

Justin Sybrandt, Michael Shtutman, Ilya Safro

Hypothesis generation is becoming a crucial time-saving technique which allows biomedical researchers to quickly discover implicit connections between important concepts. Typically, these systems operate on domain-specific fractions of public medical data. MOLIERE, in contrast, utilizes information from over 24.5 million documents. At the heart of our approach lies a multi-modal and multi-relational network of biomedical objects extracted from several heterogeneous datasets from the National Center for Biotechnology Information (NCBI). These objects include but are not limited to scientific papers, keywords, genes, proteins, diseases, and diagnoses. We model hypotheses using Latent Dirichlet Allocation applied on abstracts found near shortest paths discovered within this network, and demonstrate the effectiveness of MOLIERE by performing hypothesis generation on historical data. Our network, implementation, and resulting data are all publicly available for the broad scientific community.