Viral diseases have had a significant impact on millions of people worldwide. Utilizing time series forecasting methods allows for the estimation of cases, facilitating the control of disease spread and the allocation of necessary resources in medical facilities. Traditional estimation methods use model per dataset methodology where a model is trained only on a single dataset for a disease. However, foundation models including large language models have shown improved results by training on multiple datasets before being applied to a target dataset. Following this strategy, we aim to train a time series model using multiple datasets from diverse domains and viral diseases. We utilize a pretrained large language model and adapt it for estimating Influenza-Like Illness. We propose a novel network architecture called LLM4cast that encodes the input patches through a bidirectional encoder for rich embedding extraction and passes the encoded patches to a pretrained TinyLlama for fine-tuning. The output from TinyLlama is then flattened and projected to estimate the probable cases. The framework is trained in two stages. In the first stage, we train using time series data from diverse domains with 2.56M timesteps. The second stage involves training using domain specific time series data about viral diseases. The trained network is used in the evaluation of disease cases. The results demonstrate significant improvement in the accuracy of forecasts for foundation models compared to state-of-the-art models trained from scratch.

We propose a new framework for 2-D interpreting (features and samples) black-box machine learning models via a metamodeling technique, by which we study the output and input relationships of the underlying machine learning model. The metamodel can be estimated from data generated via a trained complex model by running the computer experiment on samples of data in the region of interest. We utilize a Gaussian process as a surrogate to capture the response surface of a complex model, in which we incorporate two parts in the process: interpolated values that are modeled by a stationary Gaussian process Z governed by a prior covariance function, and a mean function μ that captures the known trends in the underlying model. The optimization procedure for the variable importance parameter θ is to maximize the likelihood function. This θ corresponds to the correlation of individual variables with the target response. There is no need for any pre-assumed models since it depends on empirical observations. Experiments demonstrate the potential of the interpretable model through quantitative assessment of the predicted samples.

The deep learning revolution, driven especially by incredible achievements in image recognition technology, encourages a predominately data-driven approach to machine learning and data science techniques. However, the emphasis on statistical inference from big data sources alone has diminished the role of existing knowledge and well-established theory. For example, vehicle traffic patterns are highly regular on a weekly period but can deviate unexpectedly in certain situations such as inclement weather, accidents, road work, etc. A purely datadriven approach would simply ask the model to handle all the different situations, while knowledge and situation-aware systems will take advantage of existing knowledge to help guide the machine learning process. This work aims to illustrate how knowledge and situational awareness can help data scientists to build more effective models in the field of vehicle traffic forecasting. The special circumstances considered here include holidays, special weather conditions, accidents, and location awareness that facilitates transfer learning. In addition, we present a novel modeling technique, Quadratic Extreme Learning Machine, that generally improves upon the standard Extreme Learning Machine model while remaining relatively efficient. The Quadratic Extreme Learning Machine can be potentially used as an alternative to Neural Networks, which generally entails higher computational costs.

Efforts to improve early identification of aggressive high grade breast cancers, which pose the greatest risk to patient health if not detected early, are hindered by the rarity of these events. To address this problem, we proposed an accurate and efficient deep transfer learning method to handle the imbalanced data problem that is prominent in breast cancer data. In contrast to existing approaches based primarily on large image databases, we focused on structured data, which has not been commonly used for deep transfer learning. We used a number of publicly available breast cancer data sets to generate a "pre-trained" model and transfer learned concepts to predict high grade malignant tumors in patients diagnosed with breast cancer at Mayo Clinic. We compared our results with state-of-the-art techniques for addressing the problem of imbalanced learning and confirmed the superiority of the proposed method. To further demonstrate the ability of the proposed method to handle different degrees of class imbalance, a series of experiments were performed on publicly available breast cancer data under simulated class imbalanced settings. Based on the experimental results, we concluded that the proposed deep transfer learning on structured data can be used as an efficient method to handle imbalanced class problems in clinical research.

One of the biggest challenges for today's data scientists is to be able to make an informed decision among an exhaustive number of different modeling techniques. As no single algorithm can perform optimally in all cases, the context for the modeling task including the dataset characteristics plays an unsurprisingly important role in deciding which modeling algorithm to choose. In our previous work, we have presented an ontology-based automated model-selection system extending the Scala-based SCALATION data framework. In this study, we present a meta-learning based model-selection system and provide an evaluation of the system.

Key foundational components of Big Data frameworks include efficient large-scale storage and high-performance linear algebra. This paper discusses efficient implementations that utilize compression techniques inspired by columnar relational databases for improving space and time profiles for vector and matrix operations. In addition, linear algebra operations are integrated with columnar relational algebra operations both in dense and compressed forms. For several of the operations substantial speedups are obtained by operating directly on the compressed relations, vectors and matrices. Advantages of mixing and matching relational and linear algebra operations are also pointed out. Both serial and parallel implementations are provided in the ScalaTion Big Data Analytics Framework.

In the big data era, scalability has become a crucial requirement for any useful computational model. Probabilistic graphical models are very useful for mining and discovering data insights, but they are not scalable enough to be suitable for big data problems. Bayesian Networks particularly demonstrate this limitation when their data is represented using few random variables with a massive set of outcome values for each of them. With hierarchical data - data that is arranged in a treelike structure with several levels - one would expect to see hundreds of thousands or millions of values distributed over even just a small number of levels. When modeling this kind of hierarchical data across large data sets, Bayesian networks become unsuitable for representing the probability distributions for the following reasons: i) each level represents a single random variable with hundreds of thousands of values, ii) the number of levels is usually small, so there are also few random variables, and iii) the structure of the network is predefined since the dependency is modeled top-down from each parent to each of its child nodes. In this paper we propose a scalable probabilistic graphical model to overcome these limitations for massive hierarchical data. We believe the proposed model will lead to an easily-scalable, more readable, and expressive implementation for problems that require probabilistic-based solutions for massive amounts of hierarchical data. We successfully applied this model to solve two different challenging probabilistic-based problems on massive hierarchical data sets for different domains, namely, bioinformatics and latent semantic discovery over search logs.

Using caching techniques to improve response time of queries is a proven approach in many contexts. However, it is not well explored for subgraph pattern matching queries, mainly because of subtleties enforced by traditional pattern matching models. Indeed, efficient caching can greatly impact the query answering performance for massive graphs in any query engine whether it is centralized or distributed. This paper investigates the capabilities of the newly introduced pattern matching models in graph simulation family for this purpose. We propose a novel caching technique, and show how the results of a query can be used to answer the new similar queries according to the similarity measure that is introduced. Using large real-world graphs, we experimentally verify the efficiency of the proposed technique in answering subgraph pattern matching queries.

Graph pattern matching is fundamentally important to many applications such as analyzing hyper-links in the World Wide Web, mining associations in online social networks, and substructure search in biochemistry. Most existing graph pattern matching algorithms are highly computation intensive, and do not scale to extremely large graphs that characterize many emerging applications. In recent years, graph processing frameworks such as Pregel have sought to harness shared nothing clusters for processing massive graphs through a vertex-centric, Bulk Synchronous Parallel (BSP) programming model. However, developing scalable and efficient BSP-based algorithms for pattern matching is very challenging because this problem does not naturally align with a vertex-centric programming paradigm. This paper presents novel distributed algorithms based on the vertex-centric programming paradigm for a set of pattern matching models, namely, graph simulation, dual simulation and strong simulation. Our algorithms are fine-tuned to consider the challenges of pattern matching on massive data graphs. Furthermore, we introduce a new pattern matching model, called strict simulation, which outperforms strong simulation in terms of scalability while preserving its important properties. We investigate potential performance bottlenecks and propose several techniques to mitigate them. This paper also presents an extensive set of experiments involving massive graphs (millions of vertices and billions of edges) to study the effects of various parameters on the scalability and performance of the proposed algorithms. The results demonstrate that our techniques are highly effective in alleviating performance bottlenecks and yield significant scalability benefits.