Many modern machine learning (ML) methods rely on embedding models to learn vector representations (embeddings) for a set of entities (embedding tables). As increasingly diverse ML applications utilize embedding models and embedding tables continue to grow in size and number, there has been a surge in the ad-hoc development of specialized frameworks targeted to train large embedding models for specific tasks. Although the scalability issues that arise in different embedding model training tasks are similar, each of these frameworks independently reinvents and customizes storage components for specific tasks, leading to substantial duplicated engineering efforts in both development and deployment. This paper presents MLKV, an efficient, extensible, and reusable data storage framework designed to address the scalability challenges in embedding model training, specifically data stall and staleness. MLKV augments disk-based key-value storage by democratizing optimizations that were previously exclusive to individual specialized frameworks and provides easy-to-use interfaces for embedding model training tasks. Extensive experiments on open-source workloads, as well as applications in eBay's payment transaction risk detection and seller payment risk detection, show that MLKV outperforms offloading strategies built on top of industrial-strength key-value stores by 1.6-12.6 ×. MLKV is open-source at https://github.com/llm-db/MLKV.

Majority of the advancement in Deep learning (DL) has occurred in domains such as computer vision, and natural language processing, where abundant training data is available. A major obstacle in leveraging DL techniques for malware analysis is the lack of sufficiently big, labeled datasets. In this paper, we take the first steps towards building a model which can synthesize labeled dataset of malware images using GAN. Such a model can be utilized to perform data augmentation for training a classifier. Furthermore, the model can be shared publicly for community to reap benefits of dataset without sharing the original dataset. First, we show the underlying idiosyncrasies of malware images and why existing data augmentation techniques as well as traditional GAN training fail to produce quality artificial samples. Next, we propose a new method for training GAN where we explicitly embed prior domain knowledge about the dataset into the training procedure. We show improvements in training stability and sample quality assessed on different metrics. Our experiments show substantial improvement on baselines and promise for using such a generative model for malware visualization systems.

Deep learning for time series is an emerging area with close ties to industry, yet under represented in performance benchmarks for machine learning systems. In this paper, we present a landscape of deep learning applications applied to time series, and discuss the challenges and directions towards building a robust performance benchmark of deep learning workloads for time series data.

In recent years connected health has emerged as a major area of focus in the healthcare sector due to the increased number of smartphone subscribers and internet users worldwide. Combining these users with big data analytics technologies can benefit both patients and providers in a substantive way. This paper gives a high-level overview of the industry, opportunities, some of the recent developments in connected health, and a quick look at some of the challenges and social implications.

In the past few years, there has been significant developments in how machine learning can be used in various industries and research. This paper discusses the potential of utilizing machine learning technologies in healthcare and outlines various industry initiatives using machine learning initiatives in the healthcare sector.

Wireless networks and personalized mobile devices are deeply integrated and embedded in our lives. Such wide adoptions of new technologies will impact user behavior and in turn will affect network performance. It is imperative to characterize the fundamental structure of wireless user behavior in order to model, manage, leverage and design efficient mobile networks. One major challenge in characterizing user behavior stems from the significant size and complexity of user behavioral data. Without summarization and dimension reduction, the sheer amount of data does not provide much useful information. The key contribution of the paper is a novel similarity metric based on a matrix representation of mobility preferences and its decomposition. This method provides an efficient way to reduce important spatiotemporal dynamics in user mobility into a few eigen-behavior vectors. This also facilitates nodes to exchange their mobility summaries and determine their mutual similarity locally. Without any assumption on the properties of user population, we use unsupervised learning (clustering) techniques to classify WLAN users. Such a user grouping scheme based on learned user behavior is crucial for applications relying on the usage context of each mobile device (e.g., participatory sensing, social-relationship-aware message forwarding). In this study, using our systematic TRACE approach, we analyze wireless users' behavioral patterns by extensively mining wireless network logs from two major university campuses to showcase its efficacy. While our findings partly validate intuitive repetitive behavioral trends and user grouping, it is surprising to find the qualitative commonalities and striking consistency of user behavior from the two universities. We discover multimodal user behavior for more than 60 percent of the users, and there are hundreds of distinct groups with unique behavioral patterns in both campuses. The sizes of the major groups follow a power-law distribution.

Similarity search methods are widely used as kernels in various data mining and machine learning applications including those in computational biology, web search/clustering. Nearest neighbor search (NNS) algorithms are often used to retrieve similar entries, given a query. While there exist efficient techniques for exact query lookup using hashing, similarity search using exact nearest neighbors suffers from a "curse of dimensionality", i.e. for high dimensional spaces, best known solutions offer little improvement over brute force search and thus are unsuitable for large scale streaming applications. Fast solutions to the approximate NNS problem include Locality Sensitive Hashing (LSH) based techniques, which need storage polynomial in n with exponent greater than 1, and query time sublinear, but still polynomial in n, where n is the size of the database. In this work we present a new technique of solving the approximate NNS problem in Euclidean space using a Ternary Content Addressable Memory (TCAM), which needs near linear space and has O(1) query time. In fact, this method also works around the best known lower bounds in the cell probe model for the query time using a data structure near linear in the size of the data base.

In this paper we advocate a service paradigm,profile-cast, within the communication framework of delay tolerant networks (DTN) (K. Fall, 2003). This novel approach leverages the behavioral patterns of mobile network users for delivering messages to a sub-group of users as defined by theirprofiles(e.g., interest, social affiliation, etc.). We study large data sets of usermobilityprofilesand present a case-study ofmobilityprofile-castwith asimilarity-basedforwarding protocol. We show that behavior-aware protocol design has a great potential - we reduce the total number of transmissions to 45% of flooding under 92% delivery success rate, or to only 3% transmissions of flooding under 61% delivery success rate. It also leads to shorter delay (at least 30% less) as compared to a random transmission protocol.

Recent years have witnessed significant growth in the adoption of portable wireless communication and computing devices (e.g., laptops, PDAs, smart phones) and large-scale deployment of wireless networks (e.g., cellular, WLANs). We envision that future usage of mobile devices and services will be highly personalized. Users will incorporate these new technologies into their daily lives, and the way they use new devices and services will reflect their personality and lifestyle. Therefore it is imperative to study and characterize the fundamental structure of wireless user behavior in order to model, manage, leverage and design efficient mobile networks and services. In this study, using our systematic TRACE approach, we analyze wireless users' behavioral patterns by extensively mining wireless network logs from two major university campuses. We represent the data using location-preference vectors, and utilize unsupervised learning (clustering) to classify trends in user behavior using novel similarity metrics. Matrix decomposition techniques are used to identify (and differentiate between) major patterns. We discover multi-modal user behavior and hundreds of distinct groups with unique behavioral patterns in both campuses, and their sizes follow a power-law distribution. Our methods and findings might provide new directions in network management and behavior-aware network protocols and applications, to name a few.

MOTIVATION: Due to the recent advances in technology of mass spectrometry, there has been an exponential increase in the amount of data being generated in the past few years. Database searches have not been able to keep with this data explosion. Thus, speeding up the data searches becomes increasingly important in mass-spectrometry-based applications. Traditional database search methods use one-against-all comparisons of a query spectrum against a very large number of peptides generated from in silico digestion of protein sequences in a database, to filter potential candidates from this database followed by a detailed scoring and ranking of those filtered candidates. RESULTS: In this article, we show that we can avoid the one-against-all comparisons. The basic idea is to design a set of hash functions to pre-process peptides in the database such that for each query spectrum we can use the hash functions to find only a small subset of peptide sequences that are most likely to match the spectrum. The construction of each hash function is based on a random spectrum and the hash value of a peptide is the normalized shared peak counts score (cosine) between the random spectrum and the hypothetical spectrum of the peptide. To implement this idea, we first embed each peptide into a unit vector in a high-dimensional metric space. The random spectrum is represented by a random vector, and we use random vectors to construct a set of hash functions called locality sensitive hashing (LSH) for preprocessing. We demonstrate that our mapping is accurate. We show that our method can filter out >95.65% of the spectra without missing any correct sequences, or gain 111 times speedup by filtering out 99.64% of spectra while missing at most 0.19% (2 out of 1014) of the correct sequences. In addition, we show that our method can be effectively used for other mass spectra mining applications such as finding clusters of spectra efficiently and accurately. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Characterizing the diversity of microbial communities and understanding the environmental factors that influence community diversity are central tenets of microbial ecology. The development and application of cultivation independent molecular tools has allowed for rapid surveying of microbial community composition at unprecedented resolutions and frequencies. There is a growing need to discern robust patterns and relationships within these datasets which provide insight into microbial ecology. Pearson correlation coefficient (PCC) analysis is commonly used for identifying the linear relationship between two species, or species and environmental factors. However, this approach may not be able to capture more complex interactions which occur in situ; thus, alternative analyses were explored. RESULTS: In this paper we introduced local similarity analysis (LSA), which is a technique that can identify more complex dependence associations among species as well as associations between species and environmental factors without requiring significant data reduction. To illustrate its capability of identifying relationships that may not otherwise be identified by PCC, we first applied LSA to simulated data. We then applied LSA to a marine microbial observatory dataset and identified unique, significant associations that were not detected by PCC analysis. LSA results, combined with results from PCC analysis were used to construct a theoretical ecological network which allows for easy visualization of the most significant associations. Biological implications of the significant associations detected by LSA were discussed. We also identified additional applications where LSA would be beneficial. AVAILABILITY: The algorithms are implemented in Splus/R and they are available upon request from the corresponding author.

Providing approximate max-min fair bandwidth allocation among flows within a network or at a single router has been an important research problem. In this paper, we study the space complexity of fairness algorithms, and the communication complexity of distributed global fairness algorithms. We show that in order to enforce max-min fairness with bounded errors, a router must maintain per-flow state. Then we present a practical edge-marking based architecture to demonstrate the enforcement of approximate global max-min fairness for representative scenarios with multiple bottlenecks and non-responsive traffic. We validate our architecture using packet level simulations.

Scalable, energy-efficient data acquisition in large sensor network deployments such as habitat monitoring is an research important problem. In several papers [1, 2], sensor networks have been modeled as having a single sink (or base-station) that acts as the data recipient for a large number of sensors (data sources) deployed over a sensor field. The sensor network might use simple querying and data collection trees for hop-by-hop query dissemination and routing of sensor responses [1] back towards the sink. Since sensors are energy-constrained devices, we wish to minimize communication energy expenditure of these sensors.

An oblivious active queue management scheme is one which does not differentiate between packets belonging to different flows. In this paper, we study the existence and the quality of Nash equilibria imposed by oblivious AQM schemes on selfish agents. Oblivious AQM schemes are of obvious importance because of the ease of implementation and deployment, and Nash equilibrium offers valuable clues into network performance under noncooperative user behavior. Specifically, we ask the following three questions: 1) do there exist oblivious AQM schemes that impose Nash equilibria on selfish agents? 2) Are the imposed equilibria, if they exist, efficient in terms of the goodput obtained and the drop probability experienced at the equilibrium? 3) How easy is it for selfish users to reach the Nash equilibrium state? We assume that the traffic sources are Poisson but the users can control the average rate. We show that drop-tail and RED do not impose Nash equilibria. We modify RED slightly to obtain an oblivious scheme, VLRED, that imposes a Nash equilibrium, but is not efficient. We then present another AQM policy, EN-AQM, that can impose an efficient Nash equilibrium. Finally, we show that for any oblivious AQM, the Nash equilibrium imposed on selfish agents is highly sensitive as the number of agents increases, thus making it hard for the users to converge to the Nash equilibrium, and motivating the need for equilibria-aware protocols.

In heterogeneous networks such as today's Internet, TCP must be made to handle links and paths with vastly different characteristics, including longer delays and dynamically changing available bandwidth. In this research, we first analyze the performance of TCP under networks with TCP-unfriendly characteristics and introduce an architecture to tackle the performance problems in such networks.