Subscribe to the gold package and get unlimited access to Shamra Academy
Register a new userTemporally-Biased Sampling for Online Model Management
60
0
0.0
(
0
)
Added by
Yuanyuan Tian
Publication date
2018
fields
Informatics Engineering
and research's language is
English
Ask ChatGPT about the research
No Arabic abstract
To maintain the accuracy of supervised learning models in the presence of evolving data streams, we provide temporally-biased sampling schemes that weight recent data most heavily, with inclusion probabilities for a given data item decaying exponentially over time. We then periodically retrain the models on the current sample. This approach speeds up the training process relative to training on all of the data. Moreover, time-biasing lets the models adapt to recent changes in the data while -- unlike in a sliding-window approach -- still keeping some old data to ensure robustness in the face of temporary fluctuations and periodicities in the data values. In addition, the sampling-based approach allows existing analytic algorithms for static data to be applied to dynamic streaming data essentially without change. We provide and analyze both a simple sampling scheme (T-TBS) that probabilistically maintains a target sample size and a novel reservoir-based scheme (R-TBS) that is the first to provide both complete control over the decay rate and a guaranteed upper bound on the sample size, while maximizing both expected sample size and sample-size stability. The latter scheme rests on the notion of a fractional sample and, unlike T-TBS, allows for data arrival rates that are unknown and time varying. R-TBS and T-TBS are of independent interest, extending the known set of unequal-probability sampling schemes. We discuss distributed implementation strategies; experiments in Spark illuminate the performance and scalability of the algorithms, and show that our approach can increase machine learning robustness in the face of evolving data.
rate research
Read More
To maintain the accuracy of supervised learning models in the presence of evolving data streams, we provide temporally-biased sampling schemes that weight recent data most heavily, with inclusion probabilities for a given data item decaying over time according to a specified decay function. We then periodically retrain the models on the current sample. This approach speeds up the training process relative to training on all of the data. Moreover, time-biasing lets the models adapt to recent changes in the data while---unlike in a sliding-window approach---still keeping some old data to ensure robustness in the face of temporary fluctuations and periodicities in the data values. In addition, the sampling-based approach allows existing analytic algorithms for static data to be applied to dynamic streaming data essentially without change. We provide and analyze both a simple sampling scheme (T-TBS) that probabilistically maintains a target sample size and a novel reservoir-based scheme (R-TBS) that is the first to provide both control over the decay rate and a guaranteed upper bound on the sample size. If the decay function is exponential, then control over the decay rate is complete, and R-TBS maximizes both expected sample size and sample-size stability. For general decay functions, the actual item inclusion probabilities can be made arbitrarily close to the nominal probabilities, and we provide a scheme that allows a trade-off between sample footprint and sample-size stability. The R-TBS and T-TBS schemes are of independent interest, extending the known set of unequal-probability sampling schemes. We discuss distributed implementation strategies; experiments in Spark illuminate the performance and scalability of the algorithms, and show that our approach can increase machine learning robustness in the face of evolving data.
We consider the problem of online reinforcement learning for the Stochastic Shortest Path (SSP) problem modeled as an unknown MDP with an absorbing state. We propose PSRL-SSP, a simple posterior sampling-based reinforcement learning algorithm for the SSP problem. The algorithm operates in epochs. At the beginning of each epoch, a sample is drawn from the posterior distribution on the unknown model dynamics, and the optimal policy with respect to the drawn sample is followed during that epoch. An epoch completes if either the number of visits to the goal state in the current epoch exceeds that of the previous epoch, or the number of visits to any of the state-action pairs is doubled. We establish a Bayesian regret bound of $O(B_star Ssqrt{AK})$, where $B_star$ is an upper bound on the expected cost of the optimal policy, $S$ is the size of the state space, $A$ is the size of the action space, and $K$ is the number of episodes. The algorithm only requires the knowledge of the prior distribution, and has no hyper-parameters to tune. It is the first such posterior sampling algorithm and outperforms numerically previously proposed optimism-based algorithms.
In exploratory data analysis, analysts often have a need to identify histograms that possess a specific distribution, among a large class of candidate histograms, e.g., find countries whose income distribution is most similar to that of Greece. This distribution could be a new one that the user is curious about, or a known distribution from an existing histogram visualization. At present, this process of identification is brute-force, requiring the manual generation and evaluation of a large number of histograms. We present FastMatch: an end-to-end approach for interactively retrieving the histogram visualizations most similar to a user-specified target, from a large collection of histograms. The primary technical contribution underlying FastMatch is a probabilistic algorithm, HistSim, a theoretically sound sampling-based approach to identify the top-$k$ closest histograms under $ell_1$ distance. While HistSim can be used independently, within FastMatch we couple HistSim with a novel system architecture that is aware of practical considerations, employing asynchronous block-based sampling policies, building on lightweight sampling engines developed in recent work. FastMatch obtains near-perfect accuracy with up to $35times$ speedup over approaches that do not use sampling on several real-world datasets.
Data Lake (DL) is a Big Data analysis solution which ingests raw data in their native format and allows users to process these data upon usage. Data ingestion is not a simple copy and paste of data, it is a complicated and important phase to ensure that ingested data are findable, accessible, interoperable and reusable at all times. Our solution is threefold. Firstly, we propose a metadata model that includes information about external data sources, data ingestion processes, ingested data, dataset veracity and dataset security. Secondly, we present the algorithms that ensure the ingestion phase (data storage and metadata instanciation). Thirdly, we introduce a developed metadata management system whereby users can easily consult different elements stored in DL.
Probabilistic databases play a preeminent role in the processing and management of uncertain data. Recently, many database research efforts have integrated probabilistic models into databases to support tasks such as information extraction and labeling. Many of these efforts are based on batch oriented inference which inhibits a realtime workflow. One important task is entity resolution (ER). ER is the process of determining records (mentions) in a database that correspond to the same real-world entity. Traditional pairwise ER methods can lead to inconsistencies and low accuracy due to localized decisions. Leading ER systems solve this problem by collectively resolving all records using a probabilistic graphical model and Markov chain Monte Carlo (MCMC) inference. However, for large datasets this is an extremely expensive process. One key observation is that, such exhaustive ER process incurs a huge up-front cost, which is wasteful in practice because most users are interested in only a small subset of entities. In this paper, we advocate pay-as-you-go entity resolution by developing a number of query-driven collective ER techniques. We introduce two classes of SQL queries that involve ER operators --- selection-driven ER and join-driven ER. We implement novel variations of the MCMC Metropolis Hastings algorithm to generate biased samples and selectivity-based scheduling algorithms to support the two classes of ER queries. Finally, we show that query-driven ER algorithms can converge and return results within minutes over a database populated with the extraction from a newswire dataset containing 71 million mentions.
Log in to be able to interact and post comments
comments
Fetching comments
Sign in to be able to follow your search criteria
