Citation Context

...histicated statistical frameworks have been developed such as Factor Graphs [33], and the Monte Carlo Database [15]. An interesting problem that is related, but orthogonal, problem is active learning =-=[32,35]-=-, where the goal is to leverage the user feedback to interactively build a model. Technically, our goal is to solicit feedback (which can dramatically help improve the model). In fact one of our initi...

Citation Context

...ntary problem of incrementally maintaining the output of the models as the underlying data change. There has been work on maintaining data mining models incrementally, notably association rule mining =-=[29, 34, 36, 37]-=-, but not for the general class of linear classifiers. Researchers have considered scaling machine learning tool kits that contain many more algorithms than we discuss here, e.g., WekaDB [38]. These a...

Citation Context

...r techniques from the body of the paper by transforming these kernels to (low) dimensional linear spaces. The idea is based on a technique of Rahimi and Recht called random non-linear feature vectors =-=[30]-=-. Suppose that all vectors are in S d , the unit ball in d dimensions (any compact set will do). The idea is to find a (random) map z : S d → R D for some D such that for x, y ∈ S d we have z(x) T z(y...

Citation Context

...histicated statistical frameworks have been developed such as Factor Graphs [33], and the Monte Carlo Database [15]. An interesting problem that is related, but orthogonal, problem is active learning =-=[32,35]-=-, where the goal is to leverage the user feedback to interactively build a model. Technically, our goal is to solicit feedback (which can dramatically help improve the model). In fact one of our initi...

Citation Context

...ntary problem of incrementally maintaining the output of the models as the underlying data change. There has been work on maintaining data mining models incrementally, notably association rule mining =-=[29, 34, 36, 37]-=-, but not for the general class of linear classifiers. Researchers have considered scaling machine learning tool kits that contain many more algorithms than we discuss here, e.g., WekaDB [38]. These a...

Citation Context

...ntary problem of incrementally maintaining the output of the models as the underlying data change. There has been work on maintaining data mining models incrementally, notably association rule mining =-=[29, 34, 36, 37]-=-, but not for the general class of linear classifiers. Researchers have considered scaling machine learning tool kits that contain many more algorithms than we discuss here, e.g., WekaDB [38]. These a...

Citation Context

... imprecise; this observation has motivated research to treat this data as probabilistic databases [1, 2, 10] and several sophisticated statistical frameworks have been developed such as Factor Graphs =-=[33]-=-, and the Monte Carlo Database [15]. An interesting problem that is related, but orthogonal, problem is active learning [32,35], where the goal is to leverage the user feedback to interactively build ...

Citation Context

...needed from the corpus. A final example is TFICF (term frequency inverse corpus frequency) in which the term frequencies are obtained from a corpus, but explicitly not updated after each new document =-=[31]-=-. In general, an application will know what feature functions are appropriate, and so we design Hazy to be extensible in this regard, but we expect that the administrator (or a similar expert) writes ...

Citation Context

...act one of our initial motivations behind the hybrid approach is to allow active learning over large data sets. There has been research on how to optimize the queries on top of data mining predicates =-=[28]-=-; we solve the complementary problem of incrementally maintaining the output of the models as the underlying data change. There has been work on maintaining data mining models incrementally, notably a...

Citation Context

...ntary problem of incrementally maintaining the output of the models as the underlying data change. There has been work on maintaining data mining models incrementally, notably association rule mining =-=[29, 34, 36, 37]-=-, but not for the general class of linear classifiers. Researchers have considered scaling machine learning tool kits that contain many more algorithms than we discuss here, e.g., WekaDB [38]. These a...

Citation Context

..., 34, 36, 37], but not for the general class of linear classifiers. Researchers have considered scaling machine learning tool kits that contain many more algorithms than we discuss here, e.g., WekaDB =-=[38]-=-. These approaches may result in lower performance than state-ofthe-art approaches. In contrast, our goal is to take advantage of incremental facilities to achieve higher levels of performance. E. REF...