Patent Grant
8099453
The challenge of data clustering—constructing semantically meaningful groups of data instances—has been a focus of information technology (IT) field for some time. Accordingly, a number of methods for data clustering have been developed. One dilemma surrounding existing methods is based on a tradeoff between effectiveness and efficiency or scalability. The enormous amount and dimensionality of data processed by modern data mining tools call for effective and scalable unsupervised learning techniques. However, most clustering algorithms in the art are either effective or scalable, but not both. In other words, these methods either provide fairly powerful learning capabilities but are too resource-intensive for large or highly dimensional datasets, or they are useable on large datasets but produce low-quality results.
Modern resources for generation, accumulation, and storage of data have made giga- and terabyte datasets more and more common. Due to the magnitude of such tasks, as well as the time and processing power that they can consume, data mining practitioners often tend to use simpler methods in the interest of feasibility. However, such an approach sacrifices mining power and may provide unsatisfactory results. Furthermore, for very large and/or complex amounts of data, even simple methods may not be feasible. If one considers, for example, a problem of clustering one million data instances using a simple online clustering algorithm: first initialize n clusters with one data point each, then iteratively assign the rest of points into their closest clusters (in the Euclidean space). Even for small values of n (e.g. n=1000), such an algorithm may work for hours on a modern personal computer (PC). The results would however be quite unsatisfactory, especially if the data points are 100,000-dimensional vectors.
Therefore, a number of IT fields could benefit from methods and systems of data clustering that combine a powerful learning algorithm with a scalability that addresses modern dataset demands.
Features and advantages of the invention will be apparent from the detailed description which follows, taken in conjunction with the accompanying drawings, which together illustrate, by way of example, features of the invention; and, wherein:
Reference will now be made to the exemplary embodiments illustrated, and specific language will be used herein to describe the same. The present disclosure sets forth a method and system by which multi-modal data clustering may be accomplished while utilizing more powerful clustering algorithms in a readily scalable method. Parallel processing methods provide one approach to providing scalability. However, despite the long-held interest in data clustering in the data mining community, not many clustering algorithms have been parallelized, and not many software tools for parallel clustering have been built, and of these most are fairly simple. In contrast, there are two families of data clustering methods that are widely considered as very powerful: multi-modal clustering and information-theoretic clustering (ITC).
Multi-modal (or multivariate) clustering is a framework for simultaneously clustering a few types, or modalities, of data. As used herein “modality” may refer to a collection of data points that may be clustered based on a shared type. One example of a multi-modal clustering task would be to construct a clustering of Web pages, together with a clustering of words from those pages, as well as a clustering of URLs hyperlinked from those pages. It is commonly believed that multi-modal clustering is able to achieve better results than traditional, uni-modal methods. Of the multi-modal approaches, the two-modal case (often called “co-clustering” or “double clustering”) can be of particular interest.
Information-theoretic clustering (ITC) is an adequate solution to clustering highly multi-dimensional data, such as documents or genes. ITC methods perform global optimization of an information-theoretic objective function. Many global optimization methods are essentially sequential and therefore hard to parallelize. In contrast, local optimization methods may be more easily parallelizable. However, many known local optimization methods are not useful for multi-dimensional datasets.
One principle of ITC is known in the art as the Information Bottleneck (IB), in which a random variable X is clustered with respect to an interacting variable Y: the clustering {tilde over (X)} is represented as a low-bandwidth channel (i.e. a bottleneck) between the input signal X and the output signal Y. This channel is constructed to minimize the communication error while maximizing the compression:
max[I({tilde over (X)};Y)−βI({tilde over (X)};X)], (Equation 1)
where I is a mutual information (MI) and β is a Lagrange multiplier. The IB principle can be generalized to a multi-variate case. In a simple form, for clustering two variables, X and Y, a channel X{tilde over (X)}
{tilde over (Y)}
Y is constructed to optimize the objective
max[I({tilde over (X)};{tilde over (Y)})−β1I({tilde over (X)};X)−βI({tilde over (Y)};Y)]. (Equation 2)
When more than two variables are clustered, the mutual information I({tilde over (X)};{tilde over (Y)}) is generalized into its multivariate version, called multi-information. The complexity of computing multi-information grows exponentially as more variables are added, and can therefore be restrictive in practical cases even for only three variables.
Information-theoretic co-clustering (IT-CC) can be an alternative to multivariate IB for the two-variate case when the numbers of clusters |{tilde over (X)}| and |{tilde over (Y)}| are fixed. In this case, one may drop the compression constraints I({tilde over (X)};X) and I({tilde over (Y)};Y) in Equation 2 and directly minimize the information loss:
min[I(X;Y)−I({tilde over (X)};{tilde over (Y)})]=maxI({tilde over (X)};{tilde over (Y)}), (Equation 3)
when I(X;Y) is constant for a given dataset.
As discussed above, ITC methods work through global optimization and may be difficult to parallelize. Co-clustering (CC) methods like IT-CC can also provide an alternative to this problem, utilizing the approach of breaking down a global optimization task into a set of local optimizations which may be done in parallel. Embodiments of the present invention provide a further benefit through a parallelization scheme that incorporates a co-clustering method that, like IT-CC, provides a local optimization approach, while being more aggressive in optimizing an objective function.
A general embodiment of the present invention includes a method for data clustering. This method can comprise entering the data into a computer network that is configured for parallel processing of the components of an objective function. Such a network may include a master processor and an array of slave processors.
An algorithm for optimizing a uni-modal objective function, such as Equation 1 according to the IB principle, may use the procedure of: (a) as an initialization step, assigning all data points in a dataset into clusters; (b) constructing a random permutation of all the data points in which each is pulled out of its cluster and iteratively assigned to another cluster; and (c) leaving the data point in the cluster such that the objective function is maximized. A multi-modal algorithm in accordance with embodiments of the present invention can further iterate over data modalities expressed in a dataset, applying the above optimization procedure at each iteration in order to optimize a co-clustering objective function such as Equation 3. The algorithm improves clustering by continuously updating cluster memberships and exchanging data between clusters where needed. To decide whether to change a cluster membership, the algorithm directly evaluates the objective.
In a particular embodiment of the data clustering method, the method can comprise entering data into a computer network as described above, in which a master processor executing a master process prepares the data by dividing it into clusters. The clusters may then be sent to the array of slave processors for execution of an optimization procedure such as described above. To accommodate the pairwise treatment of clusters, each slave processor in the computer network may have two locations or seats associated therewith for holding a pair of clusters during an optimization step.
Parallelization of this co-clustering algorithm is allowed based on the consideration that its objective function I({tilde over (X)};{tilde over (Y)}) has the additive property over either of its arguments. That is, in each decision whether to move data point x ε{tilde over (x)} to cluster {tilde over (x)}*, only the portion of the function that corresponds to clusters {tilde over (x)} and {tilde over (x)}* is affected. Indeed, by definition:
where {tilde over (x)} and {tilde over (y)} are the corresponding clusters of x and y, respectively, p({tilde over (x)},{tilde over (y)}) is a joint probability distribution and p({tilde over (x)}) and p({tilde over (y)}) are marginal probability distributions of these clusters.
To check whether such a move increases the objective function, it is sufficient to calculate the delta between its value before the move and after the move. Similarly, only the portion of the function corresponding to the clusters in question are involved and all other terms cancel out. So for clusters {tilde over (x)} and {tilde over (x)}*:
By splitting all clusters of {tilde over (X)} into disjoint pairs, probing the moves x→{tilde over (x)}* can be performed in parallel.
Each probing of moves can be then executed using a separate processor, after which the processors can exchange their data. Since the communication is generally expensive, it is beneficial to test all elements of both clusters in a pair. If the probe shows that the objective can be increased, the element is immediately moved from its cluster into another. Using this approach, data points do not necessarily move into the cluster such that the objective function is maximized, but only increased. While such a loss might look crucial, it has been found that both approaches are comparably effective as soon as the number of optimization steps is about the same. The latter can be achieved by iterating over all the cluster pairs.
In view of this, treatment of the clustered data can comprise a master process executed by the master processor and slave process to be executed in parallel by the slave processors. A schematic summary of an exemplary embodiment of the algorithm is shown in
The optimization cycle includes parallel execution of the slave process. In general, the slave process as executed by each slave processor comprises two basic tasks: adjusting membership of the pair of clusters in the seats associated with that processor, and participating in resorting the cluster pairs in the array so that the next optimization step can be executed on pairs that are different from those in the prior step.
As discussed above, the step of adjusting membership of the clusters can involve exchanging data as needed to locally increase the objective function based on two modalities present in the data. This is done by iteratively moving data points between the clusters while evaluating the delta, such as expressed in Equation 5. It should be noted that, while the process involves exchanging data and probing the effect of each move, some iterations of this process will result in fewer data moves than others, and it is conceivable that the iteration may call for no changes in cluster membership.
The other step of the slave process accomplishes the resorting of cluster pairs so that new pairs may be subjected to the optimization step. Accordingly, after each slave processor has completed its optimization step, it is ready to send and receive clusters to and from other slave processors. To do this, it is enough for each slave processor to send only one cluster of each pair to another slave processor. This way, the communication cost of the optimization cycle may be minimized. The resorting step allows executing of the slave process on a plurality of possible cluster pairs during the course of an optimization cycle. In one embodiment, a deterministic protocol may be employed so that all possible cluster pairs are created during the cycle. In a particular embodiment, such a protocol involves a protocol illustrated in
The general movement of clusters through the array is indicated by dashed arrows. The protocol alternates sending the clusters in the upper row to the right, (with periodic boundary conditions in effect) and sending the clusters in the lower row to the left. Processor number 0 is an exception in that it always keeps the cluster in seat n in place, and therefore always sends the other cluster. This processor thereby inverts the direction in which a cluster is moving and therefore insures that all possible pairs are created. To clarify the order, referring again to the exemplary array in
In one embodiment, this protocol also insures that every pair of clusters meets exactly once. Two clusters meet at a processor if the sum of their seat numbers modulo (n−1) is 0. Every two iterations thereafter, the new seat number of every cluster will increase by 1 modulo (n−1). The sum will hence increase by 2 every two iterations. Similarly, if the sum of seat numbers modulo (n−1) is (n−2), then the clusters will meet in the next iteration. It can be seen that by adding 2 modulo (n−1) it takes at most (n−1) iterations until any two regular clusters (without the stationary one) meet, and every cluster will hit processor 0 and meet the stationary cluster from either seat (n/2)−1 or (n−2) when moving in the same direction for (n−1) iterations.
In another embodiment, the resorting may be done by having each slave processor send a cluster to another slave processor at random. This sending, while essentially random, may be done in a way so that every cluster seat remains occupied by a cluster, so as to avoid idle processors due to incomplete pairing. This stochastic protocol avoids the situation where an initial ordering of clusters may be disadvantageous is preserved through the optimization cycle. However, the stochastic protocol does not provide the guarantee of completeness provided by the deterministic protocol.
After completion of the optimization cycle, the master process may terminate its wait state and then compute the whole objective function based on the new clustering accomplished during the optimization cycle. As discussed above, the basic algorithm may be applied to multi-modal tasks by iterating the algorithm over modalities. Accordingly, in a particular embodiment of the present invention, the master process may be repeated for one or more iterations with different combinations of modalities for each iteration. In a more specific embodiment, at least one different modality is utilized for each iteration.
By breaking down a global optimization task into a set of local optimizations and parallelizing them, the embodiments of the present invention provide a clustering method that is both powerful and highly scalable. As such, it can accommodate datasets that are very large. In a particular embodiment, the method may be used for clustering from hundreds of thousands to millions of data instances, though this range is not intended to be limiting. In a more specific aspect, the method may involve from 3 to 1,000,000 clusters.
The present disclosure also sets forth a system configured for executing the processes described herein. The system can comprise a master processor, an array of slave processors, and two cluster seats associated with each slave processor. It should be understood that each processor may be separately packaged units (e.g. chips) or may instead be an individual processing unit housed on one die with other such units, as in multi-core processor chips. Also, virtual processors may constitute processors in accordance with the present invention, where said virtual processors are logical processing units embodied in a physical processor. As such, each physical processor may provide multiple processors, thereby further increasing the size of the network beyond the number of physical processors that may be available. Furthermore, the master and slave processors may be platformed and connected by any means known in the art that provide the communication speed and bandwidth appropriate for the application at hand. The parallel setup and modular nature of the system and the algorithm mean that the system can be scaled up to match the data task. As such, the slave processor array may include up to 1000 separate processors, or even many more.
The present disclosure further sets forth program code for executing the algorithms described herein, where said program code may be embodied on a computer-readable medium and therefore read by and executed on a machine. In a particular embodiment, the machine or system may be a computer network comprising a master processor, an array of slave processors, and two cluster seats associated with each slave processor. In another particular embodiment, the program code comprises a section of master process code for execution by the master processor, and a section of slave process code for execution by each slave processor.
In an exemplary implementation of the algorithm of the present invention, the communication was based on the Massage Passing Interface (MPI). The algorithm was deployed on a Hewlett Packard XC Linux cluster system consisting of 62 eight-core machines with 16 GB of RAM each.
A parallelized algorithm in accordance with the present invention (“DataLoom”) was compared to a sequential co-clustering method (SCC). As a baseline, a parallelized IT-CC algorithm was used. For simplicity, two datasets having an even number of categories were chosen. One (sanders-r) was a small collection of 1188 email messages, grouped into 30 folders. The other was the 20 Newsgroups (20NG) dataset, consisting of 19,997 postings submitted to 20 newsgroups. About 4.5% of the 20NG documents are duplications, but these were not removed for better replicability. For all four of the datasets, documents and their words were simultaneously clustered. For the email datasets, the third modality—names of email correspondents—was also clustered. For this 3-way clustering, a Clique-wise optimization scheme was used. The measure of clustering performance used was a micro-averaged accuracy measure. The results are summarized in Table 1.
The results show that sequential co-clustering greatly outperforms IT-CC, and the algorithms of the present invention demonstrate comparable performance to the powerful SCC.
The RCV1 dataset consists of 806,791 documents each of which belongs to a hierarchy of categories: the top level contains four categories, and the second level contains 55 categories. In this experiment, the top level was ignored and categories from all the lower levels were mapped onto their parents from the second level (using this scheme, 27076 documents were not assigned to any category and therefore were always considered as wrongly categorized). Stopwords and low frequency words were removed, leaving 150,032 distinct words overall. Represented as a contingency table, the resulting data contained over 120 billion entries.
As an initial step, 800 document clusters and 800 word clusters were built. The clustering precision measure of Example 1 was used. The results of this system (“DataLoom”) compared with parallelized IT-CC and parallelized double k-means is plotted in
While the forgoing examples are illustrative of the principles of the present invention in one or more particular applications, it will be apparent to those of ordinary skill in the art that numerous modifications in form, usage and details of implementation can be made without the exercise of inventive faculty, and without departing from the principles and concepts of the invention. Accordingly, it is not intended that the invention be limited, except as by the claims set forth below.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5202998 | Yanes | Apr 1993 | A |
| 6092072 | Guha et al. | Jul 2000 | A |
| 6269376 | Dhillon et al. | Jul 2001 | B1 |
| 6303297 | Lincoln et al. | Oct 2001 | B1 |
| 6633863 | Sørensen Møller et al. | Oct 2003 | B1 |
| 6839680 | Liu et al. | Jan 2005 | B1 |
| 7024049 | Bern et al. | Apr 2006 | B2 |
| 7197504 | Runkler et al. | Mar 2007 | B1 |
| 7308451 | Lamping et al. | Dec 2007 | B1 |
| 7353359 | Kalafala et al. | Apr 2008 | B2 |
| 7379890 | Myr et al. | May 2008 | B2 |
| 7383258 | Harik et al. | Jun 2008 | B2 |
| 7386522 | Bigus et al. | Jun 2008 | B1 |
| 7395235 | Dhurandhar et al. | Jul 2008 | B2 |
| 7577091 | Antal et al. | Aug 2009 | B2 |
| 7706617 | Cahill et al. | Apr 2010 | B2 |
| 7707148 | Fogel | Apr 2010 | B1 |
| 7743058 | Liu et al. | Jun 2010 | B2 |
| 7761514 | Popescu et al. | Jul 2010 | B2 |
| 7769803 | Birdwell et al. | Aug 2010 | B2 |
| 7801836 | Sureka | Sep 2010 | B2 |
| 7818272 | Mishra et al. | Oct 2010 | B1 |
| 7890445 | Ben Hur et al. | Feb 2011 | B2 |
| 20030101449 | Bentolila et al. | May 2003 | A1 |
| 20050071140 | Ben-Hur et al. | Mar 2005 | A1 |
| 20060112049 | Mehrotra et al. | May 2006 | A1 |
| 20070005292 | Jin | Jan 2007 | A1 |
| 20070198553 | Wolman | Aug 2007 | A1 |
| 20070208677 | Goldberg et al. | Sep 2007 | A1 |
| 20080114750 | Saxena et al. | May 2008 | A1 |
| 20080126176 | Iguchi | May 2008 | A1 |
| 20080168061 | Liu et al. | Jul 2008 | A1 |
| 20080294686 | Long et al. | Nov 2008 | A1 |
| Number | Date | Country | |
|---|---|---|---|
| 20100185695 A1 | Jul 2010 | US |
