arXiv subjects
Publications and source records attributed to Ping Li.
Compressed Counting (CC), based on maximally skewed stable random projections, was recently proposed for estimating the p-th frequency moments of data streams. The case p->1 is extremely useful for estimating Shannon entropy of data streams. In this study, we provide a very simple algorithm based on the sample minimum estimator and prove a much improved sample complexity bound, compared to prior results.
The Shannon entropy is a widely used summary statistic, for example, network traffic measurement, anomaly detection, neural computations, spike trains, etc. This study focuses on estimating Shannon entropy of data streams. It is known that Shannon entropy can be approximated by Reenyi entropy or Tsallis entropy, which are both functions of the p-th frequency moments and approach Shannon entropy as p->1. Compressed Counting (CC) is a new method for approximating the p-th frequency moments of data streams. Our contributions include: 1) We prove that Renyi entropy is (much) better than Tsallis entropy for approximating Shannon entropy. 2) We propose the optimal quantile estimator for CC, which considerably improves the previous estimators. 3) Our experiments demonstrate that CC is indeed highly effective approximating the moments and entropies. We also demonstrate the crucial importance of utilizing the variance-bias trade-off.
We develop abc-logitboost, based on the prior work on abc-boost and robust logitboost. Our extensive experiments on a variety of datasets demonstrate the considerable improvement of abc-logitboost over logitboost and abc-mart.
We develop the concept of ABC-Boost (Adaptive Base Class Boost) for multi-class classification and present ABC-MART, a concrete implementation of ABC-Boost. The original MART (Multiple Additive Regression Trees) algorithm has been very successful in large-scale applications. For binary classification, ABC-MART recovers MART. For multi-class classification, ABC-MART considerably improves MART, as evaluated on several public data sets.
Compressed Counting (CC)} was recently proposed for approximating the $α$th frequency moments of data streams, for $0<α\leq 2$. Under the relaxed strict-Turnstile model, CC dramatically improves the standard algorithm based on symmetric stable random projections}, especially as $α\to 1$. A direct application of CC is to estimate the entropy, which is an important summary statistic in Web/network measurement and often serves a crucial "feature" for data mining. The Rényi entropy and the Tsallis entropy are functions of the $α$th frequency moments; and both approach the Shannon entropy as $α\to 1$. A recent theoretical work suggested using the $α$th frequency moment to approximate the Shannon entropy with $α=1+δ$ and very small $|δ|$ (e.g., $<10^{-4}$). In this study, we experiment using CC to estimate frequency moments, Rényi entropy, Tsallis entropy, and Shannon entropy, on real Web crawl data. We demonstrate the variance-bias trade-off in estimating Shannon entropy and provide practical recommendations. In particular, our experiments enable us to draw some important conclusions: (1) As $α\to 1$, CC dramatically improves {\em symmetric stable random projections} in estimating frequency moments, Rényi entropy, Tsallis entropy, and Shannon entropy. The improvements appear to approach "infinity." (2) Using {\em symmetric stable random projections} and $α= 1+δ$ with very small $|δ|$ does not provide a practical algorithm because the required sample size is enormous.
Compressed Counting (CC) was recently proposed for very efficiently computing the (approximate) $α$th frequency moments of data streams, where $0<α<= 2$. Several estimators were reported including the geometric mean estimator, the harmonic mean estimator, the optimal power estimator, etc. The geometric mean estimator is particularly interesting for theoretical purposes. For example, when $α-> 1$, the complexity of CC (using the geometric mean estimator) is $O(1/ε)$, breaking the well-known large-deviation bound $O(1/ε^2)$. The case $α\approx 1$ has important applications, for example, computing entropy of data streams. For practical purposes, this study proposes the optimal quantile estimator. Compared with previous estimators, this estimator is computationally more efficient and is also more accurate when $α> 1$.
The method of stable random projections is a tool for efficiently computing the $l_α$ distances using low memory, where $0<α\leq 2$ is a tuning parameter. The method boils down to a statistical estimation task and various estimators have been proposed, based on the geometric mean, the harmonic mean, and the fractional power etc. This study proposes the optimal quantile estimator, whose main operation is selecting, which is considerably less expensive than taking fractional power, the main operation in previous estimators. Our experiments report that the optimal quantile estimator is nearly one order of magnitude more computationally efficient than previous estimators. For large-scale learning tasks in which storing and computing pairwise distances is a serious bottleneck, this estimator should be desirable. In addition to its computational advantages, the optimal quantile estimator exhibits nice theoretical properties. It is more accurate than previous estimators when $α>1$. We derive its theoretical error bounds and establish the explicit (i.e., no hidden constants) sample complexity bound.
Applications in machine learning and data mining require computing pairwise Lp distances in a data matrix A. For massive high-dimensional data, computing all pairwise distances of A can be infeasible. In fact, even storing A or all pairwise distances of A in the memory may be also infeasible. This paper proposes a simple method for p = 2, 4, 6, ... We first decompose the l_p (where p is even) distances into a sum of 2 marginal norms and p-1 ``inner products'' at different orders. Then we apply normal or sub-Gaussian random projections to approximate the resultant ``inner products,'' assuming that the marginal norms can be computed exactly by a linear scan. We propose two strategies for applying random projections. The basic projection strategy requires only one projection matrix but it is more difficult to analyze, while the alternative projection strategy requires p-1 projection matrices but its theoretical analysis is much easier. In terms of the accuracy, at least for p=4, the basic strategy is always more accurate than the alternative strategy if the data are non-negative, which is common in reality.
Counting is among the most fundamental operations in computing. For example, counting the pth frequency moment has been a very active area of research, in theoretical computer science, databases, and data mining. When p=1, the task (i.e., counting the sum) can be accomplished using a simple counter. Compressed Counting (CC) is proposed for efficiently computing the pth frequency moment of a data stream signal A_t, where 0 = 0, which includes the strict Turnstile model as a special case. For natural data streams encountered in practice, this restriction is minor. The underly technique for CC is what we call skewed stable random projections, which captures the intuition that, when p=1 a simple counter suffices, and when p = 1+/Δwith small Δ, the sample complexity of a counter system should be low (continuously as a function of Δ). We show at small Δthe sample complexity (number of projections) k = O(1/ε) instead of O(1/ε^2). Compressed Counting can serve a basic building block for other tasks in statistics and computing, for example, estimation entropies of data streams, parameter estimations using the method of moments and maximum likelihood. Finally, another contribution is an algorithm for approximating the logarithmic norm, \sum_{i=1}^D\log A_t[i], and logarithmic distance. The logarithmic distance is useful in machine learning practice with heavy-tailed data.
We propose skewed stable random projections for approximating the pth frequency moments of dynamic data streams (0 1) improves previous methods based on (symmetric) stable random projections. Our proposed method is applicable to data streams that are (a) insertion only (the cash-register model); or (b) always non-negative (the strict Turnstile model), or (c) eventually non-negative at check points. This is only a minor restriction for practical applications. Our method works particularly well when p = 1+/- Δand Δis small, which is a practically important scenario. For example, Δmay be the decay rate or interest rate, which are usually small. Of course, when Δ= 0, one can compute the 1th frequent moment (i.e., the sum) essentially error-free using a simple couter. Our method may be viewed as a ``genearlized counter'' in that it can count the total value in the future, taking in account of the effect of decaying or interest accruement. In a summary, our contributions are two-fold. (A) This is the first propsal of skewed stable random projections. (B) Based on first principle, we develop various statistical estimators for skewed stable distributions, including their variances and error (tail) probability bounds, and consequently the sample complexity bounds.
We propose a novel capacity model for complex networks against cascading failure. In this model, vertices with both higher loads and larger degrees should be paid more extra capacities, i.e. the allocation of extra capacity on vertex $i$ will be proportional to $ k_{i}^γ $, where $ k_{i}$ is the degree of vertex $i$ and $γ>0$ is a free parameter. We have applied this model on Barabási-Albert network as well as two real transportation networks, and found that under the same amount of available resource, this model can achieve better network robustness than previous models.
This paper will focus on three different aspects in improving the current practice of stable random projections. Firstly, we propose {\em very sparse stable random projections} to significantly reduce the processing and storage cost, by replacing the $α$-stable distribution with a mixture of a symmetric $α$-Pareto distribution (with probability $β$, $0<β\leq1$) and a point mass at the origin (with a probability $1-β$). This leads to a significant $\frac{1}β$-fold speedup for small $β$. Secondly, we provide an improved estimator for recovering the original $l_α$ norms from the projected data. The standard estimator is based on the (absolute) sample median, while we suggest using the geometric mean. The geometric mean estimator we propose is strictly unbiased and is easier to study. Moreover, the geometric mean estimator is more accurate, especially non-asymptotically. Thirdly, we provide an adequate answer to the basic question of how many projections (samples) are needed for achieving some pre-specified level of accuracy. \cite{Proc:Indyk_FOCS00,Article:Indyk_TKDE03} did not provide a criterion that can be used in practice. The geometric mean estimator we propose allows us to derive sharp tail bounds which can be expressed in exponential forms with constants explicitly given.
For dimension reduction in $l_1$, the method of {\em Cauchy random projections} multiplies the original data matrix $\mathbf{A} \in\mathbb{R}^{n\times D}$ with a random matrix $\mathbf{R} \in \mathbb{R}^{D\times k}$ ($k\ll\min(n,D)$) whose entries are i.i.d. samples of the standard Cauchy C(0,1). Because of the impossibility results, one can not hope to recover the pairwise $l_1$ distances in $\mathbf{A}$ from $\mathbf{B} = \mathbf{AR} \in \mathbb{R}^{n\times k}$, using linear estimators without incurring large errors. However, nonlinear estimators are still useful for certain applications in data stream computation, information retrieval, learning, and data mining. We propose three types of nonlinear estimators: the bias-corrected sample median estimator, the bias-corrected geometric mean estimator, and the bias-corrected maximum likelihood estimator. The sample median estimator and the geometric mean estimator are asymptotically (as $k\to \infty$) equivalent but the latter is more accurate at small $k$. We derive explicit tail bounds for the geometric mean estimator and establish an analog of the Johnson-Lindenstrauss (JL) lemma for dimension reduction in $l_1$, which is weaker than the classical JL lemma for dimension reduction in $l_2$. Asymptotically, both the sample median estimator and the geometric mean estimators are about 80% efficient compared to the maximum likelihood estimator (MLE). We analyze the moments of the MLE and propose approximating the distribution of the MLE by an inverse Gaussian.
Given a seminormal affine monoid M we consider several monoid properties of M and their connections to ring properties of the associated affine monoid ring K[M] over a field K. We characterize when K[M] satisfies Serre's condition (S_2) and analyze the local cohomology of K[M]. As an application we present criteria which imply that K[M] is Cohen--Macaulay and we give lower bounds for the depth of K[M]. Finally, the seminormality of an arbitrary affine monoid M is studied with characteristic p methods.