Xiaoyun Li

Ping Li, Xiaoyun Li

Consider two $D$-dimensional data vectors (e.g., embeddings): $u, v$. In many embedding-based retrieval (EBR) applications where the vectors are generated from trained models, $D=256\sim 1024$ are common. In this paper, OPORP (one permutation + one random projection) uses a variant of the ``count-sketch'' type of data structures for achieving data reduction/compression. With OPORP, we first apply a permutation on the data vectors. A random vector $r$ is generated i.i.d. with moments: $E(r_i) = 0, E(r_i^2)=1, E(r_i^3) =0, E(r_i^4)=s$. We multiply (as dot product) $r$ with all permuted data vectors. Then we break the $D$ columns into $k$ equal-length bins and aggregate (i.e., sum) the values in each bin to obtain $k$ samples from each data vector. One crucial step is to normalize the $k$ samples to the unit $l_2$ norm. We show that the estimation variance is essentially: $(s-1)A + \frac{D-k}{D-1}\frac{1}{k}\left[ (1-ρ^2)^2 -2A\right]$, where $A\geq 0$ is a function of the data ($u,v$). This formula reveals several key properties: (1) We need $s=1$. (2) The factor $\frac{D-k}{D-1}$ can be highly beneficial in reducing variances. (3) The term $\frac{1}{k}(1-ρ^2)^2$ is a substantial improvement compared with $\frac{1}{k}(1+ρ^2)$, which corresponds to the un-normalized estimator. We illustrate that by letting the $k$ in OPORP to be $k=1$ and repeat the procedure $m$ times, we exactly recover the work of ``very spars random projections'' (VSRP). This immediately leads to a normalized estimator for VSRP which substantially improves the original estimator of VSRP. In summary, with OPORP, the two key steps: (i) the normalization and (ii) the fixed-length binning scheme, have considerably improved the accuracy in estimating the cosine similarity, which is a routine (and crucial) task in modern embedding-based retrieval (EBR) applications.

Ping Li, Xiaoyun Li

In this paper, we develop a series of differential privacy (DP) algorithms from a family of random projections (RP) for general applications in machine learning, data mining, and information retrieval. Among the presented algorithms, iDP-SignRP is remarkably effective under the setting of ``individual differential privacy'' (iDP), based on sign random projections (SignRP). Also, DP-SignOPORP considerably improves existing algorithms in the literature under the standard DP setting, using ``one permutation + one random projection'' (OPORP), where OPORP is a variant of the celebrated count-sketch method with fixed-length binning and normalization. Without taking signs, among the DP-RP family, DP-OPORP achieves the best performance. Our key idea for improving DP-RP is to take only the signs, i.e., $sign(x_j) = sign\left(\sum_{i=1}^p u_i w_{ij}\right)$, of the projected data. The intuition is that the signs often remain unchanged when the original data ($u$) exhibit small changes (according to the ``neighbor'' definition in DP). In other words, the aggregation and quantization operations themselves provide good privacy protections. We develop a technique called ``smooth flipping probability'' that incorporates this intuitive privacy benefit of SignRPs and improves the standard DP bit flipping strategy. Based on this technique, we propose DP-SignOPORP which satisfies strict DP and outperforms other DP variants based on SignRP (and RP), especially when $ε$ is not very large (e.g., $ε= 5\sim10$). Moreover, if an application scenario accepts individual DP, then we immediately obtain an algorithm named iDP-SignRP which achieves excellent utilities even at small~$ε$ (e.g., $ε<0.5$).

Weijie Zhao, Xuewu Jiao, Mingqing Hu et al.

Click-Through Rate (CTR) prediction is a crucial component in the online advertising industry. In order to produce a personalized CTR prediction, an industry-level CTR prediction model commonly takes a high-dimensional (e.g., 100 or 1000 billions of features) sparse vector (that is encoded from query keywords, user portraits, etc.) as input. As a result, the model requires Terabyte scale parameters to embed the high-dimensional input. Hierarchical distributed GPU parameter server has been proposed to enable GPU with limited memory to train the massive network by leveraging CPU main memory and SSDs as secondary storage. We identify two major challenges in the existing GPU training framework for massive-scale ad models and propose a collection of optimizations to tackle these challenges: (a) the GPU, CPU, SSD rapidly communicate with each other during the training. The connections between GPUs and CPUs are non-uniform due to the hardware topology. The data communication route should be optimized according to the hardware topology; (b) GPUs in different computing nodes frequently communicates to synchronize parameters. We are required to optimize the communications so that the distributed system can become scalable. In this paper, we propose a hardware-aware training workflow that couples the hardware topology into the algorithm design. To reduce the extensive communication between computing nodes, we introduce a $k$-step model merging algorithm for the popular Adam optimizer and provide its convergence rate in non-convex optimization. To the best of our knowledge, this is the first application of $k$-step adaptive optimization method in industrial-level CTR model training. The numerical results on real-world data confirm that the optimized system design considerably reduces the training time of the massive model, with essentially no loss in accuracy.

Xiaoyun Li, Ping Li

Minwise hashing (MinHash) is a classical method for efficiently estimating the Jaccrad similarity in massive binary (0/1) data. To generate $K$ hash values for each data vector, the standard theory of MinHash requires $K$ independent permutations. Interestingly, the recent work on "circulant MinHash" (C-MinHash) has shown that merely two permutations are needed. The first permutation breaks the structure of the data and the second permutation is re-used $K$ time in a circulant manner. Surprisingly, the estimation accuracy of C-MinHash is proved to be strictly smaller than that of the original MinHash. The more recent work further demonstrates that practically only one permutation is needed. Note that C-MinHash is different from the well-known work on "One Permutation Hashing (OPH)" published in NIPS'12. OPH and its variants using different "densification" schemes are popular alternatives to the standard MinHash. The densification step is necessary in order to deal with empty bins which exist in One Permutation Hashing. In this paper, we propose to incorporate the essential ideas of C-MinHash to improve the accuracy of One Permutation Hashing. Basically, we develop a new densification method for OPH, which achieves the smallest estimation variance compared to all existing densification schemes for OPH. Our proposed method is named C-OPH (Circulant OPH). After the initial permutation (which breaks the existing structure of the data), C-OPH only needs a "shorter" permutation of length $D/K$ (instead of $D$), where $D$ is the original data dimension and $K$ is the total number of bins in OPH. This short permutation is re-used in $K$ bins in a circulant shifting manner. It can be shown that the estimation variance of the Jaccard similarity is strictly smaller than that of the existing (densified) OPH methods.

Xiaoyun Li, Ping Li

Traditional minwise hashing (MinHash) requires applying $K$ independent permutations to estimate the Jaccard similarity in massive binary (0/1) data, where $K$ can be (e.g.,) 1024 or even larger, depending on applications. The recent work on C-MinHash (Li and Li, 2021) has shown, with rigorous proofs, that only two permutations are needed. An initial permutation is applied to break whatever structures which might exist in the data, and a second permutation is re-used $K$ times to produce $K$ hashes, via a circulant shifting fashion. (Li and Li, 2021) has proved that, perhaps surprisingly, even though the $K$ hashes are correlated, the estimation variance is strictly smaller than the variance of the traditional MinHash. It has been demonstrated in (Li and Li, 2021) that the initial permutation in C-MinHash is indeed necessary. For the ease of theoretical analysis, they have used two independent permutations. In this paper, we show that one can actually simply use one permutation. That is, one single permutation is used for both the initial pre-processing step to break the structures in the data and the circulant hashing step to generate $K$ hashes. Although the theoretical analysis becomes very complicated, we are able to explicitly write down the expression for the expectation of the estimator. The new estimator is no longer unbiased but the bias is extremely small and has essentially no impact on the estimation accuracy (mean square errors). An extensive set of experiments are provided to verify our claim for using just one permutation.

Xiaoyun Li, Ping Li

Minwise hashing (MinHash) is an important and practical algorithm for generating random hashes to approximate the Jaccard (resemblance) similarity in massive binary (0/1) data. The basic theory of MinHash requires applying hundreds or even thousands of independent random permutations to each data vector in the dataset, in order to obtain reliable results for (e.g.,) building large-scale learning models or approximate near neighbor search in massive data. In this paper, we propose {\bf Circulant MinHash (C-MinHash)} and provide the surprising theoretical results that we just need \textbf{two} independent random permutations. For C-MinHash, we first conduct an initial permutation on the data vector, then we use a second permutation to generate hash values. Basically, the second permutation is re-used $K$ times via circulant shifting to produce $K$ hashes. Unlike classical MinHash, these $K$ hashes are obviously correlated, but we are able to provide rigorous proofs that we still obtain an unbiased estimate of the Jaccard similarity and the theoretical variance is uniformly smaller than that of the classical MinHash with $K$ independent permutations. The theoretical proofs of C-MinHash require some non-trivial efforts. Numerical experiments are conducted to justify the theory and demonstrate the effectiveness of C-MinHash.

Xiaoyun Li, Ping Li

The method of random projection (RP) is the standard technique in machine learning and many other areas, for dimensionality reduction, approximate near neighbor search, compressed sensing, etc. Basically, RP provides a simple and effective scheme for approximating pairwise inner products and Euclidean distances in massive data. Closely related to RP, the method of random Fourier features (RFF) has also become popular, for approximating the Gaussian kernel. RFF applies a specific nonlinear transformation on the projected data from random projections. In practice, using the (nonlinear) Gaussian kernel often leads to better performance than the linear kernel (inner product), partly due to the tuning parameter $(γ)$ introduced in the Gaussian kernel. Recently, there has been a surge of interest in studying properties of RFF. After random projections, quantization is an important step for efficient data storage, computation, and transmission. Quantization for RP has also been extensive studied in the literature. In this paper, we focus on developing quantization algorithms for RFF. The task is in a sense challenging due to the tuning parameter $γ$ in the Gaussian kernel. For example, the quantizer and the quantized data might be tied to each specific tuning parameter $γ$. Our contribution begins with an interesting discovery, that the marginal distribution of RFF is actually free of the Gaussian kernel parameter $γ$. This small finding significantly simplifies the design of the Lloyd-Max (LM) quantization scheme for RFF in that there would be only one LM quantizer for RFF (regardless of $γ$). We also develop a variant named LM$^2$-RFF quantizer, which in certain cases is more accurate. Experiments confirm that the proposed quantization schemes perform well.