1 Introduction
With the increasing popularity of social networks, there exist many interesting and difficult problems, such as friends recommendation, information propagation, etc. In this paper, we study the problem of social trust prediction, which aims to estimate the positive or negative relationship among the users based on the existing trustness information associated with them. This problem plays an important role in social networks as the system can block the invitation from someone that the user does not trust, or recommend new friends who enjoys a high reputation.
Naturally, the social trust problem can be formulated within the matrix completion framework [LibenNowell and Kleinberg2007] [Billsus and Pazzani1998]. That is, the th entry of the observed data matrix is a 1bit code implying that the th user trusts the th user if . Here, denotes the number of users. However, what we observe is only a small fraction of the entries, whose values are zero. And our goal is to estimate the missing entries according to the 1bit measurements in .
Note that the problem is illposed if no assumption is imposed on the structure of the data. To solve the problem, a number of methods are proposed. Generally, existing social trust prediction methods fall into three categories. The first category is based on similarity measures or the structural context similarity [Newman2001] [Chowdhury2010] [Katz1953] [Jeh and Widom2002], motivated by the intuition that an individual tends to trust their neighbors, or the ones with similar trusted people. The second is based on low rank matrix completion [Billsus and Pazzani1998] [Cai et al.2010] [Huang et al.2013]
, which assumes that the underlying data matrix is lowrank or can be approximate by a lowrank matrix. The third one models the problem as a binary classification one and utilizes techniques such as logistic regression
[Leskovec et al.2010].Challenges. However, there are two issues emerging in social trust prediction which are not well characterized by the algorithms in previous works. First, the value of the observed entry is either 1 or , which is analogous to the binary classification problem. But in our problem, we are handling much more complex matrix data. Fortunately, [Srebro et al.2004]
presented a maximum margin matrix factorization framework that unifies the binary problem for vector case and matrix case. The key idea in their work is a lownorm matrix factorization, which will also be utilized in this paper. Second, the locations of the entries are sampled nonuniformly, which gaps the theory and practice for a lot of matrix completion algorithms. To tackle this challenge, we suggest using the maxnorm as a convex surrogate for the rank function, which is shown to be superior to the wellknown nuclear norm when addressing the nonuniform data
[Salakhutdinov and Srebro2010].Our contributions are twofolds: 1) To the best of our knowledge, we are the first to address the social trust prediction problem by utilizing a maxnorm constrained formulation. 2) Although a maxnorm constrained problem can be solved by SDP solvers and an accurate enough solution can be achieved, we here utilize a projected gradient algorithm that is scalable to large scale datasets. We empirically show the improvement of our formulation for the nonuniform 1bit benchmarks compared to stateoftheart solvers.
2 Related Work
Social interaction is investigated intensively in the last decades. The social interaction indicates the friendship, support, enemy or disapproval as shown in Figure 1. Online users rely on the trustworthiness information to filter information, establish collaboration or build social bonds. Social networks rely on the trust information to make recommendation, attract users from other circles, or influencing public opinions. Thus, the exploration of social trust has a wide range of applications, and has emerged as an important topic in social network research. A number of methods are proposed.
One kind of methods are based on the similarity measurement. Specifically, Jaccard’s coefficient is commonly used to measure the probability two items that have a relationship. Inspired by the metric, Jeh and Widom
[Jeh and Widom2002] proposed a domainindependent similarity measurement, SimRank. [Newman2001] directly defined a score to verify the correlation between common neighbors. Some methods are based on relational data modeling, structural proximity measures and stochastic relational model [Getoor and Diehl2005] [LibenNowell and Kleinberg2007] [Yu et al.2009]. The above mentioned methods are mainly derived from the solutions of link prediction. The link prediction is oriented to networklevel prediction, whereas social trust prediction focuses on personlevel. Another class of methods are derived from the collaborative filtering methods, such as clustering techniques [Sarwar et al.2001], modelbased methods [Hofmann and Puzicha1999], and the matrix factorization models [Srebro et al.2003] [Mnih and Salakhutdinov2007]. However, the data matrix of trust has some structure properties different from the useritem matrix, such as transitivity. Meanwhile, the social trust in reality is extremely sparse. For instance, Facebook has hundreds of millions of users, but most of them have less than 1,000 friends. Besides, the people with similar personality tend to behave similarly. To sum up, the data matrix of social trust has both sparse and lowrank structure. Thus, the social trust prediction problem is especially suitable for the matrix completion model. That is the focus of our paper.The problem of matrix completion is to recover a lowrank matrix from a subset of entries [Candès and Recht2009], which is given by:
(2.1) 
where is the data matrix, is the recovered matrix, and is the index set of observed entries. Th optimization problem (2.1) is not only NPhard, but requires double exponential time complexity with the number of samples [Recht et al.2010]. To solve the above problem, one alternative is to use nuclear norm as a relaxation to the rank function:
(2.2) 
where
denotes the sum of singular values of matrix
. [Cai et al.2010] developed a firstorder procedure to solve the convex problem (2.2), namely singular value thresholding (SVT). [Jain et al.2010] minimized the rank minimization by the singular value projection (SVP) algorithm. [Keshavan et al.2010] solved the problem by first trimming each row and column with too few entries, then compute the truncated SVD of the trimmed matrix. Under certain conditions, it showed accurate recovery on the order of samples ( is the number of samples, is the rank of recovered matrix). With the rapid development of matrix completion problem, some more efficient methods have been proposed [Candès and Tao2010][Gross2011][Wang and Xu2012][Huang et al.2013].However, all the methods mentioned above use the nuclear norm as the surrogate to the rank, whose exact recovery can be guaranteed only when the data are sampled uniformly, which is not practical in real world applications. On the other hand, recent empirical on maxnorm [Srebro et al.2004] shows promising results for nonuniform data if one utilize the maxnorm as a surrogate [Salakhutdinov and Srebro2010]. Notably, for some specific problems, such as collaborative filtering, [Srebro and Shraibman2005] proved that the generalization error bound for maxnorm is better than the nuclear norm. More recently, [Shen et al.2014] reported encouraging results on the subspace recovery task (which is closely relevant to matrix completion). Since the social trust data is nonuniformly sampled, we believe that a maxnorm regularized formulation can better handle the challenge than the nuclear norm. Our formulation is also inspired by a recent theoretical study on matrix completion with 1bit measurement [Cai and Zhou2013], which established a minimax lower bound on the general sampling model and derived the optimal convergence rate in terms of Frobenius norm loss. Furthermore, there are several practical algorithms to solve maxnorm regularized or maxnorm constrained problems, see [Lee et al.2010] and [Shen et al.2014] for example.
2.1 Overview
After review of related work in Section 2, we introduce the notations and formulate the problem in Section 3. Then we give algorithm to solve the maxnorm constrained 1bit matrix completion (MMC) problem in Section 4. Meanwhile, we also provide an equivalent SDP formulation for the MMC, which can be accurately solved at the expense of efficiency. Then we report the empirical study on two benchmark datasets in Section 5. Section 6 concludes this paper and discusses possible future work.
3 Notations and Problem Setup
In this section, we introduce the notations that will be used in this paper. Capital letters such as are used for matrices and lowercase bold letters such as denotes vectors. The th row and th column of a matrix is denoted by and respectively, and the th entry is denoted by . For a vector , we use to denote its th element. We denote the norm of a vector by . For a matrix , we denote the Frobenius norm by and denotes the maximum row norm of , i.e.,
We further define the maxnorm of [Linial et al.2007],
(3.1) 
where we enumerate all possible factorizations to obtain the minimum.
Intuition on maxnorm. At a first sight, the maxnorm is hard to understand. We simply explain why it is a tighter approximation to the rank function than the nuclear norm. Again, we write the nuclear norm of as a factorization form [Recht et al.2010]:
Note that the Frobenius norm is the sum of the square of the row norm. Thus, a nuclear norm regularizer actually constrains the average of the row norm, while the maxnorm constrains the maximum of the row norm!
Given the observed data , we are interested in approximating with a lowrank matrix , which can be formulated by,
where is an index set of observed entries and is some expected rank. is a projection operator on a matrix such that if and zero otherwise. However, it is usually intractable to optimize the above program as the rank function is nonconvex and noncontinuous [Recht et al.2010]. One common approach is to use the nuclear norm as a convex surrogate to the rank function. However, it is well known that the nuclear norm cannot well handle the nonuniform data. Motivated by the recent progress in maxnorm [Salakhutdinov and Srebro2010, Cai and Zhou2013, Shen et al.2014], we use the maxnorm as an alternative convex relaxation, which gives the following formulation:
(3.2) 
where is some tunable parameter.
4 Algorithm
The maxnorm is convex and moreover, it can be solved by any SDP solver. Formally, we have the following lemma:
Lemma 4.1 ([Srebro et al.2004]).
For any matrix and , if and only if there exist and , such that is semidefinite positive and each diagonal element in and is upper bounded by .
With Lemma 4.1 on hand, one can formulate Problem 3.2 as an SDP:
(4.1) 
And this program can be solved by any SDP solver to obtain accurate enough solution.
However, SDP solvers are not scalable to large matrices. Thus, in this paper, we apply a projected gradient method to solve Problem (3.2), which is due to [Lee et al.2010]. A key technique is the reformulation of the maxnorm (3.1). Assume that the rank of the optimal solution produced by the SDP (4.1) is at most . Then we can safely factorize , with and . Combining the factorization and the definition, we obtain the following equivalent program :
(4.2) 
Note that the gradient of the objective function w.r.t. and can be easily computed. That is,
(4.3) 
Here, for simplicity we define
The inequality constraints can be addressed by a projection step. That is, when we have a new iterate at the th iteration, we can check if they violate the constraints. If not, we can proceed to the next iteration. Otherwise, we can scale the rows of and/or by and/or respectively. In this way, we have the projection operator:
(4.4) 
If we further pick the step size via the Armijo rule [Armijo and others1966], it can be shown that the sequence of will converge to a stationary point [Bertsekas1999]. The algorithm is summarized in Algorithm 1.
The benefits of applying the factorization on the maxnorm are twofolds: 1) the memory cost can be significantly reduced from of SDP to . 2) it facilitates the projected gradient algorithm, which is computationally efficient when working on large matrices (see Section 5). However, note that Problem (4.2) is nonconvex. Fortunately, [Burer and Monteiro2005] proved that as long as we pick a sufficiently large value for , then any local minimum of Eq. (4.2) is a global optimum. In Section 5, we will report the influence of on the performance. Actually, in Algorithm 1, the stopping criteria is set to be a maximum iteration. One may also check if it reaches a local minima as the stopping criteria, as discussed in [Cai and Zhou2013].
4.1 Heuristic on
The is the only tunable parameter in our algorithm. For our problem, note that the data is of 1bit measurements, i.e., for . Also note that . Thus, . So we have . However, if we choose a large , the estimation may deviate away from . We find that lead to satisfactory improvement.
5 Experiments
In this section, we empirically evaluate our method for the matrix completion performance. We will first introduce the used datasets. In the experimental settings, we present the comparative methods and evaluation metrics. Then we report encouraging results on two benchmark datasets. We also examine the influence of matrix rank
.5.1 Datasets
We conduct the experiments on two benchmark datasets: Epinions and Slashdot. In these two datasets, the users are connected by explicit positive (trust) or negative (distrust) links (i.e., the 1bit measurements in ). The first dataset contains 119,217 nodes (users) and 841,000 edges (links), 85.0% of which are positive. The Slashdot dataset contains 82,144 users and 549,202 links, and 77.4% of the edges are labeled as positive. Table 1 gives a summary description about the subset used in our experiment.
It is clear that the distribution of links are not uniform since each user has his/her individual preference and own friendship network. Following [Huang et al.2013], we select 2,000 users with the highest degrees from each dataset to form the observation matrix .
Dataset  Epinions  Slashdot 

of Users  2,000  2,000 
of Trust  171,731  68,932 
of Distrust  18,916  20,032 
5.2 Experimental Settings
Baselines. We choose four stateoftheart methods as baselines, including SVP [Jain et al.2010], SVT [Cai et al.2010], OPTSpace [Keshavan et al.2010] and RRMC [Huang et al.2013]. Since SVT and RRMC need a specified rank, we tune the rank for these methods and choose the best performance as the final result.
Evaluation Metric. Let be the index set of all observed entries. We use two evaluation metrics to measure the performance, mean average error (MAE) and root mean square error (RMSE), computed as follows:
where denotes the cardinality of .
Observed entries (%)  Methods  
SVT  OPTSpace  SVP  RRMC  MMC  
10  0.3590.004  0.2890.019  0.4500.008  0.5760.001  0.2540.003 
20  0.3940.022  0.2360.005  0.2940.002  0.5180.002  0.2120.003 
30  0.3600.057  0.2190.009  0.2480.001  0.4600.002  0.2010.002 
40  0.4100.099  0.2050.008  0.2240.001  0.4180.002  0.1930.001 
50  0.4710.129  0.1970.007  0.2100.001  0.3860.001  0.1900.002 
60  0.4760.146  0.1970.003  0.1990.001  0.3620.002  0.2060.003 
Observed entries (%)  Methods  
SVT  OPTSpace  SVP  RRMC  MMC  
10  0.5130.010  0.5300.021  0.6100.010  0.6500.001  0.4660.004 
20  0.5590.031  0.4560.005  0.4590.002  0.6060.002  0.4060.004 
30  0.5320.082  0.4220.011  0.4150.002  0.5630.002  0.3830.002 
40  0.6200.171  0.4060.015  0.3940.002  0.5330.002  0.3710.002 
50  0.7190.225  0.3980.016  0.3810.001  0.5090.001  0.3640.001 
60  0.7280.288  0.4030.009  0.3710.002  0.4910.002  0.3650.002 
Observed entries (%)  Methods  
SVT  OPTSpace  SVP  RRMC  MMC  
10  0.6790.008  0.5540.017  0.7550.005  0.7150.001  0.5460.008 
20  0.5620.004  0.4580.008  0.5820.007  0.7040.001  0.4370.008 
30  0.5130.030  0.4270.009  0.5010.003  0.6860.003  0.3950.006 
40  0.5060.041  0.3950.009  0.4600.002  0.6470.006  0.3800.009 
50  0.4950.060  0.3760.004  0.4320.002  0.6090.002  0.3660.003 
60  0.5200.065  0.3620.011  0.4130.002  0.5850.002  0.3500.002 
Observed entries (%)  Methods  
SVT  OPTSpace  SVP  RRMC  MMC  
10  0.7880.008  0.8260.021  0.8730.004  0.8290.001  0.7740.008 
20  0.7180.020  0.7040.013  0.7460.007  0.8210.001  0.6790.009 
30  0.6800.043  0.6520.006  0.6800.003  0.8070.002  0.6330.008 
40  0.6700.056  0.6200.009  0.6470.003  0.7780.005  0.6150.011 
50  0.6420.055  0.5960.009  0.6240.003  0.7490.002  0.5810.006 
60  0.6990.084  0.5770.011  0.6090.002  0.7300.002  0.5660.002 
Dataset  SVT  OPTSpace  SVP  RRMC  MMC  

MAE  RMSE  Time  MAE  RMSE  Time  MAE  RMSE  Time  MAE  RMSE  Time  MAE  RMSE  Time  
Epinions  0.466  0.618  26.76  0.288  0.531  15.26  0.450  0.610  0.84  0.576  0.651  43.11  0.262  0.481  1.92 
Slashdot  0.618  0.728  50.73  0.458  0.705  9.23  0.582  0.746  0.83  0.715  0.830  44.12  0.437  0.679  2.41 
Training and Testing. We randomly split the dataset for training and testing. In particular, the number of observation measurements
for training ranges from 10% to 60%, with step size 10%. For each split, we run all the algorithms for 20 trials, with the training data in each trail being randomly sampled. Then, we report the mean and standard deviation of MAE and RMSE over all 20 trials.
5.3 Experimental Results
We report detailed results from Table 2 to Table 5. From the results in Tables 2 and 3, we observe that MMC outperforms the other methods in terms of both evaluation metrics on the Epinions dataset most of the time. In particular, when there are few observations 10 % (which indicates a hard task), MMC obtains the RMSE of 0.466, much better than OPTSpace (0.530), SVP (0.610) and RRMC (0.650). Except on the case with 60% observed entries, OPTSpace obtains the smallest MSE with 0.197, but our algorithm is comparative with 0.206. In a nut of shell, the gap between MMC and the baselines becomes larger as the fraction of observed entries decreases.
Similarly, our method achieves the least MAE and RMSE on the Slashdot dataset (see Table 4 and 5). For instance with 30 % observed entries, MMC obtains the MAE with less than 0.4, much better than the comparative methods, such as SVT (0.513), OPTSpace (0.427), SVP (0.501) and RRMC (0.686). In terms of RMSE, in the case of 20% observed entries, the RMSE values of other methods are all above 0.7 while our method reaches 0.679. In sum, our method is superior than the comparative methods on two reallife datasets in terms of MSE and RMSE most of the time.
Since we have studied the effectiveness of our method, here we examine the computational efficiency in Table 6, which is important for practical applications. To test the time complexity of the methods, we report the averaged time cost on the Epinions dataset with 10% observed entries and Slashdot with 20% observed entries. To illustrate the tradeoff between accuracy and efficiency, we also report the MAE and RMSE. As we see, SVP is the most efficient method, whose running time is 0.84 seconds on Epinions while ours is 1.92 seconds. On Slashdot, it also achieves the best performance in terms of efficiency. However, our method enjoys a significant improvement of MAE and RMSE compared to all baselines. Also, our algorithm is orders of magnitude faster than other three baselines. This implies that MMC favors a good tradeoff between the accuracy and efficiency.
5.4 Examine The Influence of
The nonconvex reformulation (4.2) requires an explicit rank estimation on the true matrix. In this section, we investigate the influence of on the Epinions dataset as an example. The rank is chosen from [1, 5, 50 ,100, 300, 500] and the results are plot in Figure 2. We observe that the rank has little influence on the performance. This is possibly because that the actual data has a lowrank structure (close to rank one). And from [Burer and Monteiro2005], we know that if is larger than the actual rank, any local minimum of Eq. (4.2) is also a global optima.
6 Conclusion and Future Work
In this paper, we formulated the social trust prediction in the matrix completion framework. In particular, due to the special structure of the social trust problem, i.e., the measurements are 1bit and the observed entries are nonuniformly sampled, we presented a maxnorm constrained 1bit matrix completion (MMC) algorithm. Since SDP solvers are not scalable to large scale matrices, we utilized a nonconvex reformulation of the maxnorm, which facilitates an efficient projected gradient decent algorithm. We empirically examined our algorithm on two benchmark datasets. Compared to other stateoftheart matrix completion formulations, MMC consistently outperformed them, which meets with recently developed theories on maxnorm. We also studied the tradeoff between the accuracy and efficiency and observed that MMC achieved superior accuracy while keeping comparable computational efficiency.
The maxnorm has been studied for several years and in many applications, such as collaborative filtering, clustering, subspace recovery. It is empirically and theoretically shown to be superior than the popular nuclear norm. This work investigates the power of maxnorm for social trust problem and demonstrates encouraging results. It is interesting and promising to apply maxnorm as a convex surrogate to other practical problems such as face recognition, subspace clustering etc.
References
 [Armijo and others1966] Larry Armijo et al. Minimization of functions having lipschitz continuous first partial derivatives. Pacific Journal of mathematics, 16(1):1–3, 1966.
 [Bertsekas1999] Dimitri P Bertsekas. Nonlinear programming. 1999.
 [Billsus and Pazzani1998] Daniel Billsus and Michael J Pazzani. Learning collaborative information filters. In ICML, volume 98, pages 46–54, 1998.
 [Burer and Monteiro2005] Samuel Burer and Renato DC Monteiro. Local minima and convergence in lowrank semidefinite programming. Mathematical Programming, 103(3):427–444, 2005.

[Cai and Zhou2013]
Tony Cai and WenXin Zhou.
A maxnorm constrained minimization approach to 1bit matrix
completion.
The Journal of Machine Learning Research
, 14(1):3619–3647, 2013.  [Cai et al.2010] JianFeng Cai, Emmanuel J Candès, and Zuowei Shen. A singular value thresholding algorithm for matrix completion. SIAM Journal on Optimization, 20(4):1956–1982, 2010.
 [Candès and Recht2009] Emmanuel J Candès and Benjamin Recht. Exact matrix completion via convex optimization. Foundations of Computational mathematics, 9(6):717–772, 2009.
 [Candès and Tao2010] Emmanuel J Candès and Terence Tao. The power of convex relaxation: Nearoptimal matrix completion. IEEE Transactions on Information Theory, 56(5):2053–2080, 2010.
 [Chowdhury2010] Gobinda Chowdhury. Introduction to modern information retrieval. 2010.
 [Getoor and Diehl2005] Lise Getoor and Christopher P Diehl. Link mining: a survey. ACM SIGKDD Explorations Newsletter, 7(2):3–12, 2005.
 [Gross2011] David Gross. Recovering lowrank matrices from few coefficients in any basis. IEEE Transactions on Information Theory, 57(3):1548–1566, 2011.
 [Hofmann and Puzicha1999] Thomas Hofmann and Jan Puzicha. Latent class models for collaborative filtering. In IJCAI, volume 99, pages 688–693, 1999.
 [Huang et al.2013] Jin Huang, Feiping Nie, Heng Huang, YiCheng Tu, and Yu Lei. Social trust prediction using heterogeneous networks. ACM Transactions on Knowledge Discovery from Data, 7(4):17, 2013.
 [Jain et al.2010] Prateek Jain, Raghu Meka, and Inderjit S Dhillon. Guaranteed rank minimization via singular value projection. In NIPS, pages 937–945, 2010.
 [Jeh and Widom2002] Glen Jeh and Jennifer Widom. Simrank: a measure of structuralcontext similarity. In ACM KDD, pages 538–543, 2002.
 [Katz1953] Leo Katz. A new status index derived from sociometric analysis. Psychometrika, 18(1):39–43, 1953.
 [Keshavan et al.2010] Raghunandan H Keshavan, Andrea Montanari, and Sewoong Oh. Matrix completion from a few entries. Information Theory, IEEE Transactions on, 56(6):2980–2998, 2010.
 [Lee et al.2010] Jason D Lee, Ben Recht, Nathan Srebro, Joel Tropp, and Ruslan R Salakhutdinov. Practical largescale optimization for maxnorm regularization. In NIPS, pages 1297–1305, 2010.
 [Leskovec et al.2010] Jure Leskovec, Daniel Huttenlocher, and Jon Kleinberg. Predicting positive and negative links in online social networks. In WWW, pages 641–650, 2010.
 [LibenNowell and Kleinberg2007] David LibenNowell and Jon Kleinberg. The linkprediction problem for social networks. Journal of the American society for information science and technology, 58(7):1019–1031, 2007.
 [Linial et al.2007] Nati Linial, Shahar Mendelson, Gideon Schechtman, and Adi Shraibman. Complexity measures of sign matrices. Combinatorica, 27(4):439–463, 2007.
 [Mnih and Salakhutdinov2007] Andriy Mnih and Ruslan Salakhutdinov. Probabilistic matrix factorization. In NIPS, pages 1257–1264, 2007.
 [Newman2001] Mark EJ Newman. Clustering and preferential attachment in growing networks. Physical Review E, 64(2):025102, 2001.
 [Recht et al.2010] Benjamin Recht, Maryam Fazel, and Pablo A Parrilo. Guaranteed minimumrank solutions of linear matrix equations via nuclear norm minimization. SIAM review, 52(3):471–501, 2010.
 [Salakhutdinov and Srebro2010] Ruslan Salakhutdinov and Nathan Srebro. Collaborative filtering in a nonuniform world: Learning with the weighted trace norm. tc (X), 10:2, 2010.
 [Sarwar et al.2001] Badrul Sarwar, George Karypis, Joseph Konstan, and John Riedl. Itembased collaborative filtering recommendation algorithms. In WWW, pages 285–295, 2001.
 [Shen et al.2014] Jie Shen, Huan Xu, and Ping Li. Online optimization for maxnorm regularization. In NIPS, pages 1718–1726, 2014.
 [Srebro and Shraibman2005] Nathan Srebro and Adi Shraibman. Rank, tracenorm and maxnorm. In Learning Theory, pages 545–560. 2005.
 [Srebro et al.2003] Nathan Srebro, Tommi Jaakkola, et al. Weighted lowrank approximations. In ICML, volume 3, pages 720–727, 2003.
 [Srebro et al.2004] Nathan Srebro, Jason DM Rennie, and Tommi Jaakkola. Maximummargin matrix factorization. In NIPS, volume 17, pages 1329–1336, 2004.
 [Wang and Xu2012] YuXiang Wang and Huan Xu. Stability of matrix factorization for collaborative filtering. ICML, 2012.
 [Yu et al.2009] Kai Yu, John Lafferty, Shenghuo Zhu, and Yihong Gong. Largescale collaborative prediction using a nonparametric random effects model. In ICML, pages 1185–1192, 2009.
Comments
There are no comments yet.