# robust_optimization_for_nonconvex_objectives__ca90cdff.pdf

Robust Optimization for Non-Convex Objectives

Robert Chen Computer Science Harvard University

Brendan Lucier Microsoft Research

New England

Yaron Singer Computer Science Harvard University

Vasilis Syrgkanis Microsoft Research

New England

We consider robust optimization problems, where the goal is to optimize in the worst case over a class of objective functions. We develop a reduction from robust improper optimization to stochastic optimization: given an oracle that returns -approximate solutions for distributions over objectives, we compute a distribution over solutions that is -approximate in the worst case. We show that derandomizing this solution is NP-hard in general, but can be done for a broad class of statistical learning tasks. We apply our results to robust neural network training and submodular optimization. We evaluate our approach experimentally on corrupted character classiﬁcation and robust inﬂuence maximization in networks.

1 Introduction

In many learning tasks we face uncertainty about the loss we aim to optimize. Consider, for example, a classiﬁcation task such as character recognition, required to perform well under various types of distortion. In some environments, such as recognizing characters in photos, the classiﬁer must handle rotation and patterned backgrounds. In a different environment, such as low-resolution images, it is more likely to encounter noisy pixelation artifacts. Instead of training a separate classiﬁer for each possible scenario, one seeks to optimize performance in the worst case over different forms of corruption (or combinations thereof) made available to the trainer as black-boxes.

More generally, our goal is to ﬁnd a minimax solution that optimizes in the worst case over a given family of functions. Even if each individual function can be optimized effectively, it is not clear such solutions would perform well in the worst case. In many cases of interest, individual objectives are non-convex and hence state-of-the-art methods are only approximate. In stochastic optimization, where one must optimize a distribution over loss functions, approximate stochastic optimization is often straightforward, since loss functions are commonly closed under convex combination. Can approximately optimal stochastic solutions yield an approximately optimal robust solution?

In this paper we develop a reduction from robust optimization to stochastic optimization. Given an - approximate oracle for stochastic optimization we show how to implement an -approximate solution for robust optimization under a necessary extension, and illustrate its effectiveness in applications.

Main Results. Given an -approximate stochastic oracle for distributions over (potentially nonconvex) loss functions, we show how to solve -approximate robust optimization in a convexiﬁed solution space. This outcome is improper in the sense that it may lie outside the original solution space, if the space is non-convex. This can be interpreted as computing a distribution over solutions. We show that the relaxation to improper learning is necessary in general: It is NP-hard to achieve robust optimization with respect to the original outcome space, even if stochastic optimization can be solved exactly, and even if there are only polynomially many loss functions. We complement this by showing that in any statistical learning scenario where loss is convex in the predicted dependent variable, we can ﬁnd a single (deterministic) solution with matching performance guarantees.

31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.

Technical overview. Our approach employs an execution of no-regret dynamics on a zero-sum game, played between a learner equipped with an -approximate stochastic oracle, and an adversary who aims to ﬁnd a distribution over loss functions that maximizes the learner s loss. This game converges to an approximately robust solution, in which the learner and adversary settle upon an - approximate minimax solution. This convergence is subject to an additive regret term that converges at a rate of T 1/2 over T rounds of the learning dynamics.

Applications. We illustrate the power of our reduction through two main examples. We ﬁrst consider statistical learning via neural networks. Given an arbitrary training method, our reduction generates a net that optimizes robustly over a given class of loss functions. We evaluate our method experimentally on a character recognition task, where the loss functions correspond to different corruption models made available to the learner as black boxes. We verify experimentally that our approach signiﬁcantly outperforms various baselines, including optimizing for average performance and optimizing for each loss separately. We also apply our reduction to inﬂuence maximization, where the goal is to maximize a concave function (the independent cascade model of inﬂuence [11]) over a non-convex space (subsets of vertices in a network). Previous work has studied robust inﬂuence maximization directly [9, 5, 15], focusing on particular, natural classes of functions (e.g., edge weights chosen within a given range) and establishing hardness and approximation results. In comparison, our method is agnostic to the particular class of functions, and achieves a strong approximation result by returning a distribution over solutions. We evaluate our method on real and synthetic datasets, with the goal of robustly optimizing a suite of random inﬂuence instantiations. We verify experimentally that our approach signiﬁcantly outperforms natural baselines.

Related work. There has recently been a great deal of interest in robust optimization in machine learning [20, 4, 17, 21, 16]. For continuous optimization, the work that is closest to ours is perhaps that by Shalev-Shwartz and Wexler [20] and Namkoong and Duchi [17] that use robust optimization to train against convex loss functions. The main difference is that we assume a more general setting in which the loss functions are non-convex and one is only given access to the stochastic oracle. Hence, the proof techniques and general results from these papers do not apply to our setting. We note that our result generalizes these works, as they can be considered as the special case in which we have a distributional oracle whose approximation is optimal. In particular, [20, Theorem 1] applies to the realizable statistical learning setting where the oracle has small mistake bound C. Our applications require a more general framing that hold for any optimization setting with access to an approximate oracle, and approximation is in the multiplicative sense with respect to the optimal value. In submodular optimization there has been a great deal of interest in robust optimization as well [12, 13, 10, 6]. The work closest to ours is that by He and Kempe [10] who consider a slightly different objective than ours. Kempe and He s results apply to inﬂuence but do not extend to general submodular functions. Finally, we note that unlike recent work on non-convex optimization [7, 1, 8] our goal in this paper is not to optimize a non-convex function. Rather, we abstract the non-convex guarantees via the approximate stochastic oracle.

2 Robust Optimization with Approximate Stochastic Oracles

We consider the following model of optimization that is robust to objective uncertainty. There is a space X over which to optimize, and a ﬁnite set of loss functions1 L = {L1, . . . , Lm} where each Li 2 L is a function from X to [0, 1]. Intuitively, our goal is to ﬁnd some x 2 X that achieves low loss in the worst-case over loss functions in L. For x 2 X, write g(x) = maxi2[m] Li(x) for the worst-case loss of x. The minimax optimum is given by

x2X g(x) = min

i2[m] Li(x). (1)

The goal of -approximate robust optimization is to ﬁnd x such that g(x) .2

1We describe an extension to inﬁnite sets of loss functions in the full version of the paper. Our results also extend naturally to the goal of maximizing the minimum of a class of reward functions.

2This oracle framework is similar to that used by Ben-Tal et al. [3], but where the approximation is multiplicative rather than additive.

Algorithm 1 Oracle Efﬁcient Improper Robust Optimization

Input: Objectives L = {L1, . . . , Lm}, Apx stochastic oracle M, parameters T, for each time step t 2 [T] do

wt[i] / exp

Set xt = M(wt) end for Output: the uniform distribution over {x1, . . . , x T }

Given a distribution P over solutions X, write g(P) = maxi2[m] Ex P[Li(x)] for the worst-case expected loss of a solution drawn from P. A weaker version of robust approximation is improper robust optimization: ﬁnd a distribution P over X such that g(P) .

Our results take the form of reductions to an approximate stochastic oracle, which ﬁnds a solution x 2 X that approximately minimizes a given distribution over loss functions.3

Deﬁnition 1 ( -Approximate Stochastic Oracle). Given a distribution D over L, an -approximate stochastic oracle M(D) computes x 2 X such that

EL D [L(x )] min

x2X EL D [L(x)] . (2)

2.1 Improper Robust Optimization with Oracles

We ﬁrst show that, given access to an -approximate stochastic oracle, it is possible to efﬁciently implement improper -approximate robust optimization, subject to a vanishing additive loss term.

Theorem 1. Given access to an -approximate stochastic oracle, Algorithm 1 with =

2T computes a distribution P over solutions, deﬁned as a uniform distribution over a set {x1, . . . , x T }, so that

max i2[m] Ex P [Li(x)] +

Moreover, for any the distribution P computed by Algorithm 1 satisﬁes:

max i2[m] Ex P [Li(x)] (1 + ) + log(m)

Proof. We give the proof of the ﬁrst result and defer the second result to the full version of the paper. We can interpret Algorithm 1 in the following way. We deﬁne a zero-sum game between a learner and an adversary. The learner s action set is equal to X and the adversary s action set is equal to [m]. The loss of the learner when he picks x 2 X and the adversary picks i 2 [m] is deﬁned as Li(x). The corresponding payoff of the adversary is Li(x).

We will run no-regret dynamics on this zero-sum game, where at every iteration t = 1, . . . , T, the adversary will pick a distribution over functions and subsequently the learner picks a solution xt. For simpler notation we will denote with wt the probability density function on [m] associated with the distribution of the adversary. That is, wt[i] is the probability of picking function Li 2 L. The adversary picks a distribution wt based on some arbitrary no-regret learning algorithm on the m actions in L. For concreteness consider the case where the adversary picks a distribution based on the multiplicative weight updates algorithm, i.e.,

wt[i] / exp

3All our results easily extend to the case where the oracle computes a solution that is approximately optimal up to an additive error, rather than a multiplicative one. For simplicity of exposition we present the multiplicative error case as it is more in line with the literature on approximation algorithms.

Subsequently the learner picks a solution xt that is the output of the -approximate stochastic oracle on the distribution selected by the adversary at time-step t. That is,

xt = M (wt) . (7)

Write (T) =

T . By the guarantees of the no-regret algorithm for the adversary, we have that

EI wt [LI(xt)] max

Li(xt) (T). (8)

Combining the above with the guarantee of the stochastic oracle we have

i2[m] Li(x) min

EI wt [LI(x)] 1

min x2X EI wt [LI(x)]

1 EI wt [LI(xt)] (By oracle guarantee for each t)

. (By no-regret of adversary)

Thus, if we deﬁne with P to be the uniform distribution over {x1, . . . , x T }, then we have derived

max i2[m] Ex P [Li(x)] + (T) (9)

as required.

A corollary of Theorem 1 is that if the solution space X is convex and the objective functions Li 2 L are all convex functions, then we can compute a single solution x that is approximately minimax optimal. Of course, in this setting one can calculate and optimize the maximum loss directly in time proportional to |L|; this result therefore has the most bite when the set of functions is large. Corollary 2. If the space X is a convex space and each loss function Li 2 L is a convex function, then the point x = 1

t=1 xt 2 X, where {x1, . . . , x T } are the output of Algorithm 1, satisﬁes:

max i2[m] Li(x ) +

Proof. By Theorem 1, we get that if P is the uniform distribution over {x1, . . . , x T } then

max i2[m] Ex P[Li(x)] +

Since X is convex, the solution x = Ex P[x] is also part of X. Moreover, since each Li 2 L is convex, we have that Ex P[Li(x)] Li(Ex P[x]) = Li(x ). We therefore conclude

max i2[m] Li(x ) max

i2[m] Ex P[Li(x)] +

as required.

2.2 Robust Statistical Learning

Next we apply our main theorem to statistical learning. Consider regression or classiﬁcation settings where data points are pairs (z, y), z 2 Z is a vector of features, and y 2 Y is the dependent variable. The solution space X is then a space of hypotheses H, with each h 2 H a function from Z to Y. We also assume that Y is a convex subset of a ﬁnite-dimensional vector space.

We are given a set of loss functions L = {L1, . . . , Lm}, where each Li 2 L is a functional Li : H ! [0, 1]. Theorem 1 implies that, given an -approximate stochastic optimization oracle,

we can compute a distribution over T hypotheses from H that achieves an -approximate minimax guarantee. If the loss functionals are convex over hypotheses, then we can compute a single ensemble hypothesis h (possibly from a larger space of hypotheses, if H is non-convex) that achieves this guarantee. Theorem 3. Suppose that L = {L1, . . . , Lm} are convex functionals. Then the ensemble hypothesis h = 1 T

t=1 h, where {h1, . . . , h T } are the hypotheses output by Algorithm 1 given an -approximate stochastic oracle, satisﬁes

max i2[m] Li(h ) min

i2[m] Li(h) +

Proof. The proof is similar to the proof of Corollary 2.

We emphasize that the convexity condition in Theorem 3 is over the class of hypotheses, rather than over features or any natural parameterization of H (such as weights in a neural network). This is a mild condition that applies to many examples in statistical learning theory. For instance, consider the case where each loss Li(h) is the expected value of some ex-post loss function i(h(z), y) given a distribution Di over Z Y :

Li(h) = E(z,y) Di [ i(h(z), y)] . (12)

In this case, it is enough for the function i( , ) to be convex with respect to its ﬁrst argument (i.e., the predicted dependent variable). This is satisﬁed by most loss functions used in machine learning, such as multinomial logistic loss (cross-entropy loss) (ˆy, y) = P

c2[k] yc log(ˆyc) from multi-class classiﬁcation or squared loss (ˆy, y) = kˆy yk2 as used in regression. For all these settings, Theorem 3 provides a tool for improper robust learning, where the ﬁnal hypothesis h is an ensemble of T base hypotheses from H. Again, the underlying optimization problem can be arbitrarily non-convex in the natural parameters of the hypothesis space; in Section 3.1 we will show how to apply this approach to robust training of neural networks, where the stochastic oracle is simply a standard network training method. For neural networks, the fact that we achieve improper learning (as opposed to standard learning) corresponds to training a neural network with a single extra layer relative to the networks generated by the oracle.

2.3 Robust Submodular Maximization

In robust submodular maximization we are given a family of reward functions F = {f1, . . . , fm}, where each fi 2 F is a monotone submodular function from a ground set N of n elements to [0, 1]. Each function is assumed to be monotone and submodular, i.e., for any S T N, fi(S) fi(T); and for any S, T N, f(S [ T) + f(S \ T) f(S) + f(T). The goal is to select a set S N of size k whose worst-case value over i, i.e., g(S) = mini2[m] fi(S), is at least a 1/ factor of the minimax optimum = max T :|T | k mini2[m] fi(T).

This setting is a special case of our general robust optimization setting (phrased in terms of rewards rather than losses). The solution space X is equal to the set of subsets of size k among all elements in N and the set F is the set of possible objective functions. The stochastic oracle 1, instantiated in this setting, asks for the following: given a convex combination of submodular functions F(S) = Pm

i=1 w[i] fi(S), compute a set S such that F(S ) 1

max S:|S| k F(S).

Computing the maximum value set of size k is NP-hard even for a single submodular function. The following very simple greedy algorithm computes a (1 1/e)-approximate solution [19]: begin with Scur = ;, and at each iteration add to the current solution Scur the element j 2 N Scur that has the largest marginal contribution: f({j} [ Scur) f(Scur). Moreover, this approximation ratio is known to be the best possible in polynomial time [18]. Since a convex combination of monotone submodular functions is also a monotone submodular function, we immediately get that there exists a (1 1/e)-approximate stochastic oracle that can be computed in polynomial time. The algorithm is

formally given in Algorithm 2. Combining the above with Theorem 1 we get the following corollary. Corollary 4. Algorithm 1, with stochastic oracle Mgreedy, computes in time poly(T, n) a distribution P over sets of size k, deﬁned as a uniform distribution over a set {S1, . . . , ST }, such that

min i2[m] ES P [fi(S)]

(1 ) log(m)

Algorithm 2 Greedy stochastic Oracle for Submodular Maximization Mgreedy

Input: Set of elements N, objectives F = {f1, . . . , fm}, distribution over objectives w Set Scur = ; for j = 1 to k do

Let j = arg maxj2N Scur

i=1 w[i] (fi({j} [ Scur) fi(Scur)) Set Scur = {j } [ Scur end for

Figure 1: Sample MNIST image with each of the corruptions applied to it. Background Corruption Set & Shrink Corruption Set (top). Pixel Corruption Set & Mixed Corruption Set (bottom).

We show in the full version of the paper that computing a single set S that achieves a (1 1/e)- approximation to is also NP-hard. This is true even if the functions fi are additive. However, by allowing a randomized solution over sets we can achieve a constant factor approximation to in polynomial time.

Since the functions are monotone, the above result implies a simple way of constructing a single set S that is of larger size than k, which deterministically achieves a constant factor approximation to . The latter holds by simply taking the union of the sets {S1, . . . , ST } in the support of the distribution returned by Algorithm 1. We get the following bi-criterion approximation scheme.

Corollary 5. Suppose that we run the reward version of Algorithm 1, with = and for T = log(m)

2 , returning {S1, . . . , ST }. Then the set S = S1 [ . . . [ ST , which is of size at most k log(m)

2 , satisﬁes

min i2[m] fi(S )

3 Experiments4

3.1 Robust Classiﬁcation with Neural Networks

A classic application of our robust optimization framework is classiﬁcation with neural networks for corrupted or perturbed datasets. We have a data set Z of pairs (z, y) of an image z 2 Z and label y 2 Y that can be corrupted in m different ways which produces data sets Z1, . . . , Zm. The hypothesis space H is the set of all neural nets of some ﬁxed architecture and for each possible assignment of weights. We denote each such hypothesis with h( ; ) : Z ! Y for 2 Rd, with d being the number of parameters (weights) of the neural net. If we let Di be the uniform distribution over each corrupted data set Zi, then we are interested in minimizing the empirical cross-entropy (aka multinomial logistic) loss in the worst case over these different distributions Di. The latter is a special case of our robust statistical learning framework from Section 2.2.

Training a neural network is a non-convex optimization problem and we have no guarantees on its performance. We instead assume that for any given distribution D over pairs (z, y) of images and labels and for any loss function (h(z; ), y), training a neural net with stochastic gradient descent run on images drawn from D can achieve an approximation to the optimal expected loss, i.e. min 2Rd E(z,y) D [ (h(z; ), y)]. Notice that this implies an -approximate stochastic oracle for the

4Code used to implement the algorithms and run the experiments is available at https://github.com/ 12degrees/Robust-Classification/.

corrupted dataset robust training problem: for any distribution w over the different corruptions [m], the stochastic oracle asks to give an -approximation to the minimization problem:

w[i] E(z,y) Di [ (h(z; ), y)] (15)

The latter is simply another expected loss problem with distribution over images being the mixture distribution deﬁned by ﬁrst drawing a corruption index i from w and then drawing a corrupted image from distribution Di. Hence, our oracle assumption implies that SGD on this mixture is an -approximation. By linearity of expectation, an alternative way of viewing the stochastic oracle problem is that we are training a neural net on the original distribution of images, but with loss function being the weighted combination of loss functions Pm

i=1 w[i] (h(ci(z); ), y), where ci(z) is the i-th corrupted version of image z. In our experiments we implemented both of these interpretations of the stochastic oracle, which we call the Hybrid Method and Composite Method, respectively, when designing our neural network training scheme (see the full version of the paper for further details). Finally, because we use the cross-entropy loss, which is convex in the prediction of the neural net, we can also apply Theorem 3 to get that the ensemble neural net, which takes the average of the predictions of the neural nets created at each iteration of the robust optimization, will also achieve good worst-case loss (we refer to this as Ensemble Bottleneck Loss).

Experiment Setup. We use the MNIST handwritten digits data set containing 55000 training images, 5000 validation images, and 10000 test images, each image being a 28 28 pixel grayscale image. The intensities of these 576 pixels (ranging from 0 to 1) are used as input to a neural network that has 1024 nodes in its one hidden layer. The output layer uses the softmax function to give a distribution over digits 0 to 9. The activation function is Re LU and the network is trained using Gradient Descent with learning parameter 0.5 through 500 iterations of mini-batches of size 100.

In general, the corruptions can be any black-box corruption of the image. In our experiments, we consider four types of corruption (m = 4). See the full version of the paper for further details about corruptions.

Baselines. We consider three baselines: (i) Individual Corruption: for each corruption type i 2 [m], we construct an oracle that trains a neural network using the training data perturbed by corruption i, and then returns the trained network weights as t, for every t = 1, . . . , T. This gives m baselines, one for each corruption type; (ii) Even Split: this baseline alternates between training with different corruption types between iterations. In particular, call the previous m baseline oracles O1, ..., Om. Then this new baseline oracle will produce t with Oi+1, where i t mod m, for every t = 1, ..., T; (iii) Uniform Distribution: This more advanced baseline runs the robust optimization scheme with the Hybrid Method (see Appendix), but without the distribution updates. Instead, the distribution over corruption types is ﬁxed as the discrete uniform [ 1

m] over all T iterations. This allows us to check if the multiplicative weight updates in the robust optimization algorithm are providing beneﬁt.

Results. The Hybrid and Composite Methods produce results far superior to all three baseline types, with differences both substantial in magnitude and statistically signiﬁcant, as shown in Figure 2. The more sophisticated Composite Method outperforms the Hybrid Method. Increasing T improves performance, but with diminishing returns largely because for sufﬁciently large T, the distribution over corruption types has moved from the initial uniform distribution to some more optimal stable distribution (see the full version for details). All these effects are consistent across the 4 different corruption sets tested. The Ensemble Bottleneck Loss is empirically much smaller than Individual Bottleneck Loss. For the best performing algorithm, the Composite Method, the mean Ensemble Bottleneck Loss (mean Individual Bottleneck Loss) with T = 50 was 0.34 (1.31) for Background Set, 0.28 (1.30) for Shrink Set, 0.19 (1.25) for Pixel Set, and 0.33 (1.25) for Mixed Set. Thus combining the T classiﬁers obtained from robust optimization is practical for making predictions on new data.

3.2 Robust Inﬂuence Maximization

We apply the results of Section 2.3 to the robust inﬂuence maximization problem. Given a directed graph G = (V, E), the goal is to pick a seed set S of k nodes that maximize an inﬂuence function

Figure 2: Comparison of methods, showing mean of 10 independent runs and a 95% conﬁdence band. The criterion is Individual Bottleneck Loss: min[m] E P [ (h(z; ), y)], where P is uniform over all solutions i for that method. Baselines (i) and (ii) are not shown as they produce signiﬁcantly higher loss.

f G(S), where f G(S) is the expected number of individuals inﬂuenced by opinion of the members of S. We used f G(S) to be the number of nodes reachable from S (our results extend to other models).

In robust inﬂuence maximization, the goal is to maximize inﬂuence in the worst-case (Bottleneck Inﬂuence) over m functions {f1, . . . , fm}, corresponding to m graphs {G1, . . . , Gm}, for some ﬁxed seed set of size k. This is a special case of robust submodular maximization after rescaling to [0, 1].

Experiment Setup. Given a base directed graph G(V, E), we produce m graphs Gi = (V, Ei) by randomly including each edge e 2 E with some probability p. We consider two base graphs and two sets of parameters for each: (i) The Wikipedia Vote Graph [14]. In Experiment A, the parameters are |V | = 7115, |E| = 103689, m = 10, p = 0.01 and k = 10. In Experiment B, change p = 0.015 and k = 3. (ii) The Complete Directed Graph on |V | = 100 vertices. In Experiment A, the parameters are m = 50, p = 0.015 and k = 2. In Experiment B, change p = 0.01 and k = 4.

Baselines. We compared our algorithm (Section 2.3) to three baselines: (i) Uniform over Individual Greedy Solutions: Apply greedy maximization (Algorithm 2) on each graph separately, to get solutions {Sg

1, . . . , Sg

m}. Return the uniform distribution over these solutions; (ii) Greedy on Uniform Distribution over Graphs: Return the output of greedy submodular maximization (Algorithm 2) on the uniform distribution over inﬂuence functions. This can be viewed as maximizing expected inﬂuence; (iii) Uniform over Greedy Solutions on Multiple Perturbed Distributions: Generate T distributions {w

1, . . . , w

T } over the m functions, by randomly perturbing the uniform distribution. Perturbation magnitudes were chosen s.t. w

t has the same expected 1 distance from uniform as the distribution returned by robust optimization at iteration t.

Results. For both graph experiments, robust optimization outperforms all baselines on Bottleneck Inﬂuence; the difference is statistically signiﬁcant as well as large in magnitude for all T > 50 (see Figure 3). Moreover, the individual seed sets generated at each iteration t of robust optimization themselves achieve empirically good inﬂuence as well; see the full version for further details.

[1] Zeyuan Allen Zhu and Elad Hazan. Variance reduction for faster non-convex optimization. In

Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, pages 699 707, 2016.

Figure 3: Comparison for various T, showing mean Bottleneck Inﬂuence and 95% conﬁdence on 10 runs.

[2] Sanjeev Arora, Elad Hazan, and Satyen Kale. The multiplicative weights update method: a

meta-algorithm and applications. Theory of Computing, 8(6):121 164, 2012.

[3] Aharon Ben-Tal, Elad Hazan, Tomer Koren, and Shie Mannor. Oracle-based robust optimization

via online learning. Operations Research, 63(3):628 638, 2015.

[4] Sabyasachi Chatterjee, John C. Duchi, John D. Lafferty, and Yuancheng Zhu. Local minimax

complexity of stochastic convex optimization. In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 3423 3431, 2016.

[5] Wei Chen, Tian Lin, Zihan Tan, Mingfei Zhao, and Xuren Zhou. Robust inﬂuence maximization.

In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016, pages 795 804, 2016.

[6] Wei Chen, Tian Lin, Zihan Tan, Mingfei Zhao, and Xuren Zhou. Robust inﬂuence maximization.

In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016, pages 795 804, 2016.

[7] Elad Hazan, Kﬁr Y. Levy, and Shai Shalev-Shwartz. Beyond convexity: Stochastic quasi-convex

optimization. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 1594 1602, 2015.

[8] Elad Hazan, Kﬁr Yehuda Levy, and Shai Shalev-Shwartz. On graduated optimization for

stochastic non-convex problems. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, pages 1833 1841, 2016.

[9] Xinran He and David Kempe. Robust inﬂuence maximization. In Proceedings of the 22nd ACM

SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016, pages 885 894, 2016.

[10] Xinran He and David Kempe. Robust inﬂuence maximization. In Proceedings of the 22nd ACM

SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016, pages 885 894, 2016.

[11] David Kempe, Jon Kleinberg, and Éva Tardos. Maximizing the spread of inﬂuence through

a social network. In Proceedings of the Ninth ACM SIGKDD International Conference on

Knowledge Discovery and Data Mining, KDD 03, pages 137 146, New York, NY, USA, 2003. ACM.

[12] Andreas Krause, H. Brendan Mc Mahan, Carlos Guestrin, and Anupam Gupta. Selecting obser-

vations against adversarial objectives. In Advances in Neural Information Processing Systems 20, Proceedings of the Twenty-First Annual Conference on Neural Information Processing Systems, Vancouver, British Columbia, Canada, December 3-6, 2007, pages 777 784, 2007.

[13] Andreas Krause, Alex Roper, and Daniel Golovin. Randomized sensing in adversarial environ-

ments. In Proceedings of the 22nd International Joint Conference On Artiﬁcial Intelligence, Barcelona, Catalonia, Spain, July 16-22, 2011, pages 2133 2139, 2011.

[14] Jure Leskovec. Wikipedia vote network. Stanford Network Analysis Project.

[15] Meghna Lowalekar, Pradeep Varakantham, and Akshat Kumar. Robust inﬂuence maximization:

(extended abstract). In Proceedings of the 2016 International Conference on Autonomous Agents & Multiagent Systems, Singapore, May 9-13, 2016, pages 1395 1396, 2016.

[16] Yishay Mansour, Aviad Rubinstein, and Moshe Tennenholtz. Robust probabilistic inference. In

Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2015, San Diego, CA, USA, January 4-6, 2015, pages 449 460, 2015.

[17] Hongseok Namkoong and John C. Duchi. Stochastic gradient methods for distributionally

robust optimization with f-divergences. In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 2208 2216, 2016.

[18] G. L. Nemhauser and L. A. Wolsey. Best algorithms for approximating the maximum of a

submodular set function. Mathematics of Operations Research, 3(3):177 188, 1978.

[19] G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher. An analysis of approximations for maximizing

submodular set functions i. Mathematical Programming, 14(1):265 294, 1978.

[20] Shai Shalev-Shwartz and Yonatan Wexler. Minimizing the maximal loss: How and why. In

Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, pages 793 801, 2016.

[21] Jacob Steinhardt and John C. Duchi. Minimax rates for memory-bounded sparse linear regres-

sion. In Proceedings of The 28th Conference on Learning Theory, COLT 2015, Paris, France, July 3-6, 2015, pages 1564 1587, 2015.