We study the problem of attributing prediction a deep network to its input features, previously studied by several other works. identify two fundamental axioms---Sensitivity and Implementation Invariance that attribution methods ought satisfy. show they are not satisfied most known methods, which we consider be weakness those methods. use axioms guide design new method called Integrated Gradients. Our requires no modification original is extremely simple implement; it just needs few calls...

In a seminal paper, Karp, Vazirani, and Vazirani show that simple ranking algorithm achieves competitive ratio of 1-1/e for the online bipartite matching problem in standard adversarial model, where is also shown to be optimal. Their result implies random arrivals model defined by Goel Mehta, nodes arrive order, greedy 1-1/e. this we study it has at least 0.696, beating ≈ 0.632 barrier model. Our extends i.i.d. distribution Feldman et al., removing assumption known.

We design and analyze approximately revenue-maximizing auctions in general single-parameter settings. Bidders have publicly observable attributes, we assume that the valuations of indistinguishable bidders are independent draws from a common distribution. Crucially, all valuation distributions priori unknown to seller. Despite this handicap, show how obtain optimal expected revenue - nearly as large what could be obtained if were known advance under quite conditions.

For revenue and welfare maximization in single-dimensional Bayesian settings, Chawla et al. (STOC10) recently showed that sequential posted-price mechanisms (SPMs), though simple form, can perform surprisingly well compared to the optimal mechanisms. In this paper, we give a theoretical explanation of fact, based on connection notion correlation gap. Loosely speaking, for auction environments with matroid constraints, relate performance mechanism expectation monotone submodular function over...

Gradients have been used to quantify feature importance in machine learning models. Unfortunately, nonlinear deep networks, not only individual neurons but also the whole network can saturate, and as a result an important input tiny gradient. We study various observe that this phenomena is indeed widespread, across many inputs. propose examine interior gradients, which are gradients of counterfactual inputs constructed by scaling down original input. apply our method GoogleNet architecture...

We describe Analyza, a system that helps lay users explore data. Analyza has been used within two large real world systems. The first is question-and-answer feature in spreadsheet product. second provides convenient access to revenue/inventory database for sales force. Both user bases consist of who do not necessarily have coding skills, demonstrating Analyza's ability democratize discuss the key design decisions implementing this system. For instance, how mix structured and natural language...

Purpose: Hemoglobin (Hb) and red blood cell distribution width-standard deviation (RDW-SD) have clinical significance in the prognosis of heart failure (HF). Little is known regarding prognostic value Hb/RDW-SD ratio patients with HF. This study sought to investigate association between HF 3-month readmission Chinese elderly patients. Patients Methods: The present was a retrospective cohort study. A total 1816 were extracted from database. generalized linear model used explore additive...

We design an expected polynomial time, truthful in expectation, (1-1/e)-approximation mechanism for welfare maximization a fundamental class of combinatorial auctions. Our results apply to bidders with valuations that are matroid rank sums (MRS), which encompass most concrete examples submodular functions studied this context, including coverage and weighted-rank functions. approximation factor is the best possible, even known explicitly given valuations, assuming P ≠ NP. Ours first...

Most results in revenue-maximizing auction design hinge on "getting the price right" --- offering goods to bidders at a low enough encourage sale, but high garner non-trivial revenue. Getting right can be hard work, especially when seller has little or no priori information about bidders' valuations.

The problem of attributing a deep network's prediction to its \emph{input/base} features is well-studied. We introduce the notion \emph{conductance} extend attribution understanding importance \emph{hidden} units. Informally, conductance hidden unit network \emph{flow} via this unit. use understand for specific input, or over set inputs. evaluate effectiveness in multiple ways, including theoretical properties, ablation studies, and feature selection task. empirical evaluations are done...

In this paper, we first introduce a lower bound technique for the state complexity of transformations automata. Namely suggest considering class full automata in analysis, and later reducing size large alphabet via substitutions. Then apply such to complementation nondeterministic \omega-automata, obtain several results. Particularly, prove an \omega((0.76n)^n) B\"uchi complementation, which also holds almost every or determinization transformation omega-automata, optimal (\omega(nk))^n...

We consider profit maximizing (incentive compatible) mechanism design in general environments that include, e.g., position auctions (for selling advertisements on Internet search engines) and single-minded combinatorial auctions. analyze optimal envy-free pricings these settings, give economic justification for using the revenue of as a benchmark prior-free analysis. Moreover, we show pricing has simple nice structure strong connection to incentive compatible design, exploit this mechanisms...

For revenue and welfare maximization in single-dimensional Bayesian settings, Chawla et al. (STOC10) recently showed that sequential posted-price mechanisms (SPMs), though simple form, can perform surprisingly well compared to the optimal mechanisms. In this paper, we give a theoretical explanation of fact, based on connection notion correlation gap.

In the context of online ad serving, display ads may appear on different types web-pages, where each page includes several slots and therefore multiple can be shown page. The set that assigned to same needs satisfy various pre-specified constraints including exclusion constraints, diversity like. Upon arrival a user, serving system allocate current web-page respecting these per-page allocation constraints. Previous slot-based settings ignore important concept page, lead highly suboptimal...

Our study aimed to identify new-onset atrial fibrillation (NOAF) risk factors in acute myocardial infarction (AMI) patients after treatment with percutaneous coronary intervention (PCI) and investigate whether their nutritional status can be a predicting factor of NOAF.We analyzed 662 AMI PCI for NOAF occurrence during follow-up hospitalization divided them into an non-NOAF group. The patients' was assessed using the controlling (CONUT) score geriatric index (GNRI). Kaplan‒Meier analysis...

In the context of online ad serving, display ads may appear on different types web pages , where each page includes several slots and therefore multiple can be shown page. The set that assigned to same needs satisfy various prespecified constraints including exclusion constraints, diversity like. Upon arrival a user, serving system allocate current webpage respecting these per-page allocation constraints. Previous slot-based settings ignore important concept lead highly suboptimal results in...

This study aimed to investigate the potential relationship between baseline glucose lymphocyte ratio (GLR) levels and first episode of peritonitis in patients treated with peritoneal dialysis (PD). A total 314 PD were included divided into three groups based on GLR tertiles: tertile 1 (GLR ≤ 4.23); 2 (4.23 < 5.96), 3 > 5.96). The relationships analyzed using Kaplan–Meier curves multivariable Cox regression models. Competitive risk analysis, subgroup sensitivity analyses performed validate...