Variances in ad impression outcomes across demographic groups are increasingly considered to be potentially indicative of algorithmic bias in personalized ads systems. While there are many definitions of fairness that could be applicable in the context of personalized systems, we present a framework which we call the Variance Reduction System (VRS) for achieving more equitable outcomes in Meta's ads systems. VRS seeks to achieve a distribution of impressions with respect to selected protected class (PC) attributes that more closely aligns the demographics of an ad's eligible audience (a function of advertiser targeting criteria) with the audience who sees that ad, in a privacy-preserving manner. We first define metrics to quantify fairness gaps in terms of ad impression variances with respect to PC attributes including gender and estimated race. We then present the VRS for re-ranking ads in an impression variance-aware manner. We evaluate VRS via extensive simulations over different parameter choices and study the effect of the VRS on the chosen fairness metric. We finally present online A/B testing results from applying VRS to Meta's ads systems, concluding with a discussion of future work. We have deployed the VRS to all users in the US for housing ads, resulting in significant improvement in our fairness metric. VRS is the first large-scale deployed framework for pursuing fairness for multiple PC attributes in online advertising.

Even though Shapley value provides an effective explanation for a DNN model prediction, the computation relies on the enumeration of all possible input feature coalitions, which leads to the exponentially growing complexity. To address this problem, we propose a novel method SHEAR to significantly accelerate the Shapley explanation for DNN models, where only a few coalitions of input features are involved in the computation. The selection of the feature coalitions follows our proposed Shapley chain rule to minimize the absolute error from the ground-truth Shapley values, such that the computation can be both efficient and accurate. To demonstrate the effectiveness, we comprehensively evaluate SHEAR across multiple metrics including the absolute error from the ground-truth Shapley value, the faithfulness of the explanations, and running speed. The experimental results indicate SHEAR consistently outperforms state-of-the-art baseline methods across different evaluation metrics, which demonstrates its potentials in real-world applications where the computational resource is limited.

Understanding the semantic meaning of content on the web through the lens of a taxonomy has many practical advantages. However, when building large-scale content classification systems, practitioners are faced with unique challenges involving finding the best ways to leverage the scale and variety of data available on internet platforms. We present learnings from our efforts in building a content classification system for multiple document types at Facebook using Multi-modal Transformers. We empirically demonstrate the effectiveness of multi-lingual, multi-modal and cross-document type learning. We describe effective strategies for exploiting weakly supervised signals as a pre-training step and show that they lead to significant gains in downstream classification accuracy. We also discuss label collection schemes that help minimize the amount of noise in collected data.

Motivated by an application in kidney exchange, we study the following query-commit problem: we are given the set of vertices of a non-bipartite graph G. The set of edges in this graph are not known ahead of time. We can query any pair of vertices to determine if they are adjacent. If the queried edge exists, we are committed to match the two endpoints. Our objective is to maximize the size of the matching. This restriction in the amount of information available to the algorithm constraints us to implement myopic, greedy-like algorithms. A simple deterministic greedy algorithm achieves a factor 1/2 which is tight for deterministic algorithms. An important open question in this direction is to give a randomized greedy algorithm that has a significantly better approximation factor. This question was first asked almost 20 years ago by Dyer and Frieze [9] where they showed that a natural randomized strategy of picking edges uniformly at random doesn't help and has an approximation factor of 1/2 + o(1). They left it as an open question to devise a better randomized greedy algorithm. In subsequent work, Aronson, Dyer, Frieze, and Suen [2] gave a different randomized greedy algorithm and showed that it attains a factor 0.5 + o where o is 0.0000025. In this paper we propose and analyze a new randomized greedy algorithm for finding a large matching in a general graph and use it to solve the query commit problem mentioned above. We show that our algorithm attains a factor of at least 0.56, a significant improvement over 0.50000025. We also show that no randomized algorithm can have an approximation factor better than 0.7916 for the query commit problem. For another large and interesting class of randomized algorithms that we call vertex-iterative algorithms, we show that no vertex iterative algorithm can have an approximation factor better than 0.75.

We consider the online bipartite matching problem in the unknown distribution input model. We show that the Ranking algorithm of [KVV90] achieves a competitive ratio of at least 0.653. This is the first analysis to show an algorithm which breaks the natural 1 - 1/e -barrier' in the unknown distribution model (our analysis in fact works in the stricter, random order model) and answers an open question in [GM08]. We also describe a family of graphs on which Ranking does no better than 0.727 in the random order model. Finally, we show that for graphs which have k > 1 disjoint perfect matchings, Ranking achieves a competitive ratio of at least 1 - √(1/k - 1/k2 + 1/n) -- in particular Ranking achieves a factor of 1 - o(1) for graphs with ω(1) disjoint perfect matchings.

Motivated by economic thought, a recent research agenda has suggested the algorithmic study of combinatorial optimization problems under functions which satisfy the property of decreasing marginal cost. A natural first step to model such functions is to consider submodular functions. However, many fundamental problems have turned out to be extremely hard to approximate under general submodular functions, thus indicating the need for a systematic study of subclasses of submodular functions that are practically motivated and yield good approximation ratios. In this paper, we introduce and study an important subclass of submodular functions, which we call discounted price functions. These functions are succinctly representable and generalize linear(additive) price functions. We study the following fundamental combinatorial optimization problems: edge cover, spanning tree, perfect matching and $s-t$ path. We give both upper and lower bound for the approximability of these problems.

Applications in complex systems such as the Internet have spawned recent interest in studying situations involving multiple agents with their individual cost or utility functions. In this paper, we introduce an algorithmic framework for studying combinatorial problems in the presence of multiple agents with submodular cost functions. We study several fundamental covering problems (Vertex Cover, Shortest Path, Perfect Matching, and Spanning Tree) in this setting and establish tight upper and lower bounds for the approximability of these problems.

We propose REWIRED (register write inhibition by resource dedication), a technique for reducing power during high level synthesis (HLS) by selectively inhibiting the storage of function unit (FU) output data into registers. Registers are generally inferred in HLS when data produced in one clock cycle is used in a later cycle. However, when it can be established that the input registers to an FU are not changing values during a certain period, the outputs during this period can be directly read off the FU output pins without needing to store them in registers. When the life-times of such data are short, it may be possible to completely eliminate the register storage operation, thereby reducing power. We present a genetic algorithm formulation and a heuristic for maximizing the number of register stores that can be inhibited in a scheduled data flow graph (DFG) during behavioral synthesis.