We present a Multi-Task Learning (MTL) approach for improving predictions for rare (e.g., <1%) conversion events in online advertising. The conversions are classified into "rare" or "frequent" types based on historical statistics. The model learns shared representations across all signals while specializing through separate task towers for each type. The approach was tested and fully deployed to production, demonstrating consistent improvements in both offline (0.69% AUC lift) and online KPI performance metric (2% Cost per Action reduction).

Platforms design the form of presentation by which sellers are shown to the buyers. This design not only shapes the buyers' experience but also leads to different market equilibria or dynamics. One component in this design is through the platform's mediation of the search frictions experienced by the buyers for different sellers. We take a model of monopolistic competition and show that, on one hand, when all sellers have the same inspection costs, the market sees no stable price since the sellers always have incentives to undercut each other, and, on the other hand, the platform may stabilize the price by giving prominence to one seller chosen by a carefully designed mechanism. This calls to mind Amazon's Buy Box. We study natural mechanisms for choosing the prominent seller, characterize the range of equilibrium prices implementable by them, and find that in certain scenarios the buyers' surplus improves as the search friction increases.

The endowment effect , coined by Nobel Laureate Richard Thaler, posits that people tend to inflate the value of items they own. Recently, Babaioff, Dobzinski and Oren [EC'18] introduced the notion of endowed valuations --- valuations that capture the endowment effect --- and studied the stability and efficiency of combinatorial markets with endowed valuations. They showed that under a specific formulation of the endowment effect, an endowed equilibrium --- market equilibrium with respect to endowed valuations --- is guaranteed to exist in markets with submodular valuations, but fails to exist under XOS valuations. We harness the endowment effect further by introducing a general framework that captures a wide range of different formulations of the endowment effect. The different formulations are (partially) ranked from weak to strong, based on a stability-preserving order. We then provide algorithms for computing endowment equilibria with high welfare for sufficiently strong endowment effects, and non-existence results for weaker ones. Among other results, we prove the existence of endowment equilibria under XOS valuations, and show that if one can pre-pack items into irrevocable bundles then an endowment equilibrium exists for arbitrary markets.

A sequence of recent studies show that even in the simple setting of a single seller and a single buyer with additive, independent valuations over m items, the revenue-maximizing mechanism is prohibitively complex. This problem has been addressed using two main approaches: Approximation: the best of two simple mechanisms (sell each item separately, or sell all the items as one bundle) gives 1/6 of the optimal revenue [1]. Enhanced competition: running the simple VCG mechanism with additional m buyers extracts at least the optimal revenue in the original market [17]. Both approaches, however, suffer from severe drawbacks: On the one hand, losing 83% of the revenue is hardly acceptable in any application. On the other hand, attracting a linear number of new buyers may be prohibitive. We show that by combining the two approaches one can achieve the best of both worlds. Specifically, for any constant ε one can obtain a (1-ε) fraction of the optimal revenue by running simple mechanisms --- either selling each item separately or selling all items as a single bundle --- with substantially fewer additional buyers: logarithmic, constant, or even none in some cases.

Recent literature on approximately optimal revenue maximization has shown that in settings where agent valuations for items are complement free, the better of selling the items separately and bundling them together guarantees a constant fraction of the optimal revenue. However, most real-world settings involve some degree of complementarity among items. The role that complementarity plays in the trade-off of simplicity versus optimality has been an obvious missing piece of the puzzle. In “A Simple and Approximately Optimal Mechanism for a Buyer with Complements,” the authors show that the same simple selling mechanism—the better of selling separately and as a grand bundle—guarantees a $\Theta(d)$ fraction of the optimal revenue, where $d$ is a measure of the degree of complementarity. One key modeling contribution is a tractable notion of “degree of complementarity” that admits meaningful results and insights—they demonstrate that previous definitions fall short in this regard.

A seminal result of Bulow and Klemperer [1989] demonstrates the power of competition for extracting revenue: when selling a single item to n bidders whose values are drawn i.i.d. from a regular distribution, the simple welfare-maximizing VCG mechanism (in this case, a second price-auction) with one additional bidder extracts at least as much revenue in expectation as the optimal mechanism. The beauty of this theorem stems from the fact that VCG is a prior-independent mechanism, where the seller possesses no information about the distribution, and yet, by recruiting one additional bidder it performs better than any prior-dependent mechanism tailored exactly to the distribution at hand (without the additional bidder).

We study the efficiency of simple auctions in the presence of complements. Devanur et al. [2015] introduced the single-bid auction, and showed that it has a price of anarchy (PoA) of O(log m) for complement-free (i.e., subadditive) valuations. Prior to our work, no non-trivial upper bound on the PoA of single bid auctions was known for valuations exhibiting complements. We introduce a hierarchy over valuations, where levels of the hierarchy correspond to the degree of complementarity, and the PoA of the single bid auction degrades gracefully with the level of the hierarchy. This hierarchy is a refinement of the Maximum over Positive Hypergraphs (MPH) hierarchy [Feige et al. 2015], where the degree of complementarity d is captured by the maximum number of neighbours of a node in the positive hypergraph representation.We show that the price of anarchy of the single bid auction for valuations of level d of the hierarchy is O(d2 log(m/d)), where m is the number of items. We also establish an improved upper bound of O(d log m) for a subclass where every hyperedge in the positive hypergraph representation is of size at most 2 (but the degree is still d). Finally, we show that randomizing between the single bid auction and the grand bundle auction has a price of anarchy of at most O(√m) for general valuations. All of our results are derived via the smoothness framework, thus extend to coarse-correlated equilibria and to Bayes Nash equilibria.

In post-silicon functional validation, one of the most complex and time-consuming processes is the localization of an instruction that exposes a bug detected at system level. The task is particularly difficult due to the silicon's limited observability and the long time between a failure's occurrence and its detection. We propose a novel method that automates the architectural localization of post-silicon test-case failures. Our proposed tool analyzes a failing test-case, while leveraging the information derived from executing the test on an Instruction Set software Simulator (ISS), to identify a set of instructions that could lead to the faulty final state. The proposed failure localization process comprises the creation of a resource dependency graph based on the execution of the test-case on the ISS, determining a program slice of instructions that influence the faulty resources, and the reduction of the set of suspicious instructions by leveraging the knowledge of the correct resources. We evaluate our proposed solution through extensive experiments. Experimental results show that, in over 97% of all cases, our method was able to narrow down the list of suspicious instructions to under 2 instructions, on average, out of over 200. In over 59% of all cases, our method correctly reduced a test-case to a single faulty instruction.