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).

Online advertising is a driving force of Internet services today. One of the main challenges in advertising systems is finding the right balance between user experience and overall revenue.

In this work, we introduce the notion of Context-Based Prediction Models. A Context-Based Prediction Model determines the probability of a user’s action (such as a click or a conversion) solely by relying on user and contextual features, without considering any specific features of the item itself. We have identified numerous valuable applications for this modeling approach, including training an auxiliary context-based model to estimate click probability and incorporating its prediction as a feature in CTR prediction models. Our experiments indicate that this enhancement brings significant improvements in offline and online business metrics while having minimal impact on the cost of serving. Overall, our work offers a simple and scalable, yet powerful approach for enhancing the performance of large-scale commercial recommender systems, with broad implications for the field of personalized recommendations.

Online advertising systems often provide means for users to close ads and also leave feedback. Although closing ads requires additional user engagement and usually indicates a poor user experience, ad closes are not as scarce as one might expect. Recently it was shown that penalizing ads with high closing likelihood during auctions may substantially reduce the number of ad closes while maintaining a small predefined revenue loss. In this work, we focus on email since this is the property in which most ad closes occur. Using data collected from a major email provider, we present interesting insights about the interplay between ad closes in email and email-related user actions. In particular, we explore the merits of integrating information derived from user actions in email for ad-close prediction. Thorough performance evaluation reveals that incorporating such signals significantly improves ad-close prediction quality over previously reported results.

Verizon Media native advertising (also known as Yahoo Gemini native) serves billions of ad impressions daily, reaching several hundreds of millions USD in revenue yearly. Although we strive to provide the best experience for our users, there will always be some users that dislike our ads in certain cases. To address these situations Gemini native platform provides an ad close mechanism that enables users to close ads that they dislike and also to provide a reasoning for their action. Surprisingly, users do care about their ad experience and their engagement with the ad close mechanism is quite significant. While the ad close rate (ACR) is lower than the click through rate (CTR), they are of the same order of magnitude, especially on Yahoo mail properties. Since ad close events indicate bad user experience caused mostly by poor ad quality, we would like to exploit the ad close signals to improve user experience and reduce the number of ad close events while maintaining a predefined total revenue loss.

We study array codes which are based on subspaces of a linear space over a finite field, using spreads, q-Steiner systems, and subspace transversal designs. We present several constructions of such codes which are q-analogs of some known block codes, such as the Hamming and simplex codes. We examine the locality and availability of the constructed codes. In particular, we distinguish between two types of locality and availability: node versus symbol. The resulting codes have distinct symbol/node locality/availability, allowing a more efficient repair process for a single symbol stored in a storage node of a distributed storage system, compared with the repair process for the whole node.

The constant growth of machine-generated mail, which today consists of more than 90% of non-spam mail traffic, is a major contributor toinformation overload in email, where users become overwhelmed with a flood of messages from commercial entities. A large part of this traffic is often junk mail that the user would prefer not to receive. Surprisingly, nearly 95% of this traffic is in fact solicited by the users themselves in the form of subscriptions to mailing services. These subscriptions are many times unintentional. Although unsubscription option from such services is enforced by commercial laws, it is hardly actually used by users. We perform a large scale study ofunsubscribable traffic, namely, messages that provide unsubscription option to users. We consider users behavior over such traffic in Yahoo Web mail service, and demonstrate a significant gap between users low interest in this traffic, and their lack of active behavior in decreasing its load. We conjecture that the cause of this gap is the lack of an efficient and easily accessible mechanism that would help users to unsubscribe. We validate our conjecture with an online large scale experiment, where we provide users with a novel mail feature for managing unsubscribable traffic, based on personalized recommendations. The experiment demonstrates the imminent need that exists for such a mechanism.

Ever-increasing amounts of data are created and processed in internet-scale companies such as Google, Facebook, and Amazon. The efficient storage of such copious amounts of data has thus become a fundamental and acute problem in modern computing. No single machine can possibly satisfy such immense storage demands. Therefore, distributed storage systems (DSS), which rely on tens of thousands of storage nodes, are the only viable solution. Such systems are broadly used in all modern internet-scale systems. However, the design of a DSS poses a number of crucial challenges, markedly different from single-user storage systems. Such systems must be able to reconstruct the data efficiently, to overcome failure of servers, to correct errors, etc. Lots of research was done in the last few years to answer these challenges and the research is increasing in parallel to the increasing amount of stored data. The main goal of this paper is to consider codes which have two of the most important features of distributed storage systems, namely, locality and availability. Our codes are array codes which are based on subspaces of a linear space over a finite field. We present several constructions of such codes which are q-analog to some of the known block codes. Some of these codes possess independent intellectual merit. We examine the locality and availability of the constructed codes. In particular we distinguish between two types of locality and availability, node vs. symbol, locality and availability. To our knowledge this is the first time that such a distinction is given in the literature.

Batch codes, first introduced by Ishai, Kushilevitz, Ostrovsky, and Sahai, mimic a distributed storage of a set of n data items on m servers, in such a way that any batch of k data items can be retrieved by reading at most some t symbols from each server. Combinatorial batch codes, are replication-based batch codes in which each server stores a subset of the data items. In this paper, we propose a generalization of combinatorial batch codes, called multiset combinatorial batch codes (MCBCs), in which n data items are stored in m servers, such that any multiset request of k items, where any item is requested at most r times, can be retrieved by reading at most t items from each server. The setup of this new family of codes is motivated by recent work on codes which enable high availability and parallel reads in distributed storage systems. The main problem under this paradigm is to minimize the number of items stored in the servers, given the values of n, m, k, r, t, which is denoted by N(n, k, m, t; r). We first give a necessary and sufficient condition for the existence of MCBCs. Then, we present several bounds on N(n, k, m, t; r) and constructions of MCBCs. In particular, we determine the value of N(n, k, m, 1; r) for any n ≥ ⌊k - 1/r⌋ (k-1m) - (m - k + 1)A(m, 4, k - 2), where A(m, 4, k - 2) is the maximum size of a binary constant weight code of length m, distance four and weight k - 2. We also determine the exact value of N(n, k, m, 1; r) when r ϵ {k, k - 1} or k = m.

A novel technique for construction of minimum storage regenerating (MSR) codes is presented. Based on this technique, three explicit constructions of MSR codes are given. The first two constructions provide access-optimal MSR codes, with two and three parities, respectively, which attain the sub-packetization bound for access-optimal codes. The third construction provides longer MSR codes with three parities (i.e., codes with larger number of systematic nodes). This improvement is achieved at the expense of the access-optimality and the field size. In addition to a minimum storage in a node, all three constructions allow the entire data to be recovered from a minimal number of storage nodes. That is, given storage ℓ in each node, the entire stored data can be recovered from any 2 log2ℓ for two parity nodes, and either 3 log3ℓ or 4 log3ℓ for three parities. Second, in the first two constructions, a helper node accesses the minimum number of its symbols for repair of a failed node (access-optimality). The goal of this paper is to provide a construction of such optimal codes over the smallest possible finite fields. The generator matrix of these codes is based on perfect matchings of complete graphs and hypergraphs, and on a rational canonical form of matrices. For two parities, the field size is reduced by a factor of two for access-optimal codes compared to previous constructions. For three parities, in the first construction a field size of at least 6 log3ℓ +1 (or 3 log3ℓ +1 for fields with characteristic 2) is sufficient, and in the second construction the field size is larger, yet linear in log3ℓ. Both constructions with three parities provide a significant improvement over previous works due to either decreased field size or lower subpacketization.

This paper presents a new construction of high-rate minimum storage regenerating codes. In addition to a minimum storage in a node, these codes have the following two important properties: first, given storage ℓ in each node, the entire stored data can be recovered from any 2 log2ℓ (any 3 log3ℓ) nodes for two parities (for three parities, respectively); second, a helper node accesses the minimum number of its symbols for repair of a failed systematic node (access-optimality). The goal of this paper is to provide a construction of such optimal codes over the smallest possible finite fields. The generator matrix of these codes is based on perfect matchings of complete graphs and hypergraphs, and on a rational canonical form of matrices. For two parities, the field size is reduced by a factor of two for access-optimal codes compared to previous constructions. For three parities, the field size is 6 log3ℓ+1 (or 3 log3ℓ+1 for fields with characteristic 2), where only non-explicit constructions with exponential field size (in log3ℓ) were known so far.

Fractional repetition (FR) codes is a family of codes for distributed storage systems (DSS) that allow uncoded exact repairs with minimum repair bandwidth. In this work, we consider a bound on the maximum amount of data that can be stored using an FR code. Optimal FR codes which attain this bound are presented. The constructions of these FR codes are based on families of regular graphs, such as Turán graphs and graphs with large girth; and on combinatorial designs, such as transversal designs and generalized polygons. In addition, based on a connection between FR codes and batch codes, we propose a new family of codes for DSS, called fractional repetition batch codes, which allow uncoded efficient exact repairs and load balancing which can be performed by several users in parallel.

This paper presents a construction for several families of optimal binary locally repairable codes (LRCs) with small locality (2 and 3). This construction is based on various anticodes. It provides binary LRCs which attain the Cadambe-Mazumdar bound. Moreover, most of these codes are optimal with respect to the Griesmer bound.

Fractional repetition (FR) codes is a family of codes for distributed storage systems (DSSs) that allow for uncoded exact repairs having the minimum repair bandwidth. However, in contrast to minimum bandwidth regenerating (MBR) codes, where an arbitrary set of a certain size of available nodes is used for a node repair, the repairs with FR codes are table based. This usually allows to store more data compared with MBR codes. In this paper, we consider bounds on the FR capacity, which is the maximum amount of data that can be stored using an FR code. Optimal FR codes which attain these bounds are presented. The constructions of these FR codes are based on combinatorial designs and on families of regular and biregular graphs. These constructions of FR codes for given parameters raise some interesting questions in graph theory. These questions and some of their solutions are discussed in this paper. In addition, based on a connection between FR codes and batch codes, we propose a new family of codes for DSS, namely, FR batch codes, which have the properties of batch codes and FR codes simultaneously. These are the first codes for DSS which allow for uncoded efficient exact repairs and load balancing which can be performed by several users in parallel. Other concepts related to FR codes are also discussed.

This paper presents and analyzes a novel concatenated coding scheme for enabling error resilience in two distributed storage settings: one being storage using existing regenerating codes and the second being storage using locally repairable codes. The concatenated coding scheme brings together a maximum rank distance code as an outer code and either a globally regenerating or a locally repairable code as an inner code. In addition, error resilience for combination of locally repairable codes with regenerating codes is considered. This concatenated coding system is designed to handle two different types of adversarial errors: the first type includes an adversary that can replace the content of an affected node only once; while the second type studies an adversary that is capable of polluting data an unbounded number of times. The paper establishes an upper bound on the resilience capacity for a locally repairable code. This paper also proves that the proposed concatenated coding approach attains the upper bound on the resilience capacity in the presence of the first type of adversary for both minimum storage regenerating codes and locally repairable codes. Further, this paper presents mechanisms that combine the presented concatenated coding scheme with subspace signatures to achieve error resilience for the second type of errors.

This paper provides new constructions and lower bounds for subspace codes, using Ferrers diagram rank-metric codes from matchings of the complete graph and pending blocks. We present different constructions for constant dimension codes with minimum injection distance 2 or k - 1, where k is the constant dimension. Furthermore, we present a construction of new codes from old codes for any minimum distance. Then, we construct nonconstant dimension codes from these codes. Some examples of codes obtained by these constructions are the largest known codes for the given parameters.

Node failures are inevitable in distributed storage systems (DSS). To enable efficient repair when faced with such failures, two main techniques are known: Regenerating codes, i.e., codes that minimize the total repair bandwidth; and codes with locality, which minimize the number of nodes participating in the repair process. This paper focuses on regenerating codes with locality, using pre-coding based on Gabidulin codes, and presents constructions that utilize minimum bandwidth regenerating (MBR) local codes. The constructions achieve maximum resilience (i.e., optimal minimum distance) and have maximum capacity (i.e., maximum rate). Finally, the same pre-coding mechanism can be combined with a subclass of fractional-repetition codes to enable maximum resilience and repair-by-transfer simultaneously.

This paper presents coding schemes for distributed storage systems (DSS) that are secure against eavesdroppers, while simultaneously enabling efficient node repair (regeneration). Towards this, novel upper bounds on secrecy capacity for minimum storage regenerating (MSR) codes and locally repairable codes (LRCs) are derived. The eavesdropper model considered in this paper incorporates the ability to listen in on data downloaded during ℓ2node repairs in addition to content stored on ℓ1nodes. Finally, this paper presents coding schemes, based on precoding using Gabidulin codes, that achieve the upper bounds on secrecy capacity and characterize the secrecy capacity of DSS for various settings of system parameters.

This paper presents a new explicit construction for locally repairable codes (LRCs) for distributed storage systems which possess all-symbol locality and the largest possible minimum distance, or equivalently, can tolerate the maximum number of node failures. This construction, based on maximum rank distance (MRD) Gabidulin codes, provides new optimal vector and scalar LRCs. In addition, the paper also discusses mechanisms by which codes obtained using this construction can be used to construct LRCs with efficient local repair of failed nodes by combination of LRCs with regenerating codes.

This paper provides new constructive lower bounds for constant dimension codes, using Ferrers diagram rank metric codes and pending blocks. Constructions for two families of parameters of constant dimension codes are presented. The examples of codes obtained by these constructions are the largest known constant dimension codes for the given parameters.

Lifted maximum rank distance (MRD) codes, which are constant dimension codes, are considered. It is shown that a lifted MRD code can be represented in such a way that it forms a block design known as a transversal design. A slightly different representation of this design makes it similar to a$q$-analog of a transversal design. The structure of these designs is used to obtain upper bounds on the sizes of constant dimension codes which contain a lifted MRD code. Codes that attain these bounds are constructed. These codes are the largest known constant dimension codes for the given parameters. These transversal designs can also be used to derive a new family of linear codes in the Hamming space. Bounds on the minimum distance and the dimension of such codes are given.

In the above titled paper (ibid., vol. 59, no. 2, pp. 1004-1017, Feb. 2013), the matrix in Example 7 (p. 1014) is incorrect. The correct matrix is presented here.

The lifted maximum rank distance codes are considered. It is shown that these codes form structures which are transversal designs and also have some similarity to q-analog of transversal designs. Upper bounds on the sizes of codes which contain the lifted maximum rank distance codes are derived. A new construction for codes which attain one of these upper bounds is given. Low density parity check codes are derived from the lifted maximum rank distance codes and some of these codes are quasi-cyclic codes attaining the Griesmer bound.

The Grassmannian space Gq(n, k) is the set of all k-dimensional subspaces of the vector space Fqn. Recently, codes in the Grassmannian have found an application in network coding. The main goal of this paper is to present efficient enumerative encoding and decoding techniques for the Grassmannian. These coding techniques are based on two different orders for the Grassmannian induced by different representations of k-dimensional subspaces of Fqn. One enumerative coding method is based on a Ferrers diagram representation and on an order for Gq(n, k) based on this representation. The complexity of this enumerative coding is O(k5/2(n - k)5/2) digit operations. Another order of the Grassmannian is based on a combination of an identifying vector and a reduced row echelon form representation of subspaces. The complexity of the enumerative coding, based on this order, is O(nk(n - k) log n log log n) digit operations. A combination of the two methods reduces the complexity on average by a constant factor.

Coding in the projective space has received recently a lot of attention due to its application in network coding. Reduced row echelon form of the linear subspaces and Ferrers diagram can play a key role for solving coding problems in the projective space. In this paper, we propose a method to design error-correcting codes in the projective space. We use a multilevel approach to design our codes. First, we select a constant-weight code. Each codeword defines a skeleton of a basis for a subspace in reduced row echelon form. This skeleton contains a Ferrers diagram on which we design a rank-metric code. Each such rank-metric code is lifted to a constant-dimension code. The union of these codes is our final constant-dimension code. In particular, the codes constructed recently by Koetter and Kschischang are a subset of our codes. The rank-metric codes used for this construction form a new class of rank-metric codes. We present a decoding algorithm to the constructed codes in the projective space. The efficiency of the decoding depends on the efficiency of the decoding for the constant-weight codes and the rank-metric codes. Finally, we use puncturing on our final constant-dimension codes to obtain large codes in the projective space which are not constant-dimension.