Text encoders based on C-DSSM or transformers have demonstrated strong performance in many Natural Language Processing (NLP) tasks. Low latency variants of these models have also been developed in recent years in order to apply them in the field of sponsored search which has strict computational constraints. However these models are not the panacea to solve all the Natural Language Understanding (NLU) challenges as the pure semantic information in the data is not sufficient to fully identify the user intents. We propose the TextGNN model that naturally extends the strong twin tower structured encoders with the complementary graph information from user historical behaviors, which serves as a natural guide to help us better understand the intents and hence generate better language representations. The model inherits all the benefits of twin tower models such as C-DSSM and TwinBERT so that it can still be used in the low latency environment while achieving a significant performance gain than the strong encoder-only counterpart baseline models in both offline evaluations and online production system. In offline experiments, the model achieves a 0.14% overall increase in ROC-AUC with a 1% increased accuracy for long-tail low-frequency Ads, and in the online A/B testing, the model shows a 2.03% increase in Revenue Per Mille with a 2.32% decrease in Ad defect rate.

eCommerce Information Retrieval has received little attention in the academic literature, yet it is an essential component of some of the largest web sites (such as eBay, Amazon, Airbnb, Alibaba, Taobao, Target, Facebook, and others). SIGIR has for several years seen sponsorship from these kinds of organisations, who clearly value the importance of research into Information Retrieval. The purpose of this workshop is to bring together researchers and practitioners of eCommerce IR to discuss topics unique to it, to set a research agenda, and to examine how to build datasets for research into this fascinating topic. eCommerce IR is ripe for research and has a unique set of problems. For example, in eCommerce search there may be no hypertext links between documents (products); there is a click stream, but more importantly, there is often a buy stream. eCommerce problems are wide in scope and range from user interaction modalities (the kinds of search seen in when buying are different from those of web-page search (i.e. it is not clear how shopping and buying relate to the standard web-search interaction models)) through to dynamic updates of a rapidly changing collection on auction sites, and the experienceness of some products (such as Airbnb bookings). This workshop is a follow up to the "SIGIR 2017 workshop on eCommerce (ECOM17)", which was organized at SIGIR 2017, Tokyo. In the 2018 workshop, in addition to a data challenge, we will be following up on multiple aspects that were discussed in the 2017 workshop.

Push notification is a key component for E-commerce mobile applications, which has been extensively used for user growth and engagement. The effectiveness of the push notification is generally measured by message open rate. A push message can contain a recommended product, a shopping news and etc., but often only one or two items can be shown in the push message due to the limit of display space. This paper proposes a mixture model approach for predicting push message open rate for a post-purchase complementary product recommendation task. The mixture model is trained to learn latent prediction contexts, which are determined by user and item profiles, and then make open rate predictions accordingly. The item with the highest predicted open rate is then chosen to be included in the push notification message for each user. The parameters of the mixture model are optimized using an EM algorithm. A set of experiments are conducted to evaluate the proposed method live with a popular E-Commerce mobile app. The results show that the proposed method is superior than several existing solutions by a significant margin.

Many systems have been developed for machine learning at scale. Performance has steadily improved, but there has been relatively little work on explicitly defining or approaching the limits of performance. In this paper we describe the application of roofline design, an approach borrowed from computer architecture, to large-scale machine learning. In roofline design, one exposes ALU, memory, and network limits, and the constraints they imply for algorithms. Using roofline design, we have developed a system called BIDMach which has demonstrated the highest performance to date for many ML problems. On one GPU-accelerated node, it generally outperforms other single-machine toolkits and cluster toolkits running on 100s of nodes. This performance level is enabled by a relatively small number of rooflined matrix primitives. Such performance implies a dramatic reduction in the energy used to perform these calculations. Beyond matrix kernels, roofline design can be applied to the end-to-end design of machine learning algorithms which minimize memory usage to optimize speed. This approach offers a further 2x to 3x gain in performance. Roofline design can also be applied to network primitives. We describe recent work on a sparse allreduce primitive called Kylix. We have shown that Kylix approaches the practical network throughput limit for allreduce, a basic primitive for distributed machine learning. Using Kylix, we describe an efficient transformation from model-parallel to data-parallel calculations. This transformation uses a secondary storage roofline, with similar parameters to the network. Finally, we describe several deployments of these techniques on real-world problems in two large internet companies. Once again, single node rooflined design demonstrated substantial gains over alternatives on either single nodes or clusters.

Gibbs sampling is a workhorse for Bayesian inference but has several limitations when used for parameter estimation, and is often much slower than non-sampling inference methods. SAME (State Augmentation for Marginal Estimation) [15, 8] is an approach to MAP parameter estimation which gives improved parameter estimates over direct Gibbs sampling. SAME can be viewed as cooling the posterior parameter distribution and allows annealed search for the MAP parameters, often yielding very high quality estimates. But it does so at the expense of additional samples per iteration and generally slower performance. On the other hand, SAME dramatically increases the parallelism in the sampling schedule, and is an excellent match for modern (SIMD) hardware. In this paper we explore the application of SAME to graphical model inference on modern hardware. We show that combining SAME with factored sample representation (or approximation) gives throughput competitive with the fastest symbolic methods, but with potentially better quality. We describe experiments on Latent Dirichlet Allocation, achieving speeds similar to the fastest reported methods (online Variational Bayes) and lower cross-validated loss than other LDA implementations. The method is simple to implement and should be applicable to many other models.

Search advertising shows trends of vertical extension. Vertical ads, including product ads and local search ads, are proliferating at an ever increasing pace. They typically offer better ROI to advertisers as a result of better user engagement. However, campaigns and bids in vertical ads are not set at the keyword level. As a result, the matching between user query and ads suffers low recall rate and the match quality is heavily impacted by tail queries. In this paper, we propose an ad retrieval framework for retail vertical ads, based on query rewrite using personal history data to improve ad recall rate. To insure ad quality, we also present a relevance model for matching rewritten queries with user search intent, with a particular focus on tail queries. In addition, we designed and implemented a GPU-based system to accelerate the training of the relevance model to meet production performance constraints. Finally, we carry out extensive experiments on large-scale logs collected from Bing, and show significant gains in ad retrieval rate without compromising ad quality.

Allreduce is a basic building block for parallel computing. Our target here is "Big Data" processing on commodity clusters (mostly sparse power-law data). Allreduce can be used to synchronize models, to maintain distributed datasets, and to perform operations on distributed data such as sparse matrix multiply. We first review a key constraint on cluster communication, the minimum efficient packet size, which hampers the use of direct all-to-all protocols on large networks. Our allreduce network is a nested, heterogeneous-degree butterfly. We show that communication volume in lower layers is typically much less than the top layer, and total communication across all layers a small constant larger than the top layer, which is close to optimal. A chart of network communication volume across layers has a characteristic "Kylix" shape, which gives the method its name. For optimum performance, the butterfly degrees also decrease down the layers. Furthermore, to efficiently route sparse updates to the nodes that need them, the network must be nested. While the approach is amenable to various kinds of sparse data, almost all "Big Data" sets show power-law statistics, and from the properties of these, we derive methods for optimal network design. Finally, we present experiments showing with Kylix on Amazon EC2 and demonstrating significant improvements over existing systems such as PowerGraph and Hadoop.

This paper describes the BID Data Suite, a collection of hardware, software and design patterns that enable fast, large-scale data mining at very low cost. By co-designing all of these elements we achieve single-machine performance levels that equal or exceed reported cluster implementations for common benchmark problems. A key design criterion is rapid exploration of models, hence the system is interactive and primarily single-user. The elements of the suite are: (i) the data engine, a hardware design pattern that balances storage, CPU and GPU acceleration for typical data mining workloads, (ii) BIDMat, an interactive matrix library that integrates CPU and GPU acceleration and novel computational kernels (iii), BIDMach, a machine learning system that includes very efficient model optimizers, (iv) Butterfly mixing, a communication strategy that hides the latency of frequent model updates needed by fast optimizers and (v) Design patterns to improve performance of iterative update algorithms. We present several benchmark problems to show how the above elements combine to yield multiple orders-of-magnitude improvements for each problem.

Incremental model-update strategies are widely used in machine learning and data mining.By "incremental update" we refer to models that are updated many times using small subsets of the training data.Two wellknown examples are stochastic gradient and MCMC.Both provide fast sequential performance and have generated many of the best-performing methods for particular problems (logistic regression, SVM, LDA etc.).But these methods are difficult to adapt to parallel or cluster settings because of the overhead of distributing model updates through the network.Updates can be locally batched to reduce communication overhead, but convergence typically suffers as the batch size increases.In this paper we introduce and analyze butterfly mixing, an approach which interleaves communication with computation.We evaluate butterfly mixing on stochastic gradient algorithms for logistic regression and SVM, on two datasets.Results show that butterfly mix steps are fast and failure-tolerant, and overall we achieved a 3.3x speedup over full mix (AllReduce) on an Amazon EC2 cluster.

A Vehicular Sensor Network (VSN) may be used for urban environment surveillance utilizing vehicle- based sensors to provide an affordable yet good coverage for the urban area. The sensors in VSN enjoy the vehicle's steady power supply and strong computational capacity not available in traditional Wireless Sensor Network (WSN). However, the mobility of the vehicles results in highly dynamic and unpredictable network topology, leading to packet losses and distorted surveillance results. To resolve these problems, we propose a cooperative data sensing and compression approach with zero inter-sensor collaboration overhead based on sparse random projections. The algorithm provides excellent reconstruction accuracy for the sensed field, and by taking advantage of the spatial correlation of the data, enjoys much smaller communication traffic load compared to traditional sampling algorithms in wireless sensor networks. Real urban environment data sets are used in the experiments to test the reconstruction accuracy and energy efficiency under different vehicular mobility models. The results show that our approach is superior to the conventional sampling and interpolation strategy which propagates data in an uncompressed form, with 4-5dB gain in reconstruction quality and 21-55% savings in communication cost for the same sampling times.