Historically, information retrieval systems have all followed the same paradigm: information seekers frame their needs in the form of a short query, the system selects a small set of relevant results from a corpus of available documents, rank-orders the results by decreasing relevance, possibly excerpts a responsive passage for each result, and returns a list of references and excerpts to the user. Retrieval systems typically did not attempt fusing information from multiple documents into an answer and displaying that answer directly. This was largely due to available technology: at the core of each retrieval system is an index that maps lexical tokens or semantic embeddings to document identifiers. Indices are designed for retrieving responsive documents; they do not support integrating these documents into a holistic answer.

No abstract available.

Knowledge distillation is a popular technique to transfer knowledge from a large teacher model to a small student model. Typically, the student learns to imitate the teacher by minimizing the KL divergence of its output distribution with the teacher's output distribution. In this work, we argue that such a learning objective is sub-optimal because there exists a discrepancy between the teacher's output distribution and the ground truth label distribution. Therefore, forcing the student to blindly imitate the unreliable teacher output distribution leads to inferior performance. To this end, we propose a novel knowledge distillation objective PTLoss by first representing the vanilla KL-based distillation loss function via a Maclaurin series and then perturbing the leading-order terms in this series. This perturbed loss implicitly transforms the original teacher into a proxy teacher with a distribution closer to the ground truth distribution. We establish the theoretical connection between this "distribution closeness'' and the student model generalizability, which enables us to select the PTLoss's perturbation coefficients in a principled way. Extensive experiments on six public benchmark datasets demonstrate the effectiveness of PTLoss with teachers of different scales.

As Learning-to-Rank (LTR) approaches primarily seek to improve ranking quality, their output scores are not scale-calibrated by design. This fundamentally limits LTR usage in score-sensitive applications. Though a simple multi-objective approach that combines a regression and a ranking objective can effectively learn scale-calibrated scores, we argue that the two objectives are not necessarily compatible, which makes the trade-off less ideal for either of them. In this paper, we propose a practical regression compatible ranking (RCR) approach that achieves a better trade-off, where the two ranking and regression components are proved to be mutually aligned. Although the same idea applies to ranking with both binary and graded relevance, we mainly focus on binary labels in this paper. We evaluate the proposed approach on several public LTR benchmarks and show that it consistently achieves either best or competitive result in terms of both regression and ranking metrics, and significantly improves the Pareto frontiers in the context of multi-objective optimization. Furthermore, we evaluated the proposed approach on YouTube Search and found that it not only improved the ranking quality of the production pCTR model, but also brought gains to the click prediction accuracy. The proposed approach has been successfully deployed in the YouTube production system.

Bag-of-words based lexical retrieval systems are still the most commonly used methods for real-world search applications. Recently deep learning methods have shown promising results to improve this retrieval performance but are expensive to run in an online fashion, non-trivial to integrate into existing production systems, and might not generalize well in out-of-domain retrieval scenarios. Instead, we build on top of lexical retrievers by proposing a Term Weighting BERT (TW-BERT) model. TW-BERT learns to predict the weight for individual n-gram (e.g., uni-grams and bi-grams) query input terms. These inferred weights and terms can be used directly by a retrieval system to perform a query search. To optimize these term weights, TW-BERT incorporates the scoring function used by the search engine, such as BM25, to score query-document pairs. Given sample query-document pairs we can compute a ranking loss over these matching scores, optimizing the learned query term weights in an end-to-end fashion. Aligning TW-BERT with search engine scorers minimizes the changes needed to integrate it into existing production applications, whereas existing deep learning based search methods would require further infrastructure optimization and hardware requirements. The learned weights can be easily utilized by standard lexical retrievers and by other retrieval techniques such as query expansion. We show that TW-BERT improves retrieval performance over strong term weighting baselines within MSMARCO and in out-of-domain retrieval on TREC datasets.

Unbiased learning to rank (ULTR) studies the problem of mitigating various biases from implicit user feedback data such as clicks, and has been receiving considerable attention recently. A popular ULTR approach for real-world applications uses a two-tower architecture, where click modeling is factorized into a relevance tower with regular input features, and a bias tower with bias-relevant inputs such as the position of a document. A successful factorization will allow the relevance tower to be exempt from biases. In this work, we identify a critical issue that existing ULTR methods ignored - the bias tower can be confounded with the relevance tower via the underlying true relevance. In particular, the positions were determined by the logging policy, i.e., the previous production model, which would possess relevance information. We give both theoretical analysis and empirical results to show the negative effects on relevance tower due to such a correlation. We then propose two methods to mitigate the negative confounding effects by better disentangling relevance and bias. Offline empirical results on both controlled public datasets and a large-scale industry dataset show the effectiveness of the proposed approaches. We conduct a live experiment on a popular web store for four weeks, and find a significant improvement in user clicks over the baseline, which ignores the negative confounding effect.

Historically, information retrieval systems have all followed the same paradigm: information seekers frame their needs in the form of a short query, the system selects a small set of relevant results from a corpus of available documents, rank-orders the results by decreasing relevance, possibly excerpts a responsive passage for each result, and returns a list of references and excerpts to the user. Retrieval systems typically did not attempt fusing information from multiple documents into an answer and displaying that answer directly. This was largely due to available technology: at the core of each retrieval system is an index that maps lexical tokens or semantic embeddings to document identifiers. Indices are designed for retrieving responsive documents; they do not support integrating these documents into a holistic answer.

Automatic headline generation enables users to comprehend ongoing news events promptly and has recently become an important task in web mining and natural language processing. With the growing need for news headline generation, we argue that the hallucination issue, namely the generated headlines being not supported by the original news stories, is a critical challenge for the deployment of this feature in web-scale systems Meanwhile, due to the infrequency of hallucination cases and the requirement of careful reading for raters to reach the correct consensus, it is difficult to acquire a large dataset for training a model to detect such hallucinations through human curation. In this work, we present a new framework named ExHalder to address this challenge for headline hallucination detection. ExHalder adapts the knowledge from public natural language inference datasets into the news domain and learns to generate natural language sentences to explain the hallucination detection results. To evaluate the model performance, we carefully collect a dataset with more than six thousand labeled ⟨ article, headline⟩ pairs. Extensive experiments on this dataset and another six public ones demonstrate that ExHalder can identify hallucinated headlines accurately and justifies its predictions with human-readable natural language explanations.

Rax is a library for composable Learning-to-Rank (LTR) written entirely in JAX. The goal of Rax is to facilitate easy prototyping of LTR systems by leveraging the flexibility and simplicity of JAX. Rax provides a diverse set of popular ranking metrics and losses that integrate well with the rest of the JAX ecosystem. Furthermore, Rax implements a system of ranking-specific function transformations which allows fine-grained customization of ranking losses and metrics. Most notably Rax provides approx_t12n: a function transformation (t12n) that can transform any of our ranking metrics into an approximate and differentiable form that can be optimized. This provides a systematic way to directly optimize neural ranking models for ranking metrics that are not easily optimizable in other libraries. We empirically demonstrate the effectiveness of Rax by benchmarking neural models implemented using Flax and trained using Rax on two popular LTR benchmarks: WEB30K and Istella. Furthermore, we show that integrating ranking losses with T5, a large language model, can improve overall ranking performance on the MS MARCO passage ranking task. We are sharing the Rax library with the open source community as part of the larger JAX ecosystem at https://github.com/google/rax.

Learning-to-Rank (LTR) systems are ubiquitous in web applications nowadays. The existing literature mainly focuses on improving ranking performance by trying to generate the optimal order of candidate items. However, virtually all advanced ranking functions are not scale calibrated. For example, rankers have the freedom to add a constant to all item scores without changing their relative order. This property has resulted in several limitations in deploying advanced ranking methods in practice. On the one hand, it limits the use of effective ranking functions in important applications. For example, in ads ranking, predicted Click-Through Rate (pCTR) is used for ranking and is required to be calibrated for the downstream ads auction. This is a major reason that existing ads ranking methods use scale calibrated pointwise loss functions that may sacrifice ranking performance. On the other hand, popular ranking losses are translation-invariant. We rigorously show that, both theoretically and empirically, this property leads to training instability that may cause severe practical issues.

Top-K metrics such as [email protected] are frequently used to evaluate ranking performance. The traditional tree-based models such as LambdaMART, which are based on Gradient Boosted Decision Trees (GBDT), are designed to optimize [email protected] using the LambdaRank losses. Recently, there is a good amount of research interest on neural ranking models for learning-to-rank tasks. These models are fundamentally different from the decision tree models and behave differently with respect to different loss functions. For example, the most popular ranking losses used in neural models are the Softmax loss and the GumbelApproxNDCG loss. These losses do not connect to top-K metrics such as [email protected] naturally. It remains a question on how to effectively optimize [email protected] for neural ranking models. In this paper, we follow the LambdaLoss framework and design novel and theoretically sound losses for [email protected] metrics, while the original LambdaLoss paper can only do so using an unsound heuristic. We study the new losses on the LETOR benchmark datasets and show that the new losses work better than other losses for neural ranking models.

Two-tower architecture is commonly used in real-world systems for Unbiased Learning to Rank (ULTR), where a Deep Neural Network (DNN) tower models unbiased relevance predictions, while another tower models observation biases inherent in the training data like user clicks. This two-tower architecture introduces inductive biases to allow more efficient use of limited observational logs and better generalization during deployment than single-tower architecture that may learn spurious correlations between relevance predictions and biases. However, despite their popularity, it is largely neglected in the literature that existing two-tower models assume that the joint distribution of relevance prediction and observation probabilities are completely factorizable. In this work, we revisit two-tower models for ULTR. We rigorously show that the factorization assumption can be too strong for real-world user behaviors, and existing methods may easily fail under slightly milder assumptions. We then propose several novel ideas that consider a wider spectrum of user behaviors while still under the two-tower framework to maintain simplicity and generalizability. Our concerns of existing two-tower models and the effectiveness of our proposed methods are validated on both controlled synthetic and large-scale real-world datasets.

Email has been an essential communication medium for many years. As a result, the information accumulated in our mailboxes has become valuable for all of our personal and professional activities. For years, researchers have developed interfaces, models, and algorithms to facilitate email search, discovery, and organization. This tutorial brings together these diverse research directions and provides both a historical background, as well as a high-level overview of the recent advances in the field. In particular, we lay out all of the components needed in the design of email search engines, including user interfaces, indexing, document and query understanding, retrieval, ranking, evaluation, and data privacy. The tutorial also goes beyond search, presenting recent work on intelligent task assistance in email and a number of interesting future directions.

introduction Share on Graph Technologies for User Modeling and Recommendation: Introduction to the Special Issue - Part 1 Authors: Xiangnan He University of Science and Technology of China, He Fei, China University of Science and Technology of China, He Fei, ChinaSearch about this author , Zhaochun Ren Shandong University, Qingdao, China Shandong University, Qingdao, ChinaSearch about this author , Emine Yilmaz University College London, United Kingdom University College London, United KingdomSearch about this author , Marc Najork Google Research, Mountain View, CA, United States Google Research, Mountain View, CA, United StatesSearch about this author , Tat-Seng Chua National University of Singapore, Republic of Singapore, Singapore National University of Singapore, Republic of Singapore, SingaporeSearch about this author Authors Info & Claims ACM Transactions on Information SystemsVolume 40Issue 2April 2022 Article No.: 21pp 1–5https://doi.org/10.1145/3477596Online:27 September 2021Publication History 0citation282DownloadsMetricsTotal Citations0Total Downloads282Last 12 Months282Last 6 weeks7 Get Citation AlertsNew Citation Alert added!This alert has been successfully added and will be sent to:You will be notified whenever a record that you have chosen has been cited.To manage your alert preferences, click on the button below.Manage my Alerts New Citation Alert!Please log in to your account Save to BinderSave to BinderCreate a New BinderNameCancelCreateExport CitationPublisher SiteGet Access

introduction Share on Introduction to the Special Section on Graph Technologies for User Modeling and Recommendation, Part 2 Authors: Xiangnan He University of Science and Technology of China, Hefei, China University of Science and Technology of China, Hefei, ChinaSearch about this author , Zhaochun Ren Shandong University, Qingdao, China Shandong University, Qingdao, ChinaSearch about this author , Emine Yilmaz Department of Computer Science, University College London, London, United Kingdom Department of Computer Science, University College London, London, United KingdomSearch about this author , Marc Najork Google Research, Mountain View, CA, United States Google Research, Mountain View, CA, United StatesSearch about this author , Tat-Seng Chua National University of Singapore, Singapore, Republic of Singapore National University of Singapore, Singapore, Republic of SingaporeSearch about this author Authors Info & Claims ACM Transactions on Information SystemsVolume 40Issue 3July 2022 Article No.: 42pp 1–5https://doi.org/10.1145/3490180Published:14 December 2021Publication History 0citation385DownloadsMetricsTotal Citations0Total Downloads385Last 12 Months385Last 6 weeks27 Get Citation AlertsNew Citation Alert added!This alert has been successfully added and will be sent to:You will be notified whenever a record that you have chosen has been cited.To manage your alert preferences, click on the button below.Manage my Alerts New Citation Alert!Please log in to your account Save to BinderSave to BinderCreate a New BinderNameCancelCreateExport CitationPublisher SiteGet Access

Privacy preservation remains a key challenge in data mining and Natural Language Understanding (NLU). Previous research shows that the input text or even text embeddings can leak private information. This concern motivates our research on effective privacy preservation approaches for pretrained Language Models (LMs). We investigate the privacy and utility implications of applying dχ-privacy, a variant of Local Differential Privacy, to BERT fine-tuning in NLU applications. More importantly, we further propose privacy-adaptive LM pretraining methods and show that our approach can boost the utility of BERT dramatically while retaining the same level of privacy protection. We also quantify the level of privacy preservation and provide guidance on privacy configuration. Our experiments and findings lay the groundwork for future explorations of privacy-preserving NLU with pretrained LMs.

Google Chrome, one of the world's most popular web browsers, features an extension framework allowing third-party developers to enhance Chrome's functionality. Chrome extensions are distributed through the Chrome Web Store (CWS), a Google-operated online marketplace. In this paper, we describe how we developed and deployed three recommender systems for discovering relevant extensions in CWS, namely non-personalized recommendations, related extension recommendations, and personalized recommendations. Unlike most existing papers that focus on novel algorithms, this paper focuses on sharing practical experiences when building large-scale recommender systems under various real-world constraints, such as privacy constraints, data sparsity and skewness issues, and product design choices (e.g., user interface). We show how these constraints make standard approaches difficult to succeed in practice. We share success stories that turn negative live metrics to positive ones, including: 1) how we use interpretable neural models to bootstrap the systems, help identifying pipeline issues, and pave the way for more advanced models; 2) a new item-item based algorithm for related recommendations that works under highly skewed data distributions; and 3) how the previous two techniques can help bootstrapping the personalized recommendations, which significantly reduces development cycles and bypasses various real-world difficulties. All the explorations in this work are verified in live traffic on millions of users. We believe that the findings in this paper can help practitioners to build better large-scale recommender systems.

The content on the web is in a constant state of flux. New entities,issues, and ideas continuously emerge, while the semantics of the existing conversation topics gradually shift. In recent years, pre-trained language models like BERT greatly improved the state-of-the-art for a large spectrum of content understanding tasks.Therefore, in this paper, we aim to study how these language models can be adapted to better handle continuously evolving web content.In our study, we first analyze the evolution of 2013 - 2019 Twitter data, and unequivocally confirm that a BERT model trained on past tweets would heavily deteriorate when directly applied to data from later years. Then, we investigate two possible sources of the deterioration: the semantic shift of existing tokens and the sub-optimal or failed understanding of new tokens. To this end, we both explore two different vocabulary composition methods, as well as propose three sampling methods which help in efficient incremental training for BERT-like models. Compared to a new model trained from scratch offline, our incremental training (a) reduces the training costs, (b) achieves better performance on evolving content, and (c)is suitable for online deployment. The superiority of our methods is validated using two downstream tasks. We demonstrate significant improvements when incrementally evolving the model from a particular base year, on the task of Country Hashtag Prediction, as well as on the OffensEval 2019 task.

The applicability of agglomerative clustering, for inferring both hierarchical and flat clustering, is limited by its scalability. Existing scalable hierarchical clustering methods sacrifice quality for speed and often lead to over-merging of clusters. In this paper, we present a scalable, agglomerative method for hierarchical clustering that does not sacrifice quality and scales to billions of data points. We perform a detailed theoretical analysis, showing that under mild separability conditions our algorithm can not only recover the optimal flat partition but also provide a two-approximation to non-parametric DP-Means objective. This introduces a novel application of hierarchical clustering as an approximation algorithm for the non-parametric clustering objective. We additionally relate our algorithm to the classic hierarchical agglomerative clustering method. We perform extensive empirical experiments in both hierarchical and flat clustering settings and show that our proposed approach achieves state-of-the-art results on publicly available clustering benchmarks. Finally, we demonstrate our method's scalability by applying it to a dataset of 30 billion queries. Human evaluation of the discovered clusters show that our method finds better quality of clusters than the current state-of-the-art.

The milestone improvements brought about by deep representation learning and pre-training techniques have led to large performance gains across downstream NLP, IR and Vision tasks. Multimodal modeling techniques aim to leverage large high-quality visio-linguistic datasets for learning complementary information across image and text modalities. In this paper, we introduce the Wikipedia-based Image Text (WIT) Dataset to better facilitate multimodal, multilingual learning. WIT is composed of a curated set of 37.5 million entity rich image-text examples with 11.5 million unique images across 108 Wikipedia languages. Its size enables WIT to be used as a pretraining dataset for multimodal models, as we show when applied to downstream tasks such as image-text retrieval. WIT has four main and unique advantages. First, WIT is the largest multimodal dataset by the number of image-text examples by 3x (at the time of writing). Second, WIT is massively multilingual (first of its kind) with coverage over 100+ languages (each of which has at least 12K examples) and provides cross-lingual texts for many images. Third, WIT represents a more diverse set of concepts and real world entities relative to what previous datasets cover. Lastly, WIT provides a very challenging real-world test set, as we empirically illustrate using an image-text retrieval task as an example. WIT Dataset is available for download and use via a Creative Commons license here: https://github.com/google-research-datasets/wit.

Cloud-based file storage platforms such as Google Drive are widely used as a means for storing, editing and sharing personal and organizational documents. In this paper, we improve search ranking quality for cloud storage platforms by utilizing user activity logs. Different from search logs, activity logs capture general document usage activity beyond search, such as opening, editing and sharing documents. We propose to automatically learn text embeddings that are effective for search ranking from activity logs. We develop a novel co-access signal, i.e., whether two documents were accessed by a user around the same time, to train deep semantic matching models that are useful for improving the search ranking quality. We confirm that activity-trained semantic matching models can improve ranking by conducting extensive offline experimentation using Google Drive search and activity logs. To the best of our knowledge, this is the first work to examine the benefits of leveraging document usage activity at large scale for cloud storage search; as such it can shed light on using such activity in scenarios where direct collection of search-specific interactions (e.g., query and click logs) may be expensive or infeasible.

Extracting structured information from templatic documents is an important problem with the potential to automate many real-world business workflows such as payment, procurement, and payroll. The core challenge is that such documents can be laid out in virtually infinitely different ways. A good solution to this problem is one that generalizes well not only to known templates such as invoices from a known vendor, but also to unseen ones. We developed a system called Glean to tackle this problem. Given a target schema for a document type and some labeled documents of that type, Glean uses machine learning to automatically extract structured information from other documents of that type. In this paper, we describe the overall architecture of Glean, and discuss three key data management challenges : 1) managing the quality of ground truth data, 2) generating training data for the machine learning model using labeled documents, and 3) building tools that help a developer rapidly build and improve a model for a given document type. Through empirical studies on a real-world dataset, we show that these data management techniques allow us to train a model that is over 5 F1 points better than the exact same model architecture without the techniques we describe. We argue that for such information-extraction problems, designing abstractions that carefully manage the training data is at least as important as choosing a good model architecture.

Many natural language processing and information retrieval problems can be formalized as the task of semantic matching. Existing work in this area has been largely focused on matching between short texts (e.g., question answering), or between a short and a long text (e.g., ad-hoc retrieval). Semantic matching between long-form documents, which has many important applications like news recommendation, related article recommendation and document clustering, is relatively less explored and needs more research effort. In recent years, self-attention based models like Transformers and BERT have achieved state-of-the-art performance in the task of text matching. These models, however, are still limited to short text like a few sentences or one paragraph due to the quadratic computational complexity of self-attention with respect to input text length. In this paper, we address the issue by proposing the Siamese Multi-depth Transformer-based Hierarchical (SMITH) Encoder for long-form document matching. Our model contains several innovations to adapt self-attention models for longer text input. We propose a transformer based hierarchical encoder to capture the document structure information. In order to better capture sentence level semantic relations within a document, we pre-train the model with a novel masked sentence block language modeling task in addition to the masked word language modeling task used by BERT. Our experimental results on several benchmark data sets for long-form document matching show that our proposed SMITH model outperforms the previous state-of-the-art models including hierarchical attention, multi-depth attention-based hierarchical recurrent neural network, and BERT. Comparing to BERT based baselines, our model is able to increase maximum input text length from 512 to 2048. We will open source a Wikipedia based benchmark data set, code and a pre-trained model to accelerate future research on long-form document matching.

Google Drive is widely used for managing personal and work-related documents in the cloud. To help users organize their documents in Google Drive, we develop a new feature to allow users to create a set of working files for ongoing easy access, called workspace. A workspace is a cluster of documents, but unlike a typical document cluster, it contains documents that are not only topically coherent, but are also useful in the ongoing user tasks.

Although neural network models enjoy tremendous advantages in handling image and text data, tree-based models still remain competitive for learning-to-rank tasks with numerical data. A major strength of tree-based ranking models is the insensitivity to different feature scales, while neural ranking models may suffer from features with varying scales or skewed distributions. Feature transformation or normalization is a simple technique which preprocesses input features to mitigate their potential adverse impact on neural models. However, due to lack of studies, it is unclear to what extent feature transformation can benefit neural ranking models. In this paper, we aim to answer this question by providing empirical evidence for learning-to-rank tasks. First, we present a list of commonly used feature transformation techniques and perform a comparative study on multiple learning-to-rank data sets. Then we propose a mixture feature transformation mechanism which can automatically derive a mixture of basic feature transformation functions to achieve the optimal performance. Our experiments show that applying feature transformation can substantially improve the performance of neural ranking models compared to directly using the raw features. In addition, the proposed mixture transformation method can further improve the performance of the ranking model without any additional human effort.

Learning to Rank, a central problem in information retrieval, is a class of machine learning algorithms that formulate ranking as an optimization task. The objective is to learn a function that produces an ordering of a set of documents in such a way that the utility of the entire ordered list is maximized. Learning-to-rank methods do so by learning a function that computes a score for each document in the set. A ranked list is then compiled by sorting documents according to their scores. While such a deterministic mapping of scores to permutations makes sense during inference where stability of ranked lists is required, we argue that its greedy nature during training leads to less robust models. This is particularly problematic when the loss function under optimization---in agreement with ranking metrics---largely penalizes incorrect rankings and does not take into account the distribution of raw scores. In this work, we present a stochastic framework where, instead of a deterministic derivation of permutations from raw scores, permutations are sampled from a distribution defined by raw scores. Our proposed sampling method is differentiable and works well with gradient descent optimizers. We analytically study our proposed method and demonstrate when and why it leads to model robustness. We also show empirically, through experiments on publicly available learning-to-rank datasets, that the application of our proposed method to a class of ranking loss functions leads to significant model quality improvements.

The rapid growth of commercial web content has driven the development of shopping search services to help users find product offers. Due to the dynamic nature of commercial content, an effective recrawl policy is a key component in a shopping search service; it ensures that users have access to the up-to-date product details. Most of the existing strategies either relied on simple heuristics, or overlooked the resource budgets. To address this, Azar et al. [5] recently proposed an optimization strategy LambdaCrawl aiming to maximize content freshness within a given resource budget. In this paper, we demonstrate that the effectiveness of LambdaCrawl is governed in large part by how well future content change rate can be estimated. By adopting the state-of-the-art deep learning models for change rate prediction, we obtain a substantial increase of content freshness over the common LambdaCrawl implementation with change rate estimated from the past history. Moreover, we demonstrate that while LambdaCrawl is a significant advancement upon existing recrawl strategies, it can be further improved upon by a unified multi-strategy recrawl policy. To this end, we adopt the K-armed adversarial bandits algorithm that can provably optimize the overall freshness by combining multiple strategies. Empirical results over a large-scale production dataset confirm its superiority to LambdaCrawl, especially under tight resource budgets.

Learning-to-Rank deals with maximizing the utility of a list of examples presented to the user, with items of higher relevance being prioritized. It has several practical applications such as large-scale search, recommender systems, document summarization and question answering. While there is widespread support for classification and regression based learning, support for learning-to-rank in deep learning has been limited. We introduce TensorFlow Ranking, the first open source library for solving large-scale ranking problems in a deep learning framework. It is highly configurable and provides easy-to-use APIs to support different scoring mechanisms, loss functions and evaluation metrics in the learning-to-rank setting. Our library is developed on top of TensorFlow and can thus fully leverage the advantages of this platform. TensorFlow Ranking has been deployed in production systems within Google; it is highly scalable, both in training and in inference, and can be used to learn ranking models over massive amounts of user activity data, which can include heterogeneous dense and sparse features. We empirically demonstrate the effectiveness of our library in learning ranking functions for large-scale search and recommendation applications in Gmail and Google Drive. We also show that ranking models built using our model scale well for distributed training, without significant impact on metrics. The proposed library is available to the open source community, with the hope that it facilitates further academic research and industrial applications in the field of learning-to-rank.

Learning-to-Rank is a branch of supervised machine learning that seeks to produce an ordering of a list of items such that the utility of the ranked list is maximized. Unlike most machine learning techniques, however, the objective cannot be directly optimized using gradient descent methods as it is either discontinuous or flat everywhere. As such, learning-to-rank methods often optimize a loss function that either is loosely related to or upper-bounds a ranking utility instead. A notable exception is the approximation framework originally proposed by Qin et al. that facilitates a more direct approach to ranking metric optimization. We revisit that framework almost a decade later in light of recent advances in neural networks and demonstrate its superiority empirically. Through this study, we hope to show that the ideas from that work are more relevant than ever and can lay the foundation of learning-to-rank research in the age of deep neural networks.

Synonym expansion is a technique that adds related words to search queries, which may lead to more relevant documents being retrieved, thus improving recall. There is extensive prior work on synonym expansion for web search, however very few studies have tackled its application for email search. Synonym expansion for private corpora like emails poses several unique research challenges. First, the emails are not shared across users, which precludes us from directly employing query-document bipartite graphs, which are standard in web search synonym expansion. Second, user search queries are of personal nature, and may not be generalizable across users. Third, the size of the underlying corpora from which the synonyms may be mined is relatively small (i.e., user's private email inbox) compared to the size of the web corpus. Therefore, in this paper, we propose a solution tailored to the challenges of synonym expansion for email search. We formulate it as a multi-view learning problem, and propose a novel embedding-based model that joins information from multiple sources to obtain the optimal synonym candidates. To demonstrate the effectiveness of the proposed technique, we evaluate our model using both explicit human ratings as well as a live experiment using the Gmail Search service, one of the world's largest email search engines.

Presentation bias is one of the key challenges when learning from implicit feedback in search engines, as it confounds the relevance signal. While it was recently shown how counterfactual learning-to-rank (LTR) approaches \citeJoachims/etal/17a can provably overcome presentation bias when observation propensities are known, it remains to show how to effectively estimate these propensities. In this paper, we propose the first method for producing consistent propensity estimates without manual relevance judgments, disruptive interventions, or restrictive relevance modeling assumptions. First, we show how to harvest a specific type of intervention data from historic feedback logs of multiple different ranking functions, and show that this data is sufficient for consistent propensity estimation in the position-based model. Second, we propose a new extremum estimator that makes effective use of this data. In an empirical evaluation, we find that the new estimator provides superior propensity estimates in two real-world systems -- Arxiv Full-text Search and Google Drive Search. Beyond these two points, we find that the method is robust to a wide range of settings in simulation studies.

Long Short-Term Memory (LSTM) is one of the most powerful sequence models for user browsing history \citetan2016improved,korpusik2016recurrent or natural language text \citemikolov2010recurrent.Despite the strong performance, it has not gained popularity for user-facing applications, mainly owing to a large number of parameters and lack of interpretability. Recently \citetzaheer2017latent introduced latent LSTM Allocation (LLA) to address these problems by incorporating topic models with LSTM, where the topic model maps observed words in each sequence to topics that evolve using an LSTM model. In our experiments, we found the resulting model, although powerful and interpretable, to show shortcomings when applied to sequence data that exhibit multi-modes of behaviors with abrupt dynamic changes. To address this problem we introduce thLLA: a threading LLA model. thLLA has the ability to break each sequence into a set of segments and then model the dynamic in each segment using an LSTM mixture. In that way, thLLA can model abrupt changes in sequence dynamics and provides a better fit for sequence data while still being interpretable and requiring fewer parameters. In addition, thLLA uncovers hidden themes in the data via its dynamic mixture components. However, such generalization and interpretability come at a cost of complex dependence structure, for which inference would be extremely non-trivial. To remedy this, we present an efficient sampler based on particle MCMC method for inference that can draw from the joint posterior directly. Experimental results confirm the superiority of thLLA and the stability of the new inference algorithm on a variety of domains.

Existing unbiased learning-to-rank models use counterfactual inference, notably Inverse Propensity Scoring (IPS), to learn a ranking function from biased click data. They handle the click incompleteness bias, but usually assume that the clicks are noise-free, i.e., a clicked document is always assumed to be relevant. In this paper, we relax this unrealistic assumption and study click noise explicitly in the unbiased learning-to-rank setting. Specifically, we model the noise as the position-dependent trust bias and propose a noise-aware Position-Based Model, named TrustPBM, to better capture user click behavior. We propose an Expectation-Maximization algorithm to estimate both examination and trust bias from click data in TrustPBM. Furthermore, we show that it is difficult to use a pure IPS method to incorporate click noise and thus propose a novel method that combines a Bayes rule application with IPS for unbiased learning-to-rank. We evaluate our proposed methods on three personal search data sets and demonstrate that our proposed model can significantly outperform the existing unbiased learning-to-rank methods.

Web crawlers spend significant resources to maintain freshness of their crawled data. This paper describes the optimization of resources to ensure that product prices shown in ads in a context of a shopping sponsored search service are synchronized with current merchant prices. We are able to use the predictability of price changes to build a machine learned system leading to considerable resource savings for both the merchants and the crawler. We describe our solution to technical challenges due to partial observability of price history, feedback loops arising from applying machine learned models, and offers in cold start state. Empirical evaluation over large-scale product crawl data demonstrates the effectiveness of our model and confirms its robustness towards unseen data. We argue that our approach can be applicable in more general data pull settings.

Semantic text matching is one of the most important research problems in many domains, including, but not limited to, information retrieval, question answering, and recommendation. Among the different types of semantic text matching, long-document-to-long-document text matching has many applications, but has rarely been studied. Most existing approaches for semantic text matching have limited success in this setting, due to their inability to capture and distill the main ideas and topics from long-form text.

Recent studies show that an overwhelming majority of emails are machine-generated and sent by businesses to consumers. Many large email services are interested in extracting structured data from such emails to enable intelligent assistants. This allows experiences like being able to answer questions such as “What is the address of my hotel in New York?” or “When does my flight leave?”. A high-quality email classifier is a critical piece in such a system. In this paper, we argue that the rich formatting used in business-to-consumer emails contains valuable information that can be used to learn better representations. Most existing methods focus only on textual content and ignore the rich HTML structure of emails. We introduce RiSER (Richly Structured Email Representation) - an approach for incorporating both the structure and content of emails. RiSER projects the email into a vector representation by jointly encoding the HTML structure and the words in the email. We then use this representation to train a classifier. To our knowledge, this is the first description of a neural technique for combining formatting information along with the content to learn improved representations for richly formatted emails. Experimenting with a large corpus of emails received by users of Gmail, we show that RiSER outperforms strong attention-based LSTM baselines. We expect that these benefits will extend to other corpora with richly formatted documents. We also demonstrate with examples where leveraging HTML structure leads to better predictions.

In emails, information abounds. Whether it be a bill reminder, a hotel confirmation, or a shipping notification, our emails contain useful bits of information that enable a number of applications. Most of this email traffic is machine-generated, sent from a business to a human. These business-to-consumer emails are typically instantiated from a set of email templates, and discovering these templates is a key step in enabling a variety of intelligent experiences. Existing email information extraction systems typically separate information extraction into two steps: an offline template discovery process (called template induction) that is periodically run on a sample of emails, and an online email annotation process that applies discovered templates to emails as they arrive. Since information extraction requires an email's template to be known, any delay in discovering a newly created template causes missed extractions, lowering the overall extraction coverage. In this paper, we present a novel system called Crusher that discovers templates completely online, reducing template discovery delay from a week (for the existing MapReduce-based batch system) to minutes. Furthermore, Crusher has a resource consumption footprint that is significantly smaller than the existing batch system. We also report on the surprising lesson we learned that conventional stream processing systems do not present a good framework on which to build Crusher. Crusher delivers an order of magnitude more throughput than a prototype built using a stream processing engine. We hope that these lessons help designers of stream processing systems accommodate a broader range of applications like online template induction in the future.

Machine generated business-to-consumer (B2C) emails such as receipts, newsletters, and promotions constitute a large portion of users' inboxes today. These emails reflect the users' interests and often are sequentially correlated, e.g., users interested in relocating may receive a sequence of messages on housing, moving, job availability, etc. We aim to infer (and eventually serve) the users' future interests by predicting the categories of their future emails. There are many useful methods, such as recurrent neural networks, that can be applied for such predictions, but in all cases the key to better performance is an effective representation of emails and users. To this end, we propose a general framework for learning embeddings for emails and users, using as input only the sequence of B2C templates users receive and open. (A template is a B2C email stripped of all transient information related to specific users.) These learned embeddings allow us to identify both sequentially correlated emails and users with similar sequential interests. We can also use the learned embeddings either as input features or embedding initializers for email category prediction tasks. Extensive experiments with millions of fully anonymized B2C emails demonstrate that the learned embeddings can significantly improve the prediction accuracy for future email categories. We hope that this effective yet simple embedding learning framework will inspire new machine intelligence applications that will improve the users' email experience.

How to optimize ranking metrics such as Normalized Discounted Cumulative Gain (NDCG) is an important but challenging problem, because ranking metrics are either flat or discontinuous everywhere, which makes them hard to be optimized directly. Among existing approaches, LambdaRank is a novel algorithm that incorporates ranking metrics into its learning procedure. Though empirically effective, it still lacks theoretical justification. For example, the underlying loss that LambdaRank optimizes for remains unknown until now. Due to this, there is no principled way to advance the LambdaRank algorithm further. In this paper, we present LambdaLoss, a probabilistic framework for ranking metric optimization. We show that LambdaRank is a special configuration with a well-defined loss in the LambdaLoss framework, and thus provide theoretical justification for it. More importantly, the LambdaLoss framework allows us to define metric-driven loss functions that have clear connection to different ranking metrics. We show a few cases in this paper and evaluate them on three publicly available data sets. Experimental results show that our metric-driven loss functions can significantly improve the state-of-the-art learning-to-rank algorithms.

Extracting structured data from emails can enable several assistive experiences, such as reminding the user when a bill payment is due, answering queries about the departure time of a booked flight, or proactively surfacing an emailed discount coupon while the user is at that store. This paper presents Juicer, a system for extracting information from email that is serving over a billion Gmail users daily. We describe how the design of the system was informed by three key principles: scaling to a planet-wide email service, isolating the complexity to provide a simple experience for the developer, and safeguarding the privacy of users (our team and the developers we support are not allowed to view any single email). We describe the design tradeoffs made in building this system, the challenges faced and the approaches used to tackle them. We present case studies of three extraction tasks implemented on this platform---bill reminders, commercial offers, and hotel reservations---to illustrate the effectiveness of the platform despite challenges unique to each task. Finally, we outline several areas of ongoing research in large-scale machine-learned information extraction from email.

Mobile devices are pervasive, which means that users have access to web content and their personal documents at all locations, not just their home or office. Existing work has studied how locations can influence information needs, focusing on web queries. We explore whether or not location information can be helpful to users who are searching their own personal documents. We wish to study whether a users' location can predict their queries over their own personal data, so we focus on the task of query suggestion. While we find that using location directly can be helpful, it does not generalize well to novel locations. To improve this situation, we explore using semantic location: that is, rather than memorizing location-query associations, we generalize our location information to names of the closest point of interest. By using short, semantic descriptions of locations, we find that we can more robustly improve query completion and observe that users are already using locations to extend their own queries in this domain. We present a simple but effective model that can use location to predict queries for a user even before they type anything into a search box, and which learns effectively even when not all queries have location information.

A well-known challenge in learning from click data is its inherent bias and most notably position bias. Traditional click models aim to extract the ‹query, document› relevance and the estimated bias is usually discarded after relevance is extracted. In contrast, the most recent work on unbiased learning-to-rank can effectively leverage the bias and thus focuses on estimating bias rather than relevance [20, 31]. Existing approaches use search result randomization over a small percentage of production traffic to estimate the position bias. This is not desired because result randomization can negatively impact users' search experience. In this paper, we compare different schemes for result randomization (i.e., RandTopN and RandPair) and show their negative effect in personal search. Then we study how to infer such bias from regular click data without relying on randomization. We propose a regression-based Expectation-Maximization (EM) algorithm that is based on a position bias click model and that can handle highly sparse clicks in personal search. We evaluate our EM algorithm and the extracted bias in the learning-to-rank setting. Our results show that it is promising to extract position bias from regular clicks without result randomization. The extracted bias can improve the learning-to-rank algorithms significantly. In addition, we compare the pointwise and pairwise learning-to-rank models. Our results show that pairwise models are more effective in leveraging the estimated bias.

A vast majority of the emails received by people today are machine-generated by businesses communicating with consumers. While some emails originate as a result of a transaction (e.g., hotel or restaurant reservation confirmations, online purchase receipts, shipping notifications, etc.), a large fraction are commercial emails promoting an offer (a special sale, free shipping, available for a limited time, etc.). The sheer number of these promotional emails makes it difficult for users to read all these emails and decide which ones are actually interesting and actionable. In this paper, we tackle the problem of extracting information from commercial emails promoting an offer to the user. This information enables an email platform to build several new experiences that can unlock the value in these emails without the user having to navigate and read all of them. For instance, we can highlight offers that are expiring soon, or display a notification when there»s an unexpired offer from a merchant if your phone recognizes that you are at that merchant»s store. A key challenge in extracting information from such commercial emails is that they are often image-rich and contain very little text. Training a machine learning (ML) model on a rendered image-rich email and applying it to each incoming email can be prohibitively expensive. In this paper, we describe a cost-effective approach for extracting signals from both the text and image content of commercial emails in the context of Gmail, an email platform that serves over a billion users around the world. The key insight is to leverage the template structure of emails, and use off-the-shelf OCR techniques to obtain the text from images to augment the existing text features offline. Compared to a text-only approach, we show that we are able to identify 9.12% more email templates corresponding to ~5% more emails being identified as offers. Interestingly, our analysis shows that this 5% improvement in coverage is across the board, irrespective of whether the emails were sent by large merchants or small local merchants, allowing us to deliver an improved experience for everyone.

Google Drive is a cloud storage and collaboration service used by hundreds of millions of users around the world. Quick Access is a new feature in Google Drive that surfaces the most relevant documents when a user visits the home screen. Our metrics show that users locate their documents in half the time with this feature compared to previous approaches. The development of Quick Access illustrates many general challenges and constraints associated with practical machine learning such as protecting user privacy, working with data services that are not designed with machine learning in mind, and evolving product definitions. We believe that the lessons learned from this experience will be useful to practitioners tackling a wide range of applied machine learning problems.

User interaction data (e.g., click data) has proven to be a powerful signal for learning-to-rank models in web search. However, such models require observing multiple interactions across many users for the same query-document pair to achieve statistically meaningful gains. Therefore, utilizing user interaction data for improving search over personal, rather than public, content is a challenging problem. First, the documents (e.g., emails or private files) are not shared across users. Second, user search queries are of personal nature (e.g., "alice's address") and may not generalize well across users. In this paper, we propose a solution to these challenges, by projecting user queries and documents into a multi-dimensional space of fine-grained and semantically coherent attributes. We then introduce a novel parameterization technique to overcome sparsity in the multi-dimensional attribute space. Attribute parameterization enables effective usage of cross-user interactions for improving personal search quality -- which is a first such published result, to the best of our knowledge. Experiments with a dataset derived from interactions of users of one of the world's largest personal search engines demonstrate the effectiveness of the proposed attribute parameterization technique.

Email is an essential communication medium for billions of people, with most users relying on web-based email services. Two recent trends are changing the email experience: smartphones have become the primary tool for accessing online services including email, and machine learning has come of age. Smartphones have a number of compelling properties (they are location-aware, usually with us, and allow us to record and share photos and videos), but they also have a few limitations, notably limited screen size and small and tedious virtual keyboards. Over the past few years, Google researchers and engineers have leveraged machine learning to ameliorate these weaknesses, and in the process created novel experiences. In this talk, I will give three examples of machine learning improving the email experience. The first example describes how we are improving email search. Displaying the most relevant results as the query is being typed is particularly useful on smartphones due to the aforementioned limitations. Combining hand-crafted and machine-learned rankers is powerful, but training learned rankers requires a relevance-labeled training set. User privacy prohibits us from employing raters to produce relevance labels. Instead, we leverage implicit feedback (namely clicks) provided by the users themselves. Using click logs as training data in a learning-to-rank setting is intriguing, since there is a vast and continuous supply of fresh training data. However, the click stream is biased towards queries that receive more clicks -- e.g. queries for which we already return the best result in the top-ranked position. I will summarize our work on neutralizing that bias.

Click-through data has proven to be a critical resource for improving search ranking quality. Though a large amount of click data can be easily collected by search engines, various biases make it difficult to fully leverage this type of data. In the past, many click models have been proposed and successfully used to estimate the relevance for individual query-document pairs in the context of web search. These click models typically require a large quantity of clicks for each individual pair and this makes them difficult to apply in systems where click data is highly sparse due to personalized corpora and information needs, e.g., personal search. In this paper, we study the problem of how to leverage sparse click data in personal search and introduce a novel selection bias problem and address it in the learning-to-rank framework. This paper proposes a few bias estimation methods, including a novel query-dependent one that captures queries with similar results and can successfully deal with sparse data. We empirically demonstrate that learning-to-rank that accounts for query-dependent selection bias yields significant improvements in search effectiveness through online experiments with one of the world's largest personal search engines.

Query rewriting algorithms can be used as a form of query expansion, by combining the user's original query with automatically generated rewrites. Rewriting algorithms bring linguistic datasets to bear without the need for iterative relevance feedback, but most studies of rewriting have used proprietary datasets such as large-scale search logs. By contrast this paper uses readily available data, particularly ClueWeb09 link text with over 1.2 billion anchor phrases, to generate rewrites. To avoid overfitting, our initial analysis is performed using Million Query Track queries, leading us to identify three algorithms which perform well. We then test the algorithms on Web and newswire data. Results show good properties in terms of robustness and early precision.

Web-based advertising and electronic commerce, combined with the key role of search engines in driving visitors to ad-monetized and e-commerce web sites, has given rise to the phenomenon of web spam: web pages that are of little value to visitors, but that are created mainly to mislead search engines into driving traffic to target web sites. A large fraction of spam web pages is automatically generated, and some portion of these pages is generated by stitching together parts (sentences or paragraphs) of other web pages. This paper presents a scalable algorithm for detecting such "quilted" web pages. Previous work by the author and his collaborators introduced a sampling-based algorithm that was capable of detecting some, but by far not all quilted web pages in a collection. By contrast, the algorithm presented in this work identifies all quilted web pages, and it is scalable to very large corpora. We tested the algorithm on the half-billion page English-language subset of the ClueWeb09 collection, and evaluated its effectiveness in detecting web spam by manually inspecting small samples of the detected quilted pages. This manual inspection guided us in iteratively refining the algorithm to be more efficient in detecting real-world spam.

Many phenomena and artifacts such as road networks, social networks and the web can be modeled as large graphs and analyzed using graph algorithms. However, given the size of the underlying graphs, efficient implementation of basic operations such as connected component analysis, approximate shortest paths, and link-based ranking (e.g. PageRank) becomes challenging.

Previous research has suggested that people who are in the same social circle exhibit similar behaviors and tastes. The rise of social networks gives us insights into the social circles of web users, and recommendation services (including search engines, advertisement engines, and collaborative filtering engines) provide a motivation to adapt recommendations to the interests of the audience. An important primitive for supporting these applications is the ability to quantify how connected two users are in a social network. The shortest-path distance between a pair of users is an obvious candidate measure. This paper introduces a new measure of "affinity" in social networks that takes into account not only the distance between two users, but also the number of edge-disjoint paths between them, i.e. the "robustness" of their connection. Our measure is based on a sketch-based approach, and affinity queries can be answered extremely efficiently (at the expense of a one-time offline sketch computation). We compare this affinity measure against the "approximate shortest-path distance", a sketch-based distance measure with similar efficiency characteristics. Our empirical study is based on a Hotmail email exchange graph combined with demographic information and Bing query history, and a Twitter mention-graph together with the text of the underlying tweets. We found that users who are close to each other - either in terms of distance or affinity - have a higher similarity in terms of demographics, queries, and tweets.

No abstract available.

We study the fundamental problem of computing distances between nodes in large graphs such as the web graph and social networks. Our objective is to be able to answer distance queries between pairs of nodes in real time. Since the standard shortest path algorithms are expensive, our approach moves the time-consuming shortest-path computation offline, and at query time only looks up precomputed values and performs simple and fast computations on these precomputed values. More specifically, during the offline phase we compute and store a small “sketch ” for each node in the graph, and at query-time we look up the sketches of the source and destination nodes and perform a simple computation using these two sketches to estimate the distance. Categories and Subject Descriptors G.2.2 [Graph Theory]: Graph algorithms, path and circuit problems

In this paper, we attempt to improve the effectiveness and the efficiency of query-dependent link-based ranking algorithms such as HITS, MAX and SALSA. All these ranking algorithms view the results of a query as nodes in the web graph, expand the result set to include neighboring nodes, and compute scores on the induced neighborhood graph. In previous work it was shown that SALSA in particular is substantially more effective than query-independent link-based ranking algorithms such as PageRank. In this work, we show that whittling down the neighborhood graph through consistent sampling of nodes and edges makes SALSA and its cousins both faster (more efficient) and better (more effective). We offer a hypothesis as to why "less is more", i.e. why using a reduced graph improves performance.

SALSA is a link-based ranking algorithm that takes the result set of a query as input, extends the set to include additional neighboring documents in the web graph, and performs a random walk on the induced subgraph. The stationary probability distribution of this random walk, used as a relevance score, is significantly more effective for ranking purposes than popular query-independent link-based ranking algorithms such as PageRank. Unfortunately, this requires significant effort at query-time, to access the link graph and compute the stationary probability distribution. In this paper, we explore whether it is possible to perform most of the computation off-line, prior to the arrival of any queries. The off-line phase of our approach computes a "score map" for each node in the web graph by performing a SALSA-like algorithm on the neighborhood of that node and retaining the scores of the most promising nodes in the neighborhood graph. The on-line phase takes the results to a query, retrieves the score map of each result, and returns for each result a score that is the sum of the matching scores from each score map. We evaluated this algorithm on a collection of about 28,000 queries with partially labeled results, and found that it is significantly more effective than PageRank, although not quite as effective as SALSA. We also studied the trade-off between ranking effectiveness and space requirements.

No abstract available.

This paper compares the effectiveness of two well-known query-dependent link-based ranking algorithms, "Hyperlink-Induced Topic Search" (HITS) and the "Stochastic Approach for Link-Structure Analysis" (SALSA). The two algorithms are evaluated on a very large web graph induced by 463 million crawled web pages and a set of 28,043 queries and 485,656 results labeled by human judges. We employed three different performance measures - mean average precision (MAP), mean reciprocal rank (MRR), and normalized discounted cumulative gain (NDCG). We found that as an isolated feature, SALSA substantially outperforms HITS. This is quite surprising, given that the two algorithms operate over the same neighborhood graph induced by the query result set. We also studied the combination of SALSA and HITS with BM25F, a state-of-the-art text-based scoring function that incorporates anchor text. We found that the combination of SALSA and BM25F outperforms the combination of HITS and BM25F. Finally, we broke down our query set by query specificity, and found that SALSA (and to a lesser extent HITS) is most effective for general queries.

This paper describes a large-scale evaluation of the effectiveness of HITS in comparison with other link-based ranking algorithms, when used in combination with a state-of-the-art text retrieval algorithm exploiting anchor text. We quantified their effectiveness using three common performance measures: the mean reciprocal rank, the mean average precision, and the normalized discounted cumulative gain measurements. The evaluation is based on two large data sets: a breadth-first search crawl of 463 million web pages containing 17.6 billion hyperlinks and referencing 2.9 billion distinct URLs; and a set of 28,043 queries sampled from a query log, each query having on average 2,383 results, about 17 of which were labeled by judges. We found that HITS outperforms PageRank, but is about as effective as web-page in-degree. The same holds true when any of the link-based features are combined with the text retrieval algorithm. Finally, we studied the relationship between query specificity and the effectiveness of selected features, and found that link-based features perform better for general queries, whereas BM25F performs better for specific queries.

In this paper, we continue our investigations of "web spam": the injection of artificially-created pages into the web in order to influence the results from search engines, to drive traffic to certain pages for fun or profit. This paper considers some previously-undescribed techniques for automatically detecting spam pages, examines the effectiveness of these techniques in isolation and when aggregated using classification algorithms. When combined, our heuristics correctly identify 2,037 (86.2%) of the 2,364 spam pages (13.8%) in our judged collection of 17,168 pages, while misidentifying 526 spam and non-spam pages (3.1%).

Two years ago, we conducted a study on the evolution of web pages over time. In the course of that study, we discovered a large number of machine-generated "spam" web pages emanating from a handful of web servers in Germany. These spam web pages were dynamically assembled by stitching together grammatically well-formed German sentences drawn from a large collection of sentences. This discovery motivated us to develop techniques for finding other instances of such "slice and dice" generation of web pages, where pages are automatically generated by stitching together phrases drawn from a limited corpus. We applied these techniques to two data sets, a set of 151 million web pages collected in December 2002 and a set of 96 million web pages collected in June 2004. We found a number of other instances of large-scale phrase-level replication within the two data sets. This paper describes the algorithms we used to discover this type of replication, and highlights the results of our data mining.

The increasing importance of search engines to commercial web sites has given rise to a phenomenon we call "web spam", that is, web pages that exist only to mislead search engines into (mis)leading users to certain web sites. Web spam is a nuisance to users as well as search engines: users have a harder time finding the information they need, and search engines have to cope with an inflated corpus, which in turn causes their cost per query to increase. Therefore, search engines have a strong incentive to weed out spam web pages from their index.We propose that some spam web pages can be identified through statistical analysis: Certain classes of spam pages, in particular those that are machine-generated, diverge in some of their properties from the properties of web pages at large. We have examined a variety of such properties, including linkage structure, page content, and page evolution, and have found that outliers in the statistical distribution of these properties are highly likely to be caused by web spam.This paper describes the properties we have examined, gives the statistical distributions we have observed, and shows which kinds of outliers are highly correlated with web spam.

Crawling the web is deceptively simple: the basic algorithm is (a) Fetch a page (b) Parse it to extract all linked URLs (c) For all the URLs not seen before, repeat (a)-(c). However, the size of the web (estimated at over 4 billion pages) and its rate of change (estimated at 7% per week) move this plan from a trivial programming exercise to a serious algorithmic and system design challenge. Indeed, these two factors alone imply that for a reasonably fresh and complete crawl of the web, step (a) must be executed about a thousand times per second, and thus the membership test (c) must be done well over ten thousand times per second against a set too large to store in main memory. This requires a distributed architecture, which further complicates the membership test.A crucial way to speed up the test is to cache, that is, to store in main memory a (dynamic) subset of the "seen" URLs. The main goal of this paper is to carefully investigate several URL caching techniques for web crawling. We consider both practical algorithms: random replacement, static cache, LRU, and CLOCK, and theoretical limits: clairvoyant caching and infinite cache. We performed about 1,800 simulations using these algorithms with various cache sizes, using actual log data extracted from a massive 33 day web crawl that issued over one billion HTTP requests.Our main conclusion is that caching is very effective - in our setup, a cache of roughly 50,000 entries can achieve a hit rate of almost 80%. Interestingly, this cache size falls at a critical point: a substantially smaller cache is much less effective while a substantially larger cache brings little additional benefit. We conjecture that such critical points are inherent to our problem and venture an explanation for this phenomenon.

How fast does the web change? Does most of the content remain unchanged once it has been authored, or are the documents continuously updated? Do pages change a little or a lot? Is the extent of change correlated to any other property of the page? All of these questions are of interest to those who mine the web, including all the popular search engines, but few studies have been performed to date to answer them.One notable exception is a study by Cho and Garcia-Molina, who crawled a set of 720,000 pages on a daily basis over four months, and counted pages as having changed if their MD5 checksum changed. They found that 40% of all web pages in their set changed within a week, and 23% of those pages that fell into the .com domain changed daily.This paper expands on Cho and Garcia-Molina's study, both in terms of coverage and in terms of sensitivity to change. We crawled a set of 150,836,209 HTML pages once every week, over a span of 11 weeks. For each page, we recorded a checksum of the page, and a feature vector of the words on the page, plus various other data such as the page length, the HTTP status code, etc. Moreover, we pseudo-randomly selected 0.1% of all of our URLs, and saved the full text of each download of the corresponding pages.After completion of the crawl, we analyzed the degree of change of each page, and investigated which factors are correlated with change intensity. We found that the average degree of change varies widely across top-level domains, and that larger pages change more often and more severely than smaller ones.This paper describes the crawl and the data transformations we performed on the logs, and presents some statistical observations on the degree of change of different classes of pages.

Article Breadth-first crawling yields high-quality pages Share on Authors: Marc Najork Compaq Systems Research Center, 130 Lytton Avenue, Palo Alto, CA Compaq Systems Research Center, 130 Lytton Avenue, Palo Alto, CAView Profile , Janet L. Wiener Compaq Systems Research Center, 130 Lytton Avenue, Palo Alto, CA Compaq Systems Research Center, 130 Lytton Avenue, Palo Alto, CAView Profile Authors Info & Claims WWW '01: Proceedings of the 10th international conference on World Wide WebMay 2001 Pages 114–118https://doi.org/10.1145/371920.371965Online:01 April 2001Publication History 187citation1,650DownloadsMetricsTotal Citations187Total Downloads1,650Last 12 Months60Last 6 weeks4 Get Citation AlertsNew Citation Alert added!This alert has been successfully added and will be sent to:You will be notified whenever a record that you have chosen has been cited.To manage your alert preferences, click on the button below.Manage my AlertsNew Citation Alert!Please log in to your account Save to BinderSave to BinderCreate a New BinderNameCancelCreateExport CitationPublisher SiteGet Access