Web search engines often perform an initial candidate generation phase using a fast and simple ranking function, followed by subsequent reranking with more expensive rankers. Such simple ranking functions usually compute the score of a document as the sum of term-wise impact scores, and they include traditional baselines such as BM25 and Query Likelihood, as well as some recently proposed learned sparse models based on document expansion and learned impact scores. In this paper, we explore extremely fast and highly effective early termination techniques for such simple ranking functions. Our extensive experiments with a number of different ranking functions show that our methods achieve very fast response times on MSMarco V1 and V2 data while maintaining retrieval quality close to that of a safe and much slower baseline.

Top-k threshold estimation is the problem of estimating the score of the k-th highest ranking result of a search query. A good estimate can be used to speed up many common top-k query processing algorithms, and thus a number of researchers have recently studied the problem. Among the various approaches that have been proposed, quantile methods appear to give the best estimates overall at modest computational costs, followed by sampling-based methods in certain cases. In this paper, we make two main contributions. First, we study how to get even better estimates than the state of the art. Starting from quantile-based methods, we propose a series of enhancements that give improved estimates in terms of the commonly used mean under-prediction fraction (MUF). Second, we study the threshold estimation problem on recently proposed learned sparse index structures, showing that our methods also work well for these cases. Our best methods substantially narrow the gap between the state of the art and the ideal MUF of 1.0, at some additional cost in time and space.

Learned sparse retrieval systems aim to combine the effectiveness of contextualized language models with the scalability of conventional data structures such as inverted indexes. Nevertheless, the indexes generated by these systems exhibit significant deviations from the ones that use traditional retrieval models, leading to a discrepancy in the performance of existing query optimizations that were specifically developed for traditional structures. These disparities arise from structural variations in query and document statistics, including sub-word tokenization, leading to longer queries, smaller vocabularies, and different score distributions within posting lists. This paper introduces Block-Max Pruning (BMP), an innovative dynamic pruning strategy tailored for indexes arising in learned sparse retrieval environments. BMP employs a block filtering mechanism to divide the document space into small, consecutive document ranges, which are then aggregated and sorted on the fly, and fully processed only as necessary, guided by a defined safe early termination criterion or based on approximate retrieval requirements. Through rigorous experimentation, we show that BMP substantially outperforms existing dynamic pruning strategies, offering unparalleled efficiency in safe retrieval contexts and improved trade-offs between precision and efficiency in approximate retrieval tasks.

Neural information retrieval architectures based on transformers such as BERT are able to significantly improve system effectiveness over traditional sparse models such as BM25. Though highly effective, these neural approaches are very expensive to run, making them difficult to deploy under strict latency constraints. To address this limitation, recent studies have proposed new families of learned sparse models that try to match the effectiveness of learned dense models, while leveraging the traditional inverted index data structure for efficiency.

While current search engines use highly complex ranking functions with hundreds of features, they often perform an initial candidate generation step that uses a very simple ranking function to identify a limited set of promising candidates. A common approach is to use a disjunctive top-k query for this step. There are many methods for disjunctive top-k computation, but they tend to be slow for the required values of k, which are in the hundreds to thousands. We propose a new approach to safe disjunctive top-k computation that, somewhat counterintuitively, uses precomputed conjunctions of inverted lists to speed up disjunctive queries. The approach is based on a generalization of the well-known MaxScore algorithm, and utilizes recent improvements in threshold estimation techniques as well as new ideas to obtain significant improvements in performance. Our algorithms are implemented as an extension of the PISA framework for search-engine query processing, and available as open-source to support replication and follow-up work.

There has recently been significant interest in applications that require computations on massive graph structures, including scenarios where the graph is too large to be processed on a single machine. In this case, the graph needs to be partitioned into subgraphs that can be assigned to individual machines, in a process called graph or social network sharding. Given the sizes of the graphs involved, it is necessary or at least highly desirable that the partitioning itself can also be performed in a distributed manner, instead of running a sequential partitioning algorithm on a single node.We study such distributed algorithms for graph sharding, where the goal is to create subgraphs of roughly equal size that minimize the number of edges crossing subgraph boundaries. In particular, we focus on two well-known approaches that can be efficiently i mplemented i n M apReduce a nd r elated distributed computing paradigms: the Balanced Label Propagation algorithm of Ugander and Backstrom, and the method of Duong et al. based on the Bayesian Stochastic Block Modeling approach of Hofman and Wiggins. Our contributions are as follows: (1) We perform the first direct experimental comparison of the two approaches, which were independently proposed and published. (2) We propose and evaluate several enhancements of Balanced Label Propagation that result in improved graph shardings. (3) We propose and evaluate hybrid methods that perform label propagation both on individual nodes, as suggested by Ugander and Backstrom, and on stochastic blocks inferred using the approach of Duong et al.

Neural information retrieval systems typically use a cascading pipeline, in which a first-stage model retrieves a candidate set of documents and one or more subsequent stages re-rank this set using contextualized language models such as BERT. In this paper, we propose DeepImpact, a new document term-weighting scheme suitable for efficient retrieval using a standard inverted index. Compared to existing methods, DeepImpact improves impact-score modeling and tackles the vocabulary-mismatch problem. In particular, DeepImpact leverages DocT5Query to enrich the document collection and, using a contextualized language model, directly estimates the semantic importance of tokens in a document, producing a single-value representation for each token in each document. Our experiments show that DeepImpact significantly outperforms prior first-stage retrieval approaches by up to 17% on effectiveness metrics w.r.t. DocT5Query, and, when deployed in a re-ranking scenario, can reach the same effectiveness of state-of-the-art approaches with up to 5.1x speedup in efficiency.

A lot of research has focused on the efficiency of search engine query processing, and in particular on disjunctive top-k queries that return the highest scoring k results that contain at least one of the query terms. Disjunctive top-k queries over simple ranking functions are commonly used to retrieve an initial set of candidate results that are then reranked by more complex, often machine-learned rankers. Many optimized top-k algorithms have been proposed, including MaxScore, WAND, BMW, and JASS. While the fastest methods achieve impressive results on top-10 and top-100 queries, they tend to become much slower for the larger k commonly used for candidate generation. In this paper, we focus on disjunctive top-k queries for larger k. We propose new algorithms that achieve much faster query processing for values of k up to thousands or tens of thousands. Our algorithms build on top of the live-block filtering approach of Dimopoulos et al, and exploit the SIMD capabilities of modern CPUs. We also perform a detailed experimental comparison of our methods with the fastest known approaches, and release a full model implementation of our methods and of the underlying live-block mechanism, which will allows others to design and experiment with additional methods under the live-block approach.

Feature engineering is a fundamental but poorly documented component in Learning-to-Rank (LTR) search engines. Such features are commonly used to construct learning models for web and product search engines, recommender systems, and question-answering tasks. In each of these domains, there is a growing interest in the creation of open-access test collections that promote reproducible research. However, there are still few open-source software packages capable of extracting high-quality machine learning features from large text collections. Instead, most feature-based LTR research relies on "canned" test collections, which often do not expose critical details about the underlying collection or implementation details of the extracted features. Both of these are crucial to collection creation and deployment of a search engine into production. So in this regard, the experiments are rarely reproducible with new features or collections, or helpful for companies wishing to deploy LTR systems.

In the top-k threshold estimation problem, given a query q, the goal is to estimate the score of the result at rank k. A good estimate of this score can result in significant performance improvements for several query processing scenarios, including selective search, index tiering, and widely used disjunctive query processing algorithms such as MaxScore, WAND, and BMW. Several approaches have been proposed, including parametric approaches, methods using random sampling, and a recent approach based on machine learning. However, previous work fails to perform any experimental comparison between these approaches. In this paper, we address this issue by reimplementing four major approaches and comparing them in terms of estimation error, running time, likelihood of an overestimate, and end-to-end performance when applied to common classes of disjunctive top-k query processing algorithms.

Instance retrieval systems are widely used in applications such as robot navigation, medical diagnosis, and augmented reality. Blippar is a company that creates compelling augmented reality experiences or provides you with the tools to build your own. In this paper we focus on one of the company's augmented-reality applications, with which users are able to point their phone cameras at different objects in order to receive information about the objects in real time. In this paper, we provide what we believe to be the first study of forward index compression techniques for such instance retrieval systems. First, we perform an analysis of real-world data from a large-scale commercial instance retrieval system, run by Blippar focusing on augmented reality. Then we propose an entropy-based lossless compression strategy. Experiments show that our proposed Huffman-based approach outperforms a variety of other compression techniques, while also increasing overall system efficiency slightly.

An inverted index is the basic data structure used in most current large-scale information retrieval systems. It can be modeled as a collection of sorted sequences of integers. Many compression techniques for inverted indexes have been studied in the past, with some of them reaching tremendous decompression speeds through the use of SIMD instructions available on modern CPUs. While there has been some work on query processing algorithms for Graphics Processing Units (GPUs), little of it has focused on how to efficiently access compressed index structures, and we see some potential for significant improvements in decompression speed.

A lot of research has studied how to optimize inverted index structures in search engines through suitable reassignment of document identifiers. This approach was originally proposed to allow for better compression of the index, but subsequent work showed that it can also result in significant speed-ups for conjunctive queries and even certain types of disjunctive top-k algorithms. However, we do not have a good understanding of why this happens, and how we could directly optimize an index for query processing speed. As a result, existing techniques attempt to optimize for size, and treat speed increases as a welcome side-effect. In this paper, we take an initial but important step towards understanding and modeling speed increases due to document reordering. We define the problem of minimizing the cost of queries given an inverted index and a query distribution, relate it to work on adaptive set intersection, and show that it is fundamentally different from that of minimizing compressed index size. We then propose a heuristic algorithm for finding a document reordering that minimizes query processing costs under suitable cost models. Our experiments show significant increases in the speed of intersections over state-of-the-art reordering techniques.

Static index pruning techniques remove postings from inverted index structures in order to decrease index size and query processing cost, while minimizing the resulting loss in result quality. A number of authors have proposed pruning techniques that use basic properties of postings as well as results of past queries to decide what postings should be kept. However, many open questions remain, and our goal is to address some of them using a machine learning based approach that tries to predict the usefulness of a posting. In this paper, we explore the following questions: (1) How much does an approach that learns from a rich set of features outperform previous work that uses heuristic approaches or just a few features? (2) What is the relationship between index size and query processing speed in static index pruning? We show that an approach that prunes postings using a rich set of features including post-hits and doc-hits can significantly outperform previous approaches, and that there is a very pronounced trade-off between index size and query processing speed for static index pruning that has not been previously explored.

Many content-based image search and instance retrieval systems implement bag-of-visual-words strategies for candidate selection. Visual processing of an image results in hundreds of visual words that make up a document, and these words are used to build an inverted index. Query processing then consists of an initial candidate selection phase that queries the inverted index, followed by more complex reranking of the candidates using various image features. The initial phase typically uses disjunctive top-k query processing algorithms originally proposed for searching text collections. Our objective in this paper is to optimize the performance of disjunctive top-k computation for candidate selection in content-based instance retrieval systems. While there has been extensive previous work on optimizing this phase for textual search engines, we are unaware of any published work that studies this problem for instance retrieval, where both index and query data are quite different from the distributions commonly found and exploited in the textual case. Using data from a commercial large-scale instance retrieval system, we address this challenge in three steps. First, we analyze the quantitative properties of index structures and queries in the system, and discuss how they differ from the case of text retrieval. Second, we describe an optimized term-at-a-time retrieval strategy that significantly outperforms baseline term-at-a-time and document-at-a-time strategies, achieving up to 66% speed-up over the most efficient baseline. Finally, we show that due to the different properties of the data, several common safe and unsafe early termination techniques from the literature fail to provide any significant performance benefits.

Static Index Pruning is a performance optimization technique for search engines that attempts to identify and remove index postings that are unlikely to lead to top results for typical user queries. The goal is to obtain a much smaller inverted index that can quickly return results that are (almost) as good as those for the unpruned index. We make two contributions: First, we improve on previous results for pruned index size through a careful analysis of both document and query distribution characteristics. We derive an initial model based on unigram probabilities that obtains gains over previous work in some cases, and a bigram-based approach that achieves some additional improvements. We also devise a simple method for generating query logs in the absence of real-life queries, useful in modeling top results. Our second contribution is to explore, and compare to previously proposed approaches that perform pruning based on how often documents or postings appeared in top positions in the past.

Event-based online social networks, which are used to maintain interest-based groups and to distribute and organize offline events, have recently gained increasing popularity. In event-based social networks, some groups survive and thrive, while other groups fail. How to build successful groups and what factors make a “healthy” group are important open problems. We address the problem of modeling social group behavior and present detailed studies on group failure prediction by analyzing a large online event-based social network. We investigate both the statistical properties and the structural features of the social groups, and find that event features play an important role in distinguishing social groups with different topics and categories. We also observe that tightly knit communities have less average event participation, and both low level diversity and high level diversity in members' event participation will harm group activity participation. We then analyze the data of thousands of social groups collected from the Meetup platform with the goal of understanding what makes a group fail. We use two different feature selection methods in this paper and build a model to predict which groups will fail over a period of time. The experimental results show that social group failures can be predicted with high accuracy, and that member features contribute significantly to the success of social groups.

Inverted index files are commonly used to support keyword search in document collections. While the offline construction of an index can be done efficiently, its incremental update remains a hard problem, especially when the index does not completely fit in memory. We propose a novel approach for maintaining up-to-date index files on a system that constantly serves document updates and user queries. Unlike previous updating policies, we use knowledge of both the update term distribution and the query term distribution to partition the terms into functional groups. We implement two schemes for selective enforcement of contiguous layout of the data on disk, while mandating that the cost of the consolidation is less than its estimated benefit. The first is the “greedy merge” inspired by the ski-rental problem as studied in the context of competitive analysis. The second is the “opportunistic prognosticator” - by making reliable predictions, the online problem becomes suitable for offline optimizations.

In this paper, we study a 3-hop approach to distance estimation that uses two intermediate landmarks, where each landmark only stores distances to vertices in its local neighborhood and to the other landmarks. We show how to suitably represent and compress the distance data stored for each landmark, for the 2-hop and 3-hop case. Overall, we find that 3-hop methods achieve modest but promising improvement in some cases, while being comparable or slightly worse than 2-hop methods in others. Furthermore, our light compression schemes improve the practical applicability of both the 2-hop and 3-hop methods.

Current search engines use very complex ranking functions based on hundreds of features. While such functions return high-quality results, they create efficiency challenges as it is too costly to fully evaluate them on all documents in the union, or even intersection, of the query terms. To address this issue, search engines use a series of cascading rankers, starting with a very simple ranking function and then applying increasingly complex and expensive ranking functions on smaller and smaller sets of candidate results. Researchers have recently started studying several problems within this framework of query processing by cascading rankers; see, e.g., [5, 13, 17, 51].

Point-to-point distance estimation in large scale graphs is a fundamental and well studied problem with applications in many areas such as Social Search. Previous work has focused on selecting an appropriate subset of vertices as landmarks, aiming to derive distance upper or lower bounds that are as tight as possible. In order to compute a distance bound between two vertices, the proposed methods apply triangle inequalities on top of the precomputed distances between each of these vertices and the landmarks, and then use the tightest one. In this work we take a fresh look at this setting and approach it as a learning problem. As features, we use structural attributes of the vertices involved as well as the bounds described above, and we learn a function that predicts the distance between a source and a destination vertex. We conduct an extensive experimental evaluation on a variety of real-world graphs and show that the average relative prediction error of our proposed methods significantly outperforms state-of-the-art landmark-based estimates. Our method is particularily efficient when the available space is very limited.

Automatic expert assignment is a common problem encountered in both industry and academia. For example, for conference program chairs and journal editors, in order to collect "good" judgments for a paper, it is necessary for them to assign the paper to the most appropriate reviewers. Choosing appropriate reviewers of course includes a number of considerations such as expertise and authority, but also diversity and avoiding conflicts. In this paper, we explore the expert retrieval problem and implement an automatic paper-reviewer recommendation system that considers aspects of expertise, authority, and diversity. In particular, a graph is first constructed on the possible reviewers and the query paper, incorporating expertise and authority information. Then a Random Walk with Restart (RWR) [1] model is employed on the graph with a sparsity constraint, incorporating diversity information. Extensive experiments on two reviewer recommendation benchmark datasets show that the proposed method obtains performance gains over state-of-the-art reviewer recommendation systems in terms of expertise, authority, diversity, and, most importantly, relevance as judged by human experts.

A large amount of research has focused on faster methods for finding top-k results in large document collections, one of the main scalability challenges for web search engines. In this paper, we propose a method for accelerating such top-k queries that builds on and generalizes methods recently proposed by several groups of researchers based on Block-Max Indexes. In particular, we describe a system that uses a new filtering mechanism, based on a combination of block maxima and bitmaps, that radically reduces the number of documents that have to be further evaluated. Our filtering mechanism exploits the SIMD processing capabilities of current microprocessors, and it is optimized through caching policies that select and store suitable filter structures based on properties of the query load. Our experimental evaluation shows that the mechanism results in very significant speed-ups for disjunctive top-k queries under several state-of-the-art algorithms, including a speed-up of more than a factor of 2 over the fastest previously known methods.

Large web search engines use significant hardware and energy resources to process hundreds of millions of queries each day, and a lot of research has focused on how to improve query processing efficiency. One general class of optimizations called early termination techniques is used in all major engines, and essentially involves computing top results without an exhaustive traversal and scoring of all potentially relevant index entries. Recent work in [9,7] proposed several early termination algorithms for disjunctive top-k query processing, based on a new augmented index structure called Block-Max Index that enables aggressive skipping in the index.

Positional ranking functions, widely used in Web search engines, improve result quality by exploiting the positions of the query terms within documents. However, it is well known that positional indexes demand large amounts of extra space, typically about three times the space of a basic nonpositional index. Textual data, on the other hand, is needed to produce text snippets. In this paper, we study time-space trade-offs for search engines with positional ranking functions and text snippet generation. We consider both index-based and non-index based alternatives for positional data. We aim to answer the question of whether one should index positional data or not. We show that there is a wide range of practical time-space trade-offs. Moreover, we show that both position and textual data can be stored using about 71% of the space used by traditional positional indexes, with a minor increase in query time. This yields considerable space savings and outperforms, both in space and time, recent alternatives from the literature. We also propose several efficient compressed text representations for snippet generation, which are able to use about half of the space of current state-of-the-art alternatives with little impact in query processing time.

Versioned document collections are collections that contain multiple versions of each document. Important examples are Web archives, Wikipedia and other wikis, or source code and documents maintained in revision control systems. Versioned document collections can become very large, due to the need to retain past versions, but there is also a lot of redundancy between versions that can be exploited. Thus, versioned document collections are usually stored using special differential (delta) compression techniques, and a number of researchers have recently studied how to exploit this redundancy to obtain more succinct full-text index structures.

Many web search services allow users to constrain text queries to a geographic location (e.g., yoga classes near Santa Monica). Important examples include local search engines such as Google Local and location-based search services for smart phones. Several research groups have studied the efficient execution of queries mixing text and geography; their approaches usually combine inverted lists with a spatial access method such as an R-tree or space-filling curve. In this paper, we take a fresh look at this problem. We feel that previous work has often focused on the spatial aspect at the expense of performance considerations in text processing, such as inverted index access, compression, and caching. We describe new and existing approaches and discuss their different perspectives. We then compare their performance in extensive experiments on large document collections. Our results indicate that a query processor that combines state-of-the-art text processing techniques with a simple coarse-grained spatial structure can outperform existing approaches by up to two orders of magnitude. In fact, even a naive approach that first uses a simple inverted index and then filters out any documents outside the query range outperforms many previous methods.

Large search engines process thousands of queries per second over billions of documents, making query processing a major performance bottleneck. An important class of optimization techniques called early termination achieves faster query processing by avoiding the scoring of documents that are unlikely to be in the top results. We study new algorithms for early termination that outperform previous methods. In particular, we focus on safe techniques for disjunctive queries, which return the same result as an exhaustive evaluation over the disjunction of the query terms. The current state-of-the-art methods for this case, the WAND algorithm by Broder et al. [11] and the approach of Strohman and Croft [30], achieve great benefits but still leave a large performance gap between disjunctive and (even non-early terminated) conjunctive queries. We propose a new set of algorithms by introducing a simple augmented inverted index structure called a block-max index. Essentially, this is a structure that stores the maximum impact score for each block of a compressed inverted list in uncompressed form, thus enabling us to skip large parts of the lists. We show how to integrate this structure into the WAND approach, leading to considerable performance gains. We then describe extensions to a layered index organization, and to indexes with reassigned document IDs, that achieve additional gains that narrow the gap between disjunctive and conjunctive top-k query processing.

Versioned textual collections are collections that retain multiple versions of a document as it evolves over time. Important large-scale examples are Wikipedia and the web collection of the Internet Archive. Search queries over such collections often use keywords as well as temporal constraints, most commonly a time range of interest. In this paper, we study how to support such temporal range queries over versioned text. Our goal is to process these queries faster than the corresponding keyword-only queries, by exploiting the additional constraint. A simple approach might partition the index into different time ranges, and then access only the relevant parts. However, specialized inverted index compression techniques are crucial for large versioned collections, and a naive partitioning can negatively affect index size and query throughput. We show how to achieve high query throughput by using smart index partitioning techniques that take index compression into account. Experiments on over 85 million versions of Wikipedia articles show that queries can be executed in a few milliseconds on memory-based index structures, and only slightly more time on disk-based structures. We also show how to efficiently support the recently proposed stable top-k search primitive on top of our schemes.

Large web search engines are now processing billions of queries per day. Most of these queries are interactive in nature, requiring a response in fractions of a second. However, there are also a number of important scenarios where large batches of queries are submitted for various web mining and system optimization tasks that do not require an immediate response. Given the significant cost of executing search queries over billions of web pages, it is a natural question to ask if such batches of queries can be more efficiently executed than interactive queries.

Current Information Retrieval systems use inverted index structures for efficient query processing. Due to the extremely large size of many data sets, these index structures are usually kept in compressed form, and many techniques for optimizing compressed size and query processing speed have been proposed. In this paper, we focus on versioned document collections, that is, collections where each document is modified over time, resulting in multiple versions of the document. Consecutive versions of the same document are often similar, and several researchers have explored ideas for exploiting this similarity to decrease index size.

There has been a large amount of research on early termination techniques in web search and information retrieval. Such techniques return the top-k documents without scanning and evaluating the full inverted lists of the query terms. Thus, they can greatly improve query processing efficiency. However, only a limited amount of efficient top-k processing work considers the impact of term proximity, i.e., the distance between term occurrences in a document, which has recently been integrated into a number of retrieval models to improve effectiveness.

Web search engines depend on the full-text inverted index data structure. Because the query processing performance is so dependent on the size of the inverted index, a plethora of research has focused on fast end effective techniques for compressing this structure. Recently, several authors have proposed techniques for improving index compression by optimizing the assignment of document identifiers to the documents in the collection, leading to significant reduction in overall index size.

We study the problem of creating highly compressed full-text index structures for versioned document collections, that is, collections that contain multiple versions of each document. Important examples of such collections are Wikipedia or the web page archive maintained by the Internet Archive. A straightforward indexing approach would simply treat each document version as a separate document, such that index size scales linearly with the number of versions. However, several authors have recently studied approaches that exploit the significant similarities between different versions of the same document to obtain much smaller index sizes. In this paper, we propose new techniques for organizing and compressing inverted index structures for such collections. We also perform a detailed experimental comparison of new techniques and the existing techniques in the literature. Our results on an archive of the English version of Wikipedia, and on a subset of the Internet Archive collection, show significant benefits over previous approaches.

Implicit user feedback, including click-through and subsequent browsing behavior, is crucial for evaluating and improving the quality of results returned by search engines. Several recent studies [1, 2, 3, 13, 25] have used post-result browsing behavior including the sites visited, the number of clicks, and the dwell time on site in order to improve the ranking of search results. In this paper, we first study user behavior on sponsored search results (i.e., the advertisements displayed by search engines next to the organic results), and compare this behavior to that of organic results. Second, to exploit post-result user behavior for better ranking of sponsored results, we focus on identifying patterns in user behavior and predict expected on-site actions in future instances. In particular, we show how post-result behavior depends on various properties of the queries, advertisement, sites, and users, and build a classifier using properties such as these to predict certain aspects of the user behavior. Additionally, we develop a generative model to mimic trends in observed user activity using a mixture of pareto distributions. We conduct experiments based on billions of real navigation trails collected by a major search engine's browser toolbar.

Large search engines process thousands of queries per second on billions of pages, making query processing a major factor in their operating costs. This has led to a lot of research on how to improve query throughput, using techniques such as massive parallelism, caching, early termination, and inverted index compression. We focus on techniques for compressing term positions in web search engine indexes. Most previous work has focused on compressing docID and frequency data, or position information in other types of text collections. Compression of term positions in web pages is complicated by the fact that term occurrences tend to cluster within documents but not across document boundaries, making it harder to exploit clustering effects. Also, typical access patterns for position data are different from those for docID and frequency data. We perform a detailed study of a number of existing and new techniques for compressing position data in web indexes. We also study how to efficiently access position data for ranking functions that take proximity features into account.

There has been considerable past work on efficiently computing top k objects by aggregating information from multiple ranked lists of these objects. An important instance of this problem is query processing in search engines: One has to combine information from several different posting lists (rankings) of web pages (objects) to obtain the top k web pages to answer user queries. Two particularly well-studied approaches to achieve efficiency in top-k aggregation include early-termination algorithms (e.g., TA and NRA) and preaggregation of some of the input lists. However, there has been little work on a rigorous treatment of combining these approaches.

Web search engines are facing formidable performance challenges due to data sizes and query loads. The major engines have to process tens of thousands of queries per second over tens of billions of documents. To deal with this heavy workload, such engines employ massively parallel systems consisting of thousands of machines. The significant cost of operating these systems has motivated a lot of recent research into more efficient query processing mechanisms. We investigate a new way to build such high performance IR systems using graphical processing units (GPUs). GPUs were originally designed to accelerate computer graphics applications through massive on-chip parallelism. Recently a number of researchers have studied how to use GPUs for other problem domains such as databases and scientific computing. Our contribution here is to design a basic system architecture for GPU-based high-performance IR, to develop suitable algorithms for subtasks such as inverted list compression, list intersection, and top-$k$ scoring, and to show how to achieve highly efficient query processing on GPU-based systems. Our experimental results for a prototype GPU-based system on $25.2$ million web pages indicate that significant gains in query processing performance can be obtained.

Query processing is a major cost factor in operating large web search engines. In this paper, we study query result caching, one of the main techniques used to optimize query processing performance. Our first contribution is a study of result caching as a weighted caching problem. Most previous work has focused on optimizing cache hit ratios, but given that processing costs of queries can vary very significantly we argue that total cost savings also need to be considered. We describe and evaluate several algorithms for weighted result caching, and study the impact of Zipf-based query distributions on result caching. Our second and main contribution is a new set of feature-based cache eviction policies that achieve significant improvements over all previous methods, substantially narrowing the existing performance gap to the theoretically optimal (clairvoyant) method. Finally, using the same approach, we also obtain performance gains for the related problem of inverted list caching.

Web search engines use highly optimized compression schemes to decrease inverted index size and improve query throughput, and many index compression techniques have been studied in the literature. One approach taken by several recent studies first performs a renumbering of the document IDs in the collection that groups similar documents together, and then applies standard compression techniques. It is known that this can significantly improve index compression compared to a random document ordering. We study index compression and query processing techniques for such reordered indexes. Previous work has focused on determining the best possible ordering of documents. In contrast, we assume that such an ordering is already given, and focus on how to optimize compression methods and query processing for this case. We perform an extensive study of compression techniques for document IDs and present new optimizations of existing techniques which can achieve significant improvement in both compression and decompression performances. We also propose and evaluate techniques for compressing frequency values for this case. Finally, we study the effect of this approach on query processing performance. Our experiments show very significant improvements in index size and query processing speed on the TREC GOV2 collection of 25.2 million web pages.

Web search engines are facing formidable performance challenges as they need to process thousands of queries per second over billions of documents. To deal with this heavy workload, current engines use massively parallel architectures of thousands of machines that require large hardware investments.

Due to the rapid growth in the size of the web, web search engines are facing enormous performance challenges. The larger engines in particular have to be able to process tens of thousands of queries per second on tens of billions of documents, making query throughput a critical issue. To satisfy this heavy workload, search engines use a variety of performance optimizations including index compression, caching, and early termination.

Current web search engines focus on searching only themost recentsnapshot of the web. In some cases, however, it would be desirableto search over collections that include many different crawls andversions of each page. One important example of such a collectionis the Internet Archive, though there are many others. Sincethe data size of such an archive is multiple times that of a singlesnapshot, this presents us with significant performance challenges.Current engines use various techniques for index compression andoptimized query execution, but these techniques do not exploit thesignificant similarities between different versions of a page, or betweendifferent pages.In this paper, we propose a general framework for indexing andquery processing of archival collections and, more generally, anycollections with a sufficient amount of redundancy. Our approachresults in significant reductions in index size and query processingcosts on such collections, and it is orthogonal to and can be combinedwith the existing techniques. It also supports highly efficientupdates, both locally and over a network. Within this framework,we describe and evaluate different implementations that trade offindex size versus CPU cost and other factors, and discuss applicationsranging from archival web search to local search of web sites,email archives, or file systems. We present experimental resultsbased on search engine query log and a large collection consistingof multiple crawls.

Geographic web search engines allow users to constrain and order search results in an intuitive manner by focusing a query on a particular geographic region. Geographic search technology, also called local search, has recently received significant interest from major search engine companies. Academic research in this area has focused primarily on techniques for extracting geographic knowledge from the web. In this paper, we study the problem of efficient query processing in scalable geographic search engines. Query processing is a major bottleneck in standard web search engines, and the main reason for the thousands of machines used by the major engines. Geographic search engine query processing is different in that it requires a combination of text and spatial data processing techniques. We propose several algorithms for efficient query processing in geographic search engines, integrate them into an existing web search query processor, and evaluate them on large sets of real data and query traces.

Current web search engines are retrospective in that they limit users to searches against already existing pages. Prospective search engines, on the other hand, allow users to upload queries that will be applied to newly discovered pages in the future. We study and compare algorithms for efficiently matching large numbers of simple keyword queries against a stream of newly discovered pages.

While much of the data on the web is unstructured in nature, there is also a significant amount of embedded structured data, such as product information on e-commerce sites or stock data on financial sites. A large amount of research has focused on the problem of generating wrappers, i.e., software tools that allow easy and robust extraction of structured data from text and HTML sources. In many applications, such as comparison shopping, data has to be extracted from many different sources, making manual coding of a wrapper for each source impractical. On the other hand, fully automatic approaches are often not reliable enough, resulting in low quality of the extracted data.We describe a complete system for semi-automatic wrapper generation that can be trained on different data sources in a simple interactive manner. Our goal is to minimize the amount of user effort for training reliable wrappers through design of a suitable training interface that is implemented based on a powerful underlying extraction language and a set of training and ranking algorithms. Our experiments show that our system achieves reliable extraction with a very small amount of user effort.

While overall bandwidth in the internet has grown rapidly over the last few years, and an increasing number of clients enjoy broadband connectivity, many others still access the internet over much slower dialup or wireless links. To address this issue, a number of techniques for optimized delivery of web and multimedia content over slow links have been proposed, including protocol optimizations, caching, compression, and multimedia transcoding, and several large ISPs have recently begun to widely promote dialup acceleration services based on such techniques. A recent paper by Rhea, Liang, and Brewer proposed an elegant technique called value-based caching that caches substrings of files, rather than entire files, and thus avoids repeated transmission of substrings common to several pages or page versions. We propose and study a hierarchical substring caching technique that provides significant savings over this basic approach. We describe several additional techniques for minimizing overheads and perform an evaluation on a large set of real web access traces that we collected. In the second part of our work, we compare our approach to a widely studied alternative approach based on delta compression, and show how to integrate the two for best overall performance. The studied techniques are typically employed in a clientproxy environment, with each proxy serving a large number of clients, and an important aspect is how to conserve resources on the proxy while exploiting the significant memory and CPU power available on current clients.

Large web search engines have to answer thousands of queries per second with interactive response times. Due to the sizes of the data sets involved, often in the range of multiple terabytes, a single query may require the processing of hundreds of megabytes or more of index data. To keep up with this immense workload, large search engines employ clusters of hundreds or thousands of machines, and a number of techniques such as caching, index compression, and index and query pruning are used to improve scalability. In particular, two-level caching techniques cache results of repeated identical queries at the frontend, while index data for frequently used query terms are cached in each node at a lower level.We propose and evaluate a three-level caching scheme that adds an intermediate level of caching for additional performance gains. This intermediate level attempts to exploit frequently occurring pairs of terms by caching intersections or projections of the corresponding inverted lists. We propose and study several offline and online algorithms for the resulting weighted caching problem, which turns out to be surprisingly rich in structure. Our experimental evaluation based on a large web crawl and real search engine query log shows significant performance gains for the best schemes, both in isolation and in combination with the other caching levels. We also observe that a careful selection of cache admission and eviction policies is crucial for best overall performance.

The Google search engine uses a method called PageRank, together with term-based and other ranking techniques, to order search results returned to the user. PageRank uses link analysis to assign a global importance score to each web page. The PageRank scores of all the pages are usually determined off-line in a large-scale computation on the entire hyperlink graph of the web, and several recent studies have focused on improving the efficiency of this computation, which may require multiple hours on a workstation.

We study the problem of maintaining large replicated collections of files or documents in a distributed environment with limited bandwidth. This problem arises in a number of important applications, such as synchronization of data between accounts or devices, content distribution and Web caching networks, Web site mirroring, storage networks, and large scale Web search and mining. At the core of the problem lies the following challenge, called the file synchronization problem: given two versions of a file on different machines, say an outdated and a current one, how can we update the outdated version with minimum communication cost, by exploiting the significant similarity between the versions? While a popular open source tool for this problem called rsync is used in hundreds of thousands of installations, there have been only very few attempts to improve upon this tool in practice. We propose a framework for remote file synchronization and describe several new techniques that result in significant bandwidth savings. Our focus is on applications where very large collections have to be maintained over slow connections. We show that a prototype implementation of our framework and techniques achieves significant improvements over rsync. As an example application, we focus on the efficient synchronization of very large Web page collections for the purpose of search, mining, and content distribution.

Over the last few years, most major search engines have integrated link-based ranking techniques in order to provide more accurate search results. One widely known approach is the Pagerank technique, which forms the basis of the Google ranking scheme, and which assigns a global importance measure to each page based on the importance of other pages pointing to it. The main advantage of the Pagerank measure is that it is independent of the query posed by a user; this means that it can be precomputed and then used to optimize the layout of the inverted index structure accordingly. However, computing the Pagerank measure requires implementing an iterative process on a massive graph corresponding to billions of web pages and hyperlinks.In this paper, we study I/O-efficient techniques to perform this iterative computation. We derive two algorithms for Pagerank based on techniques proposed for out-of-core graph algorithms, and compare them to two existing algorithms proposed by Haveliwala. We also consider the implementation of a recently proposed topic-sensitive version of Pagerank. Our experimental results show that for very large data sets, significant improvements over previous results can be achieved on machines with moderate amounts of memory. On the other hand, at most minor improvements are possible on data sets that are only moderately larger than memory, which is the case in many practical scenarios.

Broad Web search engines as well as many more specialized search tools rely on Web crawlers to acquire large collections of pages for indexing and analysis. Such a Web crawler may interact with millions of hosts over a period of weeks or months, and thus issues of robustness, flexibility, and manageability are of major importance. In addition, I/O performance, network resources, and OS limits must be taken into account in order to achieve high performance at a reasonable cost. In this paper, we describe the design and implementation of a distributed Web crawler that runs on a network of workstations. The crawler scales to (at least) several hundred pages per second, is resilient against system crashes and other events, and can be adapted to various crawling applications. We present the software architecture of the system, discuss the, performance bottlenecks, and describe efficient techniques for achieving high performance. We also report preliminary experimental results based on a crawl of 120 million pages on 5 million hosts.

Delta compression techniques are commonly used to succinctly represent an updated version of a file with respect to an earlier one. We study the use of delta compression in a somewhat different scenario, where we wish to compress a large collection of (more or less) related files by performing a sequence of pairwise delta compressions. The problem of finding an optimal delta encoding for a collection of files by taking pairwise deltas can be reduced to the problem of computing a branching of maximum weight in a weighted directed graph, but this solution is inefficient and thus does not scale to larger file collections. This motivates us to propose a framework for cluster-based delta compression that uses text clustering techniques to prune the graph of possible pairwise delta encodings. To demonstrate the efficacy of our approach, we present experimental results on collections of Web pages. Our experiments show that cluster-based delta compression of collections provides significant improvements in compression ratio as compared to individually compressing each file or using tar+gzip, at a moderate cost in efficiency.

A large amount of research has recently focused on the graph structure (or link structure) of the World Wide Web. This structure has proven to be extremely useful for improving the performance of search engines and other tools for navigating the Web. However, since the graphs in these scenarios involve hundreds of millions of nodes and even more edges, highly space-efficient data structures are needed to fit the data in memory. A first step in this direction was done by the DEC connectivity server, which stores the graph in compressed form. We describe techniques for compressing the graph structure of the Web, and give experimental results of a prototype implementation. We attempt to exploit a variety of different sources of compressibility of these graphs and of the associated set of URLs in order to obtain good compression performance on a large Web graph.