https://www.dagstuhl.de/17181

### 01. – 05. Mai 2017, Dagstuhl-Seminar 17181

# Theory and Applications of Hashing

## Organisatoren

Martin Dietzfelbinger (TU Ilmenau, DE)

Michael Mitzenmacher (Harvard University – Cambridge, US)

Rasmus Pagh (IT University of Copenhagen, DK)

David P. Woodruff (IBM Almaden Center – San Jose, US)

## Koordinatoren

Martin Aumüller (IT University of Copenhagen, DK)

## Auskunft zu diesem Dagstuhl-Seminar erteilt

## Dokumente

Dagstuhl Report, Volume 7, Issue 5

Motivationstext

Teilnehmerliste

Gemeinsame Dokumente

## Summary

### Background

The idea of hashing was proposed in the 1950s as an efficient method for implementing symbol tables in compilers. In succeeding decades it has emerged as an algorithmic tool that goes well beyond its original purpose, providing solutions for a wide range of algorithmic problems. While the theory of hashing may appear mature, in fact, many new advances have been made in recent years. Also, the number of applications has grown to an extent that few people realize how broad the reach of hashing is, or have a comprehensive overview. The aim of this seminar was to bring together researchers with an interest in hashing methods (and more generally random mappings) from various perspectives, spanning theory and a diverse set of application areas. In this way we wanted to identify opportunities for further advancing the field.

#### Theory

The theoretical side includes aspects such as the design of hash functions, the design of algorithm and data structure primitives, and the mathematical analysis of hashing schemes. For hash function design, since Carter and Wegman proposed universal hashing there has been a fertile research agenda identifying sufficient randomness properties of hash functions for various applications, and on the other hand devising resource-efficient hash functions having these properties. While new simple and efficient hash function constructions with strong theoretical properties keep appearing (seminar participants have contributed to this), there are still many applications of hashing for which no efficient hash function construction is known. At the same time it is increasingly clear that traditional measures of hash function performance like collision probability and independence are inadequate to describe the randomness properties needed in many applications.

While hashing is interesting in its own right, it has also become a foundational building block for higher level algorithms and data structures, each of which can in turn often find uses in a variety of application spaces. Various hash-based sketches laid the groundwork for the field of streaming algorithms, by showing how approximate counting could be done effectively. More recently, hashing algorithms have provided frameworks for similarity measures for text and images, data reconciliation, and even fast sparse Fast Fourier transform algorithms. As we construct richer, more elaborate structures on top of the hashing building blocks, we stretch what we require from hashing further.

Mathematical analysis of hashing schemes, an area enriched by early work of Knuth but grown well beyond, has inspired the development of a number of combinatorial methods and results. In the 1990s it was realized that load balancing using the best of two (or more) random choices leads to a much more even distribution, often referred to as the ``power of two choices''. Another great success in this area, obtained during the last decade, has been the analysis of cuckoo hashing; several seminar participants were involved in this. On the other hand, as questions become answered, new questions arise; in the case of cuckoo hashing, the effectiveness of variants including double hashing and partial-key cuckoo hashing are not understood. Beyond the hashing schemes themselves, data structures and algorithms that make use of lower-level hashing primitives further require and benefit from mathematical analysis. The fundamental connections between hashing, sparse reconstruction algorithms, and various sketching approaches are only now starting to be realized.

#### Applications

Hashing is of course heavily used in information storage and retrieval contexts. For example, the "power of two choices" paradigm(where several seminar participants were among the pioneers) has resulted in extremely scalable and robust architectures for distributed hash tables (also known as key-value stores).

Other applications of hashing are appearing at a tremendous rate, as systems-designers and builders become accustomed to a world where approximate answers (as opposed to exact answers) are not only sufficient, they are necessary for efficiency. Indeed, hashing was one of the key methodologies for handling big data well before "big data" was even a widely used term. Since the seminal paper of Flajolet and Martin that showed how to efficiently compute an approximate count of the number of elements in a data stream, hashing has been a central tool in the design of algorithms for data streams, where the inability to store all the information that passes requires approximations. But in recent years the use of hashing has spread to many other settings where data is stored and accessible, but scalability can be achieved only by resorting to approximation algorithms similar to those developed in the setting of data streams. One early success story in this direction is the method for identifying near-duplicate web pages in Altavista using min-wise hashing. Another is HyperANF, a refined version of the Flajolet-Martin method that was used in 2012 to compute the distance distribution of the Facebook social network, making it by far the largest scale Milgram-like experiment ever performed. Finally, Bloom filters, invented around 1970 to provide a small-memory approximate representation of a set, have become a staple in systems, with countless variations expanding on its initial functionality, such as counts associated with set elements or aging out of set items as new elements are dynamically added.

In the field of machine learning, random mappings of data to lower-dimensional vectors that are easier to handle is of increasing importance for big data applications. This is true in particular since machine learning algorithms often work with kernelized feature vectors whose dimension far exceeds the size of data vectors. Designing randomized mappings that meet the criteria of machine learning applications has been an active research area in recent years. NIPS 2014 awarded one of two best paper awards a paper co-authored by seminar participant Shrivastava that presents a new asymmetric locality-sensitive hash function design for machine learning applications and shows how it leads to significant speedups.

The rich interplay between theory and practice in the field of hashing cannot be overstated. Applications drive the need for new algorithms and data structures based on hashing, as well as the need for more efficient classes of hash function families with provable theoretical guarantees. Specific implementations developed in the field often do not have theoretical guarantees on performance or accuracy, creating new theoretical problems and driving a need to make theory and practice meet.

#### Industrial relevance.

The workshop topic was highly relevant for companies dealing with big data. Three seminar participants are affiliated with Google, one has worked on hashing algorithms at AT&T for a decade, and one is affiliated with VMware. Also, one organizer was affiliated with IBM at the time of the seminar.

### Outcome of the seminar

The seminar brought together quite a few leading figures in the area of hashing, mixed with a good fraction of young researchers, some of which are behind many of the most exciting results in the area in recent years. Areas that were particularly well represented were: Analysis of multiple-choice hashing methods, hashing for high-dimensional search problems, applications of hashing in string algorithms, applications of hashing in machine learning, streaming (approximation) algorithms, high-performance hash function construction, and algorithm engineering. Many results in these areas were presented in 18 shorter talks. Four longer talks (by A. Andoni, A. McGregor, U. Wieder, and Q. Zhang) contributed background, overview over new results, and aspects of applications of hashing in industry. Open problems were discussed in an open problem session; four of them are included in this report.

The paper [1] on fillable arrays co-authored by J. Nelson was motivated by a talk given by T. Hagerup (see Section 3.7) and can thus be seen as a first direct result of the seminar.

We, the organizers, would like to thank all participants for their contributions in talks, material, and discussions, and in particular the speakers of the longer talks. Many thanks are due to the Dagstuhl staff both in the offices and in the castle for their support in making this seminar a success.

### References

- Jacob Teo Por Loong, Jelani Nelson, Huacheng Yu: Fillable arrays with constant time operations and a single bit of redundancy. CoRR abs/1709.09574 (2017)

**Summary text license**

Creative Commons BY 3.0 Unported license

Martin Dietzfelbinger, Michael Mitzenmacher, Rasmus Pagh, David P. Woodruff, and Martin Aumüller

## Classification

- Data Bases / Information Retrieval
- Data Structures / Algorithms / Complexity

## Keywords

- Hash function construction and analysis
- Hashing primitives
- Information retrieval applications
- Locality-sensitive hashing
- Data streaming applications
- Machine learning applications
- Connections to complexity theory