Embedding-based retrieval in graph-based recommendation has shown great improvements over traditional graph walk retrieval methods, and has been adopted in large-scale industry applications such as friend recommendations [16]. However, it is not without its challenges: retraining graph embeddings frequently due to changing data is slow and costly, and producing high recall of approximate nearest neighbor search (ANN) on such embeddings is challenging due to the power law distribution of the indexed users. In this work, we address theses issues by introducing a simple query expansion method in ANN, called FriendSeedSelection, where for each node query, we construct a set of 1-hop embeddings and run ANN search. We highlight our approach does not require any model-level tuning, and is inferred from the data at test-time. This design choice effectively enables our recommendation system to adapt to the changing graph distribution without frequent heavy model retraining. We also discuss how we design our system to efficiently construct such queries online to support 10k+ QPS. For friend recommendation, our method shows improvements of recall, and 11% relative friend reciprocated communication metric gains, now serving over 800 million monthly active users at Snapchat.

Friend recommendation systems in online social and professional networks such as Snapchat helps users find friends and build connections, leading to better user engagement and retention. Traditional friend recommendation systems take advantage of the principle of locality and use graph traversal to retrieve friend candidates, e.g. Friends-of-Friends (FoF). While this approach has been adopted and shown efficacy in companies with large online networks such as Linkedin and Facebook, it suffers several challenges: (i) discrete graph traversal offers limited reach in cold-start settings, (ii) it is expensive and infeasible in realtime settings beyond 1 or 2 hop requests owing to latency constraints, and (iii) it cannot well-capture the complexity of graph topology or connection strengths, forcing one to resort to other mechanisms to rank and find top-K candidates. In this paper, we proposed a new Embedding Based Retrieval (EBR) system for retrieving friend candidates, which complements the traditional FoF retrieval by retrieving candidates beyond 2-hop, and providing a natural way to rank FoF candidates. Through online A/B test, we observe statistically significant improvements in the number of friendships made with EBR as an additional retrieval source in both low- and high-density network markets. Our contributions in this work include deploying a novel retrieval system to a large-scale friend recommendation system at Snapchat, generating embeddings for billions of users using Graph Neural Networks, and building EBR infrastructure in production to support Snapchat scale.

Airbnb is an online marketplace which connects hosts and guests all over the world. Our inventory includes over 4.5 million listings, which enable the travel of over 300 million guests. The growth team at Airbnb is responsible for helping travelers find Airbnb, in part by participating in ad auctions on major search platforms such as Google and Bing. In this talk, we will describe how ad- vertising efficiently on these platforms requires solving several information retrieval and machine learning problems, including query understanding, click value estimation, and realtime pacing of our expenditure.

LinkedIn as a professional network serves the career needs of 450 Million plus members. The task of job recommendation system is to nd the suitable job among a corpus of several million jobs and serve this in real time under tight latency constraints. Job search involves nding suitable job listings given a user, query and context. Typical scoring function for both search and recommendations involves evaluating a function that matches various elds in the job description with various elds in the member pro le. This in turn translates to evaluating a function with several thousands of features to get the right ranking. In recommendations, evaluating all the jobs in the corpus for all members is not possible given the latency constraints. On the other hand, reducing the candidate set could potentially involve loss of relevant jobs. We present a way to model the underlying complex ranking function via decision trees. The branches within the decision trees are query clauses and hence the decision trees can be mapped on to real time queries. We developed an o ine framework which evaluates the quality of the decision tree with respect to latency and recall. We tested the approach on job search and recommendations on LinkedIn and A/B tests show signi cant improvements in member engagement and latency. Our techniques helped reduce job search latency by over 67% and our recommendations latency by over 55%. Our techniques show 3.5% improvement in applications from job recommendations primarily due to reduced timeouts from upstream services. As of writing the approach powers all of job search and recommendations on LinkedIn.

Online professional social networks such as LinkedIn play a key role in helping job seekers find right career opportunities and job providers reach out to potential candidates. LinkedIn's job ecosystem has been designed to serve as a marketplace for efficient matching between potential candidates and job postings, and to provide tools to connect job seekers and job providers. LinkedIn's job recommendations product is a crucial mechanism to help achieve these goals, wherein personalized sets of recommended job postings are presented for members based on the structured, context data present in their profiles.

The mission of LinkedIn is to connect the world's professionals to make them more productive and successful. LinkedIn operates the world's largest professional network on the Internet with more than 500 Million members in over 200 countries. Core to realizing the mission is to help people find jobs. In this paper, we describe how the jobs recommendations is powered by a search index and some practical challenges involved in scaling such a system.

Modern day social media search and recommender systems require complex query formulation that incorporates both user context and their explicit search queries. Users expect these systems to be fast and provide relevant results to their query and context. With millions of documents to choose from, these systems utilize a multi-pass scoring function to narrow the results and provide the most relevant ones to users. Candidate selection is required to sift through all the documents in the index and select a relevant few to be ranked by subsequent scoring functions. It becomes crucial to narrow down the document set while maintaining relevant ones in resulting set. In this tutorial we survey various candidate selection techniques and deep dive into case studies on a large scale social media platform. In the later half we provide hands-on tutorial where we explore building these candidate selection models on a real world dataset and see how to balance the tradeoff between relevance and latency.

Instant search has become a common part of the search experience in most popular search engines and social networking websites. The goal is to provide instant feedback to the user in terms of query completions (“instant suggestions”) or directly provide search results (“instant results”) as the user is typing their query. In this paper, we describe the challenges that we faced while delivering the instant search experience at LinkedIn, and present techniques that we developed to overcome them. We discuss three aspects of instant search - performance, tolerance to user errors and accuracy. On the performance side, we discuss our inverted index ordering scheme, which when combined with query rewriting and early termination techniques, helped us significantly reduce latency while maintaining good search recall. We describe our method to handle hard to spell or misspelled names using name clusters. We also present ranking methods that leverage the structured nature of queries and documents at LinkedIn. All the methods described have been fully deployed in production, and have helped us significantly improve the search experience for our members. Specifically, we increased the number of queries served via typeahead by 39.53%.

Instant search has become a common part of the search experience in most popular search engines and social networking websites. The goal is to provide instant feedback to the user in terms of query completions ("instant suggestions") or directly provide search results ("instant results") as the user is typing their query. The need for instant search has been further amplified by the proliferation of mobile devices and services like Siri and Google Now that aim to address the user's information need as quickly as possible. Examples of instant results include web queries like "weather san jose" (which directly provides the current temperature), social network queries like searching for someone's name on Facebook or LinkedIn (which directly provide the people matching the query). In each of these cases, instant search constitutes a superior user experience, as opposed to making the user complete their query before the system returns a list of results on the traditional search engine results page (SERP).

Linkedln is the largest professional network with more than 350 million members. As the member base increases, searching for experts becomes more and more challenging. In this paper, we propose an approach to address the problem of personalized expertise search on LinkedIn, particularly for exploratory search queries containing skills. In the offline phase, we introduce a collaborative filtering approach based on matrix factorization. Our approach estimates expertise scores for both the skills that members list on their profiles as well as the skills they are likely to have but do not explicitly list. In the online phase (at query time) we use expertise scores on these skills as a feature in combination with other features to rank the results. To learn the personalized ranking function, we propose a heuristic to extract training data from search logs while handling position and sample selection biases. We tested our models on two products - LinkedIn homepage and LinkedIn recruiter. A/B tests showed significant improvements in click through rates - 31% for CTR@1 for recruiter (18% for homepage) as well as downstream messages sent from search - 37% for recruiter (20% for homepage). As of writing this paper, these models serve nearly all live traffic for skills search on LinkedIn homepage as well as LinkedIn recruiter.

Clock mesh has been widely used to distribute the clock signal across the chip. Clock mesh is driven by a top-level tree and a set of mesh buffers. We present fast and efficient combinatorial algorithms to simultaneously identify the candidate locations as well as sizes of the buffers driving the clock mesh. We show that such a sizing offers a better solution than inserting buffers of uniform size across the mesh. Due to the high redundancy, a mesh architecture offers high tolerance toward variations in clock skew. However, such a redundancy comes at the expense of mesh wire length and power dissipation. Based on survivable network theory, we formulate the problem to reduce the clock mesh by retaining only those edges that are critical to maintain redundancy. Such a formulation offers designer the option to tradeoff between power and tolerance to process variations. We present efficient postprocessing techniques to reduce the size of the mesh buffers after mesh reduction. Experimental results indicate that our techniques can result inpower savings up to 28% with less than 3.3% delay penalty. We also present driver models that can help in simulating the clock mesh. Such models achieve near-HSPICE accuracy with significant speedup in run time.

Power/ground noise is a major source of VLSI circuit timing variations. This work aims to reduce clock network induced power noise by assigning different signal polarities (opposite switchings) to clock buffers in an existing buffered clock tree. Three assignment algorithms are proposed: 1) partitioning; 2) 2-coloring on minimum spanning tree; and 3) recursive min-matching. A post-processing of clock buffer sizing is performed to achieve desired clock skew. SPICE based experimental results indicate that our techniques could reduce the average peak current and average delay variations by 50% and 51%, respectively.

Nanometer VLSI systems demand robust clock distribution network design for increased process and operating condition variabilities. In this paper, we propose minimum clock distribution network augmentation for guaranteed skew yield. We present theoretical analysis results on an inserted link in a clock network, which scales down local skew and skew variation, but may not guarantee global skew and skew variation reduction in general. We propose a global minimum clock network augmentation algorithm, which inserts links simultaneously between all nearest sink pairs, apply rule-based link removal, and perform link consolidation by Steiner minimum tree construction for wirelength reduction with guaranteed clock skew yield. Our experimental results show that our proposed algorithm achieves dominant clock network augmentation solutions, e.g., an average of 16% clock skew yield improvement, 9% maximum skew reduction, and 25% reduction of clock skew variation standard deviation with identical wirelength compared with previous best clock network link insertion methods [11].

The clock distribution network is a key component of any synchronous VLSI design. High power dissipation and pressure volume temperature-induced variations in clock skew have started playing an increasingly important role in limiting the performance of the clock network. Rotary clocking is a novel technique which employs unterminated rings formed by differential transmission lines to save power and reduce skew variability. Despite its appealing advantages, rotary clocking requires flip-flop locations to match predesigned clock skew on rotary clock rings. This requirement poses a difficult chicken-and-egg problem which prevents its wide application. In this paper, we propose an integrated placement and skew scheduling methodology to break this hurdle, making rotary clocking compatible with practical design flows. A network flow based flip-flop assignment algorithm and a cost-driven skew optimization algorithm are developed. We also present an integer linear programming formulation that minimizes maximum capacitance loaded at any of the rotary rings, thereby maximizing the operating frequency. Experimental results on benchmark circuits show that our method can reduce the tapping cost (measured as the total length of the wire segments connecting the rotary rings to the clock sinks) for rotary clocking by 33%-53%

The clock distribution network is a key component on any synchronous VLSI design. As technology moves into the nanometer era, innovative clocking techniques are required to solve the power dissipation and variability issues. Rotary clocking is a novel technique which employs unterminated rings formed by differential transmission lines to save power and reduce skew variability. Despite its appealing advantages, rotary clocking requires latch locations to match pre-designed clock skew on rotary clock rings. This requirement is a difficult chicken-and-egg problem which prevents its wide application. In this work, we proposed an integrated placement and skew scheduling methodology to break this hurdle, making rotary clocking compatible with practical design flows. A network flow based latch assignment algorithm and a cost-driven skew optimization algorithm are developed. Experiments show that our method can generate chip placements which satisfy the unique requirements of rotary clocks, without sacrificing design quality. By enabling concurrent clock network and placement design, our method can also be applied in other clocking methodologies as well

Power/ground noise is a major source of VLSI circuit timing variations. This work aims to reduce clock network induced power noise by assigning different signal polarities (opposite switchings) to clock buffers in an existing buffered clock tree. Three assignment algorithms are proposed: (1) partitioning, (2) 2-coloring on minimum spanning tree and (3) recursive min-matching. A post-processing of clock buffer sizing is performed to achieve desired clock skew. SPICE based experimental results indicate that our techniques could reduce the average peak current and average delay variations by 44% and 54% respectively.

We present a fast and efficient combinatorial algorithm to simultaneously identify the candidate locations as well as the sizes of the buffers driving a clock mesh. Due to the high redundancy, a mesh architecture offers high tolerance towards variation in the clock skew. However, such a redundancy comes at the expense of mesh wire length and power dissipation. Based on survivable network theory, we formulate the problem to reduce the clock mesh by retaining only those edges that are critical to maintain redundancy. Such a formulation offers designer the option to trade-off between power and tolerance to process variations. Experimental results indicate that our techniques can result in power savings up to 28% with less than 4% delay penalty.

The synthesis of clock network in the presence of process variation is becoming a vital design issue towards the performance of digital circuits. In this paper, we propose a clock tree design algorithm which is driven by the tolerance towards process variations. We consider tolerance to process variation in various stages of clock tree synthesis which include clock skew scheduling, abstract tree generation and layout embedding. The primary objective of this work is to minimize the maximum skew violation and a layout embedding technique specifically targeting this objective is detailed. Experimental results indicate the our proposed procedure leads to significant reduction in maximum skew violation due to process variation with negligible change in wire length.

Clock skew is becoming increasingly difficult to control due to variations. Link based non-tree clock distribution is a cost-effective technique for reducing clock skew variations. However, previous works based on this technique were limited to unbuffered clock networks and neglected spatial correlations in the experimental validation. In this work, we overcome these shortcomings and make the link based non-tree approach feasible for realistic designs. The short circuit risk and multi-driver delay issues in buffered non-tree clock networks are investigated. Our approach is validated with SPICE based Monte Carlo simulations, considering spatial correlations among variations. The experimental results show that our approach can reduce the maximal skew by 47%, improve the skew yield from 15% to 73% on average with a decrease on the total wire and buffer capacitance.

Increasingly prominent variational effects impose imminent threat to the progress of VLSI technology. This work explores redundancy, which is a well-known fault tolerance technique, for variation tolerance. It is observed that delay variability can be reduced by making redundant paths distributed or less correlated. Based on this observation, a gate splitting methodology is proposed for achieving distributed redundancy. We show how to avoid short circuit and estimate delay in dual-driver nets which are caused by gate splitting. A spin-off gate placement heuristic is developed to minimize redundancy cost. Monte Carlo simulation results on benchmark circuits show that our method can improve timing yield from 59% to 72% with only 03% increase on cell area and 2.2% increase on wirelength on average.