Ranking is the most important component in a search system. Most search systems deal with large amounts of natural language data, hence an effective ranking system requires a deep understanding of text semantics. Recently, deep learning based natural language processing (deep NLP) models have generated promising results on ranking systems. BERT is one of the most successful models that learn contextual embedding, which has been applied to capture complex query-document relations for search ranking. However, this is generally done by exhaustively interacting each query word with each document word, which is inefficient for online serving in search product systems. In this paper, we investigate how to build an efficient BERT-based ranking model for industry use cases. The solution is further extended to a general ranking framework, DeText, that is open sourced and can be applied to various ranking productions. Offline and online experiments of DeText on three real-world search systems present significant improvement over state-of-the-art approaches.

Search and recommender systems share many fundamental components including language understanding, retrieval and ranking, and language generation. Building powerful search and recommender systems requires processing natural language effectively and efficiently. Recent rapid growth of deep learning technologies has presented both opportunities and challenges in this area. This tutorial offers an overview of deep learning based natural language processing (NLP) for search and recommender systems from an industry perspective. It first introduces deep learning based NLP technologies, including language understanding and language generation. Then it details how those technologies can be applied to common tasks in search and recommender systems, including query and document understanding, retrieval and ranking, and language generation. Applications in LinkedIn production systems are presented. The tutorial concludes with discussion of future trend.

Most online social networks provide a mechanism for users to broadcast messages to their personalized network through actions like shares, likes and tweets. Receiving positive feedback from the network such as likes, comments and retweets in response to such actions can provide a strong incentive for users to broadcast more often in the future. We call such feedback by the network, that influences a user to perform certain desirable future actions, social incentives. For example, after a user shares an article to her social network, receiving positive feedback such as a ''like'' from a friend can potentially encourage her to continue sharing more regularly. Typically, for every user's visit to an online social network site, good messages need to be ranked and selected by a recommender system from a large set of candidate messages (broadcasted by the user's network). In this paper, we propose a novel recommendation problem: How should we recommend messages to users to incentivize neighbors in their personal network to perform desirable actions in the future with high likelihood, without significantly hurting overall engagement for the entire system? For instance, messages could be content shared by neighbors. The goal in this case would be to encourage more content shares in the future. We call this problem social incentive optimization and study an instance of it for LinkedIn's news feed. We observe that a user who receives positive social feedback from neighbors has a higher likelihood of broadcasting more frequently. Using this observation, we develop a novel recommendation framework that incentivize users to broadcast more often, without significantly hurting overall feed engagement. We demonstrate the effectiveness of our approach through causal analysis on retrospective data and online A/B experiments.

Every year, millions of new students enter higher educational programs. Publicly available rankings of academic programs play a key role in prospective students' decisions regarding which universities to apply to and enroll in. While surveys indicate that majority of freshmen enter college to get good jobs after graduation, established methodologies for ranking universities rely on indirect indicators of career outcomes such as reputational assessments of the universities among academic peers, acceptance and graduation rates, learning environment, and availability of research funding. In addition, many of these methodologies rely on arbitrary choices of weighting factors for the different ranking indicators, and suffer from lack of analyses of statistical stability. In this paper, we addresses these challenges holistically by developing a novel methodology for ranking and recommending universities for different professions on the basis of career outcomes of professionals who graduated from those schools. Our methodology incorporates a number of techniques for achieving statistical stability, and represents a step towards personalized educational recommendations based on interests and ambitions of individuals. We have applied this methodology on LinkedIn's Economic Graph data of over 400 million professional from around the world. The resulting university rankings have been made available to the public and demonstrate that there are valuable insights to be gleaned from professional career data on LinkedIn.

User feedback like clicks and ratings on recommended items provides important information for recommender systems to predict users' interests in unseen items. Most systems rely on models trained using a single type of feedback, e.g., ratings for movie recommendation and clicks for online news recommendation. However, in addition to the primary feedback, many systems also allow users to provide other types of feedback, e.g., liking or sharing an article, or hiding all articles from a source. These additional feedback potentially provides extra information for the recommendation models. To optimize user experience and business objectives, it is important for a recommender system to use both the primary feedback and additional feedback. This paper presents an empirical study on various training methods for incorporating multiple user feedback types based on LinkedIn recommendation products. We study three important problems that we face at LinkedIn: (1) Whether to send an email based on clicks and complaints, (2) how to rank updates in LinkedIn feeds based on clicks and hides and (3) how jointly optimize for viral actions and clicks in LinkedIn feeds. Extensive offline experiments on historical data show the effectiveness of these methods in different situations. Online A/B testing results further demonstrate the impact of these methods on LinkedIn production systems.

Generalized linear model (GLM) is a widely used class of models for statistical inference and response prediction problems. For instance, in order to recommend relevant content to a user or optimize for revenue, many web companies use logistic regression models to predict the probability of the user's clicking on an item (e.g., ad, news article, job). In scenarios where the data is abundant, having a more fine-grained model at the user or item level would potentially lead to more accurate prediction, as the user's personal preferences on items and the item's specific attraction for users can be better captured. One common approach is to introduce ID-level regression coefficients in addition to the global regression coefficients in a GLM setting, and such models are called generalized linear mixed models (GLMix) in the statistical literature. However, for big data sets with a large number of ID-level coefficients, fitting a GLMix model can be computationally challenging. In this paper, we report how we successfully overcame the scalability bottleneck by applying parallelized block coordinate descent under the Bulk Synchronous Parallel (BSP) paradigm. We deployed the model in the LinkedIn job recommender system, and generated 20% to 40% more job applications for job seekers on LinkedIn.

LinkedIn dynamically delivers update activities from a user's interpersonal network to more than 300 million members in the personalized feed that ranks activities according their "relevance" to the user. This paper discloses the implementation details behind this personalized feed system at LinkedIn which can not be found from related work, and addresses the scalability and data sparsity challenges for deploying the system online. More specifically, we focus on the personalization models by generating three kinds of affinity scores: Viewer-ActivityType Affinity, Viewer-Actor Affinity, and Viewer-Actor-ActivityType Affinity. Extensive experiments based on online bucket tests (A/B experiments) and offline evaluation illustrate the effect of our personalization models in LinkedIn feed.

Users on an online social network site generate a large number of heterogeneous activities, ranging from connecting with other users, to sharing content, to updating their profiles. The set of activities within a user's network neighborhood forms a stream of updates for the user's consumption. In this paper, we report our experience with the problem of ranking activities in the LinkedIn homepage feed. In particular, we provide a taxonomy of social network activities, describe a system architecture (with a number of key components open-sourced) that supports fast iteration in model development, demonstrate a number of key factors for effective ranking, and report experimental results from extensive online bucket tests.

Predicting user affinity to items is an important problem in applications like content optimization, computational advertising, among others. While matrix factorization methods provide state-of-the-art performance when minimizing RMSE through a Gaussian response model on explicit ratings data, applying it to imbalanced binary response data presents additional challenges that we carefully study in this paper. Data in many applications usually consist of users' implicit response that is binary - clicking an item or not; the goal is to predict click rates (i.e., probabilities), which are often combined with other measures of utilities to rank items at runtime. Because of the implicit nature, such data is usually much larger than explicit rating data but often has an imbalanced distribution with a small fraction of click events, making accurate click rate prediction difficult. In this paper, we address two problems. First, we show previous techniques to estimate factor models with binary data are less accurate compared to our new approach based on adaptive rejection sampling, especially for imbalanced response. Second, we develop a parallel matrix factorization framework using Map-Reduce that scales to massive datasets. Our parallel algorithm is based on a “divide and conquer” strategy coupled with an ensemble approach. Through experiments on two benchmark data sets and a large Yahoo! Front Page Today Module data set that contains 8M users and 1B binary observations, we show that careful handling of binary response is needed to achieve good performance for click rate prediction, and that the proposed adaptive rejection sampler and the partitioning and ensemble techniques significantly improve performance.

Online social networks have become important channels for users to share content with their connections and diffuse information. Although much work has been done to identify socially influential users, the problem of finding "reputable" sharers, who share good content, has received relatively little attention. Availability of such reputation scores can be useful or various applications like recommending people to follow, procuring high quality content in a scalable way, creating a content reputation economy to incentivize high quality sharing, and many more. To estimate sharer reputation, it is intuitive to leverage data that records how recipients respond (through clicking, liking, etc.) to content items shared by a sharer. However, such data is usually biased --- it has a selection bias since the shared items can only be seen and responded to by users connected to the sharer in most social networks, and it has a response bias since the response is usually influenced by the relationship between the sharer and the recipient (which may not indicate whether the shared content is good). To correct for such biases, we propose to utilize an additional data source that provides unbiased goodness estimates for a small set of shared items, and calibrate biased social data through a novel multi-level hierarchical model that describes how the unbiased data and biased data are jointly generated according to sharer reputation scores. The unbiased data also provides the ground truth for quantitative evaluation of different methods. Experiments based on such ground-truth data show that our proposed model significantly outperforms existing methods that estimate social influence using biased social data.

Personalized article recommendation is important for news portals to improve user engagement. Existing work quantifies engagement primarily through click rates. We suggest that quality of recommendations may be improved by exploiting different types of "post-read" engagement signals like sharing, commenting, printing and e-mailing article links. Specifically, we propose a multi-faceted ranking problem for recommending articles, where each facet corresponds to a ranking task that seeks to maximize actions of a particular post-read type (e.g., ranking articles to maximize sharing actions). Our approach is to predict the probability that a user would take a post-read action on an article, so that articles can be ranked according to such probabilities. However, post-read actions are rare events --- enormous data sparsity makes the problem challenging. We meet the challenge by exploiting correlations across different post-read action types through a novel locally augmented tensor (LAT) model, so that the ranking performance of a particular action type can be improved by leveraging data from all other action types. Through extensive experiments, we show that our LAT model significantly outperforms a variety of state-of-the-art factor models, logistic regression and IR models.

Online content recommendation aims to identify trendy articles in a continuously changing dynamic content pool. Most of existing works rely on online user feedback, notably clicks, as the objective and maximize it by showing articles with highest click-through rates. Recently, click shaping was introduced to incorporate multiple objectives in a constrained optimization framework. The work showed that significant tradeoff among the competing objectives can be observed and thus it is important to consider multiple objectives. However, the proposed click shaping approach is segment-based and can only work with a few non-overlapping user segments. It remains a challenge of how to enable deep personalization in click shaping. In this paper, we tackle the challenge by proposing personalized click shaping. The main idea is to work with the Lagrangian duality formulation and explore strong convexity to connect dual and primal solutions. We show that our formulation not only allows efficient conversion from dual to primal for online personalized serving, but also enables us to solve the optimization faster by approximation. We conduct extensive experiments on a large real data set and our experimental results show that the personalized click shaping can significantly outperform the segmented one, while achieving the same ability to balance competing objectives.

User votes are important signals in community question-answering (CQA) systems. Many features of typical CQA systems, e.g. the best answer to a question, status of a user, are dependent on ratings or votes cast by the community. In a popular CQA site, Yahoo! Answers, users vote for the best answers to their questions and can also thumb up or down each individual answer. Prior work has shown that these votes provide useful predictors for content quality and user expertise, where each vote is usually assumed to carry the same weight as others. In this paper, we analyze a set of possible factors that indicate bias in user voting behavior -- these factors encompass different gaming behavior, as well as other eccentricities, e.g., votes to show appreciation of answerers. These observations suggest that votes need to be calibrated before being used to identify good answers or experts. To address this problem, we propose a general machine learning framework to calibrate such votes. Through extensive experiments based on an editorially judged CQA dataset, we show that our supervised learning method of content-agnostic vote calibration can significantly improve the performance of answer ranking and expert ranking.

Recommending interesting content to engage users is important for web portals (e.g. AOL, MSN, Yahoo!, and many others). Existing approaches typically recommend articles to optimize for a single objective, i.e., number of clicks. However a click is only the starting point of a user's journey and subsequent downstream utilities such as time-spent and revenue are important. In this paper, we call the problem of recommending links to jointly optimize for clicks and post-click downstream utilities click shaping. We propose a multi-objective programming approach in which multiple objectives are modeled in a constrained optimization framework. Such a formulation can naturally incorporate various application-driven requirements. We study several variants that model different requirements as constraints and discuss some of the subtleties involved. We conduct our experiments on a large dataset from a real system by using a newly proposed unbiased evaluation methodology [17]. Through extensive experiments we quantify the tradeoff between different objectives under various constraints. Our experimental results show interesting characteristics of different formulations and our findings may provide valuable guidance to the design of recommendation engines for web portals.

Combining correlated information from multiple contexts can significantly improve predictive accuracy in recommender problems. Such information from multiple contexts is often available in the form of several incomplete matrices spanning a set of entities like users, items, features, and so on. Existing methods simultaneously factorize these matrices by sharing a single set of factors for entities across all contexts. We show that such a strategy may introduce significant bias in estimates and propose a new model that ameliorates this issue by positing local, context-specific factors for entities. To avoid over-fitting in contexts with sparse data, the local factors are connected through a shared global model. This sharing of parameters allows information to flow across contexts through multivariate regressions among local factors, instead of enforcing exactly the same factors for an entity, everywhere. Model fitting is done in an EM framework, we show that the E-step can be fitted through a fast multi-resolution Kalman filter algorithm that ensures scalability. Experiments on benchmark and real-world Yahoo! datasets clearly illustrate the usefulness of our approach. Our model significantly improves predictive accuracy, especially in cold-start scenarios.

Reputable users are valuable assets of a web site. We focus on user reputation in a comment rating environment, where users make comments about content items and rate the comments of one another. Intuitively, a reputable user posts high quality comments and is highly rated by the user community. To our surprise, we find that the quality of a comment judged editorially is almost uncorrelated with the ratings that it receives, but can be predicted using standard text features, achieving accuracy as high as the agreement between two editors! However, extracting a pure reputation signal from ratings is difficult because of data sparseness and several confounding factors in users' voting behavior. To address these issues, we propose a novel bias-smoothed tensor model and empirically show that our model significantly outperforms a number of alternatives based on Yahoo! News, Yahoo! Buzz and Epinions datasets.

Predicting user "ratings" on items is a crucial task in recommender systems. Matrix factorization methods that computes a low-rank approximation of the incomplete user-item rating matrix provide state-of-the-art performance, especially for users and items with several past ratings (warm starts). However, it is a challenge to generalize such methods to users and items with few or no past ratings (cold starts). Prior work [4][32] have generalized matrix factorization to include both user and item features for performing better regularization of factors as well as provide a model for smooth transition from cold starts to warm starts. However, the features were incorporated via linear regression on factor estimates. In this paper, we generalize this process to allow for arbitrary regression models like decision trees, boosting, LASSO, etc. The key advantage of our approach is the ease of computing --- any new regression procedure can be incorporated by "plugging" in a standard regression routine into a few intermediate steps of our model fitting procedure. With this flexibility, one can leverage a large body of work on regression modeling, variable selection, and model interpretation. We demonstrate the usefulness of this generalization using the MovieLens and Yahoo! Buzz datasets.

We introduce a novel class of data cube, called latent-variable cube. For many data analysis tasks, data in a database can be represented as points in a multi-dimensional space. Ordinary data cubes compute aggregate functions over these "observed" data points for each cell (i.e., region) in the space, where the cells have different granularities defined by hierarchies. While useful, data cubes do not provide sufficient capability for analyzing "latent variables" that are often of interest but not directly observed in data. For example, when analyzing users' interaction with online advertisements, observed data informs whether a user clicked an ad or not. However, the real interest is often in knowing the click probabilities of ads for different user populations. In this example, click probabilities are latent variables that are not observed but have to be estimated from data. We argue that latent variables are a useful construct for a number of OLAP application scenarios. To facilitate such analyses, we propose cubes that compute aggregate functions over latent variables. Specifically, we discuss the pitfalls of common practice in scenarios where latent variables should, but are not considered; we rigorously define latent-variable cube based on Bayesian hierarchical models and provide efficient algorithms. Through extensive experiments on both simulated and real data, we show that our method is accurate and runs orders of magnitude faster than the baseline.

Recommender problems with large and dynamic item pools are ubiquitous in web applications like content optimization, online advertising and web search. Despite the availability of rich item meta-data, excess heterogeneity at the item level often requires inclusion of item-specific "factors" (or weights) in the model. However, since estimating item factors is computationally intensive, it poses a challenge for time-sensitive recommender problems where it is important to rapidly learn factors for new items (e.g., news articles, event updates, tweets) in an online fashion. In this paper, we propose a novel method called FOBFM (Fast Online Bilinear Factor Model) to learn item-specific factors quickly through online regression. The online regression for each item can be performed independently and hence the procedure is fast, scalable and easily parallelizable. However, the convergence of these independent regressions can be slow due to high dimensionality. The central idea of our approach is to use a large amount of historical data to initialize the online models based on offline features and learn linear projections that can effectively reduce the dimensionality. We estimate the rank of our linear projections by taking recourse to online model selection based on optimizing predictive likelihood. Through extensive experiments, we show that our method significantly and uniformly outperforms other competitive methods and obtains relative lifts that are in the range of 10-15% in terms of predictive log-likelihood, 200-300% for a rank correlation metric on a proprietary My Yahoo! dataset; it obtains 9% reduction in root mean squared error over the previously best method on a benchmark MovieLens dataset using a time-based train/test data split.

We propose fLDA, a novel matrix factorization method to predict ratings in recommender system applications where a "bag-of-words" representation for item meta-data is natural. Such scenarios are commonplace in web applications like content recommendation, ad targeting and web search where items are articles, ads and web pages respectively. Because of data sparseness, regularization is key to good predictive accuracy. Our method works by regularizing both user and item factors simultaneously through user features and the bag of words associated with each item. Specifically, each word in an item is associated with a discrete latent factor often referred to as the topic of the word; item topics are obtained by averaging topics across all words in an item. Then, user rating on an item is modeled as user's affinity to the item's topics where user affinity to topics (user factors) and topic assignments to words in items (item factors) are learned jointly in a supervised fashion. To avoid overfitting, user and item factors are regularized through Gaussian linear regression and Latent Dirichlet Allocation (LDA) priors respectively. We show our model is accurate, interpretable and handles both cold-start and warm-start scenarios seamlessly through a single model. The efficacy of our method is illustrated on benchmark datasets and a new dataset from Yahoo! Buzz where fLDA provides superior predictive accuracy in cold-start scenarios and is comparable to state-of-the-art methods in warm-start scenarios. As a by-product, fLDA also identifies interesting topics that explains user-item interactions. Our method also generalizes a recently proposed technique called supervised LDA (sLDA) to collaborative filtering applications. While sLDA estimates item topic vectors in a supervised fashion for a single regression, fLDA incorporates multiple regressions (one for each user) in estimating the item factors.

We propose novel multi-armed bandit (explore/exploit) schemes to maximize total clicks on a content module published regularly on Yahoo! Intuitively, one can "explore'' each candidate item by displaying it to a small fraction of user visits to estimate the item's click-through rate (CTR), and then "exploit'' high CTR items in order to maximize clicks. While bandit methods that seek to find the optimal trade-off between explore and exploit have been studied for decades, existing solutions are not satisfactory for Web content publishing applications where dynamic set of items with short lifetimes, delayed feedback and non-stationary reward (CTR) distributions are typical. In this paper, we develop a Bayesian solution and extend several existing schemes to our setting. Through extensive evaluation with nine bandit schemes, we show that our Bayesian solution is uniformly better in several scenarios. We also study the empirical characteristics of our schemes and provide useful insights on the strengths and weaknesses of each. Finally, we validate our results with a "side-by-side'' comparison of schemes through live experiments conducted on a random sample of real user visits to Yahoo!

We propose a novel latent factor model to accurately predict response for large scale dyadic data in the presence of features. Our approach is based on a model that predicts response as a multiplicative function of row and column latent factors that are estimated through separate regressions on known row and column features. In fact, our model provides a single unified framework to address both cold and warm start scenarios that are commonplace in practical applications like recommender systems, online advertising, web search, etc. We provide scalable and accurate model fitting methods based on Iterated Conditional Mode and Monte Carlo EM algorithms. We show our model induces a stochastic process on the dyadic space with kernel (covariance) given by a polynomial function of features. Methods that generalize our procedure to estimate factors in an online fashion for dynamic applications are also considered. Our method is illustrated on benchmark datasets and a novel content recommendation application that arises in the context of Yahoo! Front Page. We report significant improvements over several commonly used methods on all datasets.

We propose novel spatio-temporal models to estimate click-through rates in the context of content recommendation. We track article CTR at a fixed location over time through a dynamic Gamma-Poisson model and combine information from correlated locations through dynamic linear regressions, significantly improving on per-location model. Our models adjust for user fatigue through an exponential tilt to the first-view CTR (probability of click on first article exposure) that is based only on user-specific repeat-exposure features. We illustrate our approach on data obtained from a module (Today Module) published regularly on Yahoo! Front Page and demonstrate significant improvement over commonly used baseline methods. Large scale simulation experiments to study the performance of our models under different scenarios provide encouraging results. Throughout, all modeling assumptions are validated via rigorous exploratory data analysis.

How to mine massive datasets is a challenging problem with great potential value. Motivated by this challenge, much effort has concentrated on developing scalable versions of machine learning algorithms. However, the cost of mining large datasets is not just computational; preparing the datasets into the “right form” so that learning algorithms can be applied is usually costly, due to the human labor that is typically required and a large number of choices in data preparation, which include selecting different subsets of data and aggregating data at different granularities. We make the key observation that, for a number of practically motivated problems, these choices can be defined using database queries and analyzed in an automatic and systematic manner. Specifically, we propose a new class of data-mining problem, called bellwether analysis , in which the goal is to find a few query-defined predictors (e.g., first week sales of Peoria, IL of an item) that can be used to accurately predict the result of a target query (e.g., first year worldwide sales of the item) from a large number of queries that define candidate predictors. To make a prediction for a new item, the data needed to generate such predictors has to be collected (e.g., selling the new item in Peoria, IL for a week and collecting the sales data). A useful predictor is one that has high prediction accuracy and a low data-collection cost. We call such a cost-effective predictor a bellwether . This article introduces bellwether analysis, which integrates database query processing and predictive modeling into a single framework, and provides scalable algorithms for large datasets that cannot fit in main memory. Through a series of extensive experiments, we show that bellwethers do exist in real-world databases, and that our computation techniques achieve good efficiency on large datasets.

We describe a new content publishing system that selects articles to serve to a user, choosing from an editorially programmed pool that is frequently refreshed. It is now deployed on a major Internet portal, and selects articles to serve to hundreds of millions of user visits per day, significantly increasing the number of user clicks over the original manual approach, in which editors periodically selected articles to display. Some of the challenges we face include a dynamic content pool, short article lifetimes, non-stationary click-through rates, and extremely high traffic volumes. The fundamental problem we must solve is to quickly identify which items are popular(perhaps within different user segments), and to exploit them while they remain current. We must also explore the underlying pool constantly to identify promising alternatives, quickly discarding poor performers. Our approach is based on tracking per article performance in near real time through online models. We describe the characteristics and constraints of our application setting, discuss our design choices, and show the importance and effectiveness of coupling online models with a simple randomization procedure. We discuss the challenges encountered in a production online content-publishing environment and highlight issues that deserve careful attention. Our analysis of this application also suggests a number of future research avenues.

In this paper, we introduce a new class of data mining problems called learning from aggregate views. In contrast to the traditional problem of learning from a single table of training examples, the new goal is to learn from multiple aggregate views of the underlying data, without access to the un-aggregated data. We motivate this new problem, present a general problem framework, develop learning methods for RFA (Restriction-Free Aggregate) views defined using COUNT, SUM, AVG and STDEV, and offer theoretical and experimental results that characterize the proposed methods.

Fuzziness is an intrinsic property of knowledge retrieval. For instance, the relationship similarity is fuzzy by nature. The degree of truth of some transitive relations naturally degrades as transitivity is applied repeatedly. In this paper we propose a formal model of knowledge retrieval based on Goguen Style Fuzzy Horn Logic, a subset of Goguen's fuzzy logic system [The Logic of Inexact Concepts]. This model provides a mathematical (logical) foundation for knowledge retrieval that naturally incorporates fuzziness. By using only a subset of Goguen's logic, we have a simple but useful and efficient reasoning framework.

We propose an image retrieval methodology for a collection of similar images. By similar, we mean that one can define, for the collection, a set of dimensions, and for each of which a set of features. The dimensions are used to capture the essential characteristics of the images in the collection, and the features are for describing each image to a certain degree. We call this strategy fine-grained image retrieval to differentiate it from the more common coarse-grained retrieval, which does not assume any semantic properties on the image collection. The effectiveness of our methodology is demonstrated through an icon-based interactive retrieval system on a collection of butterfly images. This system provides the user with a friendly initial query-by-feature (QBF) interface. The user can then use query-by-example (QBE) to refine the query. In addition to presenting an outline of the methodology and the implementation on butterfly images, we also present some experimental results.