The most recent pointwise Large Language Model (LLM) rankers have achieved remarkable ranking results. However, these rankers are hindered by two major drawbacks: (1) they fail to follow a standardized comparison guidance during the ranking process, and (2) they struggle with comprehensive considerations when dealing with diverse semantics of the query and complicated info in the passages. To address these shortcomings, we propose to build a zero-shot pointwise ranker that first recruits a virtual annotation team to generate query-based criteria from various perspectives and then uses these criteria to conduct an ensemble passage evaluation. Additionally, we are among the first to explore how criteria can be generated automatically and used in text ranking tasks. Our method, tested on eight datasets from the BEIR benchmark, demonstrates that incorporating this multi-perspective criteria ensemble approach significantly enhanced the performance of pointwise LLM rankers.

Query expansion has been widely used to improve the search results of first-stage retrievers, yet its influence on second-stage, cross-encoder rankers remains under-explored. A recent study shows that current expansion techniques benefit weaker models but harm stronger rankers. In this paper, we re-examine this conclusion and raise the following question: Can query expansion improve generalization of strong cross-encoder rankers? To answer this question, we first apply popular query expansion methods to different cross-encoder rankers and verify the deteriorated zero-shot effectiveness. We identify two vital steps in the experiment: high-quality keyword generation and minimally-disruptive query modification. We show that it is possible to improve the generalization of a strong neural ranker, by generating keywords through a reasoning chain and aggregating the ranking results of each expanded query via self-consistency, reciprocal rank weighting, and fusion. Experiments on BEIR and TREC Deep Learning 2019/2020 show that the nDCG@10 scores of both MonoT5 and RankT5 following these steps are improved, which points out a direction for applying query expansion to strong cross-encoder rankers.

We propose a novel zero-shot document ranking approach based on Large Language Models (LLMs): the Setwise prompting approach.Our approach complements existing prompting approaches for LLM-based zero-shot ranking: Pointwise, Pairwise, and Listwise.Through the first-of-its-kind comparative evaluation within a consistent experimental framework and considering factors like model size, token consumption, latency, among others, we show that existing approaches are inherently characterised by trade-offs between effectiveness and efficiency.We find that while Pointwise approaches score high on efficiency, they suffer from poor effectiveness.Conversely, Pairwise approaches demonstrate superior effectiveness but incur high computational overhead.Our Setwise approach, instead, reduces the number of LLM inferences and the amount of prompt token consumption during the ranking procedure, compared to previous methods.This significantly improves the efficiency of LLM-based zero-shot ranking, while also retaining high zero-shot ranking effectiveness.We make our code and results publicly available at https://github.com/ielab/llm-rankers.

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

Pretrained language models such as BERT have been shown to be exceptionally effective for text ranking. However, there are limited studies on how to leverage more powerful sequence-to-sequence models such as T5. Existing attempts usually formulate text ranking as a classification problem and rely on postprocessing to obtain a ranked list. In this paper, we propose RankT5 and study two T5-based ranking model structures, an encoder-decoder and an encoder-only one, so that they not only can directly output ranking scores for each query-document pair, but also can be fine-tuned with pairwise or listwise ranking losses to optimize ranking performance. Our experiments show that the proposed models with ranking losses can achieve substantial ranking performance gains on different public text ranking data sets. Moreover, ranking models fine-tuned with listwise ranking losses have better zero-shot ranking performance on out-of-domain data than models fine-tuned with classification losses.

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

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

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

Interpretability of ranking models is a crucial yet relatively under-examined research area. Recent progress on this area largely focuses on generating post-hoc explanations for existing black-box ranking models. Though promising, such post-hoc methods cannot provide sufficiently accurate explanations in general, which makes them infeasible in many high-stakes scenarios, especially the ones with legal or policy constraints. Thus, building an intrinsically interpretable ranking model with transparent, self-explainable structure becomes necessary, but this remains less explored in the learning-to-rank setting.

A well-known challenge in leveraging implicit user feedback like clicks to improve real-world search services and recommender systems is its inherent bias. Most existing click models are based on the examination hypothesis in user behaviors and differ in how to model such an examination bias. However, they are constrained by assuming a simple position-based bias or enforcing a sequential order in user examination behaviors. These assumptions are insufficient to capture complex real-world user behaviors and hardly generalize to modern user interfaces (UI) in web applications (e.g., results shown in a grid view). In this work, we propose a fully data-driven neural model for the examination bias, Cross-Positional Attention (XPA), which is more flexible in fitting complex user behaviors. Our model leverages the attention mechanism to effectively capture cross-positional interactions among displayed items and is applicable to arbitrary UIs. We employ XPA in a novel neural click model that can both predict clicks and estimate relevance. Our experiments on offline synthetic data sets show that XPA is robust among different click generation processes. We further apply XPA to a large-scale real-world recommender system, showing significantly better results than baselines in online A/B experiments that involve millions of users. This validates the necessity to model more complex user behaviors than those proposed in the literature.

Nowadays, it is common to classify collections of documents into (human-generated, domain-specific) directory structures, such as email or document folders. But documents may be classified wrongly, for a multitude of reasons. Then they are outlying w.r.t. the folder they end up in. Orthogonally to this, and more specifically, two kinds of errors can occur: (O) Out-of-distribution: the document does not belong to any existing folder in the directory; and (M) Misclassification: the document belongs to another folder. It is this specific combination of issues that we address in this article, i.e., we mine text outliers from massive document directories, considering both error types. We propose a new proximity-based algorithm, which we dub kj-Nearest Neighbours (kj-NN). Our algorithm detects text outliers by exploiting semantic similarities and introduces a self-supervision mechanism that estimates the relevance of the original labels. Our approach is efficient and robust to large proportions of outliers. kj-NN also promotes the interpretability of the results by proposing alternative label names and by finding the most similar documents for each outlier. Our real-world experiments demonstrate that our approach outperforms the competitors by a large margin.

Aspect-based sentiment analysis is a substantial step towards text understanding which benefits numerous applications. Since most existing algorithms require a large amount of labeled data or substantial external language resources, applying them on a new domain or a new language is usually expensive and time-consuming. We aim to build an aspect-based sentiment analysis model from an unlabeled corpus with minimal guidance from users, i.e., only a small set of seed words for each aspect class and each sentiment class. We employ an autoencoder structure with attention to learn two dictionary matrices for aspect and sentiment respectively where each row of the dictionary serves as an embedding vector for an aspect or a sentiment class. We propose to utilize the user-given seed words to regularize the dictionary learning. In addition, we improve the model by joining the aspect and sentiment encoder in the reconstruction of sentiment in sentences. The joint structure enables sentiment embeddings in the dictionary to be tuned towards the aspect-specific sentiment words for each aspect, which benefits the classification performance. We conduct experiments on two real data sets to verify the effectiveness of our models.

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

In personal email search, user queries often impose different requirements on different aspects of the retrieved emails. For example, the query "my recent flight to the US" requires emails to be ranked based on both textual contents and recency of the email documents, while other queries such as "medical history" do not impose any constraints on the recency of the email. Recent deep learning-to-rank models for personal email search often directly concatenate dense numerical features (e.g., document age) with embedded sparse features (e.g., n-gram embeddings). In this paper, we first show with a set of experiments on synthetic datasets that direct concatenation of dense and sparse features does not lead to the optimal search performance of deep neural ranking models. To effectively incorporate both sparse and dense email features into personal email search ranking, we propose a novel neural model, SepAttn. SepAttn first builds two separate neural models to learn from sparse and dense features respectively, and then applies an attention mechanism at the prediction level to derive the final prediction from these two models. We conduct a comprehensive set of experiments on a large-scale email search dataset, and demonstrate that our SepAttn model consistently improves the search quality over the baseline models.

Early identification of patients at risk for postoperative complications can facilitate timely workups and treatments and improve health outcomes. Currently, a widely-used surgical risk calculator online web system developed by the American College of Surgeons (ACS) uses patients’ static features, e.g. gender, age, to assess the risk of postoperative complications. However, the most crucial signals that reflect the actual postoperative physical conditions of patients are usually real-time dynamic signals, including the vital signs of patients (e.g., heart rate, blood pressure) collected from postoperative monitoring. In this paper, we develop a dynamic postoperative complication risk scoring framework (DyCRS) to detect the “at-risk” patients in a real-time way based on postoperative sequential vital signs and static features. DyCRS is based on adaptations of the Hidden Markov Model (HMM) that captures hidden states as well as observable states to generate a real-time, probabilistic, complication risk score. Evaluating our model using electronic health record (EHR) on elective Colectomy surgery from a major health system, we show that DyCRS significantly outperforms the state-of-the-art ACS calculator and real-time predictors with 50.16% area under precision-recall curve (AUCPRC) gain on average in terms of detection effectiveness. In terms of earliness, our DyCRS can predict 15hrs55mins earlier on average than clinician’s diagnosis with the recall of 60% and precision of 55%. Furthermore, Our DyCRS can extract interpretable patients’ stages, which are consistent with previous medical postoperative complication studies. We believe that our contributions demonstrate significant promise for developing a more accurate, robust and interpretable postoperative complication risk scoring system, which can benefit more than 50 million annual surgeries in the US by substantially lowering adverse events and healthcare costs.

Aspect-based sentiment analysis is an important tool to understand user opinions in a fine-grained manner. Although extensively studied, developing such a tool for a specific domain remains an expensive process. Most existing methods either rely on massive labeled data for training or external language resource and tools which are not necessarily available or accurate. We propose to study the aspect-based sentiment analysis with only a small set of aspect and sentiment seed words as guidance on a target corpus. We first expand the aspect and sentiment lexicons from the given seed words by features created by frequent pattern mining. Then, we develop a generative model to characterize the aspect and sentiment mentions based on their word embedding, and infer the sentiment polarity for sentiment words accordingly. The effectiveness of our method is verified by experiments on two real world data sets.

The real-time discovery of local events (e.g., protests, disasters) has been widely recognized as a fundamental socioeconomic task. Recent studies have demonstrated that the geo-tagged tweet stream serves as an unprecedentedly valuable source for local event detection. Nevertheless, how to effectively extract local events from massive geo-tagged tweet streams in real time remains challenging. To bridge the gap, we propose a method for effective and real-time local event detection from geo-tagged tweet streams. Our method, named G eo B urst+ , first leverages a novel cross-modal authority measure to identify several pivots in the query window. Such pivots reveal different geo-topical activities and naturally attract similar tweets to form candidate events. G eo B urst+ further summarizes the continuous stream and compares the candidates against the historical summaries to pinpoint truly interesting local events. Better still, as the query window shifts, G eo B urst+ is capable of updating the event list with little time cost, thus achieving continuous monitoring of the stream. We used crowdsourcing to evaluate G eo B urst+ on two million-scale datasets and found it significantly more effective than existing methods while being orders of magnitude faster.

As a powerful representation paradigm for networked and multi-typed data, the heterogeneous information network (HIN) is ubiquitous. Meanwhile, defining proper relevance measures has always been a fundamental problem and of great pragmatic importance for network mining tasks. Inspired by our probabilistic interpretation of existing path-based relevance measures, we propose to study HIN relevance from a probabilistic perspective. We also identify, from real-world data, and propose to model cross-meta-path synergy, which is a characteristic important for defining path-based HIN relevance and has not been modeled by existing methods. A generative model is established to derive a novel path-based relevance measure, which is data-driven and tailored for each HIN. We develop an inference algorithm to find the maximum a posteriori (MAP) estimate of the model parameters, which entails non-trivial tricks. Experiments on two real-world datasets demonstrate the effectiveness of the proposed model and relevance measure.

Detecting local events (e.g., protest, disaster) at their onsets is an important task for a wide spectrum of applications, ranging from disaster control to crime monitoring and place recommendation. Recent years have witnessed growing interest in leveraging geo-tagged tweet streams for online local event detection. Nevertheless, the accuracies of existing methods still remain unsatisfactory for building reliable local event detection systems. We propose TrioVecEvent, a method that leverages multimodal embeddings to achieve accurate online local event detection. The effectiveness of TrioVecEvent is underpinned by its two-step detection scheme. First, it ensures a high coverage of the underlying local events by dividing the tweets in the query window into coherent geo-topic clusters. To generate quality geo-topic clusters, we capture short-text semantics by learning multimodal embeddings of the location, time, and text, and then perform online clustering with a novel Bayesian mixture model. Second, TrioVecEvent considers the geo-topic clusters as candidate events and extracts a set of features for classifying the candidates. Leveraging the multimodal embeddings as background knowledge, we introduce discriminative features that can well characterize local events, which enables pinpointing true local events from the candidate pool with a small amount of training data. We have used crowdsourcing to evaluate TrioVecEvent, and found that it improves the performance of the state-of-the-art method by a large margin.

The real-time discovery of local events (e.g., protests, crimes, disasters) is of great importance to various applications, such as crime monitoring, disaster alarming, and activity recommendation. While this task was nearly impossible years ago due to the lack of timely and reliable data sources, the recent explosive growth in geo-tagged tweet data brings new opportunities to it. That said, how to extract quality local events from geo-tagged tweet streams in real time remains largely unsolved so far.

Outlier or anomaly detection in large data sets is a fundamental task in data science, with broad applications. However, in real data sets with high-dimensional space, most outliers are hidden in certain dimensional combinations and are relative to a user's search space and interest. It is often more effective to give power to users and allow them to specify outlier queries flexibly, and the system will then process such mining queries efficiently. In this study, we introduce the concept of query-based outlier in heterogeneous information networks, design a query language to facilitate users to specify such queries flexibly, define a good outlier measure in heterogeneous networks, and study how to process outlier queries efficiently in large data sets. Our experiments on real data sets show that following such a methodology, interesting outliers can be defined and uncovered flexibly and effectively in large heterogeneous networks.

Crowdsourcing is the de-facto standard for gathering annotated data. While, in theory, data annotation tasks are assumed to be attempted by workers independently, in practice, data annotation tasks are often grouped into batches to be presented and annotated by workers together, in order to save on the time or cost overhead of providing instructions or necessary background. Thus, even though independence is usually assumed between annotations on data items within the same batch, in most cases, a worker's judgment on a data item can still be affected by other data items within the batch, leading to additional errors in collected labels. In this paper, we study the data annotation bias when data items are presented as batches to be judged by workers simultaneously. We propose a novel worker model to characterize the annotating behavior on data batches, and present how to train the worker model on annotation data sets. We also present a debiasing technique to remove the effect of such annotation bias from adversely affecting the accuracy of labels obtained. Our experimental results on both synthetic data and real-world data demonstrate the effectiveness of our proposed method.

Data annotation bias is found in many situations. Often it can be ignored as just another component of the noise floor. However, it is especially prevalent in crowdsourcing tasks and must be actively managed. Annotation bias on single data items has been studied with regard to data difficulty, annotator bias, etc., while annotation bias on batches of multiple data items simultaneously presented to annotators has not been studied. In this paper, we verify the existence of "in-batch annotation bias" between data items in the same batch. We propose a factor graph based batch annotation model to quantitatively capture the in-batch annotation bias, and measure the bias during a crowdsourcing annotation process of inappropriate comments in LinkedIn. We discover that annotators tend to make polarized annotations for the entire batch of data items in our task. We further leverage the batch annotation model to propose a novel batch active learning algorithm. We test the algorithm on a real crowdsourcing platform and find that it outperforms in-batch bias naïve algorithms.

Mining outliers in a heterogeneous information network is a challenging problem: It is even unclear what should be outliers in a large heterogeneous network (e.g., Outliers in the entire bibliographic network consisting of authors, titles, papers and venues). In this study, we propose an interesting class of outliers, query-based sub network outliers: Given a heterogeneous network, a user raises a query to retrieve a set of task-relevant sub networks, among which, sub network outliers are those that significantly deviate from others (e.g., Outliers of author groups among those studying "topic modeling"). We formalize this problem and propose a general framework, where one can query for finding sub network outliers with respect to different semantics. We introduce the notion of sub network similarity that captures the proximity between two sub networks by their membership distributions. We propose an outlier detection algorithm to rank all the sub networks according to their outlierness without tuning parameters. Our quantitative and qualitative experiments on both synthetic and real data sets show that the proposed method outperforms other baselines.

Social influence and influence diffusion has been widely studied in online social networks. However, most existing works on influence diffusion focus on static networks. In this paper, we study the problem of maximizing influence diffusion in a dynamic social network. Specifically, the network changes over time and the changes can be only observed by periodically probing some nodes for the update of their connections. Our goal then is to probe a subset of nodes in a social network so that the actual influence diffusion process in the network can be best uncovered with the probing nodes. We propose a novel algorithm to approximate the optimal solution. The algorithm, through probing a small portion of the network, minimizes the possible error between the observed network and the real network. We evaluate the proposed algorithm on both synthetic and real large networks. Experimental results show that our proposed algorithm achieves a better performance than several alternative algorithms.

Online social networks become a bridge to connect our physical daily life and the virtual Web space, which not only provides rich data for mining, but also brings many new challenges. In this paper, we present a novel Social Analytic Engine (SAE) for large online social networks. The key issues we pursue in the analytic engine are concerned with the following problems: 1) at the micro-level, how do people form different types of social ties and how people influence each other? 2) at the meso-level, how do people group into communities? 3) at the macro-level, what are the hottest topics in a social network and how the topics evolve over time?