The 18th edition of the ACM International Conference on Web Search and Data Mining (WSDM) was successfully held from March 10 to 14, 2025, at the Hannover Congress Centre (HCC), Hannover, Germany. WSDM 2025 was held as an in-person conference. The first day of the five-day event featured tutorials, industry day, and doctoral consortium. The next three days were the main conference days, which consisted of keynote speeches, plenary and poster sessions, and demo sessions. The last day of the event was dedicated to WSDM Cup, WSDM Day, diversity and inclusion sessions, and workshops. This brief report provides an overview of the organization, program details, and statistics from the Program Chairs and General Chairs. Further details are available at https://www.wsdm-conference.org/2025/. At the conclusion of this report, we express our deepest gratitude to the authors, volunteers, steering committee, sponsors, and supporters.

Historically, information retrieval systems have all followed the same paradigm: information seekers frame their needs in the form of a short query, the system selects a small set of relevant results from a corpus of available documents, rank-orders the results by decreasing relevance, possibly excerpts a responsive passage for each result, and returns a list of references and excerpts to the user. Retrieval systems typically did not attempt fusing information from multiple documents into an answer and displaying that answer directly. This was largely due to available technology: at the core of each retrieval system is an index that maps lexical tokens or semantic embeddings to document identifiers. Indices are designed for retrieving responsive documents; they do not support integrating these documents into a holistic answer. More recently, the coming-of-age of deep neural networks has dramatically improved the capabilities of large language models (LLMs). Trained on a large corpus of documents, these models not only memorize the vocabulary, morphology and syntax of human languages, but have shown to be able to memorize facts and relations. Generative language models, when provided with a prompt, will extend the prompt with likely completions -- an ability that can be used to extract answers to questions from the model. Two years ago, Metzler et al. argued that this ability of LLMs will allow us to rethink the search paradigm: to answer information needs directly rather that directing users to responsive primary sources. Their vision was not without controversy; the following year Shaw and Bender argued that such a system is neither feasible nor desirable. Nonetheless, the past year has seen the emergence of such systems, with offerings from established search engines and multiple new entrants to the industry. The keynote will summarize the short history of these generative information retrieval systems, and focus on the many open challenges in this emerging field: ensuring that answers are grounded, attributing answer passages to a primary source, providing nuanced answers to non-factoid-seeking questions, avoiding bias, and going beyond simple regurgitation of memorized facts. It will also touch on the changing nature of the content ecosystem. LLMs are starting to be used to generate web content. Should search engines treat such derived content equal to human-authored content? Is it possible to distinguish generated from original content? How should we view hybrid authorship where humans contribute ideas and LLMs shape these ideas into prose? And how will this parallel technical evolution of search engines and content ecosystems affect their respective business models?

Historically, information retrieval systems have all followed the same paradigm: information seekers frame their needs in the form of a short query, the system selects a small set of relevant results from a corpus of available documents, rank-orders the results by decreasing relevance, possibly excerpts a responsive passage for each result, and returns a list of references and excerpts to the user. Retrieval systems typically did not attempt fusing information from multiple documents into an answer and displaying that answer directly. This was largely due to available technology: at the core of each retrieval system is an index that maps lexical tokens or semantic embeddings to document identifiers. Indices are designed for retrieving responsive documents; they do not support integrating these documents into a holistic answer. More recently, the coming-of-age of deep neural networks has dramatically improved the capabilities of large language models (LLMs). Trained on a large corpus of documents, these models not only memorize the vocabulary, morphology and syntax of human languages, but have shown to be able to memorize facts and relations [3]. Generative language models, when provided with a prompt, will extend the prompt with likely completions - an ability that can be used to extract answers to questions from the model. Four years ago, Metzler et al. argued that this ability of LLMs will allow us to rethink the search paradigm: to answer information needs directly rather that directing users to responsive primary sources [1]. Their vision was not without controversy; the following year Shaw and Bender argued that such a system is neither feasible nor desirable [4]. Nonetheless, the past four years have seen the emergence of such systems, with offerings from established search engines and many new entrants to the industry. This keynote (an updated version of [2]) will summarize the history of these generative information retrieval systems, and focus on the many open challenges: ensuring that answers are grounded, attributing answer passages to a primary source, providing nuanced answers to non-factoid-seeking questions, avoiding bias, and going beyond simple regurgitation of memorized facts. It will also touch on the changing nature of the content ecosystem. LLMs are starting to be used to generateweb content. Should search engines treat such derived content equal to human-authored content? Is it possible to distinguish generated from original content? And how should we view hybrid authorship where humans contribute ideas and LLMs shape these ideas into prose?

Knowledge distillation is a popular technique to transfer knowledge from a large teacher model to a small student model. Typically, the student learns to imitate the teacher by minimizing the KL divergence of its output distribution with the teacher's output distribution. In this work, we argue that such a learning objective is sub-optimal because there exists a discrepancy between the teacher's output distribution and the ground truth label distribution. Therefore, forcing the student to blindly imitate the unreliable teacher output distribution leads to inferior performance. To this end, we propose a novel knowledge distillation objective PTLoss by first representing the vanilla KL-based distillation loss function via a Maclaurin series and then perturbing the leading-order terms in this series. This perturbed loss implicitly transforms the original teacher into a proxy teacher with a distribution closer to the ground truth distribution. We establish the theoretical connection between this "distribution closeness'' and the student model generalizability, which enables us to select the PTLoss's perturbation coefficients in a principled way. Extensive experiments on six public benchmark datasets demonstrate the effectiveness of PTLoss with teachers of different scales.

This article describes the history of information retrieval on personal document collections.

Automatic headline generation enables users to comprehend ongoing news events promptly and has recently become an important task in web mining and natural language processing. With the growing need for news headline generation, we argue that the hallucination issue, namely the generated headlines being not supported by the original news stories, is a critical challenge for the deployment of this feature in web-scale systems Meanwhile, due to the infrequency of hallucination cases and the requirement of careful reading for raters to reach the correct consensus, it is difficult to acquire a large dataset for training a model to detect such hallucinations through human curation. In this work, we present a new framework named ExHalder to address this challenge for headline hallucination detection. ExHalder adapts the knowledge from public natural language inference datasets into the news domain and learns to generate natural language sentences to explain the hallucination detection results. To evaluate the model performance, we carefully collect a dataset with more than six thousand labeled ⟨ article, headline⟩ pairs. Extensive experiments on this dataset and another six public ones demonstrate that ExHalder can identify hallucinated headlines accurately and justifies its predictions with human-readable natural language explanations.

Google My Business (GMB) is a platform that hosts business profiles, which will be displayed when a user issues a relevant query on Google Search or Google Maps. GMB businesses provide a wide variety of services, from home cleaning and repair, to legal consultation. However, the exact details of the service provided (a.k.a. job types), are often missing in business profiles. This places the burden of finding these details on the users. To alleviate this burden, we built a pipeline to automatically extract the job types from business websites. We share the various challenges we faced while developing this pipeline, and how we effectively addressed these challenges by (1) utilizing structured content to tackle the cold start problem for dataset collection; (2) exploiting context information to improve model performance without hurting scalability; and (3) formulating the extraction problem as a retrieval task to improve both generalizability, efficiency, and coverage. The pipeline has been deployed for over a year and is scalable enough to be periodically refreshed. The extracted job types are serving users of Google Search and Google Maps, with significant improvements in both precision and coverage.

Query-document relevance prediction is a critical problem in Information Retrieval systems. This problem has increasingly been tackled using (pretrained) transformer-based models which are finetuned using large collections of labeled data. However, in specialized domains such as e-commerce and healthcare, the viability of this approach is limited by the dearth of large in-domain data. To address this paucity, recent methods leverage these powerful models to generate high-quality task and domain-specific synthetic data. Prior work has largely explored synthetic data generation or query generation (QGen) for Question-Answering (QA) and binary (yes/no) relevance prediction, where for instance, the QGen models are given a document, and trained to generate a query relevant to that document. However in many problems, we have a more fine-grained notion of relevance than a simple yes/no label. Thus, in this work, we conduct a detailed study into how QGen approaches can be leveraged for nuanced relevance prediction. We demonstrate that -- contrary to claims from prior works -- current QGen approaches fall short of the more conventional cross-domain transfer-learning approaches. Via empirical studies spanning 3 public e-commerce benchmarks, we identify new shortcomings of existing QGen approaches -- including their inability to distinguish between different grades of relevance. To address this, we introduce label-conditioned QGen models which incorporates knowledge about the different relevance. While our experiments demonstrate that these modifications help improve performance of QGen techniques, we also find that QGen approaches struggle to capture the full nuance of the relevance label space and as a result the generated queries are not faithful to the desired relevance label.

Bag-of-words based lexical retrieval systems are still the most commonly used methods for real-world search applications. Recently deep learning methods have shown promising results to improve this retrieval performance but are expensive to run in an online fashion, non-trivial to integrate into existing production systems, and might not generalize well in out-of-domain retrieval scenarios. Instead, we build on top of lexical retrievers by proposing a Term Weighting BERT (TW-BERT) model. TW-BERT learns to predict the weight for individual n-gram (e.g., uni-grams and bi-grams) query input terms. These inferred weights and terms can be used directly by a retrieval system to perform a query search. To optimize these term weights, TW-BERT incorporates the scoring function used by the search engine, such as BM25, to score query-document pairs. Given sample query-document pairs we can compute a ranking loss over these matching scores, optimizing the learned query term weights in an end-to-end fashion. Aligning TW-BERT with search engine scorers minimizes the changes needed to integrate it into existing production applications, whereas existing deep learning based search methods would require further infrastructure optimization and hardware requirements. The learned weights can be easily utilized by standard lexical retrievers and by other retrieval techniques such as query expansion. We show that TW-BERT improves retrieval performance over strong term weighting baselines within MSMARCO and in out-of-domain retrieval on TREC datasets.

Differentiable Search Indices (DSIs) encode a corpus of documents in model parameters and use the same model to answer user queries directly. Despite the strong performance of DSI models, deploying them in situations where the corpus changes over time is computationally expensive because reindexing the corpus requires re-training the model. In this work, we introduce DSI++, a continual learning challenge for DSI to incrementally index new documents while being able to answer queries related to both previously and newly indexed documents. Across different model scales and document identifier representations, we show that continual indexing of new documents leads to considerable forgetting of previously indexed documents. We also hypothesize and verify that the model experiences forgetting events during training, leading to unstable learning. To mitigate these issues, we investigate two approaches. The first focuses on modifying the training dynamics. Flatter minima implicitly alleviate forgetting, so we optimize for flatter loss basins and show that the model stably memorizes more documents ($+12\%$). Next, we introduce a generative memory to sample pseudo-queries for documents and supplement them during continual indexing to prevent forgetting for the retrieval task. Extensive experiments on novel continual indexing benchmarks based on Natural Questions (NQ) and MS MARCO demonstrate that our proposed solution mitigates forgetting significantly. Concretely, it improves the average Hits@10 by $+21.1\%$ over competitive baselines for NQ and requires $6$ times fewer model updates compared to re-training the DSI model for incrementally indexing five corpora in a sequence.

As Learning-to-Rank (LTR) approaches primarily seek to improve ranking quality, their output scores are not scale-calibrated by design. This fundamentally limits LTR usage in score-sensitive applications. Though a simple multi-objective approach that combines a regression and a ranking objective can effectively learn scale-calibrated scores, we argue that the two objectives are not necessarily compatible, which makes the trade-off less ideal for either of them. In this paper, we propose a practical regression compatible ranking (RCR) approach that achieves a better trade-off, where the two ranking and regression components are proved to be mutually aligned. Although the same idea applies to ranking with both binary and graded relevance, we mainly focus on binary labels in this paper. We evaluate the proposed approach on several public LTR benchmarks and show that it consistently achieves either best or competitive result in terms of both regression and ranking metrics, and significantly improves the Pareto frontiers in the context of multi-objective optimization. Furthermore, we evaluated the proposed approach on YouTube Search and found that it not only improved the ranking quality of the production pCTR model, but also brought gains to the click prediction accuracy. The proposed approach has been successfully deployed in the YouTube production system.

Knowledge distillation is a popular technique to transfer knowledge from large teacher models to a small student model. Typically, the student learns to imitate the teacher by minimizing the KL divergence of its output distribution with the teacher's output distribution. In this work, we argue that such a learning objective is sub-optimal because there exists a discrepancy between the teacher's output distribution and the ground truth label distribution. Therefore, forcing the student to blindly imitate the unreliable teacher output distribution leads to inferior performance. To this end, we propose a novel knowledge distillation objective PTLoss by first representing the vanilla KL-based distillation loss function via a Maclaurin series and then perturbing the leading-order terms in this series. This perturbed loss implicitly transforms the original teacher into a proxy teacher with a distribution closer to the ground truth distribution. We establish the theoretical connection between this "distribution closeness" and the student model generalizability, which enables us to select the PTLoss's perturbation coefficients in a principled way. Extensive experiments on five datasets demonstrate PTLoss can significantly improve the distillation effectiveness for teachers of various scales.

Comparative decisions, such as picking between two cars or deciding between two hiking trails, require the users to visit multiple webpages and contrast the choices along relevant aspects. Given the impressive capabilities of pre-trained large language models [4, 11], we ask whether they can help automate such analysis. We refer to this task as extractive aspect-based contrastive summarization which involves constructing a structured summary that compares the choices along relevant aspects. In this paper, we propose a novel method called STRUM for this task that can generalize across domains without requiring any human-written summaries or fixed aspect list as supervision. Given a set of relevant input webpages, STRUM solves this problem using two pre-trained T5-based [11] large language models: first one fine-tuned for aspect and value extraction [14], and second one fine-tuned for natural language inference [13]. We showcase the abilities of our method across different domains, identify shortcomings, and discuss questions that we believe will be critical in this new line of research.

When discussing a tweet, people usually not only refer to the content it delivers, but also to the person behind the tweet. In other words, grounding the interpretation of the tweet in the context of its creator plays an important role in deciphering the true intent and the importance of the tweet. In this paper, we attempt to answer the question of how creator context should be used to advance tweet understanding. Specifically, we investigate the usefulness of different types of creator context, and examine different model structures for incorporating creator context in tweet modeling. We evaluate our tweet understanding models on a practical use case -- recommending relevant tweets to news articles. This use case already exists in popular news apps, and can also serve as a useful assistive tool for journalists. We discover that creator context is essential for tweet understanding, and can improve application metrics by a large margin. However, we also observe that not all creator contexts are equal. Creator context can be time sensitive and noisy. Careful creator context selection and deliberate model structure design play an important role in creator context effectiveness.

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.

Top-K metrics such as NDCG@K are frequently used to evaluate ranking performance. The traditional tree-based models such as LambdaMART, which are based on Gradient Boosted Decision Trees (GBDT), are designed to optimize NDCG@K using the LambdaRank losses. Recently, there is a good amount of research interest on neural ranking models for learning-to-rank tasks. These models are fundamentally different from the decision tree models and behave differently with respect to different loss functions. For example, the most popular ranking losses used in neural models are the Softmax loss and the GumbelApproxNDCG loss. These losses do not connect to top-K metrics such as NDCG@K naturally. It remains a question on how to effectively optimize NDCG@K for neural ranking models. In this paper, we follow the LambdaLoss framework and design novel and theoretically sound losses for NDCG@K metrics, while the original LambdaLoss paper can only do so using an unsound heuristic. We study the new losses on the LETOR benchmark datasets and show that the new losses work better than other losses for neural ranking models.

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.

introduction Share on Introduction to the Special Section on Graph Technologies for User Modeling and Recommendation, Part 2 Authors: Xiangnan He University of Science and Technology of China, Hefei, China University of Science and Technology of China, Hefei, ChinaSearch about this author , Zhaochun Ren Shandong University, Qingdao, China Shandong University, Qingdao, ChinaSearch about this author , Emine Yilmaz Department of Computer Science, University College London, London, United Kingdom Department of Computer Science, University College London, London, United KingdomSearch about this author , Marc Najork Google Research, Mountain View, CA, United States Google Research, Mountain View, CA, United StatesSearch about this author , Tat-Seng Chua National University of Singapore, Singapore, Republic of Singapore National University of Singapore, Singapore, Republic of SingaporeSearch about this author Authors Info & Claims ACM Transactions on Information SystemsVolume 40Issue 3July 2022 Article No.: 42pp 1–5https://doi.org/10.1145/3490180Published:14 December 2021Publication History 0citation385DownloadsMetricsTotal Citations0Total Downloads385Last 12 Months385Last 6 weeks27 Get Citation AlertsNew Citation Alert added!This alert has been successfully added and will be sent to:You will be notified whenever a record that you have chosen has been cited.To manage your alert preferences, click on the button below.Manage my Alerts New Citation Alert!Please log in to your account Save to BinderSave to BinderCreate a New BinderNameCancelCreateExport CitationPublisher SiteGet Access

Learning-to-Rank (LTR) systems are ubiquitous in web applications nowadays. The existing literature mainly focuses on improving ranking performance by trying to generate the optimal order of candidate items. However, virtually all advanced ranking functions are not scale calibrated. For example, rankers have the freedom to add a constant to all item scores without changing their relative order. This property has resulted in several limitations in deploying advanced ranking methods in practice. On the one hand, it limits the use of effective ranking functions in important applications. For example, in ads ranking, predicted Click-Through Rate (pCTR) is used for ranking and is required to be calibrated for the downstream ads auction. This is a major reason that existing ads ranking methods use scale calibrated pointwise loss functions that may sacrifice ranking performance. On the other hand, popular ranking losses are translation-invariant. We rigorously show that, both theoretically and empirically, this property leads to training instability that may cause severe practical issues. In this paper, we study how to perform scale calibration of deep ranking models to address the above concerns. We design three different formulations to calibrate ranking models through calibrated ranking losses. Unlike existing post-processing methods, our calibration is performed during training, which can resolve the training instability issue without any additional processing. We conduct experiments on the standard LTR benchmark datasets and one of the largest sponsored search ads dataset from Google. Our results show that our proposed calibrated ranking losses can achieve nearly optimal results in terms of both ranking quality and score scale calibration.