User retention is a critical objective for online platforms like Pinterest, as it strengthens user loyalty and drives growth through repeated engagement. A key indicator of retention is revisitation, i.e., when users return to view previously saved content, a behavior often sparked by personalized recommendations and user satisfaction. However, modeling and optimizing revisitation poses significant challenges. One core difficulty is accurate attribution: it is often unclear which specific user actions or content exposures trigger a revisit, since many confounding factors (e.g., content quality, user interface, notifications, or even changing user intent) can influence return behavior. Additionally, the scale and timing of revisitations introduce further complexity; users may revisit content days or even weeks after their initial interaction, requiring the system to maintain and associate extensive historical records across millions of users and sessions. These complexities render existing methods insufficient for robustly capturing and optimizing long-term revisitation. To address these gaps, we introduce a novel, lightweight, and interpretable framework for modeling revisitation behavior and optimizing long-term user retention in Pinterest’s search-based recommendation context. By defining a surrogate attribution process that links saves to subsequent revisitations, we reduce noise in the causal relationship between user actions and return visits. Our scalable event aggregation pipeline enables large-scale analysis of user revisitation patterns and enhances the ranking system’s ability to surface items with high retention value. Deployed on Pinterest’s Related Pins surface to serve 500+ million users, the framework led to a significant lift of 0.1% in active users without additional computational costs. Our data analysis reveals novel insights, such as the impact of content topics on revisitation rates; for example, users are more likely to revisit aesthetically pleasing topics.

Semantic IDs (SIDs) are hierarchical item identifiers learned via residual quantization, offering a promising solution to cold-start recommendation, yet their integration into discriminative recommendations remains challenging and lacks systematic investigation. We present Pin-SCALE, a framework for optimized end-to-end integration of SID into dense retrieval in cascading recommender systems. Pin-SCALE instantiates the principles through three technical contributions: (1) cascading pooling preserving residual quantization hierarchy that outperforms common pooling schema and attention;(2) engagement-aware tokenization that encodes collaborative signals to fully utilize both content and user engagement pattern; and (3) multi-view contrastive learning that harnesses query-item and item-item co-occurrence to align SID representations with engagement prediction. Through extensive offline analyses, we collected systematic guidance and proved the effectiveness of Pin-SCALE. Furthermore, the framework has been tested online on multiple surfaces at Pinterest including Closeup, Home Feed and Search, with different content types like organic and e-commerce. Pin-SCALE achieves substantial improvements in fresh engagements (+3.67% repins) as well as platform-level retention (+0.05% DAU). Our work has been shipped into production to serve billions of requests daily and drive discovery journey of our users, as well as provides principled guidelines for SID adoption in discriminative recommendation models at production scale.

The application to search ranking is one of the biggest machine learning success stories at Airbnb. Much of the initial gains were driven by a gradient boosted decision tree model. The gains, however, plateaued over time. This paper discusses the work done in applying neural networks in an attempt to break out of that plateau. We present our perspective not with the intention of pushing the frontier of new modeling techniques. Instead, ours is a story of the elements we found useful in applying neural networks to a real life product. Deep learning was steep learning for us. To other teams embarking on similar journeys, we hope an account of our struggles and triumphs will provide some useful pointers. Bon voyage!

We study the problem of learning query intent representation for an entity search task such as product retrieval, where a user would use a keyword query to retrieve entities based on their structured attribute value descriptions. Existing intent representation has been mostly based on the query space. These methods overlook the critical information from the entity space and the connection in between. Consequently, when such representation methods are used in intent mining from user engagement logs in entity search, they cannot fully discover the comprehensive knowledge of user preference, which is essential for improving the effectiveness of entity search and recommendation, as well as many applications such as business intelligence.

The booming of e-commerce in recent years has led to the generation of large amounts of product search log data. Product search log is a unique new data with much valuable information and knowledge about user preferences over product attributes that is often hard to obtain from other sources. While regular search logs (e.g., Web search logs) contain click-throughs for unstructured text documents (e.g., web pages), product search logs contain clickth-roughs for structured entities defined by a set of attributes and their values. For instance, a laptop can be defined by its size, color, cpu, ram, etc. Such structures in product entities offer us opportunities to mine and discover detailed useful knowledge about user preferences at the attribute level, but they also raise significant challenges for mining due to the lack of attribute-level observations.

The ability to let users search for products conveniently in product database is critical to the success of e-commerce. Although structured query languages (e.g. SQL) can be used to effectively access the product database, it is very difficult for end users to learn and use. In this paper, we study how to optimize search over structured product entities (represented by specifications) with keyword queries such as "cheap gaming laptop". One major difficulty in this problem is the vocabulary gap between the specifications of products in the database and the keywords people use in search queries. To solve the problem, we propose a novel probabilistic entity retrieval model based on query generation, where the entities would be ranked for a given keyword query based on the likelihood that a user who likes an entity would pose the query. Different ways to estimate the model parameters would lead to different variants of ranking functions. We start with simple estimates based on the specifications of entities, and then leverage user reviews and product search logs to improve the estimation. Multiple estimation algorithms are developed based on Maximum Likelihood and Maximum a Posteriori estimators. We evaluate the proposed product entity retrieval models on two newly created product search test collections. The results show that the proposed model significantly outperforms the existing retrieval models, benefiting from the modeling of attribute-level relevance. Despite the focus on product retrieval, the proposed modeling method is general and opens up many new opportunities in analyzing structured entity data with unstructured text data. We show the proposed probabilistic model can be easily adapted for many interesting applications including facet generation and review annotation.

Understanding users' search intents is critical component of modern search engines. A key limitation made by most query log analyses is the assumption that each clicked web result represents one unique intent. However, there are many search tasks, such as comparison shopping or in-depth research, where a user's intent is to explore many documents. In these cases, the assumption of a one-to-one correspondence between clicked documents and user intent breaks down.

Query spelling correction is a crucial component of modern search engines. Existing methods in the literature for search query spelling correction have two major drawbacks. First, they are unable to handle certain important types of spelling errors, such as concatenation and splitting. Second, they cannot efficiently evaluate all the candidate corrections due to the complex form of their scoring functions, and a heuristic filtering step must be applied to select a working set of top-K most promising candidates for final scoring, leading to non-optimal predictions. In this paper we address both limitations and propose a novel generalized Hidden Markov Model with discriminative training that can not only handle all the major types of spelling errors, including splitting and concatenation errors, in a single unified framework, but also efficiently evaluate all the candidate corrections to ensure the finding of a globally optimal correction. Experiments on two query spelling correction datasets demonstrate that the proposed generalized HMM is effective for correcting multiple types of spelling errors. The results also show that it significantly outperforms the current approach for generating top-K candidate corrections, making it a better first-stage filter to enable any other complex spelling correction algorithm to have access to a better working set of candidate corrections as well as to cover splitting and concatenation errors, which no existing method in academic literature can correct.

Modern search engines usually provide a query language with a set of advanced syntactic operators (e.g., plus sign to require a term's appearance, or quotation marks to require a phrase's appearance) which if used appropriately, can significantly improve the effectiveness of a plain keyword query. However, they are rarely used by ordinary users due to the intrinsic difficulties and users' lack of corpora statistics. In this paper, we propose to automatically reformulate queries that do not work well by selectively adding syntactic operators. Particularly, we propose to perform syntactic operator-based query reformulation when a retrieval system detects users encounter difficulty in search as indicated by users' behaviors such as scanning over top k documents without click-through. We frame the problem of automatic reformulation with syntactic operators as a supervised learning problem, and propose a set of effective features to represent queries with syntactic operators. Experiment results verify the effectiveness of the proposed method and its applicability as a query suggestion mechanism for search engines. As a negative feedback strategy, syntactic operator-based query reformulation also shows promising results in improving search results for difficult queries as compared with existing methods.

In this paper, we study the problem of online spelling correction for query completions. Misspelling is a common phenomenon among search engines queries. In order to help users effectively express their information needs, mechanisms for automatically correcting misspelled queries are required. Online spelling correction aims to provide spell corrected completion suggestions as a query is incrementally entered. As latency is crucial to the utility of the suggestions, such an algorithm needs to be not only accurate, but also efficient.

In this article, we address the problem of question clustering and study its use for re-ranking question search results. In question clustering we have to organize question search results into certain meaningful and condensed groups. Specifically, we propose to use a data structure consisting of question topic and question focus for modeling questions, and then cluster questions on the basis of the data structure. Experimental results show that our approach to question clustering improves the performance of question search significantly better than the approach not utilizing the topic–focus structure.

The paper is concerned with the problem of question recommendation. Specifically, given a question as query, we are to retrieve and rank other questions according to their likelihood of being good recommendations of the queried question. A good recommendation provides alternative aspects around users' interest. We tackle the problem of question recommendation in two steps: first represent questions as graphs of topic terms, and then rank recommendations on the basis of the graphs. We formalize both steps as the tree-cutting problems and then employ the MDL (Minimum Description Length) for selecting the best cuts. Experiments have been conducted with the real questions posted at Yahoo! Answers. The questions are about two domains, 'travel' and 'computers & internet'. Experimental results indicate that the use of the MDL-based tree cut model can significantly outperform the baseline methods of word-based VSM or phrase-based VSM. The results also show that the use of the MDL-based tree cut model is essential to our approach.