Lanbo Zhang

Hongjian Wang, Qi Li, Lanbo Zhang et al.

Historical features are important in ads click-through rate (CTR) prediction, because they account for past engagements between users and ads. In this paper, we study how to efficiently construct historical features through counting features. The key challenge of such problem lies in how to automatically identify counting keys. We propose a tree-based method for counting key selection. The intuition is that a decision tree naturally provides various combinations of features, which could be used as counting key candidate. In order to select personalized counting features, we train one decision tree model per user, and the counting keys are selected across different users with a frequency-based importance measure. To validate the effectiveness of proposed solution, we conduct large scale experiments on Twitter video advertising data. In both online learning and offline training settings, the automatically identified counting features outperform the manually curated counting features.

Mustafa Abdool, Malay Haldar, Prashant Ramanathan et al.

One of the long-standing questions in search systems is the role of diversity in results. From a product perspective, showing diverse results provides the user with more choice and should lead to an improved experience. However, this intuition is at odds with common machine learning approaches to ranking which directly optimize the relevance of each individual item without a holistic view of the result set. In this paper, we describe our journey in tackling the problem of diversity for Airbnb search, starting from heuristic based approaches and concluding with a novel deep learning solution that produces an embedding of the entire query context by leveraging Recurrent Neural Networks (RNNs). We hope our lessons learned will prove useful to others and motivate further research in this area.

Malay Haldar, Mustafa Abdool, Prashant Ramanathan et al.

The application of deep learning to search ranking was one of the most impactful product improvements at Airbnb. But what comes next after you launch a deep learning model? In this paper we describe the journey beyond, discussing what we refer to as the ABCs of improving search: A for architecture, B for bias and C for cold start. For architecture, we describe a new ranking neural network, focusing on the process that evolved our existing DNN beyond a fully connected two layer network. On handling positional bias in ranking, we describe a novel approach that led to one of the most significant improvements in tackling inventory that the DNN historically found challenging. To solve cold start, we describe our perspective on the problem and changes we made to improve the treatment of new listings on the platform. We hope ranking teams transitioning to deep learning will find this a practical case study of how to iterate on DNNs.

Lanbo Zhang

As structured documents with rich metadata (such as products, movies, etc.) become increasingly prevalent, searching those documents has become an important IR problem. Although advanced search interfaces are widely available, most users still prefer to use keyword-based queries to search those documents. Query keywords often imply some hidden restrictions on the desired documents, which can be represented as document facet-value pairs. To achieve high retrieval performance, it's important to be able to identify the relevant facet-value pairs hidden in a query. In this paper, we study the problem of identifying document facet-value pairs that are relevant to a keyword-based search query. We propose a machine learning approach and a set of useful features, and evaluate our approach using a movie data set from INEX.

Lanbo Zhang

This paper presents a new user feedback mechanism based on Wikipedia concepts for interactive retrieval. In this mechanism, the system presents to the user a group of Wikipedia concepts, and the user can choose those relevant to refine his/her query. To realize this mechanism, we propose methods to address two problems: 1) how to select a small number of possibly relevant Wikipedia concepts to show the user, and 2) how to re-rank retrieved documents given the user-identified Wikipedia concepts. Our methods are evaluated on three TREC data sets. The experiment results show that our methods can dramatically improve retrieval performances.

Lanbo Zhang, Yi Zhang, Qianli Xing

Existing document filtering systems learn user profiles based on user relevance feedback on documents. In some cases, users may have prior knowledge about what features are important. For example, a Spanish speaker may only want news written in Spanish, and thus a relevant document should contain the feature "Language: Spanish"; a researcher focusing on HIV knows an article with the medical subject "Subject: AIDS" is very likely to be relevant to him/her. Semi-structured documents with rich metadata are increasingly prevalent on the Internet. Motivated by the well-adopted faceted search interface in e-commerce, we study the exploitation of user prior knowledge on faceted features for semi-structured document filtering. We envision two faceted feedback mechanisms, and propose a novel user profile learning algorithm that can incorporate user feedback on features. To evaluate the proposed work, we use two data sets from the TREC filtering track, and conduct a user study on Amazon Mechanical Turk. Our experiment results show that user feedback on faceted features is useful for filtering. The proposed user profile learning algorithm can effectively learn from user feedback on both documents and features, and performs better than several existing methods.

Lanbo Zhang, Yi Zhang

Most existing content-based filtering approaches learn user profiles independently without capturing the similarity among users. Bayesian hierarchical models \cite{Zhang:Efficient} learn user profiles jointly and have the advantage of being able to borrow discriminative information from other users through a Bayesian prior. However, the standard Bayesian hierarchical models assume all user profiles are generated from the same prior. Considering the diversity of user interests, this assumption could be improved by introducing more flexibility. Besides, most existing content-based filtering approaches implicitly assume that each user profile corresponds to exactly one user interest and fail to capture a user's multiple interests (information needs). In this paper, we present a flexible Bayesian hierarchical modeling approach to model both commonality and diversity among users as well as individual users' multiple interests. We propose two models each with different assumptions, and the proposed models are called Discriminative Factored Prior Models (DFPM). In our models, each user profile is modeled as a discriminative classifier with a factored model as its prior, and different factors contribute in different levels to each user profile. Compared with existing content-based filtering models, DFPM are interesting because they can 1) borrow discriminative criteria of other users while learning a particular user profile through the factored prior; 2) trade off well between diversity and commonality among users; and 3) handle the challenging classification situation where each class contains multiple concepts. The experimental results on a dataset collected from real users on digg.com show that our models significantly outperform the baseline models of L-2 regularized logistic regression and traditional Bayesian hierarchical model with logistic regression.