Effectiveness perspectives for spatial keyword queries and personalized entity search

Abstract

The widespread use of smartphones has made efficient and effective search mechanisms essential across various domains. Whether locating points of interest (POI) in geographical areas or finding products on e-commerce platforms, the ability to search effectively significantly impacts many aspects of our daily lives. The prevalence of geo-textual data, consisting of objects associated with both geographical locations and textual descriptions, has grown significantly with the widespread adoption of smartphones equipped with GPS services. This encompasses diverse forms such as web pages with geographic information, user-generated texts with geotags, Points of Interest (POI), and multimedia data with text and location details. Accurate and efficient processing of spatial keyword queries, which involve both location and textual criteria, is crucial for enhancing user experiences across various platforms. Additionally, the surge in e-commerce over the past two decades has underscored the importance of effective product search mechanisms. This dissertation explores the challenges and opportunities in optimizing spatial keyword queries for POI searches and personalized entity searches on e-commerce platforms.

The first focus of this dissertation is to assess the effectiveness of standard spatial keyword queries. While substantial research has been conducted on spatial keyword queries over the past decades, primarily concentrating on efficiency, little attention has been devoted to addressing the effectiveness perspectives. To fill this gap, we evaluate the effectiveness of standard spatial keyword queries using two datasets with ground truth query results. Our evaluation indicates that the TkQ query, ranking objects through a weighted combination of spatial proximity and text relevance, proves most effective compared to other query types. Moreover, we find that a query-dependent weight value can achieve much better effectiveness than a uniform weight that is used in all the previous work on spatial keyword queries

Motivated by this finding, we introduce the Deep relevance with Weight learning (DrW) model to further enhance retrieval ranking effectiveness. DrW incorporates two novel ideas: a neural network architecture for learning text relevance matching through local interaction between the query and geo-textual objects, and the design of a learning-based method to learn a query-dependent weight to balance text relevance and spatial proximity. To the best of our knowledge, this is the first work proposing to learn a query-dependent weight for spatial keyword queries. Moreover, in response to the efficiency challenge encountered by the deep relevance component of DrW, we integrate DrW with a conventional method such as TkQ. Here, we employ the traditional approach to select a set of candidates and subsequently employ DrW for reranking. This strategic combination aims to harness the respective strengths of both methods, thereby achieving superior efficiency and effectiveness. Experimental results demonstrate that our model surpasses state-of-the-art methods on effectiveness, with improvements up to 32.15%, 32.34%, and 33.00% in NDCG@3, NDCG@5, and MRR, respectively.

This dissertation further explores user personalization in spatial keyword queries, particularly in the context of POI search. Analyzing real-world POI search data reveals diverse user preferences in geographical distances and regions, emphasizing the query-dependent and dynamic nature of these preferences. To address this, we propose a Query-dependent User Geographical Preference Learning model (QPL) for personalized POI search. QPL jointly models textual relevance between queries and candidate POIs, as well as the query-dependent user geographical preference. Specifically, we develop a multi-headed self-attention based method to learn query-dependent user geographical preferences for the first time. Additionally, we design a Cross Attention Layer to effectively and efficiently learn the relationship between the query and the POI region's representation. To match the query and POI by textual relevance, we introduce a new textual matching module to model the interaction between the query and POI description, incorporating both traditional lexical matching and deep semantic learning. Extensive experiments on two real-world POI search datasets demonstrate the effectiveness of our model. The results reveal that our model outperforms the best baseline by up to 28.19% in terms of NDCG@5.

Finally, we extend the investigation to structural relationship learning, specifically within the domain of personalized entity search. Existing graph embedding enhanced personalized entity search methods are mostly based on entity-relation-entity graph learning. In this dissertation, we propose to consider structural relationships in users' entity search scenarios with graph embedding by latent representation learning. We argue that explicitly modeling the structural relationship in graph embedding is essential for more accurate personalized entity search results. We propose a novel method, Graph embedding-based Structural Relationship Representation Learning (GraphSRRL), which explicitly models the structural relationship in users-queries-entities interaction. It combines three key conjunctive graph patterns to learn graph embedding for better-personalized entity search. In addition, GraphSRRL facilitates the learning of affinities between users (resp. queries or entities) in the designed geometric operation in low-dimensional latent space. We conduct extensive experiments on four datasets to evaluate GraphSRRL for personalized entity search. Experimental results show that GraphSRRL outperforms the state-of-the-art algorithm on real-world search datasets by at least 50.7% in terms of Hit@10 and 48.7% in terms of NDCG@10.

In conclusion, this dissertation advances the field of spatial keyword query processing by addressing effectiveness evaluation, enhancing traditional models, incorporating user personalization, and exploring structural relationship learning for personalized entity search. These contributions pave the way for more sophisticated and tailored location-based services and product searches on e-commerce platforms. Ultimately, these advancements enrich user interactions with geo-textual objects and entity products in diverse contexts.