Abstract Recommendation systems often contain both rich relational structures and diverse multimodal information. The multiple relations among users, items, and auxiliary entities naturally form a heterogeneous information network. A central challenge in developing scalable recommendation systems in the era of big data is efficiently identifying similar users and items across hop- n relational paths in such networks. Hashing has been widely adopted for dimensionality and data size reduction; however, existing techniques are primarily designed for directly connected (i.e., hop-1) features and rarely exploit higher-order relational information. To address this limitation, we propose two methods. First, we develop relation-aware hashing that extends locality-sensitive hashing to encode hop- n metapath semantics and builds metapath-specific hash blocks as a scalable recall layer for candidate generation. Second, we introduce a multimodal learning-to-hash model that learns binary codes from fused text, image, and temporal features, and aligns Hamming-space neighbourhoods with metapath-guided user neighbourhood graphs. By jointly leveraging both relation-aware encoding and multimodal content, the proposed approaches enable efficient neighbourhood construction and recommendation in large-scale heterogeneous networks. Extensive experiments on three real-world datasets show that our framework achieves substantial efficiency gains while delivering competitive recommendation accuracy compared with baselines.

Recommender systems (RSs) are designed to provide personalized recommendations to users. Recently, knowledge graphs (KGs) have been widely introduced in RSs to improve recommendation accuracy. In this study, however, we demonstrate that RSs do not necessarily perform worse even if the KG is downgraded to the user-item interaction graph only (or removed). We propose an evaluation framework KG4RecEval to systematically evaluate how much a KG contributes to the recommendation accuracy of a KG-based RS, using our defined metric KG utilization efficiency in recommendation (KGER). We consider the scenarios where knowledge in a KG gets completely removed, randomly distorted and decreased, and also where recommendations are for cold-start users. Our extensive experiments on four commonly used datasets and a number of state-of-the-art KG-based RSs reveal that: to remove, randomly distort or decrease knowledge does not necessarily decrease recommendation accuracy, even for cold-start users. These findings inspire us to rethink how to better utilize knowledge from existing KGs, whereby we discuss and provide insights into what characteristics of datasets and KG-based RSs may help improve KG utilization efficiency. The code and supplementary material of this article are available at: https://github.com/HotBento/KG4RecEval .

Meta-paths have been popularly used to provide explainability in recommendations. Although long/complicated meta-paths could represent complex user-item connectivity, they are not easy to interpret. This work tackles this problem by introducing a meta-path translation task. The objective is to translate a meta-path to its comparable explainable meta-paths that perform similarly in terms of recommendation but have higher explainability compared to the given one. We propose a definition of meta-path explainability to determine comparable explainable meta-paths and a meta-path grammar that allows comparable explainable meta-paths to be formed in a similar way as sentences in human languages. Based on this grammar, we propose a meta-path translation model, a sequence-to-sequence (Seq2Seq) model to translate a long and complicated meta-path to its comparable explainable meta-paths. Two novel datasets for meta-path translation were generated based on two real-world recommendation datasets. The experiments were conducted on these generated datasets. The results show that our model outperformed state-of-the-art Seq2Seq baselines regarding meta-path translation and maintained a better trade-off between accuracy and diversity/readability in predicting comparable explainable meta-paths. These results indicate that our model can effectively generate a group of explainable meta-paths as alternative explanations for those recommendations based on any given long/complicated meta-path.

Knowledge Graphs (KGs) are a crucial component of Artificial Intelligence (AI) systems, enhancing AI’s capabilities in literary analysis. However, traditional KG designs in this field have focused more on events, often ignoring character information. To tackle this issue, we created a comprehensive Human-Trait-Enhanced Knowledge Graph, HTEKG, which combines past event-centered KGs with general human traits. The HTEKG enhances query capabilities by mapping the complex relationships and traits of literary characters, thereby providing more accurate and context-relevant information. We tested our HTEKG on three typical literary comprehension methods: traditional Cypher query, integration with a BERT classifier, and integration with GPT-4, demonstrating its effectiveness in literary analysis and its adaptability to different language models.

Radiologists have an important task of diagnosing thyroid nodules present in ultra sound images. Although reporting systems exist to aid in the diagnosis process, these systems do not provide explanations about the diagnosis results. We present ThyExp -- a web based toolkit for it use by medical professionals, allowing for accurate diagnosis with explanations of thyroid nodules present in ultrasound images utilising artificial intelligence models. The proposed web-based toolkit can be easily incorporated into current medical workflows, and allows medical professionals to have the confidence of a highly accurate machine learning model with explanations to provide supplementary diagnosis data. The solution provides classification results with their probability accuracy, as well as the explanations in the form of presenting the key features or characteristics that contribute to the classification results. The experiments conducted on a real-world UK NHS hospital patient dataset demonstrate the effectiveness of the proposed approach. This toolkit can improve the trust of medical professional to understand the confidence of the model in its predictions. This toolkit can improve the trust of medical professionals in understanding the models reasoning behind its predictions.

Conversational recommender systems (CRS) use interactive discussions to recommend high-quality items to users. Two essential components in a good CRS are the recommendation module that makes pertinent product recommendations to consumers and a conversation component that creates text-based sentences with product recommendations. The most commonly used dataset to train CRS models is ReDial. In this paper, we found that using the INSPIRED dataset in place of the ReDial dataset significantly improves model performance in terms of effectiveness. Along with the INSPIRED dataset, the inclusion of historical data in the input improves efficiency. The accuracy and efficiency of the model increase when we include the historical data into the system in the form of DialoGPT corpus and Gutenberg books. The paper further extends to compare three versions of state-of-the-art knowledge graph based conversational recommender systems called KGSF â one with the INSPIRED dataset with history, one with the INSPIRED dataset without historical data and the last with the ReDial dataset without historical data, which is the original version of the KGSF model. The comparison between three versions of the KGSF model shows that the change of the dataset and the inclusion of historical data can promote the performance of this conversational recommendation system.

Recommender systems contain rich relation information. The multiple relations in a recommender system form a heterogeneous information network. How to efficiently find similar users and items based on hop-n relations in heterogeneous information networks is one significant challenge to develop scalable recommender systems in the era of big data. Hashing has been popularly used for dimensionality reduction and data size reduction. Current hashing techniques mainly focus on hashing for directly related (i.e. hop-1) features. This paper proposes to develop relation-aware hashing techniques to bridge this gap. The proposed approaches use locality sensitive hashing (LSH) and consider hop-n relations in an information network to construct user or item blocks. They help facilitate efficient neighborhood formation and recommendation making. The experiments conducted on a large-scale real-life dataset show that the proposed approaches are effective.

Recommender systems are popular for personalization in online communities. Users, items, and other affiliated information such as tags, item genres, and user friends of an online community form a heterogenous information network. User profiling is the foundation of personalized recommender systems. It provides the basis to discover knowledge about an individual user’s interests to items. Typically, users are profiled with their direct explicit or implicit ratings, which ignored the inter-connections among users, items, and other entity nodes of the information network. This paper proposes a deep reinforcement user profiling approach for recommender systems. The user profiling process is framed as a sequential decision making problem which can be solved with a Reinforcement Learning (RL) agent. The RL agent interacts with the external heterogenous information network environment and learns a decision making policy network to decide whether there is an interest or preference path between a user and an unobserved item. To effectively train the RL agent, this paper proposes a multi-iteration training process to combine both expert and data-specific knowledge to profile users, generate meta-paths, and make recommendations. The effectiveness of the proposed approaches is demonstrated in experiments conducted on three datasets.

Health claims are sentences on the food product packages to claim the nutrition and the benefits of the nutrition. Consumers in different European contexts often have difficulties understanding health claims, leading to increased confusion about and decreased trust in the food they buy. Focusing on this problem, we develop a toolkit for improving the communication of health claims for consumers. The toolkit provides (1) interactive activities to disseminate knowledge about health claims to the public, and (2) an NLP-based analysis and prediction engine that food manufacturers can use to estimate how consumers like the health claims that the manufacturers created. By using the AI-powered toolkit, consumers, manufacturers, and food safety regulators are engaged in determining the different linguistic and cultural barriers to the effective communication of health claims and formulating solutions that can be implemented on multiple levels, including regulation, enforcement, marketing, and consumer education.

In this work we propose a semantic-aware blocking framework for entity resolution (ER). The proposed framework is built using locality-sensitive hashing (LSH) techniques to efficiently unify both textual and semantic features into an ER blocking process. In order to understand how similarity metrics may affect the effectiveness of ER blocking, we study the robustness of similarity metrics and their properties in terms of LSH families. We further discuss how the semantic similarity of records can be captured, measured, and integrated with LSH techniques over multiple similarity spaces. We have evaluated our proposed framework over two real-world data sets, and compared it with the state-of-the-art blocking techniques. The experimental study shows that using a combination of semantic features and textual features can considerably improve the quality of blocking. Due to the probabilistic nature of LSH, this semantic-aware blocking framework also enables us to build fast and reliable blocking for performing entity resolution tasks in a large-scale data environment.

In this paper, we propose a semantic-aware blocking framework for entity resolution (ER). The proposed framework is built using locality-sensitive hashing (LSH) techniques, which efficiently unifies both textual and semantic features into an ER blocking process. In order to understand how similarity metrics may affect the effectiveness of ER blocking, we study the robustness of similarity metrics and their properties in terms of LSH families. Then, we present how the semantic similarity of records can be captured, measured, and integrated with LSH techniques over multiple similarity spaces. In doing so, the proposed framework can support efficient similarity searches on records in both textual and semantic similarity spaces, yielding ER blocking with improved quality. We have evaluated the proposed framework over two real-world data sets, and compared it with the state-of-the-art blocking techniques. Our experimental study shows that the combination of semantic similarity and textual similarity can considerably improve the quality of blocking. Furthermore, due to the probabilistic nature of LSH, this semantic-aware blocking framework enables us to build fast and reliable blocking for performing entity resolution tasks in a large-scale data environment.

User profiling is a key component of personalized recommender systems, and is used to generate user profiles that describe individual user interests and preferences. The increasing availability of big data is driving the urgent need for user profiling algorithms that are able to generate accurate user profiles from large-scale user behavior data. In this paper, we propose a probabilistic rating auto-encoder to perform unsupervised feature learning and generate latent user feature profiles from large-scale user rating data. Based on the generated user profiles, neighbourhood based collaborative filtering approaches have been adopted to make personalized rating predictions. The effectiveness of the proposed approach is demonstrated in experiments conducted on a real-world rating dataset from yelp.com.

Topic recommendation can help users deal with the information overload issue in micro-blogging communities. This paper proposes to use the implicit information network formed by the multiple relationships among users, topics and micro-blogs, and the temporal information of micro-blogs to find semantically and temporally relevant topics of each topic, and to profile users' time-drifting topic interests. The Content based, Nearest Neighborhood based and Matrix Factorization models are used to make personalized recommendations. The effectiveness of the proposed approaches is demonstrated in the experiments conducted on a real world dataset that collected from Twitter.com.

Item folksonomy or tag information is popularly available on the web now. However, since tags are arbitrary words given by users, they contain a lot of noise such as tag synonyms, semantic ambiguities and personal tags. Such noise brings difficulties to improve the accuracy of item recommendations. In this paper, we propose to combine item taxonomy and folksonomy to reduce the noise of tags and make personalized item recommendations. The experiments conducted on the dataset collected from Amazon.com demonstrated the effectiveness of the proposed approaches. The results suggested that the recommendation accuracy can be further improved if we consider the viewpoints and the vocabularies of both experts and users.

The social tags in web 2.0 are becoming another important information source to profile users' interests and preferences to make personalized recommendations. To solve the problem of low information sharing caused by the free-style vocabulary of tags and the long tails of the distribution of tags and items, this paper proposes an approach to integrate the social tags given by users and the item taxonomy with standard vocabulary and hierarchical structure provided by experts to make personalized recommendations. The experimental results show that the proposed approach can effectively improve the information sharing and recommendation accuracy.