Predicting the words that a child is going to learn next can be useful for boosting language acquisition, and such predictions have been shown to be possible with both neural network techniques (looking at changes in the vocabulary state over time) and graph model (looking at data pertaining to the relationships between words). However, these models do not fully capture the complexity of the language learning process of an infant when used in isolation. In this paper, we examine how a model of language acquisition for infants and young children can be constructed and adapted for use in a Spatio-Temporal Graph Convolutional Network (STGCN), taking into account the different types of linguistic relationships that occur during child language learning. We introduce a novel approach for predicting child vocabulary acquisition, and evaluate the efficacy of such a model with respect to the different types of linguistic relationships that occur during language acquisition, resulting in insightful observations on model calibration and norm selection. An evaluation of this model found that the mean accuracy of models for predicting new words when using sensorimotor relationships (0.733) and semantic relationships (0.729) were found to be superior to that observed with a 2-layer Feed-forward neural network. Furthermore, the high recall for some relationships suggested that some relationships (e.g. visual) were superior in identifying a larger proportion of relevant words that a child should subsequently learn than others (such as auditory).

Recommendation systems are crucial in navigating the vast digital market. However, user data’s dynamic and non-stationary nature often hinders their efficacy. Traditional models struggle to adapt to the evolving preferences and behaviours inherent in user interaction data, posing a significant challenge for accurate prediction and personalisation. Addressing this, we propose a novel theoretical framework, the non-stationary transformer, designed to effectively capture and leverage the temporal dynamics within data. This approach enhances the traditional transformer architecture by introducing mechanisms accounting for non-stationary elements, offering a robust and adaptable solution for multi-tasking recommendation systems. Our experimental analysis, encompassing deep learning (DL) and reinforcement learning (RL) paradigms, demonstrates the framework’s superiority over benchmark models. The empirical results confirm our proposed framework’s efficacy, which provides significant performance enhancements, approximately 8% in LogLoss reduction and up to 2% increase in F1 score with other attention-related models. It also underscores its potential applicability across accumulative reward scenarios with pure reinforcement learning models. These findings advocate adopting non-stationary transformer models to tackle the complexities of today’s recommendation tasks.

Personalized recommendation plays a crucial role in Internet platforms, providing users with tailored content based on their user models and enhancing user satisfaction and experience. To address the challenge of information overload, it is essential to analyze user needs comprehensively, considering historical behavior and interests and the user's intentions and profiles. Previous user modeling approaches for personalized recommendations have exhibited certain limitations, relying primarily on historical behavior data to infer user preferences, which results in challenges such as the cold-start problem, incomplete modeling, and limited explanation. Motivated by recent advancements in large language models (LLMs), we present a novel approach to user modeling by embracing generative user modeling using LLMs. We propose generative user modeling with chain-of-thought prompting for personalized recommendation, which utilizes LLMs to generate comprehensive and accurate user models expressed in natural language and then employs these user models to empower LLMs for personalized recommendation. Specifically, we adopt the chain-of-thought prompting method to reason about user attributes, subjective preferences, and intentions, integrating them into a holistic user model. Subsequently, we utilize the generated user models as input to LLMs and design a collection of prompts to align the LLMs with various recommendation tasks, encompassing rating prediction, sequential recommendation, direct recommendation, and explanation generation. Extensive experiments conducted on real-world datasets demonstrate the immense potential of large language models in generating natural language user models, and the adoption of generative user modeling significantly enhances the performance of LLMs across the four recommendation tasks. Our code and dataset can be found at https://github.com/kwyyangfan/GUMRec .

Ontology alignment is critical in cross-domain integration; however, it typically necessitates the involvement of a human domain-expert, which can make the task costly. Although a variety of machine-learning approaches have been proposed that can simplify this task by learning the patterns from experts, such techniques are still susceptible to domain knowledge updates that could potentially change the patterns and lead to extra expert involvement. The use of Large Language Models (LLMs) has demonstrated a general cognitive ability, which has the potential to assist ontology alignment from the cognition level, thus obviating the need for costly expert involvement. However, the process by which the output of LLMs is generated can be opaque and thus the reliability and interpretability of such models is not always predictable. This paper proposes a dialogue model, in which multiple agents negotiate the correspondence between two knowledge sets with the support from an LLM. We demonstrate that this approach not only reduces the need for the involvement of a domain expert for ontology alignment, but that the results are interpretable despite the use of LLMs.

In the evolving landscape of portfolio management (PM), the fusion of advanced machine learning techniques with traditional financial methodologies has opened new avenues for innovation. Our study introduces a cutting-edge model combining deep reinforcement learning (DRL) with a non-stationary transformer architecture. This model is designed to decode complex patterns in financial time-series data, enhancing portfolio management strategies with deeper insights and robustness. It effectively tackles the challenges of data heterogeneity and market uncertainty, key obstacles in PM. Our approach integrates key macroeconomic indicators and targeted news sentiment analysis into its framework, capturing a comprehensive picture of market dynamics. This amalgamation of varied data types addresses the multifaceted nature of financial markets, enhancing the model’s ability to navigate the complexities of asset management. Rigorous testing demonstrates the model’s efficacy, highlighting the benefits of blending diverse data sources and sophisticated algorithmic approaches in mastering the nuances of PM.

Researchers typically leverage side information, such as social networks or the knowledge graph, to overcome the sparsity and cold start problem in collaborative filtering. To tackle the limitations of existing user interest modeling, we propose a knowledge-enhanced user multi-interest modeling for recommender systems (KEMIM). First, we utilize the user-item historical interaction as the knowledge graph's head entity to create a user's explicit interests and leverage the relationship path to expand the user's potential interests through connections in the knowledge graph. Second, considering the diversity of a user's interests, we adopt an attention mechanism to learn the user's attention to each historical interaction and each potential interest. Third, we combine the user's attribute features with interests to solve the cold start problem effectively. With the knowledge graph's structural data, KEMIM could describe the features of users at a fine granularity and provide explainable recommendation results to users. In this study, we conduct an in-depth empirical evaluation across three open datasets for two different recommendation tasks: Click-Through rate (CTR) prediction and Top-K recommendation. The experimental findings demonstrate that KEMIM outperforms several state-of-the-art baselines.

If we can predict the words a child is likely to learn next, it may lay the foundations for developing a tool to assist child language acquisition, especially for children experiencing language delay. Previous studies have demonstrated vocabulary predictions using neural network techniques and graph models; however, individually these models do not fully capture the complexities of language learning in infants. In this paper, we describe a multi-relationship-layer predictive model, based on a graph neural network. Our model combines vocabulary development over time with quantified connections between words calculated from fifteen different norms, incorporating an ensemble output stage to combine the predictions from each layer. We present results from each relationship layer and the most effective ensemble arrangement.

In the evolving landscape of Portfolio Management (PM), the fusion of advanced machine 1 learning techniques with traditional financial methodologies has opened new avenues for innovation. 2 Our study introduces a cutting-edge model combining Deep Reinforcement Learning (DRL) with 3 a Non-stationary Transformer architecture. This model is specifically designed to decode complex 4 patterns in financial time series data, enhancing portfolio management strategies with deeper insights 5 and robustness. It effectively tackles the challenges of data heterogeneity and market uncertainty, 6 key obstacles in PM. Our approach integrates key macroeconomic indicators and targeted news 7 sentiment analysis into its framework, capturing a comprehensive picture of market dynamics. This 8 amalgamation of varied data types addresses the multifaceted nature of financial markets, enhancing 9 the model’s ability to navigate through the complexities of asset management. Rigorous testing 10 demonstrates the model’s efficacy, highlighting the benefits of blending diverse data sources and 11 sophisticated algorithmic approaches in mastering the nuances of PM.12

This paper introduces a reinforcement learning-based framework designed to tackle dynamic pricing challenges in e-commerce. Prior research has predominantly concentrated on algorithm selection to enhance performance in dense data scenarios. However, many of these models fail to robustly address sparse data structures, such as low-traffic products, leading to the ‘cold-start’ problem [4]. Through numerical analysis, our framework offers innovative insights derived from the design of the reward function and integrates product clustering with pre-trained learning to mitigate this issue. As a result of this optimization, the performance of predictive models on sparse data is expected to see substantial improvement.

Automatic difficulty calibration (ADC) is the application of computational techniques to estimate the difficulty levels of assessment questions before administering them. Compared to traditional difficulty calibration approaches, ADC eliminates the need for pretesting, and minimises the time and efforts typically involved in manual calibration. In recent years, Ontology-based Automatic Question Generation (OAQG) has emerged as a powerful tool to generate assessment questions effortlessly and in massive numbers with minimal human intervention. Despite these benefits, not being able to control the characteristics of generated questions hinders their suitability to be used in pedagogical settings. However, much of the research up to now has tended to focus on the quantity of the questions rather than their quality. Indeed, most current automatic question generators produce simple questions which consist of a few facts, and simply test the recall of knowledge. Furthermore, the majority of existing frameworks are mostly technical and are not supported with a strong theoretical underpinning. In this paper, we propose a novel framework to assess the complexity of ontology-based, automatically generated questions. We discuss various aspects that are involved in determining the complexity of assessment questions, and attempt to quantify important characteristics of question complexity through the use of novel ontological metrics. To further support the plausibility of our computational framework, we shed light on its consistency with theories from education and cognitive psychology. This will provide us with a solid theoretical foundation which ensures that questions are generated according to principled methods that are grounded in theories of learning and cognition. The proposed approach is agnostic to different domains and independent of the question format, therefore, is highly general and applicable to a variety of contexts.

Video recommendation is vital for a video platform, which provides its users with videos they may be interested in. In this paper, we integrate users' ratings of videos in the video platform and community and crucial information data such as video category, director/actor, predict users' preference for videos through deep neural network, which could improve the accuracy of personalized recommendation. In addition, we use weighted force-directed Graph to show the relationship among users, videos, directors, and other elements, which could display the visualization of data elements and recommended results. Extensive experiments are conducted on three video datasets, and the experimental results demonstrate that the proposed method is more effective than several other recommendation methods.

In this paper, we describe the strategies used by our team, MLFC, that led us to achieve the 2nd place in the 15th edition of the Multi-Agent Programming Contest. The scenario used in the contest is an extension of the previous edition (14th) "Agents Assemble" wherein two teams of agents move around a 2D grid and compete to assemble complex block structures. We discuss the languages and tools used during the development of our team. Then, we summarise the main strategies that were carried over from our previous participation in the 14th edition and list the limitations (if any) of using these strategies in the latest contest edition. We also developed new strategies that were made specifically for the extended scenario: cartography (determining the size of the map); formal verification of the map merging protocol (to provide assurances that it works when increasing the number of agents); plan cache (efficiently scaling the number of planners); task achievement (forming groups of agents to achieve tasks); and bullies (agents that focus on stopping agents from the opposing team). Finally, we give a brief overview of our performance in the contest and discuss what we believe were our shortcomings.

Assessment quality and validity is heavily reliant on the quality of items included in an assessment or test. Difficulty is an essential factor that can determine items and tests’ overall quality. Therefore, item difficulty prediction is extremely important in any pedagogical learning environment. Data-driven approaches to item difficulty prediction are gaining more and more prominence, as demonstrated by the recent literature. In this paper, we provide a systematic review of data-driven approaches to item difficulty prediction. Of the 148 papers that were identified that cover item difficulty prediction, 38 papers were selected for the final analysis. A classification of the different approaches used to predict item difficulty is presented, together with the current practices for item difficulty prediction with respect to the learning algorithms used, and the most influential difficulty features that were investigated.

Purpose Screening for visual problems in stroke survivors is not standardised. Visual problems that remain undetected or poorly identified can create unmet needs for stroke survivors. We report the validation of a new Vision Impairment Screening Assessment (VISA) tool intended for use by the stroke team to improve identification of visual impairment in stroke survivors.Methods We conducted a prospective case cohort comparative study in four centres to validate the VISA tool against a specialist reference vision assessment. VISA is available in print or as an app (Medicines and Healthcare products Regulatory Agency regulatory approved); these were used equally for two groups. Both VISA and the comprehensive reference vision assessment measured case history, visual acuity, eye alignment, eye movements, visual field and visual inattention. The primary outcome measure was the presence or absence of visual impairment.Results Two hundred and twenty-one stroke survivors were screened. Specialist reference vision assessment was by experienced orthoptists. Full completion of screening and reference vision assessment was achieved for 201 stroke survivors. VISA print was completed for 101 stroke survivors; VISA app was completed for 100. Sensitivity and specificity of VISA print was 97.67% and 66.67%, respectively. Overall agreement was substantial; K=0.648. Sensitivity and specificity of VISA app was 88.31% and 86.96%, respectively. Overall agreement was substantial; K=0.690. Lowest agreement was found for screening of eye movement and near visual acuity.Conclusions This validation study indicates acceptability of VISA for screening of potential visual impairment in stroke survivors. Sensitivity and specificity were high indicating the accuracy of this screening tool. VISA is available in print or as an app allowing versatile uptake across multiple stroke settings.

In the distributed ontology alignment construction problem, two agents agree upon a meaningful subset of correspondences that map between their respective ontologies. However, an agent may be tempted to manipulate the negotiation in favour of a preferred alignment by misrepresenting the weight or confidence of the exchanged correspondences. Therefore such an agreement can only be meaningful if the agents can be incentivised to be honest when revealing information. We examine this problem and model it as a novel mechanism design problem on an edge-weighted bipartite graph, where each side of the graph represents each agent's private entities, and where each agent maintains a private set of valuations associated with its candidate correspondences. The objective is to find a matching (i.e. injective or one-to-one correspondences) that maximises the agents' social welfare. We study implementations in dominant strategies, and show that they should be solved optimally if truthful mechanisms are required. A decentralised version of the greedy allocation algorithm is then studied with a first-price payment rule, proving tight bounds on the Price of Anarchy and Stability.

Background Robots are increasingly being used to support management in certain areas of healthcare education. However, the potential application of robotics in palliative care education or simulation has not been explored. Aim This collaborative project between Computer Science and Palliative Care aimed to program a robot to convey emotion in response to human interaction, in order to develop a robotics program for potential use in palliative care education. Methods The Nao robot is an autonomous, programmable humanoid robot that is controlled by a Linux-based operating system. The robot has capabilities for voice recognition and sound localisation (in-built microphones), multilingual text-to-speech synthesis (in-built speakers) and vision, which include facial and shape recognition (in-built high definition cameras). The robot was programmed by a computer scientist to convey ten emotions (relaxed, anger, withdrawn/sad, lightly crying, heavy sobbing, happy/excited, scared, tired, laughing and dancing) through its posture, movement and speech, in response to human-voiced questions and interaction. Results The robot was successfully programmed to convey the ten target emotions in response to direct questions posed by a human subject. Discussions around the robot’s displayed emotions were explored (e.g. “why are you sad?”) to assess the potential of human-computer interaction. The robot continues to acquire a growing lexicon of vocabulary, in addition to an increasing number actions and responses. The robot acts both autonomously and through direct instruction of the operator. Conclusions We have successfully programmed a robot to interact with humans and display emotional responses. This technology could potentially be used to develop innovative ways to engage individuals in discussion about palliative care issues and create opportunities to use robots for interactive educational activity. Consequently, further research can explore the potential to use robotic technology in palliative care for education, and to promote discussion with the public (e.g. children) and healthcare professionals.

The aim of the ontology alignment problem is to find meaningful correspondences between two ontologies represented as collections of entities. This problem can be modelled as a novel mechanism design problem on an edge-weighted bipartite graph, where each side of the graph holds each agent's private entities, and the objective is to maximise the agents' social welfare. Having studied implementation in dominant strategies with and without payments, we report on findings that for truthful mechanisms, these problems need to be solved optimally. We also study greedy allocation rules with a first-price payment rule, and implementation in pure, mixed and Bayesian Nash equilibria, and have found tight bounds on the price of anarchy and stability.

Software agents, acting on behalf of humans, have been identified as an important solution for future electronic markets. Such agents can make their own decisions based on given prior preferences and the market environment. These preferences can be described using Web Ontology Languages (OWL), while the market mechanism can be represented in a machine-understandable way by utilizing the technique of Semantic Web Services (SWS). Besides, SWS enables agents to automatically discover, select, compose and invoke services. To extend the dependability and interactivity of SWS, we have utilized dialogue games and the Proof-Carrying Code (PCC) to enable buyers to interact with sellers, so that desirable properties for an online auction market can be automatically certified. Our decision-making framework combines formal proofs with informal evidence collected by web services in a dialogue game between a seller and a buyer. We have implemented our approach and experimental results have demonstrated the feasibility as well as the validity of this framework as an enabler for a buyer agent to enter or not an online auction.

An ontology signature (set of entities) can express more than its constituent concept, role and individual names, since rewriting permits defined entities to be replaced by syntactically different, albeit semantically equivalent definitions. Identifying whether a given signature permits the definition of a particular entity is a well-understood problem, while determining the smallest (minimal) signature that covers a set of entities (i.e. a task signature) poses a challenge: the complete set of alternative definitions, or even just their signature, needs to be obtained, and all combinations of such definition signatures need to be explored, for each of the entities under consideration. In this paper, we present and empirically evaluate our novel approach for efficiently computing an approximation of minimal signature cover sets.