Trends and Applications in Knowledge Discovery and Data Mining

This study presents a comprehensive evaluation of Microsoft Research’s Orca 2, a small yet potent language model, in the context of Retrieval Augmented Generation (RAG). The research involved comparing Orca 2 with other significant models such as Llama-2, GPT-3.5-Turbo, and GPT-4, particularly focusing on its application in RAG. Key metrics, included faithfulness, answer relevance, overall score, and inference speed, were assessed. Experiments conducted on high-specification PCs revealed Orca 2’s exceptional performance in generating high quality responses and its efficiency on consumer-grade GPUs, underscoring its potential for scalable RAG applications. This study highlights the pivotal role of smaller, efficient models like Orca 2 in the advancement of conversational AI and their implications for various IT infrastructures. The source codes and datasets of this paper are accessible here (https://​github.​com/​inflaton/​Evaluation-of-Orca-2-for-RAG.).

Despite significant achievements in numerous real-world applications, the black-box nature hinders GNNs from being adopted in high-stake decision situations. This paper introduces an advanced interpretable graph classification approach grounded on concept-focused structural correspondence. Our method harnesses the inherent interpretability of the case-based reasoning methodology and utilizes the Earth Mover’s Distance (EMD) to determine structural similarities between graphs. Enhanced by a concept-centric node-weighting scheme, our refined EMD prioritizes nodes within frequently observed essential subgraphs. The enhanced EMD metric is pivotal to our interpretable non-parametric predictor, which utilizes it to derive predictions based on the proximity of input graphs to reference graphs. A dual-phase strategy ensures efficiency by selecting references using Euclidean distance and refining via EMD. Our framework integrates various explanation modalities catering to diverse needs for prediction explanations, elucidating the model’s decision-making processes. Empirical evaluations and a specific user study affirm our approach’s robustness and applicability.

Numerous recommendation algorithms leverage weblogs, employing strategies such as collaborative filtering, content-based filtering, and hybrid methods to provide personalized recommendations to users. Weblogs, comprised of records detailing user activities on any website, offer valuable insights into user preferences, behavior, and interests. Despite the wealth of information weblogs provide, extracting relevant features requires extensive feature engineering. The intricate nature of the data also poses a challenge for interpretation, especially for non-experts. Additionally, they often fall short of capturing visual details and contextual nuances that influence user choices. In the present study, we introduce a sophisticated and interactive recommendation framework denoted as InteraRec, which diverges from conventional approaches that exclusively depend on weblogs for recommendation generation. This framework provides recommendations by capturing high-frequency screenshots of web pages as users navigate through a website. Leveraging advanced multimodal large language models (MLLMs), we extract valuable insights into user preferences from these screenshots by generating a user profile summary. Subsequently, we employ the InteraRec framework to extract relevant information from the summary to generate optimal recommendations. Through extensive experiments, we demonstrate the remarkable effectiveness of our recommendation system in providing users with valuable and personalized offerings.

With the continuous development of technology, we have the ability to fully record all aspects of data information of every individual in the society, so how to utilize this information to create greater value is becoming more and more important. Compared with the traditional static community detection, the study of dynamic community detection is more in line with the real situation in the society. Thus, in this paper, a method that can utilize the information of diversified dynamic community networks is proposed, i.e., Dynamic Community Detection Method based on Multidimensional Feature Information of Community (Dcdmf), which utilizes neural networks with strong learning and adaptive capabilities, the ability to automatically extract useful features and process complex data, and the ability to process the graph nodes and the data between the nodes of the dynamic community network, and the ability to real-time adjust the current community representation data based on historical information, and record the current community representation data for the next moment of community data. The experimental results in the paper show that the method has a certain degree of effectiveness.

With the advent of social network technology, the influence of collective opinions has significantly impacted business, marketing, and fundraising. Particularly in the blockchain space, Initial Coin Offerings (ICOs) gain substantial exposure across various online platforms. Yet, the intricate relationships among these elements remain largely unexplored. This study aims to investigate the relationships between social media sentiment, engagement metrics, and ICO success. We hypothesize a positive correlation between favorable sentiment in ICO-related tweets and overall project success. Additionally, we recognize social media engagement indicators (mentions, retweets, likes, follower counts) as critical factors affecting ICO performance. Employing machine learning techniques, we conduct sentiment analysis on tweets, discerning emotional nuances and categorizing expressions as positive or negative. Employing established classification methods, we further analyze engagement data to reveal its impact on ICO interest and awareness. Our research findings offer insights into the predictive potential of social media strategies for ICO success and underscore the importance of investor sentiment and engagement in the volatile cryptocurrency landscape. These insights provide actionable guidance for aspiring crypto founders in formulating effective business development strategies.

The source codes and datasets of this paper are accessible at GitHub: https://​github.​com/​inflaton/​Success-Indicators-of-Initial-Coin-Offerings.

Session-based recommendation (SBR) is a complex endeavor focused on predicting a user’s next interesting item based on his sessions (i.e., short interaction sequences). The existing SBR models usually learn only one aspect of the sessions, either the static information (e.g., spatial structure of the graph, node similarity) or the dynamic information (e.g., temporal information, position information), so that the rich information embedded in session can’t be fully exploited. A new enhanced graph neural network model based on both static and dynamic information (called EGNN-SDI) is proposed, which constructs and uses a global graph that cooperates with an undirected and a directed session graph to learn the global and local static information, as well as the dynamic information within sessions. Based on this, we propose a new node encoding layer called SDI.

In short, we use GCN and GGNN separately to learn static and dynamic information, respectively. An inverse position matrix is also introduced to learn the relative positional information within the session. By using linear combination and attention mechanisms, the enhanced item representation enables the generation of more accurate session representations for the next item to be recommended. Evaluation experiments are performed on three widely used datasets, consistently showcasing the superiority of EGNN-SDI compared to existing baseline models. Our model’s implementation can be accessed via https://​anonymous.​4open.​science/​r/​EGNN-SDI-4B8C.

Depth exploration and display of the potential correlation of innovation point can be helpful for relevant work such as field innovation discovery and field literature innovation evaluation. First, on the basis of the concept of academic innovation chain, the construction method of academic innovation chain based on multi-level clustering is proposed. Second, combining the text feature mining algorithms of tf-idf, LDA, doc2vec and the Kmeans text clustering algorithm, 639 literatures in the field of “knowledge element” are taken as examples clustering from the three levels of word frequency, topic and semantic. Final, fusion rule method with ALBERT pre-training model to extract the innovation points of literature, then the construction of academic innovation chain is realized. The academic innovation chain connects the originally isolated innovation point linearly. It provides certain references for the research of innovation evaluation and innovation metrics.

While manual transcription tools exist, music enthusiasts, including amateur singers, still encounter challenges when transcribing performances into sheet music. This paper addresses the complex task of translating music audio into music sheets, particularly challenging in the intricate field of choral arrangements where multiple voices intertwine. We propose DLVS4Audio2Sheet, a novel method leveraging advanced deep learning models, Open-Unmix and Band-Split Recurrent Neural Networks (BSRNN), for vocal separation. DLVS4Audio2Sheet segments choral audio into individual vocal sections and selects the optimal model for further processing, aiming towards audio into music sheet conversion. We evaluate DLVS4Audio2Sheet’s performance using these deep learning algorithms and assess its effectiveness in producing isolated vocals suitable for notated scoring music conversion. By ensuring superior vocal separation quality through model selection, DLVS4Audio2Sheet enhances audio into music sheet conversion. This research contributes to the advancement of music technology by thoroughly exploring state-of-the-art models, methodologies, and techniques for converting choral audio into music sheets. Code and datasets are available at: https://​github.​com/​DevGoliath/​DLVS4Audio2Sheet​.

Stress and depression have emerged as prevalent challenges in contemporary society, deeply intertwined with the complexities of modern life. This paper delves into the multifaceted nature of these phenomena, exploring their intricate relationship with various socio-cultural, technological, and environmental factors through the application of neurosymbolic AI to social media content. Through a quantitative and qualitative analysis of results, we elucidate the profound impact of technological advancements on information processing, work culture, and social dynamics, highlighting the role of digital connectivity in exacerbating stressors. Economic pressures and social isolation further compound these challenges, contributing to a pervasive sense of unease and disconnection. Environmental stressors, including climate change, add another layer of complexity, fostering existential concerns about the future. Moreover, the persistent stigma surrounding mental health perpetuates a cycle of silence and suffering, hindering access to support and resources. Addressing these issues necessitates a holistic approach, encompassing societal changes, policy interventions, and individual coping strategies.

Over the past decade, Time Series Classification (TSC) has gained an increasing attention. While various methods were explored, deep learning – particularly through Convolutional Neural Networks (CNNs) –stands out as an effective approach. However, due to the limited availability of training data, defining a foundation model for TSC that overcomes the overfitting problem is still a challenging task. The UCR archive, encompassing a wide spectrum of datasets ranging from motion recognition to ECG-based heart disease detection, serves as a prime example for exploring this issue in diverse TSC scenarios. In this paper, we address the overfitting challenge by introducing pre-trained domain foundation models. A key aspect of our methodology is a novel pretext task that spans multiple datasets. This task is designed to identify the originating dataset of each time series sample, with the goal of creating flexible convolution filters that can be applied across different datasets. The research process consists of two phases: a pre-training phase where the model acquires general features through the pretext task, and a subsequent fine-tuning phase for specific dataset classifications. Our extensive experiments on the UCR archive demonstrate that this pre-training strategy significantly outperforms the conventional training approach without pre-training. This strategy effectively reduces overfitting in small datasets and provides an efficient route for adapting these models to new datasets, thus advancing the capabilities of deep learning in TSC.

Recommendation systems play a pivotal role in diverse real-world scenarios, offering personalized suggestions to users. Despite their significance, the cold-start problem poses a formidable challenge for both conventional recommendation systems and sequential recommendations. The entry of new users or items into the system inhibits accurate recommendations due to the lack of prior interactions. To tackle this issue, researchers have delved into employing meta-learning techniques. However, predicting the time interval of interactions for new users remains a persistent challenge. This paper presents an innovative approach to forecasting the next item and the associated time interval of new user and item interactions. Our method leverages meta-learning techniques within the context of a dynamic graph structure. It showcases superior performance when compared to previous methods using three benchmark datasets, effectively addressing the cold-start problem. The instructive experiments underscore the efficacy of our proposed method in handling the next item and time interval prediction, thereby contributing to the advancement of sequential recommendation systems.

In time-series machine learning, the challenge of obtaining labeled data has spurred interest in using unlabeled data for model training. Current research primarily focuses on deep multi-task learning, emphasizing the hard parameter-sharing approach. Importantly, when correlations between tasks are weak, indiscriminate parameter sharing can lead to learning interference. Consequently, we introduce a novel framework called DPS, which separates training into dependency-learning and parameter-sharing phases. This structure allows the model to manage knowledge sharing between tasks dynamically. Additionally, we introduce a loss function to align neuron functionalities across tasks, addressing learning interference. Through experiments on real-world datasets, we demonstrate the superiority of DPS over baselines. Moreover, our results shed light on the impacts of the two designed training phases, validating that DPS consistently ensures a degree of learning stability.

We present a study on the non-stationary stochastic multi-armed bandit (MAB) problem, which is relevant for addressing real-world challenges related to sequential decision-making. Our work involves a thorough analysis of state-of-the-art algorithms in dynamically changing environments. To address the limitations of existing methods, we propose the Concept Drift Adaptive Bandit (CDAB) framework, which aims to capture and predict potential future concept drift patterns in reward distribution, allowing for better adaptation in non-stationary environments. We conduct extensive numerical experiments to evaluate the effectiveness of the CDAB approach in comparison to both stationary and non-stationary state-of-the-art baselines. Our experiments involve testing on both artificial datasets and real-world data under different types of changing environments. The results show that the CDAB approach exhibits strong empirical performance, outperforming existing methods in all versions tested.