Pub Date : 2025-01-31DOI: 10.1016/j.dss.2025.114402
David Martens , James Hinns , Camille Dams , Mark Vergouwen , Theodoros Evgeniou
Existing Explainable AI (XAI) approaches, such as the widely used SHAP values or counterfactual (CF) explanations, are arguably often too technical for users to understand and act upon. To enhance comprehension of explanations of AI decisions and the overall user experience, we introduce XAIstories, which leverage Large Language Models (LLMs) to provide narratives about how AI predictions are made: SHAPstories based on SHAP and CFstories on CF explanations. We study the impact of our approach on users’ experience and understanding of AI predictions. Our results are striking: over 90% of the surveyed general audience finds the narratives generated by SHAPstories convincing, and over 78% for CFstories, in a tabular data experiment. More than 75% of the respondents in an image experiment find CFstories more or equally convincing as their own crafted stories. We also find that the generated stories help users to more accurately summarize and understand AI decisions than they do when only SHAP values are provided. The results indicate that combining LLM generated stories with current XAI methods is a promising and impactful research direction.
{"title":"Tell me a story! Narrative-driven XAI with Large Language Models","authors":"David Martens , James Hinns , Camille Dams , Mark Vergouwen , Theodoros Evgeniou","doi":"10.1016/j.dss.2025.114402","DOIUrl":"10.1016/j.dss.2025.114402","url":null,"abstract":"<div><div>Existing Explainable AI (XAI) approaches, such as the widely used SHAP values or counterfactual (CF) explanations, are arguably often too technical for users to understand and act upon. To enhance comprehension of explanations of AI decisions and the overall user experience, we introduce XAIstories, which leverage Large Language Models (LLMs) to provide narratives about how AI predictions are made: SHAPstories based on SHAP and CFstories on CF explanations. We study the impact of our approach on users’ experience and understanding of AI predictions. Our results are striking: over 90% of the surveyed general audience finds the narratives generated by SHAPstories convincing, and over 78% for CFstories, in a tabular data experiment. More than 75% of the respondents in an image experiment find CFstories more or equally convincing as their own crafted stories. We also find that the generated stories help users to more accurately summarize and understand AI decisions than they do when only SHAP values are provided. The results indicate that combining LLM generated stories with current XAI methods is a promising and impactful research direction.</div></div>","PeriodicalId":55181,"journal":{"name":"Decision Support Systems","volume":"191 ","pages":"Article 114402"},"PeriodicalIF":6.7,"publicationDate":"2025-01-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143193315","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-01-29DOI: 10.1016/j.dss.2025.114406
Hai Wei , Ying Yang , Yu-Wang Chen
Customers' perception of review helpfulness entails a cognitive reasoning process influenced by the contextual information of reviews including product descriptions and review neighbors. Current studies on helpfulness prediction primarily focus on static features of individual reviews, neglecting the dynamic interaction among products, reviews and their contextual neighbors. To address this gap, we propose a theory-driven deep learning model for multimodal review helpfulness prediction (DeepMRHP-MCR). The model can collectively simulate human cognitive processes when voting on whether a review is helpful. Specifically, this study presents a multi-level cognitive reasoning mechanism that reconciles the inconsistencies among product descriptions, reviews and their neighbors from the modality, individual and contextual level, respectively. A case study is conducted on the real-world datasets collected from Amazon.com. Empirical results show that the proposed model can improve the quality and interpretability of review prediction process, and present a deep comprehension of consumer's cognitive decision-making process when evaluating reviews.
{"title":"Multimodal review helpfulness prediction with a multi-level cognitive reasoning mechanism: A theory-driven graph learning model","authors":"Hai Wei , Ying Yang , Yu-Wang Chen","doi":"10.1016/j.dss.2025.114406","DOIUrl":"10.1016/j.dss.2025.114406","url":null,"abstract":"<div><div>Customers' perception of review helpfulness entails a cognitive reasoning process influenced by the contextual information of reviews including product descriptions and review neighbors. Current studies on helpfulness prediction primarily focus on static features of individual reviews, neglecting the dynamic interaction among products, reviews and their contextual neighbors. To address this gap, we propose a theory-driven deep learning model for multimodal review helpfulness prediction (DeepMRHP-MCR). The model can collectively simulate human cognitive processes when voting on whether a review is helpful. Specifically, this study presents a multi-level cognitive reasoning mechanism that reconciles the inconsistencies among product descriptions, reviews and their neighbors from the modality, individual and contextual level, respectively. A case study is conducted on the real-world datasets collected from <span><span>Amazon.com</span><svg><path></path></svg></span>. Empirical results show that the proposed model can improve the quality and interpretability of review prediction process, and present a deep comprehension of consumer's cognitive decision-making process when evaluating reviews.</div></div>","PeriodicalId":55181,"journal":{"name":"Decision Support Systems","volume":"191 ","pages":"Article 114406"},"PeriodicalIF":6.7,"publicationDate":"2025-01-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143193316","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-01-28DOI: 10.1016/j.dss.2025.114409
Xuelong Chen , Yiping Chen , Guojie Yin
The anonymity and convenience of social media platforms enable the public to express and even vent themselves, which drives a surge of cyberviolence behaviors (CVB). Recent advances in machine learning, especially in deep learning, have drastically benefited CVB detection. However, despite the wide use of state-of-the-art deep-learning models, previous studies only analyzed each post/comment for the presence of (obfuscated) abusive text, which is not comprehensive and exact because the content posted online may not necessarily include negative words. In complex and conflicting situations, people may overlook implicit violence, leading to failures in situational judgment. Herein, we designed a well-grounded and explainable deep-learning framework based on the theory of planned behavior (TPB) to explore the motivations behind CVB to better detect it. Specifically, we constructed a systematic and comprehensive suite of computable features grounded in TPB and then proposed a novel model named Multilevel and Multiattribute Embedding CVB detection model considering Dual-view Contextual Information. Our framework detected implicit and explicit CVB with macro F1 scores of >88.67 %, outperforming state-of-the-art methods. We further provided differentiated strategies according to the scale and distribution of different classes of CVB and proposed related management implications. Our study sheds light on platform operations in managing online content and mitigating the risk of governance cost wastage and deterioration of the cyber ecosystem.
{"title":"Exploring the motivations behind behavior: A theory-driven deep-learning framework for cyberviolence behavior detection","authors":"Xuelong Chen , Yiping Chen , Guojie Yin","doi":"10.1016/j.dss.2025.114409","DOIUrl":"10.1016/j.dss.2025.114409","url":null,"abstract":"<div><div>The anonymity and convenience of social media platforms enable the public to express and even vent themselves, which drives a surge of cyberviolence behaviors (CVB). Recent advances in machine learning, especially in deep learning, have drastically benefited CVB detection. However, despite the wide use of state-of-the-art deep-learning models, previous studies only analyzed each post/comment for the presence of (obfuscated) abusive text, which is not comprehensive and exact because the content posted online may not necessarily include negative words. In complex and conflicting situations, people may overlook implicit violence, leading to failures in situational judgment. Herein, we designed a well-grounded and explainable deep-learning framework based on the theory of planned behavior (TPB) to explore the motivations behind CVB to better detect it. Specifically, we constructed a systematic and comprehensive suite of computable features grounded in TPB and then proposed a novel model named <strong>M</strong>ultilevel and <strong>M</strong>ultiattribute <strong>E</strong>mbedding CVB detection model considering <strong>D</strong>ual-view <strong>C</strong>ontextual <strong>I</strong>nformation. Our framework detected implicit and explicit CVB with macro F1 scores of >88.67 %, outperforming state-of-the-art methods. We further provided differentiated strategies according to the scale and distribution of different classes of CVB and proposed related management implications. Our study sheds light on platform operations in managing online content and mitigating the risk of governance cost wastage and deterioration of the cyber ecosystem.</div></div>","PeriodicalId":55181,"journal":{"name":"Decision Support Systems","volume":"190 ","pages":"Article 114409"},"PeriodicalIF":6.7,"publicationDate":"2025-01-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143160447","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-01-27DOI: 10.1016/j.dss.2025.114405
Yuxin Cai , Xiayu Chen , Shaobo Wei
Online question and answer (Q&A) communities rely on the general audience or the question asker to determine the best answer. However, limited attention has been directed toward understanding the influence of the general audience-favored answer (i.e., most popular answer) on the question asker-selected best answer (i.e., best answer). This study examines whether and how the general audience-favored answer influences the question asker-selected best answer. Drawing upon uncertainty reduction theory (URT), this paper investigates how three uncertainty reduction strategies (i.e. question askers' tenure, social closeness between the question asker and question answerer, and cultural tightness of the question asker's region) moderate the relationship between the general audience-favored answer and the question asker-selected best answer. To test the theoretical model, we used a dataset from an online Q&A community comprising 161,695 observations. Our results reveal that the general audience-favored answer more likely leads to the question asker-selected best answer. Furthermore, we find that the question asker's tenure and the social closeness between the question asker and question answerer negatively moderate the above relationship, while the cultural tightness of question asker's region positively moderates the above relationship. This research offers a new perspective on the mechanisms through which the general audience-favored answer leads to question asker-selected best answer. By identifying the critical roles of uncertainty reduction strategies during the best answer selection process, our research provides valuable insights for online Q&A community managers to optimize user engagement and satisfaction.
{"title":"Does the most popular answer lead to the best answer: The moderating roles of tenure, social closeness, and cultural tightness","authors":"Yuxin Cai , Xiayu Chen , Shaobo Wei","doi":"10.1016/j.dss.2025.114405","DOIUrl":"10.1016/j.dss.2025.114405","url":null,"abstract":"<div><div>Online question and answer (Q&A) communities rely on the general audience or the question asker to determine the best answer. However, limited attention has been directed toward understanding the influence of the general audience-favored answer (i.e., most popular answer) on the question asker-selected best answer (i.e., best answer). This study examines whether and how the general audience-favored answer influences the question asker-selected best answer. Drawing upon uncertainty reduction theory (URT), this paper investigates how three uncertainty reduction strategies (i.e. question askers' tenure, social closeness between the question asker and question answerer, and cultural tightness of the question asker's region) moderate the relationship between the general audience-favored answer and the question asker-selected best answer. To test the theoretical model, we used a dataset from an online Q&A community comprising 161,695 observations. Our results reveal that the general audience-favored answer more likely leads to the question asker-selected best answer. Furthermore, we find that the question asker's tenure and the social closeness between the question asker and question answerer negatively moderate the above relationship, while the cultural tightness of question asker's region positively moderates the above relationship. This research offers a new perspective on the mechanisms through which the general audience-favored answer leads to question asker-selected best answer. By identifying the critical roles of uncertainty reduction strategies during the best answer selection process, our research provides valuable insights for online Q&A community managers to optimize user engagement and satisfaction.</div></div>","PeriodicalId":55181,"journal":{"name":"Decision Support Systems","volume":"191 ","pages":"Article 114405"},"PeriodicalIF":6.7,"publicationDate":"2025-01-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143193314","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-01-25DOI: 10.1016/j.dss.2025.114403
Alfred Z. Liu , Angela Xia Liu , Kexin Zhao
This research investigates the disclosure of IT-related risk factors in U.S. public firms' periodic SEC filings. Drawing upon the Resource-Based View theory, we propose that a firm's IT capability determines the disclosure of its overall IT-related risk factors. We employ a machine learning-enhanced dictionary that captures emerging IT keywords from newly filed corporate reports to quantify the scope and specificity of such disclosures. Our findings indicate that IT capability enhances IT-related risk factor disclosures in general and specifically in response to adverse IT events, such as data breaches. We also find that disclosing IT-related risk factors reduces a firm's perceived risk and enhances its shareholder value. Our research underscores the critical yet under-researched role of IT capability in shaping disclosures of IT-related risk factors and highlights such disclosures' informational value to investors.
{"title":"Disclosure of IT-related risk factors in corporate filings","authors":"Alfred Z. Liu , Angela Xia Liu , Kexin Zhao","doi":"10.1016/j.dss.2025.114403","DOIUrl":"10.1016/j.dss.2025.114403","url":null,"abstract":"<div><div>This research investigates the disclosure of IT-related risk factors in U.S. public firms' periodic SEC filings. Drawing upon the Resource-Based View theory, we propose that a firm's IT capability determines the disclosure of its overall IT-related risk factors. We employ a machine learning-enhanced dictionary that captures emerging IT keywords from newly filed corporate reports to quantify the scope and specificity of such disclosures. Our findings indicate that IT capability enhances IT-related risk factor disclosures in general and specifically in response to adverse IT events, such as data breaches. We also find that disclosing IT-related risk factors reduces a firm's perceived risk and enhances its shareholder value. Our research underscores the critical yet under-researched role of IT capability in shaping disclosures of IT-related risk factors and highlights such disclosures' informational value to investors.</div></div>","PeriodicalId":55181,"journal":{"name":"Decision Support Systems","volume":"190 ","pages":"Article 114403"},"PeriodicalIF":6.7,"publicationDate":"2025-01-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143160450","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-01-23DOI: 10.1016/j.dss.2025.114407
Lin Yuan , Chaoyue Gao , Alvin Chung Man Leung , Qiang Ye
As a nascent market in recent years, the NFT market has been widely scrutinized for its significant market bubble. To help investors make more informed trading decisions, several NFT marketplaces have introduced features that display the rarity information of NFTs directly on their interfaces. Existing literature on the rarity effect suggests that this feature generally increases trading activity. However, in the unique context of the NFT marketplace, its impact on user trading behavior remains an open question. This study focuses on the event where Rarible began displaying rarity information for profile picture (PFP) NFTs on its platform. Utilizing the theoretical perspective of dual process theory, we conceptualize the introduction of the rarity label as enhanced digital embeddedness. By using other NFT collections on the platform that have rarity information but do not display rarity rank labels as the control group, this study employs a rigorous Difference-in-Differences design. We find that this event leads to a decrease in both trading volume and trading price, primarily for lower-ranked NFTs, small-size collections, recent NFTs rather than top-ranked, large-size collections, established NFTs. Additional time-varying analysis also explains the asynchronous changes in price and trading volume. This study enriches the literature on the NFT marketplaces and the rarity effect, extends the application of dual process theory, and provides practical decision support for market regulators, managers, and platform users.
{"title":"Enhanced digital embeddedness and bubble mitigation in NFT marketplaces: The impact of rarity rank on user trading behavior","authors":"Lin Yuan , Chaoyue Gao , Alvin Chung Man Leung , Qiang Ye","doi":"10.1016/j.dss.2025.114407","DOIUrl":"10.1016/j.dss.2025.114407","url":null,"abstract":"<div><div>As a nascent market in recent years, the NFT market has been widely scrutinized for its significant market bubble. To help investors make more informed trading decisions, several NFT marketplaces have introduced features that display the rarity information of NFTs directly on their interfaces. Existing literature on the rarity effect suggests that this feature generally increases trading activity. However, in the unique context of the NFT marketplace, its impact on user trading behavior remains an open question. This study focuses on the event where Rarible began displaying rarity information for profile picture (PFP) NFTs on its platform. Utilizing the theoretical perspective of dual process theory, we conceptualize the introduction of the rarity label as enhanced digital embeddedness. By using other NFT collections on the platform that have rarity information but do not display rarity rank labels as the control group, this study employs a rigorous Difference-in-Differences design. We find that this event leads to a decrease in both trading volume and trading price, primarily for lower-ranked NFTs, small-size collections, recent NFTs rather than top-ranked, large-size collections, established NFTs. Additional time-varying analysis also explains the asynchronous changes in price and trading volume. This study enriches the literature on the NFT marketplaces and the rarity effect, extends the application of dual process theory, and provides practical decision support for market regulators, managers, and platform users.</div></div>","PeriodicalId":55181,"journal":{"name":"Decision Support Systems","volume":"190 ","pages":"Article 114407"},"PeriodicalIF":6.7,"publicationDate":"2025-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143160440","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-01-23DOI: 10.1016/j.dss.2025.114404
Shaoze Cui , Ruize Gao , Junwei Kuang , Liang Yang , Huaxin Qiu , Xiaowen Wei
In the domain of medical services, patients are frequently readmitted shortly after discharge due to inadequate discharge planning or relapses of their illnesses. Such occurrences not only deplete valuable medical resources but also compromise patient satisfaction with the medical care they receive. To address this issue, we propose an interpretable imbalance ensemble classification method incorporating multi-view perturbation to evaluate the risk of patient readmission. Our study introduces a novel multi-view perturbation technique to bolster the model's generalization capabilities. Furthermore, we propose a more robust ensemble strategy based on Evidential Reasoning (EVR) rules, which enhances the stability of the ensemble learning model's fusion outcomes. Additionally, recognizing the impact of sensitive parameters on model performance, we present a parameter optimization approach utilizing the Differential Evolution (DE) algorithm, which balances model predictive accuracy and computational efficiency within the fitness function. Empirical results using real-world medical data indicate that our proposed method accurately identifies patients at high risk of readmission and surpasses current state-of-the-art methods in risk assessment.
{"title":"An interpretable imbalance ensemble classification method for readmission risk assessment incorporating multi-view perturbation and SHAP analysis","authors":"Shaoze Cui , Ruize Gao , Junwei Kuang , Liang Yang , Huaxin Qiu , Xiaowen Wei","doi":"10.1016/j.dss.2025.114404","DOIUrl":"10.1016/j.dss.2025.114404","url":null,"abstract":"<div><div>In the domain of medical services, patients are frequently readmitted shortly after discharge due to inadequate discharge planning or relapses of their illnesses. Such occurrences not only deplete valuable medical resources but also compromise patient satisfaction with the medical care they receive. To address this issue, we propose an interpretable imbalance ensemble classification method incorporating multi-view perturbation to evaluate the risk of patient readmission. Our study introduces a novel multi-view perturbation technique to bolster the model's generalization capabilities. Furthermore, we propose a more robust ensemble strategy based on Evidential Reasoning (EVR) rules, which enhances the stability of the ensemble learning model's fusion outcomes. Additionally, recognizing the impact of sensitive parameters on model performance, we present a parameter optimization approach utilizing the Differential Evolution (DE) algorithm, which balances model predictive accuracy and computational efficiency within the fitness function. Empirical results using real-world medical data indicate that our proposed method accurately identifies patients at high risk of readmission and surpasses current state-of-the-art methods in risk assessment.</div></div>","PeriodicalId":55181,"journal":{"name":"Decision Support Systems","volume":"190 ","pages":"Article 114404"},"PeriodicalIF":6.7,"publicationDate":"2025-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143160449","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-01-18DOI: 10.1016/j.dss.2025.114401
Haoran Wang , Zhen-Song Chen , Mingjie Fang , Yilong Wang , Feng Liu
Flash sales are a widely adopted e-commerce marketing strategy that operate over a brief period, offering limited-time discounts, special promotions, or clearance items to create a sense of urgency and promote rapid sales. This study proposes panoramic sales insight (PSI), a multimodal revenue forecasting framework designed to improve the accuracy of revenue predictions for flash sales. Using historical flash sales data from the fast fashion retailer Shein, the proposed PSI framework integrates both structured and unstructured data, utilizing a text–image fusion module to fuse features from product images and text descriptions and a deep neural network to forecast revenue. The text features are extracted using bidirectional encoder representations from transformers (BERT), the product image features are extracted using a vision transformer (ViT), and review keyword extraction is conducted using Fumeus. Multimodal fusion then integrates these features to deliver accurate revenue forecasting. Controlled experiments evaluate the performance of each module within the PSI framework, while ablation analysis confirms the robustness of PSI. This study provides valuable insights for managers, enabling more accurate revenue forecasting and improving the effectiveness of flash sales.
{"title":"Panoramic sales insight: Using multimodal fusion to improve the effectiveness of flash sales","authors":"Haoran Wang , Zhen-Song Chen , Mingjie Fang , Yilong Wang , Feng Liu","doi":"10.1016/j.dss.2025.114401","DOIUrl":"10.1016/j.dss.2025.114401","url":null,"abstract":"<div><div>Flash sales are a widely adopted e-commerce marketing strategy that operate over a brief period, offering limited-time discounts, special promotions, or clearance items to create a sense of urgency and promote rapid sales. This study proposes panoramic sales insight (PSI), a multimodal revenue forecasting framework designed to improve the accuracy of revenue predictions for flash sales. Using historical flash sales data from the fast fashion retailer Shein, the proposed PSI framework integrates both structured and unstructured data, utilizing a text–image fusion module to fuse features from product images and text descriptions and a deep neural network to forecast revenue. The text features are extracted using bidirectional encoder representations from transformers (BERT), the product image features are extracted using a vision transformer (ViT), and review keyword extraction is conducted using Fumeus. Multimodal fusion then integrates these features to deliver accurate revenue forecasting. Controlled experiments evaluate the performance of each module within the PSI framework, while ablation analysis confirms the robustness of PSI. This study provides valuable insights for managers, enabling more accurate revenue forecasting and improving the effectiveness of flash sales.</div></div>","PeriodicalId":55181,"journal":{"name":"Decision Support Systems","volume":"190 ","pages":"Article 114401"},"PeriodicalIF":6.7,"publicationDate":"2025-01-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143160448","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-01-03DOI: 10.1016/j.dss.2025.114400
Xicheng Yin , Jing Li , Kevin Zhu , Wei Wang , Hongwei Wang
Given the “profit-seeking” behavior of task solvers and the “quality-seeking” focus of solution seekers in knowledge-intensive crowdsourcing contests, task recommender systems must manage the trade-off between their respective benefits. This study proposes a multitask deep learning model with a multigate hybrid expert structure to jointly model solver preference and ability, thereby balancing bilateral benefits. The knowledge source for participation and performance prediction tasks are grounded in expectancy theory and performance theory, respectively. Linear and deep neural network (DNN) modules are integrated to enhance both memorization and generalization capabilities. By incorporating gating networks, the model effectively captures correlations between the two prediction tasks, balances intertask weights, and allows each task to learn features in different ways using linear and DNN modules. Additionally, our method addresses sample selection bias and data sparsity issues through feature transfer learning, leveraging the sequential pattern between participation and winning. Cross-validation experiments on Kaggle data demonstrate the model effectiveness, provide data-driven decision support for task recommendation and resource allocation in knowledge-intensive crowdsourcing platforms.
{"title":"Willing and able: Task recommendation with a trade-off of the bilateral benefits for knowledge-intensive crowdsourcing","authors":"Xicheng Yin , Jing Li , Kevin Zhu , Wei Wang , Hongwei Wang","doi":"10.1016/j.dss.2025.114400","DOIUrl":"10.1016/j.dss.2025.114400","url":null,"abstract":"<div><div>Given the “profit-seeking” behavior of task solvers and the “quality-seeking” focus of solution seekers in knowledge-intensive crowdsourcing contests, task recommender systems must manage the trade-off between their respective benefits. This study proposes a multitask deep learning model with a multigate hybrid expert structure to jointly model solver preference and ability, thereby balancing bilateral benefits. The knowledge source for participation and performance prediction tasks are grounded in expectancy theory and performance theory, respectively. Linear and deep neural network (DNN) modules are integrated to enhance both memorization and generalization capabilities. By incorporating gating networks, the model effectively captures correlations between the two prediction tasks, balances intertask weights, and allows each task to learn features in different ways using linear and DNN modules. Additionally, our method addresses sample selection bias and data sparsity issues through feature transfer learning, leveraging the sequential pattern between participation and winning. Cross-validation experiments on Kaggle data demonstrate the model effectiveness, provide data-driven decision support for task recommendation and resource allocation in knowledge-intensive crowdsourcing platforms.</div></div>","PeriodicalId":55181,"journal":{"name":"Decision Support Systems","volume":"190 ","pages":"Article 114400"},"PeriodicalIF":6.7,"publicationDate":"2025-01-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143160451","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-12-24DOI: 10.1016/j.dss.2024.114378
Christopher Bockel-Rickermann , Sam Verboven , Tim Verdonck , Wouter Verbeke
Personal loan pricing requires accurate estimates of individual customer behavior, such as the willingness to take out a loan at a given price, the “bid response”. This is challenging due to the nonlinearity of responses hindering the discretionary definition of models, as well as the confoundedness of observational training data. This paper investigates the application of data-driven and machine learning (ML) methods to estimate individual bid responses. We argue that framing bid response modeling as a problem of causal inference is crucial for accurate modeling and understanding of challenging factors. We test established ML algorithms and state-of-the-art causal ML methods on a dataset on mortgage loan applications in Belgium and investigate the effects of different levels of confounding in the data. Our results demonstrate that methods that address confounding can improve bid response estimation, especially when established non-causal methods are negatively affected.
{"title":"Can causal machine learning reveal individual bid responses of bank customers? — A study on mortgage loan applications in Belgium","authors":"Christopher Bockel-Rickermann , Sam Verboven , Tim Verdonck , Wouter Verbeke","doi":"10.1016/j.dss.2024.114378","DOIUrl":"10.1016/j.dss.2024.114378","url":null,"abstract":"<div><div>Personal loan pricing requires accurate estimates of individual customer behavior, such as the willingness to take out a loan at a given price, the “bid response”. This is challenging due to the nonlinearity of responses hindering the discretionary definition of models, as well as the confoundedness of observational training data. This paper investigates the application of data-driven and machine learning (ML) methods to estimate individual bid responses. We argue that framing bid response modeling as a problem of causal inference is crucial for accurate modeling and understanding of challenging factors. We test established ML algorithms and state-of-the-art causal ML methods on a dataset on mortgage loan applications in Belgium and investigate the effects of different levels of confounding in the data. Our results demonstrate that methods that address confounding can improve bid response estimation, especially when established non-causal methods are negatively affected.</div></div>","PeriodicalId":55181,"journal":{"name":"Decision Support Systems","volume":"190 ","pages":"Article 114378"},"PeriodicalIF":6.7,"publicationDate":"2024-12-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143160439","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号