Tackling the Accuracy-Interpretability Trade-off: Interpretable Deep Learning Models for Satellite Image-based Real Estate Appraisal

2023 | ACM Transactions on Management Information Systems | Citations: 0

Authors: Kucklick, Jan-Peter; Müller, Oliver

Abstract: Deep learning models fuel many modern decision support systems, because they typ ... Expand

Abstract: Deep learning models fuel many modern decision support systems, because they typically provide high predictive performance. Among other domains, deep learning is used in real-estate appraisal, where it allows extending the analysis from hard facts only (e.g., size, age) to also consider more implicit information about the location or appearance of houses in the form of image data. However, one downside of deep learning models is their intransparent mechanic of decision making, which leads to a trade-off between accuracy and interpretability. This limits their applicability for tasks where a justification of the decision is necessary. Therefore, in this article, we first combine different perspectives on interpretability into a multi-dimensional framework for a socio-technical perspective on explainable artificial intelligence. Second, we measure the performance gains of using multi-view deep learning, which leverages additional image data (satellite images) for real estate appraisal. Third, we propose and test a novel post hoc explainability method called Grad-Ram. This modified version of Grad-Cam mitigates the intransparency of convolutional neural networks for predicting continuous outcome variables. With this, we try to reduce the accuracy-interpretability trade-off of multi-view deep learning models. Our proposed network architecture outperforms traditional hedonic regression models by 34% in terms of MAE. Furthermore, we find that the used satellite images are the second most important predictor after square feet in our model and that the network learns interpretable patterns about the neighborhood structure and density. Collapse

Topics: Bing explainable artificial intelligence shapley additive explanation video conference price management

Methods: principal component analysis data transformation experiment artificial neural network regression analysis method

Theories: game theory

Generative Artificial Intelligence in Information Systems Education: Challenges, Consequences, and Responses

2023 | Communications of the Association for Information Systems | Citations: 0

Authors: Van Slyke, Craig; Johnson, Richard; Sarabadani, Jalal

Abstract: ChatGPT, an interactive, generative artificial intelligence (AI) system, was in ... Expand

Abstract: ChatGPT, an interactive, generative artificial intelligence (AI) system, was introduced in late 2022, quickly becoming one of the most rapidly adopted technologies in history. The rapid emergence of ChatGPT and similar AI tools, such as Google's Bard, and GPT-enabled Bing from Microsoft have led to intense discussions about how they will affect various aspects of society, including higher education. Information systems (IS) education will not escape the impact of AI tools. Our goal for this paper is to develop a better understanding of the range of possible impacts of ChatGPT on IS education and to describe how IS educators might respond to these potential impacts. To that end, we discuss challenges for IS education brought on by generative AI tools, and discuss potential future scenarios based on the emergence of such tools, ranging from AI having little impact on IS education to AI serving as competition for IS educators. We examine the challenges and consequences of each scenario. We also discuss potential responses, ranging from doing nothing to embracing AI tools as legitimate learning aids. We then provide several specific recommendations that will allow IS educators to effectively respond to the rise of AI tools. Collapse

Topics: artificial intelligence IS education IT career generative AI disruptive innovation

Methods: chatGPT experiment case study

From Apathy to Algoactivism: Worker Resistance to Algorithmic Control in Food Delivery Platforms

2023 | International Conference on Business Informatics | Citations: 0

Authors: Weber, Matthias; Jong, Alexander De; Remus, Ulrich

Abstract: Platforms in the gig economy rely on algorithmic control to manage their workfo ... Expand

Abstract: Platforms in the gig economy rely on algorithmic control to manage their workforce, but recent scientific evidence has shown that workers have begun to resist this control. Due to lacking focus and limited empirical data, the phenomenon of worker resistance to algorithmic control is still insufficiently understood. Based on a topic modeling approach with over 2 million text documents extracted from Reddit forums of different food-delivery platforms, we identify 14 resistance actions showing how food-delivery workers resist algorithmic control. Our study contributes to current research by expanding the understanding of resistance to algorithmic control in the gig economy, showing what resistant actions workers take, and discussing the concepts of individual opacity and collective knowledge as possible escalators and de-escalators of this resistance. Collapse

Topics: platform worker resistance to change Reddit search engine sensemaking

Methods: topic model qualitative coding theory development computational algorithm focus group

Toward a Method for Reviewing Software Artifacts from Practice

2022 | DESRIST | Citations: 0

Authors: Gnewuch, Ulrich; Maedche, Alexander

Abstract: Solving real-world problems with innovative and novel artifacts is at the core o ... Expand

Abstract: Solving real-world problems with innovative and novel artifacts is at the core of design science research (DSR). Given DSR’s emphasis on a strong connection to the real-world, artifacts for solving a particular problem may not only be described in extant literature, but also exist in practice. This is particularly the case for software artifacts. Therefore, DSR scholars need to explore the state of the art and demonstrate the novelty of their software artifact relative to existing artifacts in research and practice. However, while methodological guidance for conducting literature reviews is abundant, there is little guidance on how to review software artifacts from practice. This paper takes a first step toward addressing this gap by proposing and illustrating a seven-step method for reviewing software artifacts from practice. Our research provides actionable guidance for DSR scholars on how to support the claim that their software artifact constitutes a substantial contribution to knowledge. Collapse

Topics: digital dashboard mobile application data visualization website mobile health

Methods: design artifact literature study synthesis

A Rollercoaster of Emotions – A Semantic Analysis of Fundraising Campaigns over the Course of the Covid-19 Pandemic

2022 | Americas Conference on Information Systems | Citations: 0

Authors: Grieger, Nicole; Seutter, Janina; Kundisch, Dennis

Abstract: Donation-based crowdfunding platforms are increasingly important for those in n ... Expand

Abstract: Donation-based crowdfunding platforms are increasingly important for those in need, especially in times of crisis. When fundraisers face financial difficulties in addition to the original challenges underpinning their appeal, we can expect to find highly emotional textual descriptions in charitable crowdfunding campaigns. On top of any existing challenges, the COVID-19 pandemic has induced feelings of helplessness and uncertainty. Hence, we pose the question of whether, and to what extent, this pandemic has influenced the topics of donation campaigns and the emotionality expressed in them. Conducting an exploratory analysis on a dataset from GoFundMe, we combine a quantitative approach of dictionary-based sentiment analysis with exploratory topic modeling. Key findings indicate that negative emotions (i.e., negatively associated words) in the campaign descriptions have increased sharply at the onset of the pandemic, and that the topics of pandemic-related campaigns have spread from health to other aspects of life, such as personal difficulties. Collapse

Topics: pandemic crowdsourcing crowdfunding covid 19 social media

Methods: topic model sentiment analysis regression analysis method linear regression analysis literature study

Second Screening—The Influence of Concurrent TV Consumption on Online Shopping Behavior

2022 | Information Systems Research | Citations: 0

Authors: Hinz, Oliver; Hill, Shawndra; Sharma, Amit

Abstract: The practice of using a “second screen” while concurrently watching television ( ... Expand

Abstract: The practice of using a “second screen” while concurrently watching television (TV) has become a widespread phenomenon. Whereas work on multitasking suggests that TV consumption may lead to lower online sales, research in the area of impulse buying suggests the opposite. In this paper, we study the impact of concurrent TV consumption on online shopping behavior and try to understand this conflict. Our first study analyzes the impact of TV consumption on online shopping behavior at the aggregate level using a panel of 100,000 U.S. customers over a period of 24 months. We address potential endogeneity problems with an instrumental variable approach. Our second study proposes a novel approach to realizing individual-level analysis over thousands of consumers and products. Examining massive amounts of data, we identify a sample of consumers who watched a certain TV show and who were simultaneously doing online shopping at Amazon.com. We then compare their shopping behavior on the day of the TV show to their shopping behavior one week prior. Both studies reveal that whereas TV consumption is correlated with a higher likelihood of online shopping in general, the causal effect of TV watching differs with respect to the complexity of the products purchased. If a TV program appeals to a large TV audience, then this results in fewer immediate sales of high-complexity products (1% increase in TV consumption leads to −2.2% sales) and more sales of low-complexity products (1% increase in TV consumption leads to +8.8% sales). Collapse

Topics: electronic commerce Bing advertising management digital service website

Methods: robustness check descriptive statistic instrumental variables estimation statistical hypothesis test survey

Lantern to the Underworld - Following User Actions Using Bing URL Parameters When Child Exploitation Terms Are Suggested by the Search Engine

2022 | HICSS | Citations: 0

Authors: Spear, Nicholas; Hawbecker, David; Mcelyea, Benjamin

Abstract: During a 2016 criminal investigation overseen by the authors, a discovery was ma ... Expand

Abstract: During a 2016 criminal investigation overseen by the authors, a discovery was made that the Bing search engine was suggesting search terms known to be associated with child exploitation materials. This was in response to a non-contraband search by a subject, leading to increasingly explicit suggestions. This information led to the discovery that, with a user’s browser history files, the specific actions taken by a user could be isolated and tracked using Bing’s own unique URL parameters. While the suggestion of contraband terms by the Bing search engine has since ceased, there has been little to no research conducted on analyzing a user’s browser activity in order to determine their specific actions behind the keyboard through the use of these unique URL parameters. The purpose of this paper is to document Bing’s URL parameters related to image searches (specifically the FORM parameter) in order to detail how a user’s actions during searching may be determined by an analyst. The authors also provide a new tool for creating a timeline of Bing image search events when an analyst possesses a user’s browser history files. Collapse

Topics: Bing search engine Google+ accounting artificial intelligence

Methods: case study machine learning

An Ontology of Emotion Process to Support Sentiment Analysis

2022 | Journal of the Association for Information Systems | Citations: 1

Authors: Storey, Veda; Park, Eun Hee

Abstract: Sentiment analysis is used to mine text data from many sources, including blogs ... Expand

Abstract: Sentiment analysis is used to mine text data from many sources, including blogs, support forums, and social media, in order to extract customers' opinions and attitudes. The results can be used to make important assessments about a customer's attitude toward a company and if and how a company should respond. However, much research on sentiment analysis uses simple classification, where the polarity of a text that is mined is classified as positive, negative, or neutral. This research creates an ontology of emotion process to support sentiment analysis, with an emphasis on obtaining a more fine-grained assessment of sentiment than polarity. The ontology is grounded in a theory of emotion process and consists of concepts that capture the generation of emotion all the way from the occurrence of an event to the resulting behaviors of the person expressing the sentiment. It includes two lexicons: one for affect and one for appraisal. The ontology is applied to posts obtained from customer support forums of large companies to show its applicability in a multilevel evaluation. Doing so provides an example of a complete ontology assessment effort. Collapse

Topics: knowledge representation online review Bing evaluation criteria website

Methods: ontological modelling design artifact sentiment analysis experiment aspect-based sentiment analysis

QUANTIFYING THE IMPACT OF LOCATION DATA FOR REAL ESTATE APPRAISAL – A GIS-BASED DEEP LEARNING APPROACH

2021 | European Conference On Information Systems | Citations: 0

Authors: Kucklick, Jan-Peter; Müller, Jennifer; Beverungen, Daniel; Mueller, Oliver

Abstract: Appraising real estate is challenging since a property's characteristics, locat ... Expand

Abstract: Appraising real estate is challenging since a property's characteristics, location, and neighborhood make up for substantial shares of its price. Current Computer-Assisted Mass Appraisal (CAMA) approaches rely on hedonic models and structured real estate data. Recent work proposed to include geospatial or image data to improve CAMA's predictive accuracy. However, the value of these data sources and their preparation costs remain unclear. In this study, we use Geographic Information Systems (GIS) and Convolutional Neural Networks (CNNs) to appraise 71,000 properties in Philadelphia. We find that geospatial data requires substantial preprocessing but increases CAMA's predictive accuracy by 12%, while image data come with higher computational complexity and an improvement of 16%. Using both datatypes is complex to optimize but offers increased robustness towards extreme outliers. Beyond real estate appraisal, our study is one of the first to quantify the value of geospatial data for data-driven services in the information systems discipline. Collapse

Topics: geographic information system IS discipline decision support Bing price management

Methods: machine learning artificial neural network random forest classification computational algorithm linear regression analysis

Do All Roads Lead to Rome? Exploring the Relationship Between Social Referrals, Referral Propensity and Stickiness to Video-on-Demand Websites

2021 | Business & Information Systems Engineering | Citations: 0

Authors: Köster, Antonia; Matt, Christian; Hess, Thomas

Abstract: Content website providers have two main goals: They seek to attract consumers an ... Expand

Abstract: Content website providers have two main goals: They seek to attract consumers and to keep them on their websites as long as possible. To reach potential consumers, they can utilize several online channels, such as paid search results or advertisements on social media, all of which usually require a substantial marketing budget. However, with rising user numbers of online communication tools, website providers increasingly integrate social sharing buttons on their websites to encourage existing consumers to facilitate referrals to their social networks. While little is known about this social form of guiding consumers to a content website, the study proposes that the way in which consumers reach a website is related to their stickiness to the website and their propensity to refer content to others. By using a unique clickstream data set of a video-on-demand website, the study compares consumers referred by their social network to those consumers arriving at the website via organic search or social media advertisements in terms of stickiness to the website (e.g., visit length, number of page views, video starts) and referral likelihood. The results show that consumers referred through social referrals spend more time on the website, view more pages, and start more videos than consumers who respond to social media advertisements, but less than those coming through organic search. Concerning referral propensity, the results indicate that consumers attracted to a website through social referrals are more likely to refer content to others than those who came through organic search or social media advertisements. The study offers direct insights to managers and recommends an increase in their efforts to promote social referrals on their websites. Collapse

Topics: website social media search engine advertising management Facebook

Methods: logistic regression regression analysis method descriptive statistic robustness check statistical hypothesis test