Credit risk estimation is crucial for financial institutions to minimize defaults and maximize profitable opportunities. Traditional credit scoring models, such as Logistic Regression, offer high interpretability but may lack predictive performance, while complex models like Random Forest provide better accuracy but lack transparency. This paper introduces spline-rule ensembles as a novel approach in credit scoring, combining strengths of tree ensembles and linear models to obtain a high-performing, structurally interpretable model. Three variants using different tree generation methods are benchmarked against their conventional rule ensemble counterparts and other classifiers. Results indicate that spline-rule ensembles outperform traditional interpretable classifiers, and compete favorably with state-of-the-art models. Additionally, spline-rule ensembles with rules generated from Boosting generally perform better than those with rules from Random Forest and Bagging. A meta-learner identifies factors driving their superior performance, and a case study highlights their interpretability advantage over tree ensembles.

The formation of political coalitions is a complex negotiation driven by both concrete policy objectives and deep-seated ideological convictions. While Large Language Models (LLMs) open new avenues for computational political science, the neutrality and helpfulness biases instilled by Reinforcement Learning from Human Feedback (RLHF) prevent them from sustaining steadfast partisan behaviour. We present a multi-agent framework that reconciles factual grounding with ideological alignment by combining Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Retrieval-Augmented Generation (RAG): DPO instils aggressive party-specific personas, while a per-party RAG pipeline keeps each agent bounded to its official manifesto. We operationalize the framework on the 2019 Flemish election, deploying the partisan agents in a hub-and-spoke negotiation arbitrated by a formateur. To make the emergent negotiation interpretable, we introduce a Multi-Layered Information Lineage Topology (MILT) that traces every clause in the final agreement back to its manifesto origin and classifies it into five provenance states, a Coalition Influence Score (CIS) that aggregates these traceable contributions to identify which party shaped the agreement, and a real-world grounding pass that benchmarks each simulated provision against the historically adopted coalition agreement. Across three independent simulations the framework yields a stable winner and ranking (N-VA ahead of CD&V and Open Vld), and manifesto-anchored lineage reliably predicts real-world materialization whereas hallucinated content does not. The result is a transparent, scalable testbed for the ex-ante exploration of party compatibility and formateur-mediated compromise.

The BERTopic framework leverages transformer embeddings and hierarchical clustering to extract latent topics from unstructured text corpora. While effective, it often struggles with social media data, which tends to be noisy and sparse, resulting in an excessive number of overlapping topics. Recent work explored the use of large language models for end-to-end topic modelling. However, these approaches typically require significant computational overhead, limiting their scalability in big data contexts. In this work, we propose a framework that combines BERTopic for topic generation with large language models for topic reduction. The method first generates an initial set of topics and constructs a representation for each. These representations are then provided as input to the language model, which iteratively identifies and merges semantically similar topics. We evaluate the approach across three Twitter/X datasets and four different language models. Our method outperforms the baseline approach in enhancing topic diversity and, in many cases, coherence, with some sensitivity to dataset characteristics and initial parameter selection.

BERTopic is a state-of-the-art topic modeling framework that generates topics by clustering contextualized document embeddings. However, its default clustering algorithm, HDBSCAN, often generates an excessive number of topics, which hinders meaningful comparisons or applications in downstream tasks. While different topic reduction methods exist, literature lacks a comparison of these methods and their impact on the quality of the reduced topics. This study offers an in-depth exploration of (1) topic reduction methods for BERTopic, including existing approaches for direct or indirect reduction, and (2) novel techniques that leverage large language models (LLMs), either by using the LLM-generated topic labels to create topic embeddings or by directly prompting the model for identifying overlapping topics. A comparative study of these topic reduction methods is performed, evaluating their performance in terms of coherence and diversity across multiple social media and web datasets. Our findings indicate that indirect topic reduction (e.g., agglomerative clustering) yields more diverse topics, though sometimes at the cost of coherence, compared to direct topic reduction (increasing the minimum cluster size). LLM prompting outperforms agglomerative clustering in both coherence and diversity, though at a higher computational cost. Therefore, we recommend selecting a reduction method based on dataset characteristics, computational resources, and the desired balance between diversity and coherence.

Understanding how customers react to changes in service encounters during crisis situations, is extremely valuable for companies. Guided by the four fundamental premises of customer experience (Becker and Jaakkola, 2020) and drawing on the Touchpoints-Context-Qualities (TCQ)-framework (De Keyser et al., 2020) and Stimulus-Organism-Response (S-O-R) Theory (Mehrabian and Russell, 1974), this study aims to better understand (1) the customer experience during a crisis, and (2) how outside versus within firm-controlled touchpoints, (3) contextual factors (here, individual and firm/industry) and (4) a firm’s service adaptation (e.g., firms adapt to regulatory measures) and service transformation (e.g., firms transform their business by exploring new service delivery models) strategies shape customer experiences and coping behaviors. Using social media data (here, tweets) in the context of the COVID-19 pandemic, our qualitative and quantitative findings reveal that (1) customers’ affective responses are not restricted to negative experience (here, fear, anger, sadness, and disgust), but also include positive experiences (here, joy, surprise, acceptance, and anticipation), (2) the government as third-party affects the customer experience, (3) service adaptions and service transformations evoke different affective responses, and (4) both positive and negative experiences influence customers’ coping behaviors, but positive experiences exert a greater influence on the inward coping behaviors (here, active coping, planning, and positive reinterpretation), while negative experiences exert a greater influence on the outward coping behaviors (here, seeking social support for instrumental reasons, and venting). Our findings have implications for managers and public policy makers that want to manage the customer experience during crisis situations.

Unverified rumor detection recently received considerable academic attention due to the societal impact resulting from this potential misinformation. Previous work in this area mainly focused on textual features using a limited number of data sets and candidate algorithms, and completely disregarded model explainability. This study aims to come up with a more comprehensive social media rumor detection methodology. First, we investigate which machine or deep learning algorithm is best suited to classify tweets into rumors and non-rumors using both textual and structured features. Next, we interpret these rumor detection models with the LIME method and assess the quality of the explanations via fidelity and stability. To ensure the robustness of our methodology, it is benchmarked across the well-known PHEME data sets and two novel data sets, which are made publicly available. The results indicate that machine learners perform best on small data sets, while transformer architectures show the highest predictive accuracy for larger data sets. Unfortunately, these high accuracy transformer models are incompatible with LIME, which results in low fidelity. Moreover, our study shows that all LIME explanations are unstable across folds. Based on these results, we argue to evaluate explanation quality using fidelity and stability before explanation deployment. Our results further demonstrate that apparent model-agnostic explanations such as LIME do not seem to be completely model-agnostic and should be used with caution. (c) 2023 Elsevier B.V. All rights reserved.

Using mobile phone data to improve spatiotemporal crime rates and crime predictions This article assesses whether the ambient population is a more suitable population-at-risk measure for crime types with mobile targets than residential population. Specifically, the potential use of ambient population as a crime rate denominator and predictor for predictive policing models is evaluated, using mobile phone data (n = 595,858,852 raw data points) as a proxy. The results show that ambient population correlates more strongly with crime than residential population. Crime rates based on ambient population designate different problem areas than crime rates based on residential population. The prediction performance of predictive policing models can be improved by using ambient population instead of residential population. These findings support that ambient population is a more suitable population-at-risk measure, as it better reflects the underlying dynamics in spatiotemporal crime trends.

The purpose of this paper is to enhance current practices in business-to-business (B2B) customer churn prediction modelling. Following the recent trend from accuracy-based to profit-driven evaluation business-to-customer churn prediction, we present a novel expected maximum profit measure for B2B customer churn (EMPB), which is used to demonstrate how current practices are suboptimal due to large discrepancies in customer value. To directly incorporate the heterogeneity of customer values and profit concerns of the company, we propose an instance-dependent profit maximizing classifier based on gradient boosting, named B2Boost. The main innovation of B2Boost is the fact that it considers these differences and incorporates them into the model construction by maximizing the objective function in terms of the EMPB. The results indicate that the expected maximal profit gains made in our analyses are substantial. This study arguments towards both deploying models based on customer-specific profitability differences, as well as evaluating based on our instance-dependent EMPB measure.

Calving is one of the most critical moments during the life of a cow and their calves. Timely supervision is therefore crucial for animal welfare as well as the farm economics. In this study, we propose a framework to predict calving within 24 h, 12 h, 6 h, 3 h and 1 h of dairy cows using sequential sensor data. In particular, data were extracted from 2363 cows coming from 8 commercial farms between August 2016 and November 2020. Two sensors attached to the neck and leg of each cow measured rumination, eating, lying, standup, walking and inactive behavior on a minute basis. A novel methodology was used to impute the missing values in the sensor sequences by leveraging the observed values of all the behavioral activities recorded by the sensors. A deep learning model was then used to predict the moment of calving on an hourly basis using the imputed sensor sequences. Results show that 65% of the calvings within 24 h can be detected with a precision of 77%, while 57% of calvings occurring within 3 h can be identified with a precision equal to 49%. Moreover, we find that using the missing value imputations significantly improves the predictive performance for observations containing up to 60% of missing values. The framework proposed in this study can be used by farmers to optimize their calving management and hence improve animal monitoring.

Condition-based maintenance (CBM) is becoming more commonplace within the petrochemical industry. While we find that previous research leveraging machine learning has provided high accuracy in the predictive aspect of machine breakdowns, the diagnostic aspect of these approaches is often lacking. This paper implements a supervised machine learning approach, with the goal of both prediction and diagnosis of machinery breakdowns, emphasizing the latter. To achieve this, it uses an XGBoost model trained on a combination of sensor and report data, and enriches the model with Shapley values for diagnostic insights. We show that this combination of statistical methods, combined with a proper data treatment, can be used to great effect and can vastly improve the diagnostic value of machine learning approaches. The insights that follow from the analysis can subsequently be leveraged by plant operators in CBM strategies or root-cause analyses. (c) 2021 Elsevier Ltd. All rights reserved.

ABSTRACT The main purpose of this paper is to evaluate the feasibility of predicting whether yes or no a Facebook user has self‐reported to have watched a given movie genre. Therefore, we apply a data analytical framework that (1) builds and evaluates several predictive models explaining self‐declared movie watching behavior, and (2) provides insight into the importance of the predictors and their relationship with self‐reported movie watching behavior. For the first outcome, we benchmark several algorithms (logistic regression, random forest, adaptive boosting, rotation forest, and naive Bayes) and evaluate their performance using the area under the receiver operating characteristic curve. For the second outcome, we evaluate variable importance and build partial dependence plots using information‐fusion sensitivity analysis for different movie genres. To gather the data, we developed a custom native Facebook app. We resampled our dataset to make it representative of the general Facebook population with respect to age and gender. The results indicate that adaptive boosting outperforms all other algorithms. Time‐ and frequency‐based variables related to media (movies, videos, and music) consumption constitute the list of top variables. To the best of our knowledge, this study is the first to fit predictive models of self‐reported movie watching behavior and provide insights into the relationships that govern these models. Our models can be used as a decision tool for movie producers to target potential movie‐watchers and market their movies more efficiently.

Distributed stream processing frameworks have gained widespread adoption in the last decade because they abstract away the complexity of parallel processing. One of their key features is built-in fault tolerance. In this work, we dive deeper into the implementation, performance, and efficiency of this critical feature for four state-of-the-art frameworks. We include the established Spark Streaming and Flink frameworks and the more novel Spark Structured Streaming and Kafka Streams frameworks. We test the behavior under different types of faults and settings: master failure with and without high-availability setups, driver failures for Spark frameworks, worker failure with or without exactly-once semantics, application and task failures. We highlight differences in behavior during these failures on several aspects, e.g., whether there is an outage, downtime, recovery time, data loss, duplicate processing, accuracy, and the cost and behavior of different message delivery guarantees. Our results highlight the impact of framework design on the speed of fault recovery and explain how different use cases may benefit from different approaches. Due to their task-based scheduling approach, the Spark frameworks can recover within 30 seconds and in most cases without necessitating an application restart. Kafka Streams has only a few seconds of downtime, but is slower at catching up on delays. Finally, Flink can offer end-to-end exactly-once semantics at a low cost but requires job restarts for most failures leading to high recovery times of around 50 seconds.

This paper aims to determine the power of social media data (Facebook and Twitter) in predicting box office sales, which platforms, data types and variables are the most important and why. To do so, we compare several models based on movie data, Facebook data, and Twitter data. We benchmark these model comparisons using various prediction algorithms. Next, we apply information-fusion sensitivity analysis to evaluate which variables are driving the predictive performance. Our analysis shows that social media data significantly increases the predictive power of traditional box office prediction models. Facebook data clearly outperform Twitter data and including user-generated content next to marketer-generated always improves predictive power. Our sensitivity analysis reveals that volume and valence based combination variables pertaining to Facebook comments are the most important variables. Furthermore, we provide an in-depth analysis of the potential mechanisms driving differential predictive ability of Facebook and Twitter. Our findings suggest that Twitter has less of an impact on box office sales than Facebook because Twitter users have less source credibility than Facebook users. Our results are important for practitioners, marketers and academics who want to employ social media data for box office sales predictions.

This article assesses whether ambient population is a more suitable population-at-risk measure for crime types with mobile targets than residential population for the purpose of intelligence-led policing applications. Specifically, the potential use of ambient population as a crime rate denominator and predictor for predictive policing models is evaluated, using mobile phone data (with a total of 9,397,473 data points) as a proxy. The results show that ambient population correlates more strongly with crime than residential population. Crime rates based on ambient population designate different problem areas than crime rates based on residential population. The prediction performance of predictive policing models can be improved by using ambient population instead of residential population. These findings support that ambient population is a more suitable population-at-risk measure, as it better reflects the underlying dynamics in spatiotemporal crime trends. Its use has therefore much as-of-yet unused potential not only for criminal research and theory testing, but also for intelligence-led policy and practice.

More and more use cases require fast, accurate, and reliable processing of large volumes of data. To do this, a distributed stream processing framework is needed which can distribute the load over several machines. In this work, we study and benchmark the scalability of stream processing jobs in four popular frameworks: Flink, Kafka Streams, Spark Streaming, and Structured Streaming. Besides that, we determine the factors that influence the performance and efficiency of scaling processing jobs with distinct characteristics. We evaluate horizontal, as well as vertical scalability. Our results show how the scaling efficiency is impacted by many factors including the initial cluster layout and direction of scaling, the pipeline design, the framework design, resource allocation, and data characteristics. Finally, we give some recommendations on how practitioners should undertake to scale their clusters.

Hosts list their accommodations on Airbnb aspiring to attract guests. Extant research on the drivers of guests' booking behaviour has solely considered structured information on the Airbnb platform, thereby omitting the rich information provided in the unstructured textual listing description. This work adds to the stream of research on Airbnb demand determinants by identifying the latent topics used in these unstructured descriptions as drivers of listing demand. Both our empirical model and follow-up experimental study indicate that Airbnb guests value unique accommodation aspects of which hosts can convince their potential guests by using the textual description. Guests especially value enthusiastic home experiences and a unique local city guide accompanying the listing. However, when hotel-like properties are conveyed in the description, prospective guests are dissuaded.

Existing lactation models predict milk yields based on a fixed amount of observed milk production in early lactation. In contrast, this study proposes a model to predict the entire lactation curve of dairy cows by leveraging historical milk yield information observed in the preceding cycle. More specifically, we present a deep learning framework to encode the model inputs, predict the latent representation of the milk yield sequences and generate the corresponding lactation curves. Results show that the proposed framework outperforms the baseline models and that during the first 26 days of lactation, the model’s predictions are more accurate than those of a state-of-the-art lactation model which is able to leverage the observed milk yields. As a result, the framework presented in this study allows farmers to increase their forecast horizon with respect to predicting its herd’s total production and hence facilitates optimal herd management. Additionally, the model can be used to compare a cow’s actual and expected milk yield over the entire course of the lactation cycle. This in turn can help to accelerate disease detection and enhance current animal monitoring systems. Finally, as the model incorporates the impact of health and reproduction events as well as herd management on the cow’s productivity, future earnings and costs can be estimated more accurately.

The increasing need for real-time insights in data sparked the development of multiple stream processing frameworks. Several benchmarking studies were conducted in an effort to form guidelines for identifying the most appropriate framework for a use case. In this article, we extend this research and present the results gathered. In addition to Spark Streaming and Flink, we also include the emerging frameworks Structured Streaming and Kafka Streams. We define four workloads with custom parameter tuning. Each of these is optimized for a certain metric or for measuring performance under specific scenarios such as bursty workloads. We analyze the relationship between latency, throughput, and resource consumption and we measure the performance impact of adding different common operations to the pipeline. To ensure correct latency measurements, we use a single Kafka broker. Our results show that the latency disadvantages of using a micro-batch system are most apparent for stateless operations. With more complex pipelines, customized implementations can give event-driven frameworks a large latency advantage. Due to its micro-batch architecture, Structured Streaming can handle very high throughput at the cost of high latency. Under tight latency SLAs, Flink sustains the highest throughput. Additionally, Flink shows the least performance degradation when confronted with periodic bursts of data. When a burst of data needs to be processed right after startup, however, micro-batch systems catch up faster while event-driven systems output the first events sooner.

In this research, we propose a novel data-centric way of optimizing a catalytic cracking unit. We first design a soft sensor to predict catalyst saturation levels within a Fluid Catalytic Cracking Unit (FCCU). To achieve this, we implement an established method and combine it with modern algorithms for accurate and robust results. The input for this model is data from a number of sensors throughout the refinery, combined with laboratory data. Catalyst saturation level is measured by way of manual refraction analysis and lookup tables. These manual measurements were combined with laboratory data to provide training input for our soft sensor models. Subsequently, we utilize this new soft sensor model in an input mix optimization in order to continuously optimize the use of the catalyst within the FCCU. This model leads to a higher product yield, less catalyst consumption, and a more efficient process. This proposed optimization pipeline can be introduced as smart process control tying into the development towards Industry 4.0.

The purpose of this study is to demonstrate the value of Facebook data in predicting first-time donation behavior. More specifically, we provide evidence that Facebook data can be used as a valuable data source for nonprofit organizations in acquiring new donors. To do so, we evaluate three different dimensionality reduction techniques (i.e., singular value decomposition, non-negative matrix factorization, and latent Dirichlet allocation) over seven classification techniques (i.e., logistic regression, k-nearest neighbors, bagged trees, random forest, adaboost, extreme gradient boosting, and artificial neural networks) using five times twofold cross-validation. Next, we assess what type of Facebook data and which predictors are most important. The results indicate that we can predict first-time donation behavior based on Facebook data with high predictive performance. Our benchmark indicates that the combination of singular value decomposition and logistic regression outperforms all other analytical methodologies with an area under the receiver operating characteristic of 0.72 and a top decile lift of 3.33. The results show that Facebook pages and categories of Facebook pages are the most important data types. The most important predictors are dimensions related to age, education, residence, materialism, responsible consumption, and interest in nonprofits. The presented acquisition models can be used by nonprofit organizations to implement a one-to-one targeted marketing campaign towards Facebook fans. To the best of our knowledge, our study is the first to determine the predictive value of Facebook data for nonprofits in a real-life acquisition context.