arXiv subjects

Selim Balcisoy

Publications and source records attributed to Selim Balcisoy.

5 recordsLinked to original sources

arXiv2025

Textured Word-As-Image illustration

In this paper, we propose a novel fully automatic pipeline to generate text images that are legible and strongly aligned to the desired semantic concept taken from the users' inputs. In our method, users are able to put three inputs into the system, including a semantic concept, a word, and a letter. The semantic concept will be used to change the shape of the input letter and generate the texture based on the pre-defined prompt using stable diffusion models. Our pipeline maps the texture on a text image in a way that preserves the readability of the whole output while preserving legibility. The system also provides real-time adjustments for the user to change the scale of the texture and apply it to the text image. User evaluations demonstrate that our method effectively represents semantic meaning without compromising legibility, making it a robust and innovative tool for graphic design, logo creation, and artistic typography.

cs.GR↗

arXiv2025

Instruction-Based Fine-tuning of Open-Source LLMs for Predicting Customer Purchase Behaviors

In this study, the performance of various predictive models, including probabilistic baseline, CNN, LSTM, and finetuned LLMs, in forecasting merchant categories from financial transaction data have been evaluated. Utilizing datasets from Bank A for training and Bank B for testing, the superior predictive capabilities of the fine-tuned Mistral Instruct model, which was trained using customer data converted into natural language format have been demonstrated. The methodology of this study involves instruction fine-tuning Mistral via LoRA (LowRank Adaptation of Large Language Models) to adapt its vast pre-trained knowledge to the specific domain of financial transactions. The Mistral model significantly outperforms traditional sequential models, achieving higher F1 scores in the three key merchant categories of bank transaction data (grocery, clothing, and gas stations) that is crucial for targeted marketing campaigns. This performance is attributed to the model's enhanced semantic understanding and adaptability which enables it to better manage minority classes and predict transaction categories with greater accuracy. These findings highlight the potential of LLMs in predicting human behavior.

cs.IR↗

arXiv2022

One City, Two Tales: Using Mobility Networks to Understand Neighborhood Resilience and Fragility during the COVID-19 Pandemic

What predicts a neighborhood's resilience and adaptability to essential public health policies and shelter-in-place regulations that prevent the harmful spread of COVID-19? To answer this question, in this paper we present a novel application of human mobility patterns and human behavior in a network setting. We analyze mobility data in New York City over two years, from January 2019 to December 2020, and create weekly mobility networks between Census Block Groups by aggregating Point of Interest level visit patterns. Our results suggest that both the socioeconomic and geographic attributes of neighborhoods significantly predict neighborhood adaptability to the shelter-in-place policies active at that time. That is, our findings and simulation results reveal that in addition to factors such as race, education, and income, geographical attributes such as access to amenities in a neighborhood that satisfy community needs were equally important factors for predicting neighborhood adaptability and the spread of COVID-19. The results of our study provide insights that can enhance urban planning strategies that contribute to pandemic alleviation efforts, which in turn may help urban areas become more resilient to exogenous shocks such as the COVID-19 pandemic.

physics.soc-ph↗

arXiv2022

Finding Proper Time Intervals for Dynamic Network Extraction

Extracting a proper dynamic network for modelling a time-dependent complex system is an important issue. Building a correct model is related to finding out critical time points where a system exhibits considerable change. In this work, we propose to measure network similarity to detect proper time intervals. We develop three similarity metrics, node, link, and neighborhood similarities, for any consecutive snapshots of a dynamic network. Rather than a label or a user-defined threshold, we use statistically expected values of proposed similarities under a null-model to state whether the system changes critically. We experimented on two different data sets with different temporal dynamics: The Wi-Fi access points logs of a university campus and Enron emails. Results show that, first, proposed similarities reflect similar signal trends with network topological properties with less noisy signals, and their scores are scale invariant. Second, proposed similarities generate better signals than adjacency correlation with optimal noise and diversity. Third, using statistically expected values allows us to find different time intervals for a system, leading to the extraction of non-redundant snapshots for dynamic network modelling.

cs.SI↗

arXiv2018

Twitter Reveals: Using Twitter Analytics to Predict Public Protests

The right to protest is perceived as one of the primary civil rights. Citizens participate in mass demonstrations to express themselves and exercise their democratic rights. However, because of the large number of participants, protests may lead to violence and destruction, and hence can be costly. Thus, it is important to predict such demonstrations in advance to safeguard against such damages. Recent research has shown that about 75 percent of protests that are regarded as legal, are planned in advance. Twitter, the prominent micro-blogging website, has been used as a tool by protestors for planning, organizing, and announcing many of the recent protests worldwide such as those that led to the Arab Spring, Britain riots, and those against Mr. Trump after the presidential election in the U.S. In this paper, we aim to predict protests by means of machine learning algorithms. In particular, we consider the case of protests against the then-president-elect Mr. Trump after the results of the presidential election were announced in November 2016. We first identify the hashtags calling for demonstration from Trending Topics on Twitter, and download the corresponding tweets. We then apply four machine learning algorithms to make predictions. Our findings indicate that Twitter can be used as a powerful tool for predicting future protests with an average prediction accuracy of over 75 percent (up to 100 percent). We further validate our model by predicting the protests held in the U.S. airports after President Trump's executive order banning citizens of seven Muslim countries from entering the U.S. An important contribution of our study is the inclusion of event specific features for prediction purposes which helps to achieve high levels of accuracy.

cs.SI↗