Collaborative filtering (CF) is a cornerstone of recommender systems and plays a relevant role in many modern applications. CF uses user-item interaction data to discover future preferences. Although deep learning models have shown promise in CF, Embarrassingly Shallow Autoencoders for Sparse Data (EASE) has gained attention for its outstanding ranking accuracy provided by its closed-form solution. EASE relies primarily on relationships among items to fit a full-rank highdimensional linear mapping. We hypothesize that this design limits its capacity to capture similar but not equivalent fine-grained user relationships, consequently limiting its recommendation accuracy. This paper introduces an alternative formulation based on EASE that takes advantage of user-user information. Furthermore, we propose a hybrid model that combines both user-user and item-item adjacency distributions. Our experiments reveal that the proposed models outperform EASE on well-known recommendation benchmarks, highlighting the significance of including the user alternative in shallow autoencoder studies for CF.
In this paper, we propose Attention-based Counterfactual Explanation (AB-CF), a novel model that generates post-hoc counter-factual explanations for multivariate time series classification that narrow the attention to a few important segments. We validated our model using seven real-world time-series datasets from the UEA repository. Our experimental results show the superiority of AB-CF in terms of validity, proximity, sparsity, contiguity, and efficiency compared with other competing state-of-the-art baselines.
In sequential pattern mining, the time intervals between each pair of items are usually ignored or mined using additional constraints such as gaps. However, a large part of time-related knowledge is lost during the original proposition of sequential pattern mining. In many applications, time-related information is valuable and its preservation is necessary. In this paper, we present two methods for exploiting time intervals between items and evaluate them in a classification task using a real-world customer intent dataset.
Solar flare prediction is a central problem in space weather forecasting. Existing solar flare prediction tools are mainly dependent on the GOES classification system, and models commonly use a proxy of maximum (peak) X-ray flux measurement over a particular prediction window to label instances. However, the background X-ray flux dramatically fluctuates over a solar cycle and often misleads both flare detection and flare prediction models during solar minimum, leading to an increase in false alarms. We aim to enhance the accuracy of flare prediction methods by introducing novel labeling regimes that integrate relative increases and cumulative measurements over prediction windows. Our results show that the data-driven labels can offer more precise prediction capabilities and complement the existing efforts.
In data science, data pre-processing and data exploration require various convoluted steps such as creating variables, merging data sets, filtering records, value transformation, value replacement and normalization. By analyzing the source code behind analytic pipelines, it is possible to infer the nature of how data objects are used and related to each other. To the best of our knowledge, there is scarce research on analyzing data science source code to provide a data-centric view. On the other hand, two important diagrams have proven to be essential to manage database and software development projects: (1) Entity-Relationship (ER) diagrams (to understand data structure and data interrelationships) and (2) flow diagrams (to capture main processing steps). These two diagrams have historically been used separately, complementing each other. In this work, we defend the idea that these two diagrams should be combined in a unified view of data pre-processing and data exploration. Heeding such motivation, we propose a hybrid diagram called FLOWER (FLOW+ER) that combines modern UML notation with data flow symbols, in order to understand complex data pipelines embedded in source code (most commonly Python). The goal of FLOWER is to assist data scientists by providing a reverse-engineered analytic view, with a data-centric angle. We present a preliminary demonstration of the concept of FLOWER, where it is incorporated into a prototype that traces a representative data pipeline and automatically builds a diagram capturing data relationships and data flow.
