A Classification Technique Based on Exploratory Data Analysis for Activity Recognition
Riku Shinohara, Huakun Liu, Monica Perusquía-Hernández, Naoya Isoyama, Hideaki Uchiyama, Kiyoshi Kiyokawa · 2024
In this paper, we first analyzed the nurse activity dataset provided in the “4th Nurse Care Activity Recognition Challenge 2022” to design a behavior prediction model. Understanding the data characteristics and applying appropriate processing to create predictive behavior models is meaningful rather than simply feeding the data into end-to-end machine learning. The dataset was analyzed from various perspectives to investigate its characteristics. As a result, we avoided using the acceleration dataset due to the incomplete and unbalanced data. We propose a predictive model using only care record data. We create a heatmap for each nurse using the activity type id and date columns of the nurse data. We cluster the nurses into three groups using the heatmap. We used k-means for clustering. The predictive model takes user id and date as input. It outputs 1-28 activity labels corresponding to the input date data. The provided training data and test data periods are used for the training data, and the provided test data periods are used for the test data. We evaluate each user individually accuracy and F1. When averaged out for each accuracy and F1 are 0.73, 0.34. We are able to create the model with high precision using the simple method based on EDA without using machine learning or deep learning models.