Meta-Decomposition: Dynamic Segmentation Approach Selection in IoT-based Activity Recognition

TL;DR

This work targets biases introduced by segmentation in IoT-based activity recognition. It reframes segmentation as a data decomposition problem with a decomposer, resolutions, and a composer, and introduces meta-decomposition to dynamically select segmentation strategies over time. Empirical results across multiple public datasets demonstrate that adaptive, meta-learned decomposition reduces biases and improves recognition performance, provided the composition step is properly accounted for. The approach offers a principled path to bias-resilient, environment-adaptive IoT activity recognition and lays groundwork for more advanced meta-learning in segmentation design.

Abstract

Internet of Things (IoT) devices generate heterogeneous data over time; and relying solely on individual data points is inadequate for accurate analysis. Segmentation is a common preprocessing step in many IoT applications, including IoT-based activity recognition, aiming to address the limitations of individual events and streamline the process. However, this step introduces at least two families of uncontrollable biases. The first is caused by the changes made by the segmentation process on the initial problem space, such as dividing the input data into 60 seconds windows. The second category of biases results from the segmentation process itself, including the fixation of the segmentation method and its parameters. To address these biases, we propose to redefine the segmentation problem as a special case of a decomposition problem, including three key components: a decomposer, resolutions, and a composer. The inclusion of the composer task in the segmentation process facilitates an assessment of the relationship between the original problem and the problem after the segmentation. Therefore, It leads to an improvement in the evaluation process and, consequently, in the selection of the appropriate segmentation method. Then, we formally introduce our novel meta-decomposition or learning-to-decompose approach. It reduces the segmentation biases by considering the segmentation as a hyperparameter to be optimized by the outer learning problem. Therefore, meta-decomposition improves the overall system performance by dynamically selecting the appropriate segmentation method without including the mentioned biases. Extensive experiments on four real-world datasets demonstrate the effectiveness of our proposal.

Figures (18)

• Figure 1: (a) shows the conventional segmentation approach, which creates a set of segments in data preparation and then treats them as individual instances to be inputted into the ML system (abstract schema from Kumar2023). (b) illustrates the proposed formulation of segmentation as a data decomposition problem, including the decomposer, the ML model, and the composer. (c) provides an overview of our proposed meta-decomposition model. It can dynamically select the proper decomposition sub tasks from multiple decomposition algorithms.
• Figure 2: CASAS Aruba Cook2012, Home1 and Home2 datasets Krishnan2014 sensors configuration. Around 70% of the activities are unlabeled.
• Figure 3: Orange4Home Dataset Configuration - This dataset represents a two-floor home with 236 sensors, including 83 binary, 55 integer, 67 float, and 31 categorical sensors. These sensors capture data related to motion, switches, humidity, water consumption, luminosity, temperature, weather conditions, and heater settings over a 5-month period.
• Figure 4: Orange4Home Activity Durations: This box-plot depicts the duration of various activities within the Orange4Home dataset. The y-axis measures time in seconds on a logarithmic scale, showcasing the diversity in average activity durations. Activities are listed on the x-axis, with durations ranging from 10 seconds (e.g., 'going down' activity) to approximately 3 hours (e.g., 'computing' activity).
• Figure 5: Sensor Event Frequency for the activities in the Aruba Dataset. For each activity, a hit map is shown which represents the frequency and distribution of sensor events in the Y-axis. The absence of events is shown in white, while darker tones indicate higher event frequencies. On the legend, the number of sensor's heats are shown. These visual patterns reveal significant differences in the occurrence and timing of sensor-triggered activities, such as between 'Meal Preparation' and 'Wash Dishes'. Pre-segmented data showcases distinct patterns, reducing ambiguity and enhancing the recognition performance of activities. Although valuable for experimental analysis, the practicality of pre-segmented data in real-world scenarios may be limited. Similar observation exists in the other datasets that are represented in the appendix.