TL;DR
This paper tackles the challenge of building accurate usage profiles for smart appliances by proposing a novel daily pattern-based probabilistic model. Crucially, it introduces an algorithm to automatically discover the optimal representation of days of the week, moving beyond the standard seven-day approach to improve prediction accuracy.
What Problem Does This Paper Address?
The paper addresses the under-explored problem of generating "usage profiles" for smart appliances. These profiles are essential for intelligent systems to predict how a machine will be used, enabling benefits like energy efficiency (e.g., turning off unused devices), enhanced user comfort (e.g., pre-heating a room), and improved safety (e.g., anomaly detection). While many classification methods exist, usage profiling demands precise probability estimations, not just accurate class prediction. A key challenge is that historical usage data is time-series and often exhibits complex temporal patterns (daily, weekly) that are interrupted by unpredictable events, making it difficult to build robust models from limited data.
Methods
The authors propose a daily pattern-based probability model for usage profiling. This model represents a machine as a set of services, and for each service, it defines "day-types" with associated probabilities of service usage across different time frames within a day.
The core innovation lies in finding the optimal days-of-the-week representation. Instead of rigidly using seven distinct days (Sunday-Saturday), they propose an algorithm that uses hierarchical clustering and the Conditional Minimum Description Length (CMDL) score to group similar days into "day-types." The algorithm works by:
- Calculating the conditional mutual information between days of the week and hours of the day, conditioned on usage, to measure their dissimilarity.
- Using this dissimilarity to build a dendrogram via hierarchical clustering.
- Adapting the CMDL score to find the "best cut" of the dendrogram, which represents the optimal number of day-types and their assignments. This cut minimizes a score that balances model complexity and data fit.
- The resulting optimal representation is then used to train Bayesian Network Classifiers (BNCs) for usage profiling.
Datasets
The methodology was tested on a real-life dataset of office printer usage collected over three years (2009-2012) from 53 printers at KU Leuven. After filtering for printers with sufficient usage history, 21 printers were used for analysis. The data logs consist of (time, printer, user, pages, copies) tuples, with time discretized into 1-hour frames. The "service" defined for printers was the existence of any print job within an hour.
Key Insights and Ideas
- Beyond the Seven-Day Week: The most significant insight is that a fixed seven-day representation is often suboptimal for usage profiling. Grouping similar days (e.g., weekdays, weekends, or even more granular groupings) can lead to more robust probability estimations due to increased sample size per cluster.
- Probabilistic Precision Matters: For smart appliance control, knowing how likely something is to happen (e.g., a 10% vs. 40% chance of usage) is more critical than a simple binary classification. This paper prioritizes this probabilistic accuracy.
- CMDL as a Smart Clustering Criterion: Using CMDL to find the optimal number of clusters for days of the week is a clever way to balance the desire for more data per cluster against the risk of losing important distinctions.
- Office Printers as a Proxy: While the study uses office printers, the methodology is presented as generic and applicable to various smart appliances and environments, with the understanding that the optimal representation might differ.
Key Results
- The proposed algorithm successfully identified optimal days-of-the-week representations for office printers, often resulting in fewer than seven distinct day-types (e.g., (1233331) for printer 38, meaning Sunday/Saturday are one type, Monday another, and Tue-Fri a third).
- Models trained using the optimal representation consistently showed better conditional log-likelihood (CLL) scores compared to models using the standard seven-day representation. This indicates improved probability estimation.
- Classification accuracy (CA) was similar between the two approaches, reinforcing the idea that the benefit lies in probabilistic accuracy rather than raw classification performance.
- The paper provides a table (Table I) showing validation results for 8 printers, consistently demonstrating better CLL for the optimal representation.
Limitations and Future Work
- Constant Usage Behavior Assumption: The current model assumes each day within a cluster has constant usage behavior. The authors acknowledge that this is a simplification and that real-world usage is influenced by factors like public holidays and seasons.
- Generalizing Beyond Printers: While tested on printers, the authors aim to expand the model to incorporate holidays and seasons to create more realistic probability models for diverse smart home scenarios.
- Refined Day-Type Definitions: The paper suggests that for specific applications, services could be further refined (e.g., including user-specific requests or coffee types for a coffee maker), leading to more granular profiles.
Summary updated