Hello everyone,
The Brick by Brick Challenge 2024 focused on automating sensor metadata labelling in smart buildings using time-series data. The Brick by Brick Challenge 2024 brought together 313+ participants from 40+ countries, resulting in an incredible 1,870+ submissions!
Letβs meet the winners and see how they tackled this problem!
MEET THE WINNERS
1st Place: Team yddm
Prize: 5,000 AUD + 2,500 AUD Travel Grant
Members: @mike_q (Chengfeng Qiu), @zoe1113(Jiahui Zhou), @leo2025vv (Yongfeng Liao)
Approach:
Team yddm developed a hierarchical feature extraction framework to capture the complex temporal dynamics in time-series building data. They transformed the multi-label classification task into a 91-class single-label problem, simplifying the process while maintaining high precision. The team used CatBoost, a gradient-boosting algorithm specifically designed for categorical data, which proved highly effective in managing the intricacies of time-series data. Their approach supported energy-efficient building management and contributed to the standardisation of data formats for practical deployment. The model achieved exceptional performance, winning first place in the Brick by Brick 2024 competition, demonstrating its potential for real-world applications in sustainable building operations.
Background:
Chengfeng Qiu works at NetEase and is an expert AI researcher, having won top prizes at NIPS, ICDM, and WWW. His expertise includes multimodal data, time-series analysis, and large language models, particularly in risk control algorithm development.
2nd Place: Team xiaobaiLan
Prize: 3,000 AUD + 2,000 AUD Travel Grant
Members: @xiaobaiLan (Meilan Xu), @20N (Zheng Wen), @wangbing106 (Bing Wang)
Approach:
The team applied multi-domain feature engineering, hierarchical probability calibration, and model optimisation to classify 94 sub-classes from irregular time-series data. The raw sensor data was segmented into daily intervals, and 78 features were extracted across statistical, temporal, spectral, and peak domains. The multi-label task was converted into a 91-class problem using label concatenation. A 64th-root probability calibration method was used to preserve label hierarchy and reduce extreme probabilities. Temporal difference features were the most discriminative.
The model, built with XGBoost, achieved macro F1 scores between 0.7366 and 0.7391 in validation. It recorded a macro F1 score of 0.6127 on the public leaderboard and an overall score of 0.544 across both public and private test sets. Prediction remapping (-1 to 0.1, 1 to 0.9) was applied to align with competition metrics. Training and inference were completed within 14 hours. The approach supports scalable, real-world building energy management and enables fine-grained equipment status classification for sustainable operations.
Background:
Meilan Xu is a database system administrator interested in data mining competitions across diverse industries.
3rd Place (1): Team NaiveBaes
Prize: 1,000 AUD + 1,500 AUD Travel Grant
Members: @joe_zhao7 (Haokai Zhao), @jonasmacken (Jonas Macken), @leocd (Leo Dinendra), @RuiyuanY (Ruiyuan Yang), @xenonhe (Yaqing He)
Approach:
NaiveBaes used feature engineering, hierarchical label structuring, data augmentation, and ensemble learning. They extracted features from statistical, temporal, and spectral domains. The multi-label task was divided into five tiers of multi-class classification to leverage label dependencies, with outputs from higher tiers used as inputs for lower tiers.
Data augmentation expanded the training set by 28 times and the test set by 2 to 3 times. The models, Random Forest and XGBoost, were trained using K-Fold cross-validation. Final predictions were aggregated using soft voting to balance precision and recall.
Background:
The team consists of postgraduate students from UNSW Sydney with backgrounds in AI, data science, and engineering. They collaborated through the universityβs AI Study Club.
3rd Place (2): Team bram
Prize: 1,000 AUD + 1,500 AUD Travel Grant
Member: @bram (Bram Steenwinckel)
Approach:
Bram converted the multi-label task into a multiclass classification problem by identifying 91 unique label combinations. He extracted 337 features from statistical, frequency, and temporal domains, including resampling the sensor data at different intervals to capture finer details. These features were reduced to the 40 most influential ones through feature selection techniques. He used an ensemble extra-trees classifier, which builds multiple decision trees with added randomness, to improve both the accuracy and stability of predictions. This well-structured approach not only delivered a strong performance on the competition leaderboard but also provided a clear pathway for integrating automated classification into smarter, energy-efficient building management systems.
Background:
Bram is a postdoctoral researcher at IDLab (Ghent University β imec), focusing on hybrid AI and the Semantic Web. He applies academic research to practical use cases in healthcare and predictive maintenance and is a Kaggle Competition Master.
3rd Place (3): Team chan_jun_hao
Prize: 1,000 AUD + 1,500 AUD Travel Grant
Member: @chan_jun_hao (Chan Jun Hao)
Approach:
Chan Jun Hao developed a multi-label time-series classification framework for building metadata labelling, with an emphasis on ratio- and correlation-based features that generalise well across buildings. His strategy involved:
- Minimal data cleaning, removing only abrupt drops to zero beyond a Z-score threshold, and treating missing values as informative signals by retaining them to reflect offline periods;
- Custom label-tuned class weights, using a tailored scaling function that slightly boosts minority labels without overly penalising major ones, to balance rare vs. common labels while maintaining both precision and recall;
- Relative feature focus, favouring ratio- and correlation-based features over absolute values to avoid generalisation issues due to varying building baselines. This approach demonstrated strong generalisation across buildings in the dataset.
Background:
With a background in architecture and current work in AI, Chan Jun Haoβs unique blend of domain expertise enabled him to approach the challenge with both technical depth and real-world insight.
A big thank you to all participants for your incredible efforts in the **Brick by Brick Challenge
Have feedback? Let us know: https://forms.gle/K7qh3pt1WBDyW5sEA
Join a new exciting challenge: AIcrowd | Challenges