Bayesian Belief-based Rule Extraction
In data mining, the relationship between variables is extracted from the association of events in a record. For example, consider a shopping receipt as a record of acquired items and each item as an event. Therefore, data mining looks for the association of products that are usually purchased together.
In recent years, there has been a large body of research in solving the problem of data mining for large datasets, i.e., BigData. In addition, the study of data mining time series data has become relevant as the temporal dimension plays an important role in the association of events. For example, associations of sensors and actuators in buildings, or the process of patient care as a function of time.
Whether mining in records or time series, there has not been a drastic change in the core principles of the algorithms used to find the relationships. At the core, all data mining algorithms use thresholds on confidence and support as a mean to weed out false relationships. As algorithm complexity increases, the practical meaning of the thresholds gets lost and expert knowledge about the problem is harder to incorporate, resulting in the need for time-consuming parameter sweeps.
Our general approach
We focus on time series data mining to extract functional and spatial relationships between variables. Furthermore, we are developing a new rule mining algorithm based on Bayesian belief. The algorithm uses as a discriminating method the change in the conditional probability of the variables, updated after every observation. This process is called the change in belief.
To convert each variable’s time series into minable structures, we first extract events from each variable’s time series. An event is usually associated with a significant change in value. Then all events are collated, according to their timestamps, into a single time series. Using an observation window as the relationship search scope, we apply our mining algorithm to extract relationships in the form of rules. The extracted rules can then be used to create hierarchical trees, variable grouping, spatial locations among etc.
Bayesian Rule Mining
Current methodologies for associative rue mining depend on the idea of thresholds on support or confidence to determine which rules are useful. In contrast, we propose the idea of increasing belief as rule selection criteria. An initial proof of concept was first demonstrated by Dr. Luis I. Lopera in his dissertation work Mining Functional and Structural Relationships of Context Variables in Smart-Buildings. The Bayesian rule extraction (BRE) algorithm uses the iterative application of the Bayes theorem to distinguish useful rules from frequent associations. BRE is especially useful for extracting rare rules which will not otherwise pass the support or confidence thresholds.
"Predicting Anxiety in Routine Palliative Care Using Bayesian-Inspired Association Rule Mining", Digital Health, Frontiers, 2021.,
"Mining Functional and Structural Relationships of Context Variables in Smart-Buildings", PhD Thesis, opus4.kobv.de, 2018.,
"Data Mining-based Localisation of Spatial Low-resolution Sensors in Commercial Buildings", Proceedings of the 3rd ACM International Conference on Systems for Energy-Efficient Built Environments (BuildSys '16), ACM, November 2016.,
"Mining Device Data to Auto-commission Buildings: Poster Abstract", Proceedings of the 3rd ACM International Conference on Systems for Energy-Efficient Built Environments (BuildSys '16), ACM, November 2016.,
"Mining hierarchical relations in building management variables", Pervasive and Mobile Computing, ScienceDirect, February 2016.,
"Mining relations and physical grouping of building-embedded sensors and actuators", Proceedings of the International Conference on Pervasive Computing and Communications (PerCom '15), IEEE, March 2015.,