KI-LEARN: Knowledge-Intensive Learning

The current methodology for applying machine learning can be depicted as follows:

A machine learning researcher confronts a set of data. Typically, there is a large gap between the raw data and the desired "prediction target" (i.e., the variable whose value we wish to predict). The ML researcher bridges this gap by defining a set of features. To do this, he or she interviews experts in the domain to understand how the raw data relates to the prediction target. He or she applies ML expertise to choose a set of features that reflect this domain knowledge and that narrow the gap. Then one or more machine learning algorithms are applied to build a classifier or predictor.

This methodology has many problems. It is difficult to apply, because you need someone with a PhD in machine learning or statistics to design the features. It is difficult to maintain. If new sources of data become available, the feature design process must be repeated. In addition, the rationale behind the set of features is typically not captured in any formal or informal way.

We propose a new methodology that replaces the hand-crafting of features by the hand-crafting of contextual knowledge, as shown below:

In this methodology, contextual knowledge is captured in the form of an object-relational model and a collection of qualitative causal relationships. This information is then automatically transformed to design the features for a learning system and a set of constraints that can be incorporated into the learning system. The hope is that the explicit contextual knowledge will be easier for someone without a PhD in machine learning to maintain. We also believe that if the contextual knowledge is made explicit, then it can be incorporated into the learning algorithms to constraint the parameter values of the fitted predictive models. This should enable fast learning from small samples, and thus provide higher-performance systems than can be obtained with the current methodology.

In our methodology, we envision the following steps which result in a probabilistic relational model (PRM) that can be fit to the available data.

• Build object-relational model of the domain. This model describes the objects, their properties, and their relationships. It can be annotated to indicate which properties are observed and which are the targets to be predicted.

• Indicate qualitative causal relationships. The object-relational model is then augmented with qualitative causal relationships. We are currently exploring extensions of Ken Forbus' Qualitative Process Theory as our representation language.

• Introduce aggregations. If a variable has a large and variable number of input causal links, then those links must be aggregated to provide a compact parameterization of a predictive model. By analyzing the kinds of qualitative relationships, we believe it will be possible to automatically choose good aggregations.

• Parameterize the Probabilistic Relational Model. Each property in the object-relational model can be converted to a random variable in a probabilistic relational model. The conditional probability distribution (CPD) of each such variable must be parameterized in a way that is compact and that reflects constraints arising from the qualitative relationships (such as monotonicity, saturation, etc.).

• Unroll the Probabilistic Relational Model and fit to the data. Our plan is to use existing Bayesian network toolkits to fit the model to the data, such as Intel's Probabilistic Networks Library PNL.

We are developing and testing this methodology in three application domains: (a) TaskTracer (the intelligent desktop), (b) predicting the spread of West Nile Virus, and (c) predicting grasshopper populations in Eastern Oregon.

This project is in the early stages, so we have no technical results to report at this time.

Financial Support

This project has been funded by the following grants and contracts:

• NSF SGER Grant IIS 0335525: Exploiting Contextual Knowledge to Design Input Representations for Machine Learning
• DARPA Contract HR0011-04-1-0005: KI-LEARN: Knowledge-Intensive Learning Methods for Knowledge-Rich/Data-Poor Domains

The views expressed on this page are those of the principal investigators and do not necessarily reflect the views of the National Science Foundation or the Defense Advanced Research Projects Agency.