There are two themes of agent and DM interaction and integration in the litera- ture [23]: DM for agents, referred to as mining-driven agents [113]; and agents for DM, referred to as agent-driven DM, commonly known as Multi-Agent Data Min- ing (MADM). The former concerns issues of transforming the discovered knowl- edge, extracted by DM, into the inference mechanisms or simply the behaviours of agents and multi-agent systems; as well as the arguable challenge of generating intelligence from data while transferring it to a separate, possibly autonomous, software entity. A FIPA-compliant multi-agent platform based on mining-driven agents (Agent Academy) that offers facilities for design, implementation and de- ployment of multi-agent systems is proposed in [113]. The authors describe the Agent Academy as an attempt to develop a framework through which users can create an agent community having the ability to train and retrain its own agents using DM techniques.
MADM, rather than mining-driven agent systems, is the focus of this thesis. It is concerned with the use of agent and MAS to perform DM activities. The contribution of this section is to provide a broad review of prominent MADM approaches in the literature and a discussion of the benefits that individual agent- based DM architectures can provide. This section is not concerned with particular DM techniques. It is however concerned with the collaborative work of distributed software in the design of multi-agent systems directed at DM.
Several systems have been developed for MADM. These systems can be cat- egorised, according to their strategy of learning, into three types:
2. Meta-learning
3. Hybrid-learning
Each is described in some further detail in the following subsections.
2.3.1
Central-learning Strategy
A central learning strategy is when all the data can be gathered at a central site and a single model built. The only requirement is to be able to move the data to the central location in order to merge it and then apply sequential DM algorithms.
This strategy is appropriate when the quantity of geographically distributed data is small and there are no issues of privacy and security that would prohibit the movement of the data. For large volumes of data, the strategy is generally very expensive but also more accurate in its DM results [13, 20] when compared to alternative approaches. The process of gathering data in general is not simply a merging step; it depends on the original distribution. For example, different records may be placed in different sites, different attributes of the same records may be distributed across different sites, or different tables can be placed at differ- ent sites; therefore when gathering data it is necessary to adopt some appropriate merging strategy. However, as noted previously, in many cases this strategy is infeasible because of security and privacy of data requirements and for the reasons outlined in [96].
One of the earliest references to MADM that adopted a central-learning strat- egy can be found in Kargupta et al. [62] who describe a parallel DM system (PADMA) that uses software agents for local data accessing and analysis, and a web based interface for interactive data visualisation. PADMA has been used in medical applications.
Peng et al. [92] presented an interesting comparison between single-agent and multi-agent text classification using the central-learning strategy in terms of a number of criteria including: response time, quality of classification, and economic/privacy considerations. Their results indicated, not unexpectedly, in favour of a multi-agent approach.
2.3.2
Meta-learning Strategy
Meta-learning is the process of automatic induction of correlations between tasks and solving strategies, based on a domain characterisation. Meta-learning meth- ods have been widely used within DM [122, 21], particularly in the area of clas- sification and regression.
With respect to classification the meta-learning strategy offers a way to mine classifiers from homogeneously distributed data. It follows three main steps. The first is to generate base classifiers at each site using classifier learning algorithms. The second step is to collect the base classifiers at a central site. The third step is to generate the final classifier (meta-classifier) from meta-level data via a combiner or an arbiter. Copies of the classifier agent will exist, or be deployed, on nodes in the network being used (see for example [95]).
One of the most popular MADM approaches that adopt the meta-learning strategy is the METAL project [82] whose emphasis is on helping the user to obtain a ranking of suitable DM algorithms through an on-line advisory system. Gorodetsky et al. [44] correctly consider that the core problem in MADM is not the DM algorithms themselves (in many case these are well understood), but the most appropriate mechanisms to allow agents to collaborate. Gorodetsky et al. present a MADM system to achieve DDM and, specifically, classification. A more recent system, proposed in [76], uses the MAGE middle-ware [104] to build an execution engine that uses a directed acyclic graph to formalise the representation of the KDD process. In [93] a multi-agent system for KDD (AgentDiscover) has been proposed. It uses task-based reasoning [24] for problem solving on a multi- agent platform. Perhaps the most mature agent-based meta-learning systems are: JAM [111], and BODHI [63]. JAM and BODHI are both intended for data classification.
2.3.3
Hybrid-learning Strategy
A hybrid learning strategy is a technique that combines local and remote learning for model building [46]. An example of a hybrid learning framework is Papyrus [12]. Papyrus is designed to support both learning strategies. In contrast to JAM and BODHI, Papyrus can not only move models from site to site, but can also move data when that strategy is desired. Papyrus is a specialised system which
is designed for clustering. These systems are reviewed in detail in [70].