CHAPTER 3. DEFINING DIVERSITY 47
We have now reviewed the state of the field with regard to explanations of what the
term ‘diversity of errors’ means, and why it can lead to improved ensemble performance. In
conclusion, the community puts great stead in the concept of classification error “diversity”,
though it is still an ill-defined concept. The lack of a definition for diversity has not stopped
researchers attempting to achieve it. So how do you make an effective, well-performing
ensemble?
3.2 Towards A Taxonomy of Methods for Creating Diversity
We have determined in the previous section that in both regression and classification contexts, the correlation between the individual predictor outputs has a definite effect on the
overall ensemble error, though for classification it is not yet formalised in the literature. In
this section we attempt to move towards a possible way to understand the many different
methods which researchers use to create an ensemble exhibiting error diversity; we show
that these schemes can be categorised along three “diversity-creation” axes.
In constructing the ensemble, we may choose to take information about diversity into
account, or we may not; i.e. we may or may not explicitly try to optimise some metric of
diversity during building the ensemble. We make a distinction between these two types,
explicit and implicit diversity methods. A technique such as Bagging is an implicit method,
it randomly samples the training patterns to produce different sets for each network; at no
point is a measurement taken to ensure diversity will emerge. Boosting is an explicit method,
it directly manipulates the training data distributions to ensure some form of diversity in
the base set of classifiers (although it is obviously not guaranteed to be the ‘right’ form of
diversity).
During learning, a function approximator follows a trajectory in hypothesis space. We
would like the networks in our ensemble to occupy different points in that hypothesis space.
While implicit methods rely on randomness to generate diverse trajectories in the hypothesis
space, explicit methods deterministically choose different paths in the space. In addition to
this high level dichotomy, there are several other possible dimensions for ensuring diversity
in the ensemble.
Sharkey [123] proposed that a neural network ensemble could be made to exhibit diversity by influencing one of four things: the initial weights, the training data used, the
CHAPTER 3. DEFINING DIVERSITY 47We have now reviewed the state of the field with regard to explanations of what theterm ‘diversity of errors’ means, and why it can lead to improved ensemble performance. Inconclusion, the community puts great stead in the concept of classification error “diversity”,though it is still an ill-defined concept. The lack of a definition for diversity has not stoppedresearchers attempting to achieve it. So how do you make an effective, well-performingensemble?3.2 Towards A Taxonomy of Methods for Creating DiversityWe have determined in the previous section that in both regression and classification contexts, the correlation between the individual predictor outputs has a definite effect on theoverall ensemble error, though for classification it is not yet formalised in the literature. Inthis section we attempt to move towards a possible way to understand the many differentmethods which researchers use to create an ensemble exhibiting error diversity; we showthat these schemes can be categorised along three “diversity-creation” axes.In constructing the ensemble, we may choose to take information about diversity intoaccount, or we may not; i.e. we may or may not explicitly try to optimise some metric ofdiversity during building the ensemble. We make a distinction between these two types,explicit and implicit diversity methods. A technique such as Bagging is an implicit method,it randomly samples the training patterns to produce different sets for each network; at nopoint is a measurement taken to ensure diversity will emerge. Boosting is an explicit method,it directly manipulates the training data distributions to ensure some form of diversity inthe base set of classifiers (although it is obviously not guaranteed to be the ‘right’ form ofdiversity).During learning, a function approximator follows a trajectory in hypothesis space. Wewould like the networks in our ensemble to occupy different points in that hypothesis space.While implicit methods rely on randomness to generate diverse trajectories in the hypothesisspace, explicit methods deterministically choose different paths in the space. In addition tothis high level dichotomy, there are several other possible dimensions for ensuring diversityin the ensemble.Sharkey [123] proposed that a neural network ensemble could be made to exhibit diversity by influencing one of four things: the initial weights, the training data used, the
การแปล กรุณารอสักครู่..
