Transcript KBS和KM
Discovering Interesting Patterns for Investment Decision Making with GLOWER-A Genetic Learner Overlaid With Entropy Reduction Advisor： Dr. Hsu Graduate：Min-Hong Lin IDSL seminar 2001/12/11 Outline Motivation Objective Genetic Search for Rules Genetic Rule Discovery Comparison of GLOWER to Tree and Rule Induction Conclusions Comments Motivation Prediction in financial domains is difficult to model. The dimensionality is high. Relationships among variables are weak and nonlinear. Variable interactions can be significant. Objective Describe and evaluate several variations of a new genetic learning algorithm (GLOWER) on a variety of data sets. Comparison of GLOWER variants. Comparison of GLOWER to Tree and Rule Induction. Genetic Search for Rules Genetic algorithms (Packard, 1989; Goldberg, 1989, Holland, 1992) have been shown to be well suited to learning rule-like patterns. They have the ability to search large spaces for pattern without resorting to heuristics that are biased against term interactions. Limitations of Genetic Search Lack of speed Randomness in creating the initial population They can be myopic after find one good solution Benefits of Genetic Search Scour a search space thoroughly Allow arbitrary fitness functions in the search The failings of greedy search Tree Induction algorithms are currently among the most widely used techniques in data mining. But greedy search conducted by TI algorithms can overlook good patterns. See next example An example An example(Cont’d) Entropy Reduction Tree induction algorithms typically determine split points based on a heuristic such as entropy reduction Entropy Reduction(cont’d) The entropy of a cluster i can be computed using the standard formula: The gain from a split is computed based on the difference between the entropy of the parent cluster and the entropy of the child clusters resulting from the split. An example(Cont’d) Evaluation of partial models Two commonly used metrics used to measure the goodness of individual rules are confidence and support. If N is the total number of examples in a data set, then, for a rule form A à B: Genetic Rule Discovery Representation: Gene and Chromosome Genetic Rule Discovery Process An initial population of patterns (chromosomes) is created randomly. Each chromosome is evaluated and ranked. The higher-ranked chromosomes are selected to participate in “mating” and “mutation” to produce new offspring. Mating involves exchanging parts of chromosomes (genes) between pairs. (Crossover) Mutating involves changing the value of a gene Genetic Rule Discovery Process(Cont’d) By repeating these steps over and over, the search converges on populations of better patterns, as measured using some fitness function. Genetic Rule Discovery (Cont’d) Selection (or reproduction) Probability of survival Selection Selection Crossover Crossover Mutation Genetic Rule Discovery (Cont’d) Genetic Rule Discovery (Cont’d) Two problems A1 is the dominant pattern in the data set, chromosomes typically gather in the neighborhood of A1. If such a pattern dominates the early populations, the algorithm is likely to overlook the other patterns. Approach(inductive strengthening ) Sequential Niching:SN Data Reduction:DR Sequential Niching:SN Niching schemes have been described for dealing with multimodal problems (Goldberg and Richardson (1987), Oei et al. (1991). inductive strengthening (Provost & Buchanan, 1992): placing stronger restrictions on the search based on the rules that have already been induced. Genetic search can perform inductive strengthening by niching sequentially. After the search converges on a high-fitness schema, the evaluation function can be modified to penalize patterns corresponding to it. Sequential Niching:SN(Cont’d) Data Reduction:DR The most commonly used heuristic for inductive strengthening is the covering heuristic. we can apply the covering heuristic to genetic search. Instead of penalizing an area as classified when calculating the fitness function, the data corresponding to the area are removed from further consideration. Compare DR to SN Compare DR to SN(Cont’d) The SN is more likely to produce non-overlapping rules whereas the DR produces a hierarchy of rules Compare DR to SN(Cont’d) Comparison of GLOWER variants Sampling Design GLOWER Variants sequential niching sequential niching plus entropy reduction data reduction plus entropy reduction Results Comparison of GLOWER to Tree and Rule Induction To predict “earnings surprises” Date set : a history of earnings forecasts and actual reported earnings for S&P500 stocks between 1990/1~1998/9 60 independent variables were chosen based on fundamental and technical analysis. Results Results(Cont’d) GLOWER is more thorough in its search than RL which is more thorough than TI GLOWER outperforms in both confidence and support for predicting positive surprises. For negative surprises, the confidence levels are similar, but GLOWER is better in support. GA is much more suited to capturing symmetry in the problem space TI and RL find it harder to predict a negative earnings surprise than a positive one. Discussion We are able to combine the speed of the traditional methods with the more comprehensive search of the GA. GA employ useful heuristics for achieving higher levels of support while maintaining high accuracy. The GA is about two or three orders of magnitude slower than a TI algorithm. Explainability is an important consideration in using knowledge discovery methods. Conclusions Genetic learning algorithms are more thorough in their search than other rule learning algorithms for hard problems. Genetic algorithms have the flexibility to accommodate variable fitness functions. It provides comparison and of the heuristics used by different rule-mining techniques. Comments GLOWER has the ability to uncover interesting patterns for difficult problems. Maybe the time complexity should be considered in the experiment.