Information
-
Patent Application
-
20030069652
-
Publication Number
20030069652
-
Date Filed
October 05, 200122 years ago
-
Date Published
April 10, 200321 years ago
-
CPC
-
US Classifications
-
International Classifications
- G05B013/02
- G06F015/18
- G06N003/08
Abstract
A regulation method for decision tree construction is described wherein decision rules can be automatically adjusted between crisp and soft decisions. Starting with a conventional decision tree, additional statistics are stored at the terminal and non-terminal nodes during training and used during application to new samples. The regulation process allows automatic determination of tree structure. It also allows incremental updating of a regulation decision tree with graceful change to classification performance characteristics. Compound regulation decision trees are described for use to update the decision structure when new training input samples include new classes. Methods for pruning regulation decision trees, for focusing regulation decision trees, for determining optimal depth and regulation parameters and for determining optimal sample weighting are taught.
Description
TECHNICAL FIELD
[0001] This invention is related to intelligent systems, classifications and decision making and more particularly to regulating the quality of decision rules automatically.
BACKGROUND OF THE INVENTION
[0002] Many intelligent system applications such as data mining, automatic process control, automatic target recognition, intelligent database search, data warehousing, and machine vision perform decision making using rules derived from offline training or online learning. Prior art approaches can be used for decision rule generation. These include knowledge acquisition methods in expert systems, statistical discriminate analysis, Bayesian decision theory, Bayes belief networks, fuzzy systems, artificial neural networks, genetic algorithms, etc.
[0003] Several of the approaches are capable of generating complicated decision rules to optimize decisions for the training data and yield superior re-substitution (test on training data) results. In simple applications, almost all above approaches could result in reasonable performance. However, due to the dynamic nature of many applications, unforeseen conditions or data are often encountered online that challenge the decision rules created without the benefits of the new information. Decision rules specifically optimized for the earlier training data may fail on the new data. Thus, they lack robustness.
[0004] To overcome the difficulty of non-robust performance, prior art approaches divide available data into training and testing sets. They use the training set to generate decision rules and use the test set to evaluate the robustness of the decision rules generated from the training set. This approach could improve the robustness of the decision rules. However, it is inefficient since it generates decision rules from only partial data and most of them cannot use new data to update decision rules incrementally.
[0005] A Decision tree is a popular prior art set of decision rules. A typical decision tree classifier makes crisp decisions. That is, it makes decisions following a definitive path of decision structure and assigns a class unequivocally to an input sample. This method supports applications with discontinuous decision boundaries well and is desirable in classification applications where context switching is required around decision boundaries. However, in applications that require generalization or in applications where the training data cannot accurately predict decision boundaries or when the input samples are subject to noise and therefore perturb around the decision boundaries, a smooth decision around a decision boundary is desirable and more robust.
[0006] Most of the decision methodologies such as decision trees are not designed to allow for incremental update. There is no easy way to incrementally update a decision rule using new training samples after the tree is constructed. Alternatively, completely new rules are constructed when new samples are available. However, the new rules may have very different performance characteristics from the old ones. This is not desirable in critical applications where performance characteristics should be stable and update learning should change the performance characteristic gracefully.
OBJECTS AND ADVANTAGES
[0007] It is an object of this invention to provide a method for regulating the quality of decision rules automatically. Such regulation provides a balance between crisp and soft decisions and allows for the adjustment of the balancing point to match an application.
[0008] It is a further object to provide automatic optimization of decision rules using all training samples.
[0009] It is an object of the invention to allow for automatic optimization of hierarchical decision structures generated from other approaches.
[0010] It is an object of the invention to allow graceful incremental updating of decision structures using newly acquired samples from existing and new classes.
[0011] It is an object of the invention to improve regulation tree accuracy by using multiple stages of trees focusing toward the detailed decisions near decision boundaries through further stages of processing.
SUMMARY OF THE INVENTION
[0012] This invention describes new methods for regulating hierarchic decisions in intelligent systems. Storing additional statistics acquired during training for both terminal and non-terminal nodes alters a conventional decision tree structure. These statistics are used to create smooth decision boundaries with the smoothness adjustable and optimizable through the choice of sample weights and regulation parameters. Methods for trimming regulation trees and focusing them arise through selection of weights, tree depth, regulation parameters, and evaluation of accuracies obtained.
[0013] Methods for incrementally increasing the number of training samples for an already trained regulation decision tree provide graceful change to classification characteristics. Methods for construction of compound trees support the incremental addition of new trained classes and new training samples.
BRIEF DESCRIPTION OF THE DRAWINGS
[0014] The preferred embodiments and other aspects of the invention will become apparent from the following detailed description of the invention when read in conjunction with the accompanying drawings which are provided for the purpose of describing embodiments of the invention and not for limiting same, in which:
[0015]
FIG. 1 shows the construction process for a regulation decision tree;
[0016]
FIG. 2 shows the computational structure of a typical binary decision tree classifier;
[0017]
FIG. 3 shows the non-terminal node likelihood value determination process;
[0018]
FIG. 4 shows a regulation tree with its associated likelihood values for an input sample Xinput;
[0019]
FIG. 5 shows the process for automatically determining the appropriate depth of a tree and the appropriate regulation parameter;
[0020]
FIG. 6 shows the compound tree new tree construction flow;
[0021]
FIG. 7 shows regulation tree incremental updating with new samples;
[0022]
FIG. 8 shows a compound tree classification flow and confidence calculation;
[0023]
FIG. 9 shows a method for pruning a regulation tree;
[0024]
FIG. 10 shows a method for constructing a focusing tree from a regulation tree.
DETAILED DESCRIPTION OF THE INVENTION
[0025] This invention provides a method to regulate the quality of decision rules automatically by providing an adjustable balance between crisp and soft decisions to match an application. Methods for incremental updating of the decision structure by additional learning samples from the same classes and for adding new classes are taught. A preferred embodiment of the invention can be constructed using a decision tree hierarchical decision structure.
[0026] I. Decision Tree
[0027] A decision tree makes a decision using hierarchical decision structures implemented as a tree. (Breiman L., Friedman J. H., Olshen R. A. and Stone C. J., “Classification and Regression Trees”, Chapman and Hall/CRC, 1984 pp18-58) A tree consists of at least one non-terminal node and at least as many terminal nodes as the number of decision outcomes to be decided. Each outcome has associated at least one terminal node, and the non-terminal nodes represent various collections of mixed outcomes. The root node represents the entire collection of outcomes into which a new sample may be decided.
[0028] Almost all decision trees are binary decision trees where each non-terminal node branches out into two descending nodes.
[0029] Without loss of generality, we use binary decision tree classifiers in the descriptions of this invention. Those skilled of the art should recognize that the method of this invention is applicable to other type of decision trees or other types of parametric or non-parametric decision methods.
[0030] I.1 Binary Decision Tree Classifier
[0031]
FIG. 2 shows the computational structure for a typical binary decision tree classifier. A binary decision tree has two descendant paths for each non-terminal node. There is a decision rule associated with each non-terminal node to determine the descendant path for a sample at the node, until a terminal node is reached. The decision outcomes of a decision tree classifier are classes associated with the input data. Examples of classes include disease types, defect classifications, weather patterns, etc.
[0032] In the classification mode, an unknown sample enters the decision tree at the root node and the decision rule associated with the root node is applied to the sample's feature vector to determine the descendant path that the sample will follow. This process is repeated through descending non-terminal nodes until a terminal node is reached. Every terminal node has an associated class to which the sample is assigned.
[0033] The commonly used decision rule for each non-terminal node is a thresholding value for a discrimination function associated with the node. In one embodiment of the invention, if the node's discrimination function value is less than or equal to a threshold value, the left child is taken; otherwise, the right child is taken for the descendant path. Feature thresholding is the simplest and most easily understandable discrimination function. Other discrimination functions such as Fisher linear decision function, Bayes linear decision function, Bayes quadratic decision function and other single stage decision rules can also be used. Note that it is harder to interpret a decision tree when multiple features are involved in a discrimination function.
[0034] I.2 Binary Decision Tree Construction Procedure
[0035] Given a set of training samples, a binary decision tree can be constructed automatically by a divide and conquer procedure. All training samples are used to construct the root node. Each subsequent node is trained by a subset of the training samples. The decision tree construction procedure is as follows:
[0036] 1. For a given node n with associated training sample set Un, sort the samples in an ascending order according to their discrimination function values, i.e. f(Xkn)≦f(Xk+1n). In the case of a feature thresholding method, the sorting is performed for each of the available features so that both feature and threshold value selection can be accomplished simultaneously.
[0037] 2. A set of candidate thresholds Tn is defined by
1
[0038] 3. For each partition at a candidate threshold, calculate the following parameters:
[0039] the weighted number of class c samples assigned to LEFT, NLc, and the weighted number of class c samples assigned to RIGHT, NRc.
2
[0040] Where wic is the weighting factor for sample i belonging to class c.
[0041] The total weighted number of samples assigned to LEFT and RIGHT by the partition:
3
[0042] Evaluation functions to be used for the partition selection at node n include:
[0043] Purity (Entropy):
4
[0044] Where
5
[0045] Purity has the maximum value when the training samples are completely separated in LEFT and RIGHT.
[0046] Another criteria that can be used for an evaluation function is the probability of correct classification
[0047] 4. Select the partition for node n as the one that maximizes the evaluation function.
[0048] 5. Check the following stopping criteria (OR conditions):
[0049] Reached the maximum allowable level of tree: L=log2 N−1 or a user specified limit. Where N is the total number of training samples.
[0050] χ2 value is smaller than an allowable limit.
[0051] Where
6
[0052] and Ncn is the weighted number of samples of class c at node n. Nn is the total weighted number of samples at node n.
[0053] Nn value is smaller than an allowed limit.
[0054] Type I error>limit. Where type I error is the probability that a sample whose true class is in LEFT yet it is classified as RIGHT class
[0055] Type II error>limit. Where type II error is the probability that a sample whose true class is in RIGHT yet it is classified as LEFT class
[0056] 6. If none of the stopping criteria is met, assign node n as a non-terminal node and use the step 4 selected partition for this node
[0057] 7. If at least one of the stopping criteria is met, assign node n as a terminal node, and assign the most probable class from its associated training samples as the classification outcome for this node.
[0058] II. Decision Tree Regulation
[0059] A typical decision tree classifier makes crisp decisions. That is, it makes decisions following a definitive path of the decision structure and assigns a class unequivocally to an input sample. This method supports discontinuous decision boundaries well and is desirable in classification applications where context switching is required around decision boundaries. However, in applications that require generalization or in applications where the training samples cannot accurately predict decision boundaries or when the input samples are subject to noise and therefore perturb around the decision boundaries, a smooth decision around the decision boundary is desirable and more robust.
[0060] This invention provides a decision tree regulation method and new decision structure that supports soft decisions. It provides a balance between crisp and soft decisions. It allows for the adjustment of the balancing point to match an application.
[0061] II.1 Regulation Tree Construction
[0062] A regulation tree can be derived from an existing decision tree as shown in FIG. 1. A regulation tree has the same structure as the existing decision tree. It derives and stores additional statistics at each node. The regulation tree construction process inputs a decision tree and a set of training samples 100. In a preferred embodiment of the invention, the regulation tree construction process includes the following steps:
[0063] 1. For each non-terminal node n, determine the distance-to-threshold (din) values for each of the training sample i associated with this node 102. The weighted mean (μdn) and standard deviation (σdn) for the distance values are derived from the training sample distance values 104 and stored in the node for the classification of new samples. The weighting factors are the weights associated with each training sample. Weights can be associated with samples on a variety of basis such as with the confidence of representation or accuracy of data acquisition, significance to a class determination, or other emphasis criteria. Equal weights can be applied if no additional information is available. Weights can also be automatically determined by a process such as described for tree focusing in section V. In one embodiment of the invention, a simple method accumulates the weighted distance value using the following rule:
Distancen(k+1)=Distancen(k)+wk+1*dnk+1
Weightn(k+1)=Weightn(k)+wk+1
SquareDistancen(k+1)=SquareDistancen(k)+wk+1*dn2k+1
[0064] After the accumulation of all training samples associated with each node (Nn), the mean distance for node μdn is determined by
μdn=Distancen(Nn)/Weightn(Nn)
[0065] and the mean square distance for node n s2dn is determined by
s
2
d
n
=SquareDistance(Nn)/Weight(Nn)
[0066] The distance standard deviation σdn value is determined by
σdn={square root}{square root over (sd2n−(μdn)2)}
[0067] To avoid the problem of outliers, an alternative embodiment of the invention can be derived by ranking di and find the weighted median value (mdn) and the range between the 10% and 90% values of the distance (Rn10%-90%). The mean distance and the distance standard deviation can be estimated from mdn and Rn10%-90% as follows:
μdn=mdn
σdn=Rn10%-90%/2.56
[0068] Those skilled in the art should recognize that other methods could be used to estimate the mean and standard deviation values. For example, different percentiles can be used to measure the range for the standard deviation estimation. Furthermore, the weights for each sample can all be equal to one (set equally).
[0069] 2. For a terminal node n having Nn weighted training samples. Denote Ncn as the weighted class c training sample count at this terminal node n. In one embodiment of the invention, the likelihood value for class c at terminal node n is determined 112 as:
L
n
class
—
c
=N
c
n
/N
n
[0070] An alterative method for likelihood value calculation is:
L
n
class
—
c
=N
c
n
/(Nn+β).
[0071] Where β is a regulation parameter that weakens the likelihood values for terminal nodes having a small number of training samples. In a preferred embodiment there is one value of β that is common to all terminal nodes 110. The likelihood values are stored for each terminal node of the tree.
[0072] II.2 Regulation Tree Application
[0073] Given an input sample, Xinput, the likelihood values at each non-terminal node n are determined first. To determine the likelihood value for a non-terminal node n, the distance to threshold value for Xinput is calculated as:
d
input
n
=f
n
(Xinput)−Tn
[0074] In one embodiment of the invention, the likelihood value determination process is shown in FIG. 2. In FIG. 2 non-terminal nodes 200, 202, 204, 206 are shown as clear ovals, terminal nodes 208, 210, 212, 214, 216 are blacked out ovals. The decision rule for a node n is indicated as fn(.). A threshold value for a node n is indicated as Tn. If distance value, dinputn, is zero or negative, the likelihood value for descending through the left branch, Lleft, is one in a crisp tree. In the regulation tree the probability that the sample could descend through the right branch is calculated by the rule as follows:
7
[0075] Pnright 300 is indicated diagrammatically as a shaded area in FIG. 3. FIG. 3 indicates an assumed Gaussian example probability density function separately derived for each node n from the set of training samples. The Gaussian example is used to illustrate a preferred embodiment and is not limiting. Other distribution functions or empirical models can also be used. Actual statistics at each node should be determined during training. For the example non-terminal node n having standard deviation σdn 304 and mean μdn 306 wherein the area 300 under the probability density curve 302 is a function of the distance to threshold dinputn 308. The probability of being a sample that should descend through the left branch is:
P
n
left
=1−Pnright
[0076] The likelihood value, Lleft, of the regulation tree for descending through the left branch can be calculated as a linear combination of the crisp tree value (i.e. 1) and Pnleft by a regulation parameter α. That is,
L
n
left
=0.5+α+(0.5−α)Pnleft
[0077] and
L
n
right
=1−Lnleft
[0078] In a preferred embodiment there is the same value of α for every non-terminal node in the regulation decision tree 108. If distance value, dninput, is positive, the likelihood value for descending through the left branch, Lnleft, and the likelihood value for descending through the right branch, Lnright can be similarly determined.
[0079] Note that the α value regulates an adjustable condition between the crisp tree and the probabilistic tree. When α=0.5, the tree reverts to the original crisp tree. When α=0, the tree averages the original crisp tree and a complete probabilistic tree with equal weight. When α=−0.5, the tree is a complete probabilistic tree.
[0080] Since a sample (Xinput) most likely has non-zero branch likelihood values for each of the terminal nodes, in the preferred embodiment of the invention the confidence value corresponds to class c for Xinput, Confidencec(Xinput), is (see also FIG. 4):
8
[0081]
FIG. 4 shows the example data for confidence determination diagrammatically for a tree having 3 non-terminal nodes 400, 402, 404 and 4 terminal nodes 406, 408, 410, 412 for classifying input samples into two classes, Class1 or Class2. For any input sample Xinput the likelihood values that it will pass down any branch of the tree can be determined as illustrated in Section II.2. For example, in FIG. 4 the likelyhood that it will pass down branch 414 has been determined to be L1_left(Xinput). In the FIG. 4 example, these likelihood determinations are labeled for each branch of the tree for the particular input sample. Determine that the new sample Xinput is in a particular class as:
Confidence_class1(Xinput)=L4_class1*L2_left(Xinput)*L1_left(Xinput)+L5_class1*L2_right(Xinput)*L1_left(Xinput)+L6_class1*L3_left(Xinput)*L1_right(Xinput)+L7_class1*L3_right(Xinput)*L1_right(Xinput)
Confidence_class2(Xinput)=L4_class2*L2_left(Xinput)*L1_left(Xinput)+L5_class2*L2_right(Xinput)*L1_left(Xinput)+L6_class2*L3_left(Xinput)*L1_right(Xinput)+L7_class2*L3_right(Xinput)*L1_right(Xinput)
[0082] where the likelihood value for each class is known for each of the terminal nodes and depicted for node 4 (406) as L4_Class1 and L4_Class2, node 5 (408) as L5_Class1 and L5_Class2, node 6 (410) as L6_Class1 and L6_Class2 and node 7 (412) as L7_Class1 and L7_Class2. The associated Likelihood values determined for a particular input sample Xinput that are referred to in the equations above are L1_left(Xinput) (414), L1_right(Xinput) (416), L2_left(Xinput) (418), L2_right(Xinput) (420), L3_left(Xinput) (422), L3_right(Xinput) (424).
[0083] II.3 Automatic Tree Regulation Process
[0084] The recursive partitioning method of constructing a decision tree often results in a very complex tree that over-fits the data. To automatically determine the appropriate depth of a tree and the appropriate regulation parameter α, this invention includes a tree regulation process. In one embodiment of the invention, the process (also shown in FIG. 5) is described in the following steps:
[0085] 1. From the minimum allowable tree depth 500 to the maximum allowable tree depth 508, 516
[0086] a. Construct a decision tree up to the given depth 502
[0087] b. Derive the corresponding regulation tree 504
[0088] c. Determine the projected tree accuracy and its associated regulation parameter α value 506
[0089] 2. Select the optimal depth that yields the highest projected tree accuracy 510
[0090] 3. Use the regulation parameter value corresponding to the optimal depth as the final regulation parameter α value 512
[0091] In the preferred embodiment of the invention, determine the projected tree accuracy for a regulation tree for all training samples at α values ranging from 0 through 0.5 in 0.05 increments. For a given tree depth d, this results in accuracy values Ad0, Ad0.05, Ad0.1, through Ad0.5. Accuracy is calculated for a particular tree depth and α as follows:
9
[0092] Ci is defined to be the true class of a sample i. The accuracy values are determined using the result for classification of samples combined using their sample weighting values (Wi) and including global class weights (WCi) for emphasis of particular classes within the accuracy classification. Those skilled in the art should recognize that WCi can be set to an equal value (or 1) for all classes. In the preferred embodiment of the invention, the projected tree accuracy Ad is calculated as:
A
d
=μdA−Rd/9
[0093] Where μdA is the trim mean (i.e. trim 2 outliers from the 11 Adα values) of the 9 medium Adα values and Rd is the range of the remaining 9 medium Adα values.
[0094] An alternative embodiment of the invention, the projected tree accuracy Ad can be calculated as:
A
d
=μdA−Rd/9(1+t(d−dmin)/dmax−dmin))
[0095] Where dmin is the minimum allowable depth, dmax is the maximum allowable depth, and t is a parameter that penalizes deeper trees. In the preferred embodiment of the invention, 0.1 is used for t. Other values can be used in alternative embodiments. The α whose Adα value is closest to Ad is selected as the regulation parameter for the regulation tree at depth d.
[0096] III. Regulation Tree Update Learning
[0097] There is no easy way to incrementally update a decision tree classifier using new training samples after the tree is constructed. However, computer speed and storage capability have improved dramatically so that it has become practical to store all training samples and reconstruct a new tree using old and new samples when new samples are available. However, the new tree may have very different performance characteristics for certain types of samples after it is re-built. This is not desirable in critical applications where performance characteristics should be well understood and updated learning should only change this characteristic gracefully. The regulation tree methodology of this invention offers the unique advantage of graceful incremental update for the classifier. Using new samples from existing classes, regulation tree statistics can be incrementally updated to gracefully optimize performance for new samples. Both terminal node and non-terminal node statistics can be updated by the new samples of existing classes. When new samples are from new classes, a compound tree structure can be constructed to handle new classes yet maintain stable performance to samples from existing classes.
[0098] III.1 Terminal Node Update for New Samples from Existing Classes
[0099]
FIG. 7 shows the steps involved in updating the terminal nodes for new samples from existing classes. To facilitate updated learning, the total weighted training sample count, Nn, and the weighted training sample counts for each class c, Ncn, are stored in each terminal node. To update the terminal node statistics, a new training sample i is classified using a crisp decision method (702). Therefore, only one terminal node result is associated with the new training sample. Let the true class of the new training sample i be c and its associated terminal node be n. In one embodiment of the invention, the terminal node update process updates only terminal node n. Nn and Ncn are updated by the following rules (704):
N
n
=N
n
+w
i
[0100] and
N
c
n
=N
c
n
+w
i
[0101] After updating, the terminal node likelihood values of each class are updated using the rules disclosed in section II.1.2 (706).
[0102] III.2 Non-terminal Node Update for new Samples from Existing Classes
[0103] Given a new training sample i, each non-terminal node visited by the new training sample when using the crisp decision method should be updated. To support the update, the accumulated distance (Distance), weight (Weight) and square distance (SquareDistance) values or a weighted distance histogram are stored at each non-terminal node of the regulation tree.
[0104] To update a non-terminal node n, determine the distance-to-threshold (di)n values for the training sample i. In the case of Distancen, Weightn and SquareDistancen approach, the Distancen, Weightn and SquareDistancen values are updated as follows:
Distancen(new)=Distancen(old)+wi*din
Weightn(new)=Weightn(old)+wi
SquareDistancen(new)=SquareDistancen(old)+wi*dn2i
[0105] The updated distance weighted mean (μndi) and standard deviation (σndi) are:
μnd=Distancen(new)/Weightn(new)
s
n2
d
=SquareDistancen(new)/Weightn(new)
10
[0106] In the case of a weighted distance histogram approach, the weighted histogram is updated by the new sample din and wi. In one embodiment of the invention, the weighted median point (mnd) and the range between the 10% and 90% values of the distance (Rn10%-90%) can be determined from the updated histogram. The mean distance and the distance standard deviation are updated as follows:
μnd=mnd
σnd=Rn10%-90%/2.56
[0107] II.3 Regulation Tree Update for New Classes
[0108] In this invention new samples from new classes are incorporated by creating a compound regulation tree structure. This new structure (shown in FIG. 6) provides graceful update of the overall classifier yet maintains stable performance for the existing classes. The compound regulation tree structure includes multiple trees that are trained from an expanding training data base 600 that includes old and new training samples. The first tree 602 is the original regulation tree that is subject only to terminal and non-terminal node update learning of new samples from existing classes. The other trees 604, 608, 610, 612 correspond to new classes that are sequentially trained into the decision system through new class addition. Each added tree incorporates a new class.
[0109] The preferred embodiment for new class update learning flow is shown in FIG. 6. FIG. 6 assumes four new classes are update learned into the compound regulation tree structure.
[0110] III.3.1 New Tree Construction
[0111] Update learning of the first new class (new class 1) triggers the generation of the second tree 604 using the samples available for the new class 1 update learning along with the existing training data.
[0112] In one embodiment of the invention, to enable new tree construction, the system stores all training data with their weights 600. For new samples from new classes, the system appends the samples with their class label into the training data. The system checks the new class sample size. If the sample size is greater than the minimal required sample size for a new class, a new tree is constructed for all existing classes and the new class.
[0113] Similarly, trees are generated for new class 2608 through new class 4612 as shown in FIG. 6.
[0114] III.3.2 Compound Tree Update
[0115] For a compound tree with 4 new classes, a sample belonging to the original classes will update all trees 613, 614, 616, 618, 620. A sample from new class 1 will update all trees except for the tree for original classes 614, 616, 618, 620. A sample from new class 2 will update all trees except for the tree for original classes and the tree for new class 1616, 618, 620. Similarly, a sample from new class 3 will update the tree for new class 3 and the tree for new class 4618, 620. A sample from new class 4 will update only the tree for new class 4620. In one embodiment of the invention, the same terminal and non-terminal node update methods as described in sections III.1 and III.2 are used for the update.
[0116] III.3.3 Compound Tree Application
[0117] The compound tree classification flow is shown in FIG. 8. In this preferred embodiment, trees for 4 new classes are generated with a total of 5 trees 802, 804, 806, 808, 810. For a sample 800 to be applied, all trees are applied and the results are combined. To combine the results from all trees, the resulting confidence value for new class4 (C4) 811 from the tree for new class4 is used as the confidence value for new class4. The resulting confidence value for new class3 (C3) from the tree for new class3 times (1-C4), i.e. (1-C4)C3, 809 is used as the confidence value for new class3. Similarly, the resulting confidence value for new class2 (C2) from the tree for new class2 times (1-C4)(1-C3), i.e. (1-C4)(1-C3)C2, 807 is used as the confidence value for new class2. The resulting confidence value for new class1 (C1) from the tree for new class1 times (1-C4)(1-C3)(1-C2), i.e. (1-C4)(1-C3)(1-C2)C1, 805 is used value for new class1. Finally, the resulting confidence value for an original classi(ci) from the tree for existing classes times (1-C4)(1-C3)(1-C2)(1-C1), i.e. (1-C4)(1-C3)(1-C2)(1-C1)ci, 803 is used as the confidence value for an original classi.
[0118] IV. Regulation Tree Pruning
[0119] To avoid the over-fitting of data, a decision tree is often pruned (Quinlan, J. Ross, “Programs for Machine Learning”, Morgan Kaufmann 1993 pp 35-42). Discarding one or more sub-trees and replacing them with terminal nodes usually simplify decision trees. The class associated with a terminal node is the one with the most frequent (after weighting) class from the training samples. In the prior art approach there are two methods of tree pruning: Cost-complexity pruning and Reduced-error pruning. These methods evaluate the error criteria for the samples that fall into a non-terminal node that precedes a pair of terminal nodes. If the error criteria result is in favor of the combined nodes, the sub-tree is pruned. These methods perform pruning using partial data and are not reliable due to data variations.
[0120] In the preferred embodiment of this invention, tree pruning is accomplished by the regulation tree method. The regulation tree method uses all training samples to determine the tree pruning status for each sub-tree. This better use of the training data usually achieves more robust results. As shown in FIG. 9, the regulation tree pruning procedure includes the following steps:
[0121] 1. Beginning with a regulation tree or compound regulation tree 900;
[0122] 2. For each non-terminal node of the tree having two descending terminal nodes,
[0123] a. Determine the accuracy to all training samples under the following two conditions 902:
[0124] i. Consider the two terminal nodes separately and result in accuracy a1
[0125] ii. Combine the two terminal nodes into one node and result in accuracy a2;
[0126] b. If a1>a2, no pruning of the terminal node pairs 904;
[0127] c. Otherwise, prune the terminal nodes by combing the two terminal nodes and converting the associated non-terminal nodes into one terminal node as follows 906:
[0128] Total sample count: Ncombine=Nfirst—node+Nsecond—node and For each class c, update sample count by Nccombine=Ncfirst—node+Ncsecond—node
[0129] 3. If changes occur in the previous iteration, repeat process 1908. Otherwise, the pruning process is completed 910. Those skilled in the art should recognize that the pruning procedure can be conducted on a subset of tree nodes. Therefore, Step 3 may not have to be carried out.
[0130] V. Focusing Tree
[0131] Boosting is a technique for improving the performance of learning algorithms by manipulating the distribution of the training examples. (Freund, Y., R. E. Schapie, Iyer, R., Singer, Y. 1998 “An Efficient Boosting Algorithm for Combining Preferences”, Machine Learning: Proceedings of the Fifteenth International Conference 1998).
[0132] In a preferred embodiment, regulation tree accuracy is improved by multiple stages of trees focusing toward the detailed decisions near the decision boundaries through further stages of processing. This is accomplished by increasing weights to the training samples that are close to decision boundaries and decreasing weights for samples that are far away from the decision boundaries and also by constructing an additional regulation tree (called a focusing tree) using the new weights.
[0133] In a preferred embodiment of the invention, the discrimination merit of a regulation tree to an input sample Xinput of class c is calculated as
11
[0134] Where Confidencec(Xinput) is the confidence value for the correct classification as class c. Confidencem(Xinput) is the highest confidence value for an incorrect classification and m is the class corresponding to the incorrect classification. t is set to 0.001 in a preferred embodiment of the invention.
[0135] A weight update factor Kinput is determined as
12
[0136] Where Klow and Khigh are the lower and upper bounds of the weight update factor. The weight for the sample Xinput is adjusted by the following rule:
w
input
(new)=Kinput*winput(old)
[0137] In a preferred embodiment of the invention, a focusing tree of a regulation tree is generated by the following procedure (see FIG. 10):
[0138] 1. Starting with a regulation tree or a compound regulation tree having training samples 1000;
[0139] 2. For each training sample,
[0140] a. determine its discrimination merit 1002
[0141] b. derive the weight update factor 1004
[0142] c. generate the new weight 1006
[0143] 3. Generate a new regulation tree using the new weight 1008 of all training samples. This is the focusing tree 1010
[0144] The process can be repeated to generate multi-stage focusing trees.
[0145] In a preferred embodiment of the invention, the classification reliability of a regulation tree to an input sample Xinput is calculated by
13
[0146] Where class c1 yields the highest confidence value and class c2 yields the second highest confidence value for Xinput. p is set to 0.001 in a preferred embodiment of the invention.
[0147] In the application phase of a two-stage focusing tree, the following procedure applies:
[0148] 1. Classify the input sample by the first regulation tree
[0149] 2. If the classification reliability>threshold, use the current result as the final result and stop
[0150] 3. Else, Classify the input sample by the second (focus) regulation tree and use the new result as the final result
[0151] This process is repeated multiple times for multi-stage focusing trees.
[0152] The invention has been described herein in considerable detail in order to comply with the Patent Statutes and to provide those skilled in the art with the information needed to apply the novel principles and to construct and use such specialized components as are required. However, it is to be understood that the inventions can be carried out by specifically different classifiers and devices, and that various modifications, both as to the classification details and operating procedures, can be accomplished without departing from the scope of the invention itself.
Claims
- 1. Method for regulation tree construction comprising the following steps:
(a) input a decision tree; (b) input a set of training samples; (c) determine statistics from the training samples for at least one non-terminal node; (d) determine statistics from the training samples for at least one terminal node. (e) select regulation parameters
- 2. The method of claim 1 wherein the statistics for at least one non-terminal node include mean distance.
- 3. The method of claim 1 wherein the statistics for at least one non-terminal node include distance standard deviation.
- 4. The method of claim 1 wherein the statistics for at least one terminal node include the likelihood value for a class.
- 5. A regulation tree application method comprises the following steps:
(a) input a sample; (b) determine the likelihood values for at least one non-terminal node; (c) determine the likelihood value for a branch to at least one terminal node; (d) determine the confidence value for at least one class.
- 6. The method of claim 5 wherein the likelihood values for at least one non-terminal node comprises a likelihood value for descending through the left branch and a likelihood value for descending through the right branch.
- 7. The method of claim 5 wherein the confidence value for a class c is determined by the following formula:
- 8. An automatic tree regulation method comprises the following steps:
(a) determine the projected tree accuracies for a plurality of depths and a plurality of regulation parameter values; (b) select the optimal depth that yields the highest projected tree accuracy; (c) use the optimal regulation parameter value for the optimal depth
- 9. The method of claim 8 wherein the projected tree accuracy is determined by the following steps:
(a) construct a regulation tree up to a given depth; (b) determine the projected tree accuracy; (c) determine a regulation parameter value based on projected tree accuracy.
- 10. A regulation tree terminal node update learning method comprises the following steps:
(a) input a training sample; (b) input the true class of the training sample; (c) classify the input training sample using a crisp decision method to determine its associated terminal node; (d) update terminal node statistics
- 11. The method of claim 10 wherein the terminal node statistics include Nn and Ncn.
- 12. A regulation tree non-terminal node update learning method comprises the following steps:
(a) input a training sample; (b) input the true class of the training sample; (c) classify the input training sample using a crisp decision method to determine its association with at least one non-terminal node; (d) update non-terminal node statistics.
- 13. The method of claim 12 wherein the non-terminal node statistics include mean distance and the distance standard deviation.
- 14. A regulation tree update learning method for new classes comprises the following steps:
(a) perform new regulation tree construction; (b) perform a compound tree update.
- 15. The method of claim 14 wherein the new tree construction further comprises the following steps:
(a) input at least one sample from a new class; (b) check to confirm the sample size is greater than the minimal required sample size for the new class; (c) construct a new compound tree for all existing classes and the new class.
- 16. The method of claim 14 wherein the compound tree update further comprises the following steps:
(a) input a new sample and its class; (b) update all trees trained to include the input class.
- 17. A compound tree application method comprises the following steps:
(a) input a sample to be applied; (b) apply the sample to all trees; (c) combine the results from all trees.
- 18. A regulation tree pruning method comprises the following steps:
(a) input a set of training samples; (b) for a non-terminal node of the tree having two descending terminal nodes, determine the accuracies for identical training samples under two separate nodes or combined node conditions; (c) if combined node accuracy is greater than the two separate node accuracy, prune the terminal nodes by combing the two terminal nodes and converting the associated non-terminal nodes into one terminal node.
- 19. The method of claim 18 wherein combing the two terminal nodes combines their sample counts.
- 20. A focusing tree construction method comprises the following steps:
(a) input a set of training samples; (b) generate a new weight for each training sample; (c) generate a new tree using the new weight.
- 21. The method of claim 20 wherein generate new weight for each training sample further comprises the following steps:
(d) determine the sample's discrimination merit; (e) derive the sample's weight update factor; (f) generate the new weight for the sample.
- 22. A focusing tree application method comprises the following steps:
(a) input a sample to be applied; (b) classify the input sample by the first tree; (c) if the classification reliability>threshold, use the current result as the final result and stop; (d) else, classify the input sample by the focus tree and use the new result as the final result.
- 23. The method of claim 22 wherein the classification reliability to an input sample Xinput is calculated by
- 24. The method of claim 20 is repeated multiple times to create multi-stage focusing trees.