1AI::DecisionTree(3) User Contributed Perl Documentation AI::DecisionTree(3)
2
3
4
6 AI::DecisionTree - Automatically Learns Decision Trees
7
9 version 0.11
10
12 use AI::DecisionTree;
13 my $dtree = new AI::DecisionTree;
14
15 # A set of training data for deciding whether to play tennis
16 $dtree->add_instance
17 (attributes => {outlook => 'sunny',
18 temperature => 'hot',
19 humidity => 'high'},
20 result => 'no');
21
22 $dtree->add_instance
23 (attributes => {outlook => 'overcast',
24 temperature => 'hot',
25 humidity => 'normal'},
26 result => 'yes');
27
28 ... repeat for several more instances, then:
29 $dtree->train;
30
31 # Find results for unseen instances
32 my $result = $dtree->get_result
33 (attributes => {outlook => 'sunny',
34 temperature => 'hot',
35 humidity => 'normal'});
36
38 The "AI::DecisionTree" module automatically creates so-called "decision
39 trees" to explain a set of training data. A decision tree is a kind of
40 categorizer that use a flowchart-like process for categorizing new
41 instances. For instance, a learned decision tree might look like the
42 following, which classifies for the concept "play tennis":
43
44 OUTLOOK
45 / | \
46 / | \
47 / | \
48 sunny/ overcast \rainy
49 / | \
50 HUMIDITY | WIND
51 / \ *no* / \
52 / \ / \
53 high/ \normal / \
54 / \ strong/ \weak
55 *no* *yes* / \
56 *no* *yes*
57
58 (This example, and the inspiration for the "AI::DecisionTree" module,
59 come directly from Tom Mitchell's excellent book "Machine Learning",
60 available from McGraw Hill.)
61
62 A decision tree like this one can be learned from training data, and
63 then applied to previously unseen data to obtain results that are
64 consistent with the training data.
65
66 The usual goal of a decision tree is to somehow encapsulate the
67 training data in the smallest possible tree. This is motivated by an
68 "Occam's Razor" philosophy, in which the simplest possible explanation
69 for a set of phenomena should be preferred over other explanations.
70 Also, small trees will make decisions faster than large trees, and they
71 are much easier for a human to look at and understand. One of the
72 biggest reasons for using a decision tree instead of many other machine
73 learning techniques is that a decision tree is a much more scrutable
74 decision maker than, say, a neural network.
75
76 The current implementation of this module uses an extremely simple
77 method for creating the decision tree based on the training instances.
78 It uses an Information Gain metric (based on expected reduction in
79 entropy) to select the "most informative" attribute at each node in the
80 tree. This is essentially the ID3 algorithm, developed by J. R.
81 Quinlan in 1986. The idea is that the attribute with the highest
82 Information Gain will (probably) be the best attribute to split the
83 tree on at each point if we're interested in making small trees.
84
86 Building and Querying the Tree
87 new(...parameters...)
88 Creates a new decision tree object and returns it. Accepts the
89 following parameters:
90
91 noise_mode
92 Controls the behavior of the "train()" method when "noisy" data
93 is encountered. Here "noisy" means that two or more training
94 instances contradict each other, such that they have identical
95 attributes but different results.
96
97 If "noise_mode" is set to "fatal" (the default), the "train()"
98 method will throw an exception (die). If "noise_mode" is set
99 to "pick_best", the most frequent result at each noisy node
100 will be selected.
101
102 prune
103 A boolean "prune" parameter which specifies whether the tree
104 should be pruned after training. This is usually a good idea,
105 so the default is to prune. Currently we prune using a simple
106 minimum-description-length criterion.
107
108 verbose
109 If set to a true value, some status information will be output
110 while training a decision tree. Default is false.
111
112 purge
113 If set to a true value, the "do_purge()" method will be invoked
114 during "train()". The default is true.
115
116 max_depth
117 Controls the maximum depth of the tree that will be created
118 during "train()". The default is 0, which means that trees of
119 unlimited depth can be constructed.
120
121 add_instance(attributes => \%hash, result => $string, name => $string)
122 Adds a training instance to the set of instances which will be used
123 to form the tree. An "attributes" parameter specifies a hash of
124 attribute-value pairs for the instance, and a "result" parameter
125 specifies the result.
126
127 An optional "name" parameter lets you give a unique name to each
128 training instance. This can be used in coordination with the
129 "set_results()" method below.
130
131 train()
132 Builds the decision tree from the list of training instances. If a
133 numeric "max_depth" parameter is supplied, the maximum tree depth
134 can be controlled (see also the "new()" method).
135
136 get_result(attributes => \%hash)
137 Returns the most likely result (from the set of all results given
138 to "add_instance()") for the set of attribute values given. An
139 "attributes" parameter specifies a hash of attribute-value pairs
140 for the instance. If the decision tree doesn't have enough
141 information to find a result, it will return "undef".
142
143 do_purge()
144 Purges training instances and their associated information from the
145 DecisionTree object. This can save memory after training, and
146 since the training instances are implemented as C structs, this
147 turns the DecisionTree object into a pure-perl data structure that
148 can be more easily saved with "Storable.pm", for instance.
149
150 purge()
151 Returns true or false depending on the value of the tree's "purge"
152 property. An optional boolean argument sets the property.
153
154 copy_instances(from => $other_tree)
155 Allows two trees to share the same set of training instances. More
156 commonly, this lets you train one tree, then re-use its instances
157 in another tree (possibly changing the instance "result" values
158 using "set_results()"), which is much faster than re-populating the
159 second tree's instances from scratch.
160
161 set_results(\%results)
162 Given a hash that relates instance names to instance result values,
163 change the result values as specified.
164
165 Tree Introspection
166 instances()
167 Returns a reference to an array of the training instances used to
168 build this tree.
169
170 nodes()
171 Returns the number of nodes in the trained decision tree.
172
173 depth()
174 Returns the depth of the tree. This is the maximum number of
175 decisions that would need to be made to classify an unseen
176 instance, i.e. the length of the longest path from the tree's root
177 to a leaf. A tree with a single node would have a depth of zero.
178
179 rule_tree()
180 Returns a data structure representing the decision tree. For
181 instance, for the tree diagram above, the following data structure
182 is returned:
183
184 [ 'outlook', {
185 'rain' => [ 'wind', {
186 'strong' => 'no',
187 'weak' => 'yes',
188 } ],
189 'sunny' => [ 'humidity', {
190 'normal' => 'yes',
191 'high' => 'no',
192 } ],
193 'overcast' => 'yes',
194 } ]
195
196 This is slightly remniscent of how XML::Parser returns the parsed
197 XML tree.
198
199 Note that while the ordering in the hashes is unpredictable, the
200 nesting is in the order in which the criteria will be checked at
201 decision-making time.
202
203 rule_statements()
204 Returns a list of strings that describe the tree in rule-form. For
205 instance, for the tree diagram above, the following list would be
206 returned (though not necessarily in this order - the order is
207 unpredictable):
208
209 if outlook='rain' and wind='strong' -> 'no'
210 if outlook='rain' and wind='weak' -> 'yes'
211 if outlook='sunny' and humidity='normal' -> 'yes'
212 if outlook='sunny' and humidity='high' -> 'no'
213 if outlook='overcast' -> 'yes'
214
215 This can be helpful for scrutinizing the structure of a tree.
216
217 Note that while the order of the rules is unpredictable, the order
218 of criteria within each rule reflects the order in which the
219 criteria will be checked at decision-making time.
220
221 as_graphviz()
222 Returns a "GraphViz" object representing the tree. Requires that
223 the GraphViz module is already installed, of course. The object
224 returned will allow you to create PNGs, GIFs, image maps, or
225 whatever graphical representation of your tree you might want.
226
227 A "leaf_colors" argument can specify a fill color for each leaf
228 node in the tree. The keys of the hash should be the same as the
229 strings appearing as the "result" parameters given to
230 "add_instance()", and the values should be any GraphViz-style color
231 specification.
232
233 Any additional arguments given to "as_graphviz()" will be passed on
234 to GraphViz's "new()" method. See the GraphViz docs for more info.
235
237 A few limitations exist in the current version. All of them could be
238 removed in future versions - especially with your help. =)
239
240 No continuous attributes
241 In the current implementation, only discrete-valued attributes are
242 supported. This means that an attribute like "temperature" can
243 have values like "cool", "medium", and "hot", but using actual
244 temperatures like 87 or 62.3 is not going to work. This is because
245 the values would split the data too finely - the tree-building
246 process would probably think that it could make all its decisions
247 based on the exact temperature value alone, ignoring all other
248 attributes, because each temperature would have only been seen once
249 in the training data.
250
251 The usual way to deal with this problem is for the tree-building
252 process to figure out how to place the continuous attribute values
253 into a set of bins (like "cool", "medium", and "hot") and then
254 build the tree based on these bin values. Future versions of
255 "AI::DecisionTree" may provide support for this. For now, you have
256 to do it yourself.
257
259 All the stuff in the LIMITATIONS section. Also, revisit the pruning
260 algorithm to see how it can be improved.
261
263 Ken Williams, ken@mathforum.org
264
266 Mitchell, Tom (1997). Machine Learning. McGraw-Hill. pp 52-80.
267
268 Quinlan, J. R. (1986). Induction of decision trees. Machine Learning,
269 1(1), pp 81-106.
270
271 perl, GraphViz
272
273
274
275perl v5.30.1 2020-01-29 AI::DecisionTree(3)