1 | /*************************************************************************/ |
---|
2 | /* */ |
---|
3 | /* Routines to manage tree growth, pruning and evaluation */ |
---|
4 | /* ------------------------------------------------------ */ |
---|
5 | /* */ |
---|
6 | /*************************************************************************/ |
---|
7 | |
---|
8 | #include "defns.i" |
---|
9 | #include "types.i" |
---|
10 | #include "extern.i" |
---|
11 | |
---|
12 | ItemNo *TargetClassFreq; |
---|
13 | Tree *Raw; |
---|
14 | extern Tree *Pruned; |
---|
15 | |
---|
16 | /*************************************************************************/ |
---|
17 | /* */ |
---|
18 | /* Grow and prune a single tree from all data */ |
---|
19 | /* */ |
---|
20 | /*************************************************************************/ |
---|
21 | |
---|
22 | OneTree() |
---|
23 | /* --------- */ |
---|
24 | { |
---|
25 | Tree FormTree(), CopyTree(); |
---|
26 | Boolean Prune(); |
---|
27 | |
---|
28 | InitialiseTreeData(); |
---|
29 | InitialiseWeights(); |
---|
30 | |
---|
31 | Raw = (Tree *) calloc(1, sizeof(Tree)); |
---|
32 | Pruned = (Tree *) calloc(1, sizeof(Tree)); |
---|
33 | |
---|
34 | AllKnown = true; |
---|
35 | |
---|
36 | #pragma omp parallel default(shared) |
---|
37 | { |
---|
38 | #pragma omp single nowait |
---|
39 | { |
---|
40 | Raw[0] = FormTree(0, MaxItem); |
---|
41 | } |
---|
42 | } |
---|
43 | //printf("\n"); |
---|
44 | //PrintTree(Raw[0]); |
---|
45 | |
---|
46 | SaveTree(Raw[0], ".unpruned"); |
---|
47 | |
---|
48 | Pruned[0] = CopyTree(Raw[0]); |
---|
49 | if (Prune(Pruned[0])) { |
---|
50 | //printf("\nSimplified "); |
---|
51 | //PrintTree(Pruned[0]); |
---|
52 | } |
---|
53 | } |
---|
54 | |
---|
55 | /*************************************************************************/ |
---|
56 | /* */ |
---|
57 | /* Grow and prune TRIALS trees and select the best of them */ |
---|
58 | /* */ |
---|
59 | /*************************************************************************/ |
---|
60 | |
---|
61 | short BestTree() |
---|
62 | /* -------- */ |
---|
63 | { |
---|
64 | Tree CopyTree(), Iterate(); |
---|
65 | Boolean Prune(); |
---|
66 | short t, Best = 0; |
---|
67 | |
---|
68 | InitialiseTreeData(); |
---|
69 | |
---|
70 | TargetClassFreq = (ItemNo *) calloc(MaxClass + 1, sizeof(ItemNo)); |
---|
71 | |
---|
72 | Raw = (Tree *) calloc(TRIALS, sizeof(Tree)); |
---|
73 | Pruned = (Tree *) calloc(TRIALS, sizeof(Tree)); |
---|
74 | |
---|
75 | /* If necessary, set initial size of window to 20% (or twice |
---|
76 | the sqrt, if this is larger) of the number of data items, |
---|
77 | and the maximum number of items that can be added to the |
---|
78 | window at each iteration to 20% of the initial window size */ |
---|
79 | |
---|
80 | if (!WINDOW) { |
---|
81 | WINDOW = Max(2 * sqrt(MaxItem+1.0), (MaxItem+1) / 5); |
---|
82 | } |
---|
83 | |
---|
84 | if (!INCREMENT) { |
---|
85 | INCREMENT = Max(WINDOW / 5, 1); |
---|
86 | } |
---|
87 | |
---|
88 | FormTarget(WINDOW); |
---|
89 | |
---|
90 | /* Form set of trees by iteration and prune */ |
---|
91 | |
---|
92 | ForEach(t, 0, TRIALS-1 ) { |
---|
93 | FormInitialWindow(); |
---|
94 | |
---|
95 | printf("\n--------\nTrial %d\n--------\n\n", t); |
---|
96 | |
---|
97 | Raw[t] = Iterate(WINDOW, INCREMENT); |
---|
98 | printf("\n"); |
---|
99 | PrintTree(Raw[t]); |
---|
100 | |
---|
101 | SaveTree(Raw[t], ".unpruned"); |
---|
102 | |
---|
103 | Pruned[t] = CopyTree(Raw[t]); |
---|
104 | if (Prune(Pruned[t])) { |
---|
105 | printf("\nSimplified "); |
---|
106 | PrintTree(Pruned[t]); |
---|
107 | } |
---|
108 | |
---|
109 | if (Pruned[t]->Errors < Pruned[Best]->Errors) { |
---|
110 | Best = t; |
---|
111 | } |
---|
112 | } |
---|
113 | printf("\n--------\n"); |
---|
114 | |
---|
115 | return Best; |
---|
116 | } |
---|
117 | |
---|
118 | /*************************************************************************/ |
---|
119 | /* */ |
---|
120 | /* The windowing approach seems to work best when the class */ |
---|
121 | /* distribution of the initial window is as close to uniform as */ |
---|
122 | /* possible. FormTarget generates this initial target distribution, */ |
---|
123 | /* setting up a TargetClassFreq value for each class. */ |
---|
124 | /* */ |
---|
125 | /*************************************************************************/ |
---|
126 | |
---|
127 | FormTarget(Size) |
---|
128 | /* ----------- */ |
---|
129 | ItemNo Size; { |
---|
130 | ItemNo i, *ClassFreq; |
---|
131 | ClassNo c, Smallest, ClassesLeft = 0; |
---|
132 | |
---|
133 | ClassFreq = (ItemNo *) calloc(MaxClass + 1, sizeof(ItemNo)); |
---|
134 | |
---|
135 | /* Generate the class frequency distribution */ |
---|
136 | |
---|
137 | ForEach(i, 0, MaxItem) { |
---|
138 | ClassFreq[Class(Item[i])]++; |
---|
139 | } |
---|
140 | |
---|
141 | /* Calculate the no. of classes of which there are items */ |
---|
142 | |
---|
143 | ForEach(c, 0, MaxClass) { |
---|
144 | if (ClassFreq[c]) { |
---|
145 | ClassesLeft++; |
---|
146 | } else { |
---|
147 | TargetClassFreq[c] = 0; |
---|
148 | } |
---|
149 | } |
---|
150 | |
---|
151 | while (ClassesLeft) { |
---|
152 | /* Find least common class of which there are some items */ |
---|
153 | |
---|
154 | Smallest = -1; |
---|
155 | ForEach(c, 0, MaxClass) { |
---|
156 | if (ClassFreq[c] && (Smallest < 0 || ClassFreq[c] |
---|
157 | < ClassFreq[Smallest])) { |
---|
158 | Smallest = c; |
---|
159 | } |
---|
160 | } |
---|
161 | |
---|
162 | /* Allocate the no. of items of this class to use in the window */ |
---|
163 | |
---|
164 | TargetClassFreq[Smallest] |
---|
165 | = Min(ClassFreq[Smallest], Round(Size/ClassesLeft)); |
---|
166 | |
---|
167 | ClassFreq[Smallest] = 0; |
---|
168 | |
---|
169 | Size -= TargetClassFreq[Smallest]; |
---|
170 | ClassesLeft--; |
---|
171 | } |
---|
172 | |
---|
173 | cfree(ClassFreq); |
---|
174 | } |
---|
175 | |
---|
176 | /*************************************************************************/ |
---|
177 | /* */ |
---|
178 | /* Form initial window, attempting to obtain the target class profile */ |
---|
179 | /* in TargetClassFreq. This is done by placing the targeted number */ |
---|
180 | /* of items of each class at the beginning of the set of data items. */ |
---|
181 | /* */ |
---|
182 | /*************************************************************************/ |
---|
183 | |
---|
184 | FormInitialWindow() |
---|
185 | /* ------------------- */ |
---|
186 | { |
---|
187 | ItemNo i, Start = 0, More; |
---|
188 | ClassNo c; |
---|
189 | void Swap(); |
---|
190 | |
---|
191 | Shuffle(); |
---|
192 | |
---|
193 | ForEach(c, 0, MaxClass) { |
---|
194 | More = TargetClassFreq[c]; |
---|
195 | |
---|
196 | for (i = Start; More; i++) { |
---|
197 | if (Class(Item[i]) == c) { |
---|
198 | Swap(Start, i); |
---|
199 | Start++; |
---|
200 | More--; |
---|
201 | } |
---|
202 | } |
---|
203 | } |
---|
204 | } |
---|
205 | |
---|
206 | /*************************************************************************/ |
---|
207 | /* */ |
---|
208 | /* Shuffle the data items randomly */ |
---|
209 | /* */ |
---|
210 | /*************************************************************************/ |
---|
211 | |
---|
212 | Shuffle() |
---|
213 | /* ------- */ |
---|
214 | { |
---|
215 | ItemNo This, Alt, Left; |
---|
216 | Description Hold; |
---|
217 | |
---|
218 | This = 0; |
---|
219 | for (Left = MaxItem + 1; Left;) { |
---|
220 | Alt = This + (Left--) * Random; |
---|
221 | Hold = Item[This]; |
---|
222 | Item[This++] = Item[Alt]; |
---|
223 | Item[Alt] = Hold; |
---|
224 | } |
---|
225 | } |
---|
226 | |
---|
227 | /*************************************************************************/ |
---|
228 | /* */ |
---|
229 | /* Grow a tree iteratively with initial window size Window and */ |
---|
230 | /* initial window increment IncExceptions. */ |
---|
231 | /* */ |
---|
232 | /* Construct a classifier tree using the data items in the */ |
---|
233 | /* window, then test for the successful classification of other */ |
---|
234 | /* data items by this tree. If there are misclassified items, */ |
---|
235 | /* put them immediately after the items in the window, increase */ |
---|
236 | /* the size of the window and build another classifier tree, and */ |
---|
237 | /* so on until we have a tree which successfully classifies all */ |
---|
238 | /* of the test items or no improvement is apparent. */ |
---|
239 | /* */ |
---|
240 | /* On completion, return the tree which produced the least errors. */ |
---|
241 | /* */ |
---|
242 | /*************************************************************************/ |
---|
243 | |
---|
244 | Tree Iterate(Window, IncExceptions) |
---|
245 | /* ------- */ |
---|
246 | ItemNo Window, IncExceptions; { |
---|
247 | Tree Classifier, BestClassifier = Nil, FormTree(); |
---|
248 | ItemNo i, Errors, TotalErrors, BestTotalErrors = MaxItem + 1, Exceptions, |
---|
249 | Additions; |
---|
250 | ClassNo Assigned, Category(); |
---|
251 | short Cycle = 0; |
---|
252 | void Swap(); |
---|
253 | |
---|
254 | printf("Cycle Tree -----Cases----"); |
---|
255 | printf(" -----------------Errors-----------------\n"); |
---|
256 | printf(" size window other"); |
---|
257 | printf(" window rate other rate total rate\n"); |
---|
258 | printf("----- ---- ------ ------"); |
---|
259 | printf(" ------ ---- ------ ---- ------ ----\n"); |
---|
260 | |
---|
261 | do { |
---|
262 | /* Build a classifier tree with the first Window items */ |
---|
263 | |
---|
264 | InitialiseWeights(); |
---|
265 | AllKnown = true; |
---|
266 | Classifier = FormTree(0, Window - 1); |
---|
267 | |
---|
268 | /* Error analysis */ |
---|
269 | |
---|
270 | Errors = Round(Classifier->Errors); |
---|
271 | |
---|
272 | /* Move all items that are incorrectly classified by the |
---|
273 | classifier tree to immediately after the items in the |
---|
274 | current window. */ |
---|
275 | |
---|
276 | Exceptions = Window; |
---|
277 | ForEach(i, Window, MaxItem) { |
---|
278 | Assigned = Category(Item[i], Classifier); |
---|
279 | if (Assigned != Class(Item[i])) { |
---|
280 | Swap(Exceptions, i); |
---|
281 | Exceptions++; |
---|
282 | } |
---|
283 | } |
---|
284 | Exceptions -= Window; |
---|
285 | TotalErrors = Errors + Exceptions; |
---|
286 | |
---|
287 | /* Print error analysis */ |
---|
288 | |
---|
289 | printf("%3d %7d %8d %6d %8d%5.1f%% %6d%5.1f%% %6d%5.1f%%\n", |
---|
290 | ++Cycle, TreeSize(Classifier), Window, MaxItem - Window + 1, |
---|
291 | Errors, 100 * (float) Errors / Window, Exceptions, 100 |
---|
292 | * Exceptions / (MaxItem - Window + 1.001), TotalErrors, |
---|
293 | 100 * TotalErrors / (MaxItem + 1.0)); |
---|
294 | |
---|
295 | /* Keep track of the most successful classifier tree so far */ |
---|
296 | |
---|
297 | if (!BestClassifier || TotalErrors < BestTotalErrors) { |
---|
298 | if (BestClassifier) |
---|
299 | ReleaseTree(BestClassifier); |
---|
300 | BestClassifier = Classifier; |
---|
301 | BestTotalErrors = TotalErrors; |
---|
302 | } else { |
---|
303 | ReleaseTree(Classifier); |
---|
304 | } |
---|
305 | |
---|
306 | /* Increment window size */ |
---|
307 | |
---|
308 | Additions = Min(Exceptions, IncExceptions); |
---|
309 | Window = Min(Window + Max(Additions, Exceptions / 2), MaxItem + 1); |
---|
310 | } while (Exceptions); |
---|
311 | |
---|
312 | return BestClassifier; |
---|
313 | } |
---|
314 | |
---|
315 | /*************************************************************************/ |
---|
316 | /* */ |
---|
317 | /* Print report of errors for each of the trials */ |
---|
318 | /* */ |
---|
319 | /*************************************************************************/ |
---|
320 | |
---|
321 | Evaluate(CMInfo, Saved) |
---|
322 | /* -------- */ |
---|
323 | Boolean CMInfo;short Saved; { |
---|
324 | ClassNo RealClass, PrunedClass, Category(); |
---|
325 | short t; |
---|
326 | ItemNo *ConfusionMat, i, RawErrors, PrunedErrors; |
---|
327 | |
---|
328 | if (CMInfo) { |
---|
329 | ConfusionMat = (ItemNo *) calloc((MaxClass + 1) * (MaxClass + 1), |
---|
330 | sizeof(ItemNo)); |
---|
331 | } |
---|
332 | |
---|
333 | printf("\n"); |
---|
334 | |
---|
335 | if (TRIALS > 1) { |
---|
336 | printf("Trial\t Before Pruning After Pruning\n"); |
---|
337 | printf("-----\t---------------- ---------------------------\n"); |
---|
338 | } else { |
---|
339 | printf("\t Before Pruning After Pruning\n"); |
---|
340 | printf("\t---------------- ---------------------------\n"); |
---|
341 | } |
---|
342 | printf("\tSize Errors Size Errors Estimate\n\n"); |
---|
343 | |
---|
344 | ForEach(t, 0, TRIALS-1) { |
---|
345 | RawErrors = PrunedErrors = 0; |
---|
346 | |
---|
347 | ForEach(i, 0, MaxItem) { |
---|
348 | RealClass = Class(Item[i]); |
---|
349 | |
---|
350 | if (Category(Item[i], Raw[t]) != RealClass) |
---|
351 | RawErrors++; |
---|
352 | |
---|
353 | PrunedClass = Category(Item[i], Pruned[t]); |
---|
354 | |
---|
355 | if (PrunedClass != RealClass) |
---|
356 | PrunedErrors++; |
---|
357 | |
---|
358 | if (CMInfo && t == Saved) { |
---|
359 | ConfusionMat[RealClass * (MaxClass + 1) + PrunedClass]++; |
---|
360 | } |
---|
361 | } |
---|
362 | |
---|
363 | if (TRIALS > 1) { |
---|
364 | printf("%4d", t); |
---|
365 | } |
---|
366 | |
---|
367 | printf("\t%4d %3d(%4.1f%%) %4d %3d(%4.1f%%) (%4.1f%%)%s\n", |
---|
368 | TreeSize(Raw[t]), RawErrors, 100.0 * RawErrors |
---|
369 | / (MaxItem + 1.0), TreeSize(Pruned[t]), PrunedErrors, |
---|
370 | 100.0 * PrunedErrors / (MaxItem + 1.0), 100 * Pruned[t]->Errors |
---|
371 | / Pruned[t]->Items, (t == Saved ? " <<" : "")); |
---|
372 | } |
---|
373 | |
---|
374 | if (CMInfo) { |
---|
375 | PrintConfusionMatrix(ConfusionMat); |
---|
376 | free(ConfusionMat); |
---|
377 | } |
---|
378 | } |
---|