1 | /*************************************************************************/ |
---|
2 | /* */ |
---|
3 | /* Classify items interactively using a decision tree */ |
---|
4 | /* -------------------------------------------------- */ |
---|
5 | /* */ |
---|
6 | /*************************************************************************/ |
---|
7 | |
---|
8 | |
---|
9 | #include "defns.i" |
---|
10 | #include "types.i" |
---|
11 | |
---|
12 | |
---|
13 | /* External data -- see c4.5.c for meanings */ |
---|
14 | |
---|
15 | short MaxAtt, MaxClass, MaxDiscrVal; |
---|
16 | |
---|
17 | ItemNo MaxItem; |
---|
18 | |
---|
19 | Description *Item; |
---|
20 | |
---|
21 | DiscrValue *MaxAttVal; |
---|
22 | |
---|
23 | String *ClassName, |
---|
24 | *AttName, |
---|
25 | **AttValName, |
---|
26 | FileName = "DF"; |
---|
27 | |
---|
28 | char *SpecialStatus; |
---|
29 | |
---|
30 | short VERBOSITY = 0, |
---|
31 | TRACE = 0; |
---|
32 | |
---|
33 | |
---|
34 | /* The interview module uses a more complex description of an |
---|
35 | case called a "Range Description". The value of an |
---|
36 | attribute is given by |
---|
37 | - lower and upper bounds (continuous attribute) |
---|
38 | - probability of each possible value (discrete attribute) */ |
---|
39 | |
---|
40 | |
---|
41 | typedef struct ValRange *RangeDescRec; |
---|
42 | |
---|
43 | struct ValRange |
---|
44 | { |
---|
45 | Boolean Known, /* is range known? */ |
---|
46 | Asked; /* has it been asked? */ |
---|
47 | float LowerBound, /* lower bound given */ |
---|
48 | UpperBound, /* upper ditto */ |
---|
49 | *Probability; /* prior prob of each discr value */ |
---|
50 | }; |
---|
51 | |
---|
52 | RangeDescRec RangeDesc; |
---|
53 | |
---|
54 | Tree DecisionTree, /* tree being used */ |
---|
55 | GetTree(); |
---|
56 | |
---|
57 | float *LowClassSum, /* accumulated lower estimates */ |
---|
58 | *ClassSum = Nil; /* accumulated central estimates */ |
---|
59 | |
---|
60 | #define Fuzz 0.01 /* minimum weight */ |
---|
61 | |
---|
62 | |
---|
63 | |
---|
64 | /*************************************************************************/ |
---|
65 | /* */ |
---|
66 | /* Classify the extended case description in RangeDesc using the */ |
---|
67 | /* given subtree, by adjusting the values ClassSum and LowClassSum */ |
---|
68 | /* for each class, indicating the likelihood of the case being */ |
---|
69 | /* of that class. */ |
---|
70 | /* */ |
---|
71 | /*************************************************************************/ |
---|
72 | |
---|
73 | |
---|
74 | ClassifyCase(Subtree, Weight) |
---|
75 | /* ------------ */ |
---|
76 | Tree Subtree; |
---|
77 | float Weight; |
---|
78 | { |
---|
79 | DiscrValue v; |
---|
80 | float BranchWeight, Area(), Interpolate(); |
---|
81 | Attribute a; |
---|
82 | short s; |
---|
83 | ClassNo c; |
---|
84 | |
---|
85 | /* A leaf */ |
---|
86 | |
---|
87 | if ( ! Subtree->NodeType ) |
---|
88 | { |
---|
89 | Verbosity(1) |
---|
90 | printf("\tClass %s weight %g cases %g\n", |
---|
91 | ClassName[Subtree->Leaf], Weight, Subtree->Items); |
---|
92 | |
---|
93 | if ( Subtree->Items > 0 ) |
---|
94 | { |
---|
95 | /* Adjust class sum of ALL classes, but adjust low class sum |
---|
96 | of leaf class only */ |
---|
97 | |
---|
98 | ForEach(c, 0, MaxClass) |
---|
99 | { |
---|
100 | ClassSum[c] += Weight * Subtree->ClassDist[c] / Subtree->Items; |
---|
101 | } |
---|
102 | |
---|
103 | LowClassSum[Subtree->Leaf] += |
---|
104 | Weight * (1 - Subtree->Errors / Subtree->Items); |
---|
105 | } |
---|
106 | else |
---|
107 | { |
---|
108 | ClassSum[Subtree->Leaf] += Weight; |
---|
109 | } |
---|
110 | |
---|
111 | return; |
---|
112 | } |
---|
113 | |
---|
114 | a = Subtree->Tested; |
---|
115 | |
---|
116 | CheckValue(a, Subtree); |
---|
117 | |
---|
118 | /* Unknown value */ |
---|
119 | |
---|
120 | if ( ! RangeDesc[a].Known ) |
---|
121 | { |
---|
122 | ForEach(v, 1, Subtree->Forks) |
---|
123 | { |
---|
124 | ClassifyCase(Subtree->Branch[v], |
---|
125 | (Weight * Subtree->Branch[v]->Items) / Subtree->Items); |
---|
126 | } |
---|
127 | return; |
---|
128 | } |
---|
129 | |
---|
130 | /* Known value */ |
---|
131 | |
---|
132 | switch ( Subtree->NodeType ) |
---|
133 | { |
---|
134 | case BrDiscr: /* test of discrete attribute */ |
---|
135 | |
---|
136 | ForEach(v, 1, MaxAttVal[a]) |
---|
137 | { |
---|
138 | BranchWeight = RangeDesc[a].Probability[v]; |
---|
139 | if ( BranchWeight > 0 ) |
---|
140 | { |
---|
141 | Verbosity(1) |
---|
142 | printf("\tWeight %g: test att %s (val %s = %g)\n", |
---|
143 | Weight, AttName[a], AttValName[a][v], |
---|
144 | BranchWeight); |
---|
145 | |
---|
146 | ClassifyCase(Subtree->Branch[v], Weight * BranchWeight); |
---|
147 | } |
---|
148 | } |
---|
149 | break; |
---|
150 | |
---|
151 | case ThreshContin: /* test of continuous attribute */ |
---|
152 | |
---|
153 | BranchWeight = |
---|
154 | RangeDesc[a].UpperBound <= Subtree->Lower ? 1.0 : |
---|
155 | RangeDesc[a].LowerBound > Subtree->Upper ? 0.0 : |
---|
156 | RangeDesc[a].LowerBound != RangeDesc[a].UpperBound ? |
---|
157 | (Area(Subtree, RangeDesc[a].LowerBound) - |
---|
158 | Area(Subtree, RangeDesc[a].UpperBound)) / |
---|
159 | (RangeDesc[a].UpperBound - RangeDesc[a].LowerBound) : |
---|
160 | Interpolate(Subtree, RangeDesc[a].LowerBound) ; |
---|
161 | |
---|
162 | Verbosity(1) |
---|
163 | printf("\tWeight %g: test att %s (branch weight=%g)\n", |
---|
164 | Weight, AttName[a], BranchWeight); |
---|
165 | |
---|
166 | if ( BranchWeight > Fuzz ) |
---|
167 | { |
---|
168 | ClassifyCase(Subtree->Branch[1], Weight * BranchWeight); |
---|
169 | } |
---|
170 | if ( BranchWeight < 1-Fuzz ) |
---|
171 | { |
---|
172 | ClassifyCase(Subtree->Branch[2], Weight * (1 - BranchWeight)); |
---|
173 | } |
---|
174 | break; |
---|
175 | |
---|
176 | case BrSubset: /* subset test on discrete attribute */ |
---|
177 | |
---|
178 | ForEach(s, 1, Subtree->Forks) |
---|
179 | { |
---|
180 | BranchWeight = 0.0; |
---|
181 | ForEach(v, 1, MaxAttVal[a]) |
---|
182 | { |
---|
183 | if ( In(v, Subtree->Subset[s]) ) |
---|
184 | { |
---|
185 | BranchWeight += RangeDesc[a].Probability[v]; |
---|
186 | } |
---|
187 | } |
---|
188 | if ( BranchWeight > 0 ) |
---|
189 | { |
---|
190 | Verbosity(1) |
---|
191 | printf("\tWeight %g: test att %s (val %s = %g)\n", |
---|
192 | Weight, AttName[a], AttValName[a][v], |
---|
193 | BranchWeight); |
---|
194 | |
---|
195 | ClassifyCase(Subtree->Branch[s], Weight * BranchWeight); |
---|
196 | } |
---|
197 | } |
---|
198 | break; |
---|
199 | } |
---|
200 | } |
---|
201 | |
---|
202 | |
---|
203 | |
---|
204 | /*************************************************************************/ |
---|
205 | /* */ |
---|
206 | /* Interpolate a single value between Lower, Cut and Upper */ |
---|
207 | /* */ |
---|
208 | /*************************************************************************/ |
---|
209 | |
---|
210 | |
---|
211 | float Interpolate(t, v) |
---|
212 | /* ---- */ |
---|
213 | Tree t; |
---|
214 | float v; |
---|
215 | { |
---|
216 | float Sum=Epsilon; |
---|
217 | |
---|
218 | if ( v <= t->Lower ) |
---|
219 | { |
---|
220 | return 1.0; |
---|
221 | } |
---|
222 | |
---|
223 | if ( v <= t->Cut ) |
---|
224 | { |
---|
225 | return 1 - 0.5 * (v - t->Lower) / (t->Cut - t->Lower + Epsilon); |
---|
226 | } |
---|
227 | |
---|
228 | if ( v < t->Upper ) |
---|
229 | { |
---|
230 | return 0.5 - 0.5 * (v - t->Cut) / (t->Upper - t->Cut + Epsilon); |
---|
231 | } |
---|
232 | |
---|
233 | return 0.0; |
---|
234 | } |
---|
235 | |
---|
236 | |
---|
237 | |
---|
238 | /*************************************************************************/ |
---|
239 | /* */ |
---|
240 | /* Compute the area under a soft threshold curve to the right of a */ |
---|
241 | /* given value. */ |
---|
242 | /* */ |
---|
243 | /*************************************************************************/ |
---|
244 | |
---|
245 | |
---|
246 | float Area(t, v) |
---|
247 | /* ---- */ |
---|
248 | Tree t; |
---|
249 | float v; |
---|
250 | { |
---|
251 | float Sum=Epsilon, F; |
---|
252 | |
---|
253 | if ( v < t->Lower ) |
---|
254 | { |
---|
255 | Sum += t->Lower - v; |
---|
256 | v = t->Lower; |
---|
257 | } |
---|
258 | |
---|
259 | if ( v < t->Cut ) |
---|
260 | { |
---|
261 | F = (t->Cut - v ) / (t->Cut - t->Lower + Epsilon); |
---|
262 | |
---|
263 | Sum += 0.5 * (t->Cut - v) + 0.25 * F * (t->Cut - v); |
---|
264 | v = t->Cut; |
---|
265 | } |
---|
266 | |
---|
267 | if ( v < t->Upper ) |
---|
268 | { |
---|
269 | F = (t->Upper - v ) / (t->Upper - t->Cut + Epsilon); |
---|
270 | |
---|
271 | Sum += 0.25 * (t->Upper - v) * F; |
---|
272 | } |
---|
273 | |
---|
274 | Verbosity(1) printf("lower=%g cut=%g upper=%g area=%g\n", |
---|
275 | t->Lower, t->Cut, t->Upper, Sum); |
---|
276 | |
---|
277 | return Sum; |
---|
278 | } |
---|
279 | |
---|
280 | |
---|
281 | |
---|
282 | /*************************************************************************/ |
---|
283 | /* */ |
---|
284 | /* Process a single case */ |
---|
285 | /* */ |
---|
286 | /*************************************************************************/ |
---|
287 | |
---|
288 | |
---|
289 | InterpretTree() |
---|
290 | /* ------------- */ |
---|
291 | { |
---|
292 | ClassNo c, BestClass; |
---|
293 | float Uncertainty=1.0; |
---|
294 | char Reply; |
---|
295 | Attribute a; |
---|
296 | |
---|
297 | /* Initialise */ |
---|
298 | |
---|
299 | ForEach(a, 0, MaxAtt) |
---|
300 | { |
---|
301 | RangeDesc[a].Asked = false; |
---|
302 | } |
---|
303 | |
---|
304 | if ( ! ClassSum ) |
---|
305 | { |
---|
306 | /* The first time through .. allocate class sums */ |
---|
307 | |
---|
308 | ClassSum = (float *) malloc((MaxClass+1) * sizeof(float)); |
---|
309 | LowClassSum = (float *) malloc((MaxClass+1) * sizeof(float)); |
---|
310 | |
---|
311 | printf("\n"); |
---|
312 | } |
---|
313 | else |
---|
314 | { |
---|
315 | printf("\n-------------------------------------------\n\n"); |
---|
316 | } |
---|
317 | |
---|
318 | ForEach(c, 0, MaxClass) |
---|
319 | { |
---|
320 | LowClassSum[c] = ClassSum[c] = 0; |
---|
321 | } |
---|
322 | |
---|
323 | /* Find the likelihood of an item's being of each class */ |
---|
324 | |
---|
325 | ClassifyCase(DecisionTree, 1.0); |
---|
326 | |
---|
327 | /* Find the best class and show decision made */ |
---|
328 | |
---|
329 | BestClass = 0; |
---|
330 | ForEach(c, 0, MaxClass) |
---|
331 | { |
---|
332 | Verbosity(1) printf("class %d weight %.2f\n", c, ClassSum[c]); |
---|
333 | |
---|
334 | Uncertainty -= LowClassSum[c]; |
---|
335 | if ( ClassSum[c] > ClassSum[BestClass] ) BestClass = c; |
---|
336 | } |
---|
337 | |
---|
338 | printf("\nDecision:\n"); |
---|
339 | Decision(BestClass, ClassSum[BestClass], |
---|
340 | LowClassSum[BestClass], |
---|
341 | Uncertainty + LowClassSum[BestClass]); |
---|
342 | |
---|
343 | /* Show the other significant classes, if more than two classes */ |
---|
344 | |
---|
345 | if ( MaxClass > 1 ) |
---|
346 | { |
---|
347 | while ( true ) |
---|
348 | { |
---|
349 | ClassSum[BestClass] = 0; |
---|
350 | BestClass = 0; |
---|
351 | ForEach(c, 0, MaxClass) |
---|
352 | { |
---|
353 | if ( ClassSum[c] > ClassSum[BestClass] ) BestClass = c; |
---|
354 | } |
---|
355 | |
---|
356 | if ( ClassSum[BestClass] < Fuzz ) break; |
---|
357 | |
---|
358 | Decision(BestClass, ClassSum[BestClass], |
---|
359 | LowClassSum[BestClass], |
---|
360 | Uncertainty + LowClassSum[BestClass]); |
---|
361 | } |
---|
362 | } |
---|
363 | |
---|
364 | /* Prompt for what to do next */ |
---|
365 | |
---|
366 | while ( true ) |
---|
367 | { |
---|
368 | printf("\nRetry, new case or quit [r,n,q]: "); |
---|
369 | Reply = getchar(); |
---|
370 | SkipLine(Reply); |
---|
371 | switch ( Reply ) |
---|
372 | { |
---|
373 | case 'r': return; |
---|
374 | case 'n': Clear(); return; |
---|
375 | case 'q': exit(0); |
---|
376 | default: printf("Please enter 'r', 'n' or 'q'"); |
---|
377 | } |
---|
378 | } |
---|
379 | } |
---|
380 | |
---|
381 | |
---|
382 | |
---|
383 | /*************************************************************************/ |
---|
384 | /* */ |
---|
385 | /* Print the chosen class with certainty factor and range */ |
---|
386 | /* */ |
---|
387 | /*************************************************************************/ |
---|
388 | |
---|
389 | |
---|
390 | Decision(c, p, lb, ub) |
---|
391 | /* -------- */ |
---|
392 | ClassNo c; |
---|
393 | float p, lb, ub; |
---|
394 | { |
---|
395 | printf("\t%s", ClassName[c]); |
---|
396 | |
---|
397 | if ( p < 1-Fuzz || lb < ub - Fuzz ) |
---|
398 | { |
---|
399 | printf(" CF = %.2f", p); |
---|
400 | if ( lb < ub - Fuzz ) |
---|
401 | { |
---|
402 | printf(" [ %.2f - %.2f ]", lb, ub); |
---|
403 | } |
---|
404 | } |
---|
405 | |
---|
406 | printf("\n"); |
---|
407 | } |
---|
408 | |
---|
409 | |
---|
410 | |
---|
411 | /*************************************************************************/ |
---|
412 | /* */ |
---|
413 | /* Main routine for classifying items using a decision tree */ |
---|
414 | /* */ |
---|
415 | /*************************************************************************/ |
---|
416 | |
---|
417 | |
---|
418 | main(Argc, Argv) |
---|
419 | /* ---- */ |
---|
420 | int Argc; |
---|
421 | char *Argv[]; |
---|
422 | { |
---|
423 | int o; |
---|
424 | extern char *optarg; |
---|
425 | extern int optind; |
---|
426 | Attribute a; |
---|
427 | |
---|
428 | PrintHeader("decision tree interpreter"); |
---|
429 | |
---|
430 | /* Process options */ |
---|
431 | |
---|
432 | while ( (o = getopt(Argc, Argv, "tvf:")) != EOF ) |
---|
433 | { |
---|
434 | switch (o) |
---|
435 | { |
---|
436 | case 't': TRACE = 1; |
---|
437 | break; |
---|
438 | case 'v': VERBOSITY = 1; |
---|
439 | break; |
---|
440 | case 'f': FileName = optarg; |
---|
441 | break; |
---|
442 | case '?': printf("unrecognised option\n"); |
---|
443 | exit(1); |
---|
444 | } |
---|
445 | } |
---|
446 | |
---|
447 | /* Initialise */ |
---|
448 | |
---|
449 | GetNames(); |
---|
450 | |
---|
451 | DecisionTree = GetTree(".tree"); |
---|
452 | if ( TRACE ) PrintTree(DecisionTree); |
---|
453 | |
---|
454 | /* Allocate value ranges */ |
---|
455 | |
---|
456 | RangeDesc = (struct ValRange *) calloc(MaxAtt+1, sizeof(struct ValRange)); |
---|
457 | |
---|
458 | ForEach(a, 0, MaxAtt) |
---|
459 | { |
---|
460 | if ( MaxAttVal[a] ) |
---|
461 | { |
---|
462 | RangeDesc[a].Probability = |
---|
463 | (float *) calloc(MaxAttVal[a]+1, sizeof(float)); |
---|
464 | } |
---|
465 | } |
---|
466 | |
---|
467 | /* Consult */ |
---|
468 | |
---|
469 | Clear(); |
---|
470 | while ( true ) |
---|
471 | { |
---|
472 | InterpretTree(); |
---|
473 | } |
---|
474 | } |
---|