File size: 86,778 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 |
{
"paper_id": "O08-6001",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T08:02:05.635056Z"
},
"title": "Feature Weighting Random Forest for Detection of Hidden Web Search Interfaces",
"authors": [
{
"first": "Yunming",
"middle": [],
"last": "Ye",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Harbin Institute of Technology",
"location": {
"country": "China"
}
},
"email": "yeyunming@hit.edu.cn"
},
{
"first": "Hongbo",
"middle": [],
"last": "Li",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Harbin Institute of Technology",
"location": {
"country": "China"
}
},
"email": ""
},
{
"first": "Xiaobai",
"middle": [],
"last": "Deng",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Harbin Institute of Technology",
"location": {
"country": "China"
}
},
"email": ""
},
{
"first": "Joshua",
"middle": [
"Zhexue"
],
"last": "Huang",
"suffix": "",
"affiliation": {},
"email": "jhuang@eti.hku.hk"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Search interface detection is an essential task for extracting information from the hidden Web. The challenge for this task is that search interface data is represented in high-dimensional and sparse features with many missing values. This paper presents a new multi-classifier ensemble approach to solving this problem. In this approach, we have extended the random forest algorithm with a weighted feature selection method to build the individual classifiers. With this improved random forest algorithm (IRFA), each classifier can be learned from a weighted subset of the feature space so that the ensemble of decision trees can fully exploit the useful features of search interface patterns. We have compared our ensemble approach with other well-known classification algorithms, such as SVM, C4.5, Na\u00efve Bayes, and original random forest algorithm (RFA). The experimental results have shown that our method is more effective in detecting search interfaces of the hidden Web.",
"pdf_parse": {
"paper_id": "O08-6001",
"_pdf_hash": "",
"abstract": [
{
"text": "Search interface detection is an essential task for extracting information from the hidden Web. The challenge for this task is that search interface data is represented in high-dimensional and sparse features with many missing values. This paper presents a new multi-classifier ensemble approach to solving this problem. In this approach, we have extended the random forest algorithm with a weighted feature selection method to build the individual classifiers. With this improved random forest algorithm (IRFA), each classifier can be learned from a weighted subset of the feature space so that the ensemble of decision trees can fully exploit the useful features of search interface patterns. We have compared our ensemble approach with other well-known classification algorithms, such as SVM, C4.5, Na\u00efve Bayes, and original random forest algorithm (RFA). The experimental results have shown that our method is more effective in detecting search interfaces of the hidden Web.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Hidden Web refers to the Web pages that are dynamically generated from searchable structured or unstructured databases. Different from the Publicly Indexable Web that is accessible through static hyperlinks, the pages in a hidden Web can only be obtained through queries submitted via the search interface to the databases containing data about the hidden Web. Search interfaces are usually encoded as HTML forms that need to be filled out and submitted by users to obtain information. On the Web, there are many different HTML forms, and many of them are not search interfaces (He, Patel, Zhang, & Chang, 2007) . To extract useful information from hidden Web pages, effectively detecting the search interfaces is an essential step since the interface is the only entrance to the hidden Web. Therefore, we first need to find the entrance to the hidden database. The entrance (i.e., search forms) is mixed with lots of non-search forms in HTML pages. Thus, it is very important to distinguish the two types of forms in order to enable the Hidden Web crawler to locate the entrance and extract information further.",
"cite_spans": [
{
"start": 578,
"end": 611,
"text": "(He, Patel, Zhang, & Chang, 2007)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "Information extraction from the hidden Web has been a hot research topic (BrightPlanet.com, 2000) since the term \"Hidden Web\" was first coined (Florescu, Levy, & Mendelzon, 1998) . Most previous work, however, has been focused on the problems of automatic query generation (Ntoulas, Zerfos, & Cho, 2005) , form filling (Cavelee, Liu, & Probe, 2004) , and wrapper generation for extracting structured information (Wang & Lochovsky, 2003) , where detecting search interface was performed manually or by some heuristic methods. Using heuristic rules to find search forms is the simplest and most effective method (Raghavan & Garcia-Molina, 2001; Lage, Silva, Golgher, & Laender, 2004) . As hidden Web sites adopt different search forms, though, it is time-consuming to compose different rules for different search forms. Employing machine learning and information retrieval techniques to learn classification models from the content of search forms and using the models to classify different forms automatically is a more desirable approach with respect to scalability and robustness. This approach regards search interface detection as a two-class classification problem. One example of this is using decision trees to build form classification models to detect search interface (Cope, Craswell, & Hawking, 2003) .",
"cite_spans": [
{
"start": 73,
"end": 97,
"text": "(BrightPlanet.com, 2000)",
"ref_id": "BIBREF2"
},
{
"start": 143,
"end": 178,
"text": "(Florescu, Levy, & Mendelzon, 1998)",
"ref_id": "BIBREF6"
},
{
"start": 273,
"end": 303,
"text": "(Ntoulas, Zerfos, & Cho, 2005)",
"ref_id": "BIBREF14"
},
{
"start": 319,
"end": 348,
"text": "(Cavelee, Liu, & Probe, 2004)",
"ref_id": null
},
{
"start": 412,
"end": 436,
"text": "(Wang & Lochovsky, 2003)",
"ref_id": "BIBREF17"
},
{
"start": 610,
"end": 642,
"text": "(Raghavan & Garcia-Molina, 2001;",
"ref_id": "BIBREF16"
},
{
"start": 643,
"end": 681,
"text": "Lage, Silva, Golgher, & Laender, 2004)",
"ref_id": "BIBREF11"
},
{
"start": 1277,
"end": 1310,
"text": "(Cope, Craswell, & Hawking, 2003)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "Automatic search interface detection was first explored by Raghavan and Garcia-Molina in their hidden Web crawler HiWE (Hidden Web Exposer) (Raghavan & Garcia-Molina, 2001 ). Their crawling system used heuristic rules to detect the search entrance to the hidden database. Juliano P. Lage (Lage, Silva, Golgher, & Laender, 2004) used two heuristic rules to perform detection tasks. The first heuristic was the same as HiWE. The second one was to check whether the form contains the \"password\" HTML element or not. The disadvantage of this method, however, lies in that it does not have auto-leaning capability. Moreover, it is not robust and scalable to diverse hidden Web databases because the rules are too simple to match different form structures. Cope et al. (2003) used a decision tree classification algorithm to detect search interfaces. This method usually generates long rules due to the large size of the feature space in the training set (the number of training samples is too small compared to the number of features). Therefore, it is prone to overfitting, and the classification precision is not satisfying. Zhang et al. (2004) presented a best-effort parsing framework to address the problem of understanding Web search interfaces. The authors transformed search interfaces into a visual language under the hypothesis that automatic construction of search interfaces is guided by a hidden syntax. This hypothesis enables parsing as a principled framework to understand the semantic model of the visual language. The experimental results testified the effectiveness of their approach.",
"cite_spans": [
{
"start": 140,
"end": 171,
"text": "(Raghavan & Garcia-Molina, 2001",
"ref_id": "BIBREF16"
},
{
"start": 288,
"end": 327,
"text": "(Lage, Silva, Golgher, & Laender, 2004)",
"ref_id": "BIBREF11"
},
{
"start": 751,
"end": 769,
"text": "Cope et al. (2003)",
"ref_id": "BIBREF4"
},
{
"start": 1122,
"end": 1141,
"text": "Zhang et al. (2004)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "To summarize, little previous work has addressed the special characteristics of the search interface detection domain, for instance, large and diverse features, small size of training samples with many missing values, etc. The high dimension and sparse data of search interfaces present a tricky problem for the traditional single classifier approach. As collecting training samples (i.e., HTML forms) is costly, the training data set is usually small, while the number of features in the learning space is relatively large, due to multi-type features existing in forms. It's difficult for a single classifier to fully exploit the rich feature space (very sparse in the data matrix). For many classification methods, the single classifier tends to be overfitting. To attack this problem, we propose a multi-classifier ensemble approach in this paper.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "Our method is based on the random forest algorithm. A random forest model consists of a set of decision trees that are constructed by bootstrapping the training data. In our approach, we develop a weighted feature selection method to select a subset of features for each decision tree in the tree induction process. Classification is made by aggregating predictions of individual decision trees in the forest. Since each classifier is learned from a subset of the feature space, the ensemble approach can fully exploit the useful features in search forms. We have conducted experiments on several real data sets. The experimental results have shown that our random forest approach improves the classification accuracy in search interface detection.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "The contributions of this paper can be summarized as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "1. We explored the random forest approach to attacking the problem of detecting search interfaces from the sparse feature space of hidden Webs where specific feature extraction and representation techniques were used.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "2. We extended the random forest algorithm with a weighted feature selection method to select a subset of features for each decision tree. The new algorithm can automatically remove the noisy features in search forms so that decision tree classifiers can be learned from more representative subsets of the feature space.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "3. We conducted experiments on real data sets to compare the improved random forest algorithm with other well-known classification algorithms, such as SVM and C4.5. The experimental results have shown that the new method is more effective in detecting search interfaces of the hidden Web.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "The rest of this paper is organized as follows. In Section 2, we formalize the detecting search interface problem as a form classification problem and present the feature extraction techniques. Section 3 describes the improved random forest algorithm for form classification. Experimental results and analysis are presented in Section 4. Section 5 concludes this paper and presents our future work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "Search interface detection is a process of distinguishing the search forms of the hidden Web from non-search forms. It is a two-class classification problem in machine learning. This section describes how to extract form features from HTML pages and discusses the characteristics of the data matrix for form classification. HTML forms contain two kinds of features: one is the attributes of forms and elements, and the other is the statistics of those attributes. A form mainly contains four kinds of elements, that is, \"INPUT\", \"SELECT\", \"LABEL\" and \"TEXTAREA\", which are the children elements of \"FORM\" element. Element \"INPUT\" contains several types, such as \"text\", \"hidden\", etc. The hierarchy of a form is shown in Figure 1 . All of these elements contain a set of attributes, such as \"name\", \"value\", \"size\", etc. The attributes of \"form\" elements are \"method\", \"action\", and \"name\". Attribute \"method\" indicates the method for the form to submit query data, such as \"POST\" or \"GET\". Attribute \"action\" indicates the address of the corresponding server of the form, and attribute \"name\" indicates the name of the form. Some elements and attributes can be removed because they are not useful for form classification, for instance \"option\", \"size\", \"width\", etc. Besides, the statistics about the number of elements or attributes in each element can also be computed as important features. Figure 2 shows an example of a search form. The form contains three elements: one \"SELECT\" element and two \"INPUT\" elements. There are three attributes in the \"SELECT\" element : \"name\" with value \"and\", \"size\" with value \"1\", and \"width\" with value \"50\". It also contains several \"OPTION\" elements. The \"size\" and \"width\" attributes, along with the \"OPTION\" elements are not useful for form classification, so they can be removed. The corresponding HTML codes are shown in Figure 3 . Figure 5 . According the feature extraction rules, the useful form elements in this form include \"INPUT\", \"LABEL\", and \"FORM\", which can be used to compose the feature space. Elements \"TABEL\", \"FONT\", and \"TR\" can be removed because they are not useful. There are some important differences between the features of search form and non-search form. First, the number of \"INPUT\", \"SELECT\", and \"TEXTAREA\" elements in search forms is larger than that in non-search forms. Second, the value of the \"method\" attribute in \"FORM\" elements is always set as \"POST\" in search forms, while it is always set as \"GET\" in non-search forms. Moreover, the elements' values of search forms often contain some keywords such as \"search\", \"find\", or other words that have the same meaning as \"search\". These differences, however, are not the only decisive factors. There are other features that can be explored by classification algorithms.",
"cite_spans": [],
"ref_spans": [
{
"start": 721,
"end": 729,
"text": "Figure 1",
"ref_id": "FIGREF1"
},
{
"start": 1395,
"end": 1403,
"text": "Figure 2",
"ref_id": "FIGREF2"
},
{
"start": 1868,
"end": 1876,
"text": "Figure 3",
"ref_id": null
},
{
"start": 1879,
"end": 1887,
"text": "Figure 5",
"ref_id": "FIGREF5"
}
],
"eq_spans": [],
"section": "Feature Extraction for Form Classification",
"sec_num": "2."
},
{
"text": "According to the major differences, six kinds of rules are used in the feature extraction process as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Extraction Rules",
"sec_num": "2.1"
},
{
"text": "1. Extract the \"name\" attribute values from \"input\", \"select\", \"textarea\", and \"label\" elements;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Extraction Rules",
"sec_num": "2.1"
},
{
"text": "2. Extract the \"value\" attribute value from \"input\", \"textarea\", and \"label\" elements;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Extraction Rules",
"sec_num": "2.1"
},
{
"text": "3. Extract the \"name\" and \"method\" attribute values from \"form\" elements;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Extraction Rules",
"sec_num": "2.1"
},
{
"text": "4. Extract the words that appear between slashes(/) in the \"action\" attributes of the \"form\" elements;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Extraction Rules",
"sec_num": "2.1"
},
{
"text": "5. Extract the words that appear between slashes(/) in the \"src\" and \"alt\" attributes of the \"input-image\" element;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Extraction Rules",
"sec_num": "2.1"
},
{
"text": "6. Calculate the number of \"input\", \"select\", \"label\", and \"textarea\" elements in each form.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Extraction Rules",
"sec_num": "2.1"
},
{
"text": "The next step is to standardize the value of the features that are extracted from the forms.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Extraction Rules",
"sec_num": "2.1"
},
{
"text": "First, all strings are transformed into lowercases; then the string type values are aggregated and mapped to specific enumerating values. For instance, the values of \"search\", \"find\", or \"srch\" are mapped to \"search\".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Weighting Random Forest for Detection of 393 Hidden Web Search Interfaces",
"sec_num": null
},
{
"text": "The extracted features and the labels of forms are used to compose the data matrix for the classification algorithm. The formalized data matrix is shown in Table 1 . ",
"cite_spans": [],
"ref_spans": [
{
"start": 156,
"end": 163,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "The Sparse Data Matrix",
"sec_num": "2.2"
},
{
"text": "class 1 t 2 t i t m t 1 c 11 a 12 a 1i a 1m a j c 1 j a 2 j a ji a jm a n c 1 n a 2 n a ni a nm a The set { } 1 2 m T t t t = ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Sparse Data Matrix",
"sec_num": "2.2"
},
{
"text": ", in Table 1 represents the names of the form features. For form classification, the label of a form is represented as an element in the set { } C yes no = , , while \"yes\" indicates that the form is a search interface of hidden Web and \"no\" indicates a non-search interface. Each row is a sample form. ji a represents the value of feature i t in the j th form, and j c indicates the class of the j th form. Table 2 illustrates two examples of a search form and a non-search form as shown in Figure 2 and Figure 4 .",
"cite_spans": [],
"ref_spans": [
{
"start": 5,
"end": 12,
"text": "Table 1",
"ref_id": "TABREF1"
},
{
"start": 407,
"end": 414,
"text": "Table 2",
"ref_id": "TABREF2"
},
{
"start": 491,
"end": 499,
"text": "Figure 2",
"ref_id": "FIGREF2"
},
{
"start": 504,
"end": 512,
"text": "Figure 4",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "The Sparse Data Matrix",
"sec_num": "2.2"
},
{
"text": "The expression of i t is a four-tuple of \"element name\"-\"type\"-\"attribute name\"-\"sequence number\". The \"element name\" contains six values: \"FORM\", \"SELECT\", \"INPUT\", \"TEXT AREA\", \"LABEL\", and their statistics. For element, \"INPUT\", the value of \"type\" can be \"text\", \"hidden\", and so on. \"attribute name\" has six options: \"name\", \"value\", \"src\", \"alt\", \"method\", and \"action\". Sequence number represents the sequence of the features in the form.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Sparse Data Matrix",
"sec_num": "2.2"
},
{
"text": "As illustrated in Table 2 , the combination of \"element name\",\"type\",\"attribute name\", and \"sequence number\" has many unique alternatives. This will result in a high-dimensional feature space for form classification. Furthermore, since each form has just a few features, the data matrix for classification is very sparse and there are many missing values and noisy features. This problem presents a big challenge for search form detection. ",
"cite_spans": [],
"ref_spans": [
{
"start": 18,
"end": 25,
"text": "Table 2",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "The Sparse Data Matrix",
"sec_num": "2.2"
},
{
"text": "This section presents an improved random forest algorithm, which extends the classical random forest method with a feature weighting technique. We describe the basic random forest classification approach in Subsection 3.1 and our new algorithm in Subsection 3.2.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Weighting Random Forest Algorithm",
"sec_num": "3."
},
{
"text": "Random Forest (RFA) (Ho, 1998; Breiman, 2001 ) is an ensemble of unpruned classification or regression trees, which is induced from bootstrapping samples of the training set, using random feature selection in the tree induction process. Prediction is made by aggregating the predictions of the ensemble. Random Forest grows many classification trees. To classify a new object from an input vector, it passes the sample vector to each of the trees in the forest. Each tree gives a classification decision. All the classification results of individual trees are combined to choose the classification having the most votes over all the classification trees in the forest.",
"cite_spans": [
{
"start": 20,
"end": 30,
"text": "(Ho, 1998;",
"ref_id": "BIBREF9"
},
{
"start": 31,
"end": 44,
"text": "Breiman, 2001",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Random Forest Algorithm",
"sec_num": "3.1"
},
{
"text": "Random forest generally exhibits a substantial performance improvement over single tree classifiers, such as CART (Breiman, Friedman, & Olshen, 1984) and C4.5 (Quinlan, 1993) . It presents a good solution for classification of sparse data sets. Since basic RFA selects features randomly, it's easy to select unimportant or noisy features, especially when there are many noisy features in the training data. This may lead to bad classification results. As discussed in previous sections, the data matrix for form classification contains many missing values. It's necessary to enhance basic RFA so that the performance can be improved in search form classification.",
"cite_spans": [
{
"start": 114,
"end": 149,
"text": "(Breiman, Friedman, & Olshen, 1984)",
"ref_id": "BIBREF1"
},
{
"start": 159,
"end": 174,
"text": "(Quinlan, 1993)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Random Forest Algorithm",
"sec_num": "3.1"
},
{
"text": "Due to the sparse feature space, there are a lot of missing values in the training data set. The features with too many missing values become less important and can be treated as noisy features. Random selection of features often obtains many unimportant or noisy features, which leads to bad trees in the forest. To avoid this, we extend basic RFA, using a weighting scheme in feature selection to replace random selection. We use 2 \u03c7 statistic to measure the importance of features (Larson, 1982) . The 2 \u03c7 statistic of a feature A against the class feature is computed as follows. ",
"cite_spans": [
{
"start": 484,
"end": 498,
"text": "(Larson, 1982)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Improved Random Forest with Weighted Feature Selection",
"sec_num": "3.2"
},
{
"text": "If class i c has the highest probability, i c is the class of x . Kittler has provided a more profound explanation of this method (Kittler, Hatef, Duin, & Matas, 1998) . The pseudo-code of the new algorithm (IRFA) is given in 1 Algorithm .",
"cite_spans": [
{
"start": 130,
"end": 167,
"text": "(Kittler, Hatef, Duin, & Matas, 1998)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Improved Random Forest with Weighted Feature Selection",
"sec_num": "3.2"
},
{
"text": "Step 1 is to compute a weight for each feature according to (1). Step 2 sorts the features in descending order of feature weights. Step 3 selects n\u2032 features from the entire feature set according to a given feature selection rate \u03b2 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Improved Random Forest with Weighted Feature Selection",
"sec_num": "3.2"
},
{
"text": "Step 4 learns individual classifiers from the selected training samples (and selected features). Selection of training samples employs the bootstrapping method. The method of sampling without replacement is used to select t features from n\u2032 features, where 2 log 1 t n = + \u23a2 \u23a5 \u23a3 \u23a6 . After each iteration, the learned decision tree classifier will be added to forest M * . After forest M * is grown,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Improved Random Forest with Weighted Feature Selection",
"sec_num": "3.2"
},
{
"text": "Step 5 classifies the unlabeled instances based on (4).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Improved Random Forest with Weighted Feature Selection",
"sec_num": "3.2"
},
{
"text": "The pseudo-code of the feature weighting random forest algorithm Input: ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Algorithm 1",
"sec_num": null
},
{
"text": "-D :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Algorithm 1",
"sec_num": null
},
{
"text": "The computational complexity of RFA (Breiman, 2001) The computational cost depends on three factors: the number of decision trees k , the number of features n , and the number of training samples d . We will discuss how to select the number of decision trees k and the number of features n to balance between classification accuracy and computational cost in Section 4. ",
"cite_spans": [
{
"start": 36,
"end": 51,
"text": "(Breiman, 2001)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The computational complexity",
"sec_num": "3.3"
},
{
"text": "We used two Web page collections in our experiments. One was taken from project Metaquerier 1 , and the other was created by crawling the website Search Engine Guide 2 with a Web crawler implemented in Java. The two collections represent a pseudo-random crawling of the Web. A HTML parser was developed to extract the HTML forms and their context features from these two collections. The extracted forms were used to compose the final data sets for experiments. The forms are selected from data set 1 and data set 2.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Sets",
"sec_num": "4.1"
},
{
"text": "We manually classified the extracted forms into search forms (i.e., real search interface of hidden Web) and non-search forms. Three classification data sets were constructed from the classified forms, as shown in Table 3 . The three data sets have a variety of sample distributions. Data set 1 and Data set 2 cover different domains, while Data set 3 is a mixture of the two domains. The three data sets also have different feature types and feature numbers. The average number of features in the data sets is over 200, while the average number of samples is less than 140. The matrices of the three data sets were quite sparse, and the number of features was quite large.",
"cite_spans": [],
"ref_spans": [
{
"start": 214,
"end": 221,
"text": "Table 3",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Figure 6. The domain distribution of non-search forms",
"sec_num": null
},
{
"text": "To further test the robustness of our method, the non-search forms in the data sets were made of a variety of forms, including registration forms, login forms, network investigation forms, etc. Figure 6 shows the distribution of different non-search forms.",
"cite_spans": [],
"ref_spans": [
{
"start": 194,
"end": 202,
"text": "Figure 6",
"ref_id": null
}
],
"eq_spans": [],
"section": "Figure 6. The domain distribution of non-search forms",
"sec_num": null
},
{
"text": "We first carried out experiments to compare our random forest method with four well-known classification algorithms, i.e., Support Vector Machine (SVM), C4.5, Na\u00efve Bayes,and Random Forest Algorithm (RFA) implemented in Weka 3 . We also implemented our algorithm (IRFA) as a plug-in of Weka and conducted all experiments in this environment to make a fair comparison. We conducted the standard 10-fold classification experiments on the three data sets. The evaluation metrics used were precision and computation time. The number of trees was set to 100 for IRFA and parameter \u03b2 set to 0.5 . The final experimental results are shown in Table 4 . We can see that IRFA showed significant improvement over the other four algorithms. The result of C4.5 was better than SVM and Na\u00efve Bayes. This was due to the fact that there were a lot of missing values in the data sets and SVM and Na\u00efve Bayes did not perform well in this kind of sparse data. The high dimensionality in the training sets, however, causes an overfitting problem to C4.5 because the single decision tree could become very complex. RFA and IRFA can avoid this problem by selecting different subsets of features to build individual decision trees. Compared with RFA, IRFA uses features that are more correlated to the class label feature, so the accuracy of each individual tree is improved. Therefore, our method got better performance than RFA. Since IRFA needed to compute the 2 \u03c7 values for features, it",
"cite_spans": [],
"ref_spans": [
{
"start": 635,
"end": 642,
"text": "Table 4",
"ref_id": "TABREF6"
}
],
"eq_spans": [],
"section": "Comparison Experiments",
"sec_num": "4.2"
},
{
"text": "Hidden Web Search Interfaces took more computational time. This extra overhead, however, is worthwhile and acceptable in real applications because the training process is offline and not executed frequently.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison Experiments",
"sec_num": "4.2"
},
{
"text": "When building each decision tree, IRFA only selects a subset of the original features. The number of selected features is controlled by the selection rate \u03b2 . Different selection rates result in different classification precisions and computational costs. We carried out experiments on the three data sets with 0.1, 0.2, , 1.0",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Selection of the Number of Features",
"sec_num": "4.3"
},
{
"text": "\u03b2 = \u2026",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Selection of the Number of Features",
"sec_num": "4.3"
},
{
"text": ". Figure 7 plots selection rate \u03b2 against precision, while Figure 8 is \u03b2 against computational cost. Figure 7 shows that when 0.5 \u03b2 < , the precision increases greatly as the selection rate increases. The reason is that a larger selection rate increases the number of features to be selected. When 0.5 0.8 \u03b2 \u2264 \u2264 , the classification performance becomes relatively stable. This means that the forest has selected enough discriminative features. When 0.8 \u03b2 > , the precision will decrease as the selection rate increases. This can be explained by the idea that having too large a selection rate will increase the possibility of selecting noisy features. Most experiments have shown that 0.5 \u03b2 = was a good setting. Figure 8 shows that the computational time of IRFA increases linearly as the feature selection rate increases. This property indicates that IRFA is scalable to large high-dimensional data. ",
"cite_spans": [],
"ref_spans": [
{
"start": 2,
"end": 10,
"text": "Figure 7",
"ref_id": null
},
{
"start": 59,
"end": 67,
"text": "Figure 8",
"ref_id": null
},
{
"start": 101,
"end": 109,
"text": "Figure 7",
"ref_id": null
},
{
"start": 713,
"end": 721,
"text": "Figure 8",
"ref_id": null
}
],
"eq_spans": [],
"section": "Selection of the Number of Features",
"sec_num": "4.3"
},
{
"text": "Another interesting issue is whether the performance of IRFA highly depends on the size of a forest (number of decision trees in the forest). Since a large number of trees lead to a considerable computational cost, we need to find a good tradeoff between classification precision and computational cost. We performed experiments on the three data sets with different tree numbers ( 10, 25, 50, 75, 100, 125, 150, 200) k = . The feature selection rate \u03b2 was set to 0.5 in all experiments. The experimental results are shown in Figure 9 and Figure 10 . Figure 9 plots the precision against the number of trees k . The results show that if the number of trees is too small, the classification performance of the forest will be unstable. If the number of trees is too large, however, the computational cost for generating a forest will be very high. The classification precision becomes stable when 75 150 k < <",
"cite_spans": [
{
"start": 380,
"end": 385,
"text": "( 10,",
"ref_id": null
},
{
"start": 386,
"end": 389,
"text": "25,",
"ref_id": null
},
{
"start": 390,
"end": 393,
"text": "50,",
"ref_id": null
},
{
"start": 394,
"end": 397,
"text": "75,",
"ref_id": null
},
{
"start": 398,
"end": 402,
"text": "100,",
"ref_id": null
},
{
"start": 403,
"end": 407,
"text": "125,",
"ref_id": null
},
{
"start": 408,
"end": 412,
"text": "150,",
"ref_id": null
},
{
"start": 413,
"end": 417,
"text": "200)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 526,
"end": 534,
"text": "Figure 9",
"ref_id": "FIGREF9"
},
{
"start": 539,
"end": 548,
"text": "Figure 10",
"ref_id": "FIGREF1"
},
{
"start": 551,
"end": 559,
"text": "Figure 9",
"ref_id": "FIGREF9"
}
],
"eq_spans": [],
"section": "Selection of the Number of Trees",
"sec_num": "4.4"
},
{
"text": ". The near-optimal precision can be obtained when k is set to 100. Moreover, as shown in Figure 10 , with the increase of the number of trees, the computational time increases linearly as well. Therefore, in most situations, 100 k = is a good balance between classification accuracy and computational cost. ",
"cite_spans": [],
"ref_spans": [
{
"start": 89,
"end": 98,
"text": "Figure 10",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Selection of the Number of Trees",
"sec_num": "4.4"
},
{
"text": "The representation of search forms in different domains varies, so it is necessary to investigate if the performance of IRFA can be consistent in different domains. An experiment was designed to classify search interfaces from four different domains. We used the Web collection provided by Metaquerier 4 , which contains the information about Books, Movies, Airfares, and Jobs. The number of trees for IRFA was set to 100, and parameter \u03b2 was set to 0.5.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison in Different Domains",
"sec_num": "4.5"
},
{
"text": "The experimental results are shown in Table 5 . We can see the obvious variance of accuracy with regard to different domains. The detection accuracies for Books and Jobs domains are higher than those of Movies and Airfares. This was due to the different HTML structures of the search interfaces, as the search interfaces of Movies and Airfares domains are more complex and some of them require more than one step to get the content. The results imply that more formalized and simpler interfaces are more easily recognized. From the results, we also observe that the classification performance of IRFA was very stable in various domains. Even though the precision of SVM on the Movies data set was a little better than IRFA, it became the worst in other domains. Compared with other algorithms, IRFA was the most stable. ",
"cite_spans": [],
"ref_spans": [
{
"start": 38,
"end": 45,
"text": "Table 5",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Comparison in Different Domains",
"sec_num": "4.5"
},
{
"text": "We conducted experiments to demonstrate the performance of IRFA with different feature selection schemes. In this section, we used four commonly used feature selection functions mentioned in (Dash & Liu, 1997) .",
"cite_spans": [
{
"start": 191,
"end": 209,
"text": "(Dash & Liu, 1997)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison with Different Feature Selection Schemes",
"sec_num": "4.6"
},
{
"text": "1. Random selection, which randomly selects the features using sampling without replacement, is the simplest feature selection method. Random feature selection is the method used in the original random forest (Breiman, 2001 ).",
"cite_spans": [
{
"start": 209,
"end": 223,
"text": "(Breiman, 2001",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison with Different Feature Selection Schemes",
"sec_num": "4.6"
},
{
"text": "2. Information gain is defined as the difference between the original information requirement and the new requirement (Han & Kamber, 2007) . The information gain value of a feature X is attained by computing the difference between the prior uncertainty and expected posterior uncertainty using X . Feature X is preferred to feature Y if the information gain from feature X is greater than that from feature Y (Dash & Liu, 1997) . In this experiment, the information gain from feature X is set to feature X as its weight instead of Chi-square value in IRFA.",
"cite_spans": [
{
"start": 118,
"end": 138,
"text": "(Han & Kamber, 2007)",
"ref_id": "BIBREF7"
},
{
"start": 409,
"end": 427,
"text": "(Dash & Liu, 1997)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison with Different Feature Selection Schemes",
"sec_num": "4.6"
},
{
"text": "3. Gain Ratio is an extension of information gain. It attempts to overcome the problem that the information gain measure is biased toward tests with many outcomes (i.e., it prefers to select attributes having a large number of values) (Han & Kamber, 2007) . The process of embedding gain ratio into IRFA is similar to information gain in our experiment.",
"cite_spans": [
{
"start": 235,
"end": 255,
"text": "(Han & Kamber, 2007)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison with Different Feature Selection Schemes",
"sec_num": "4.6"
},
{
"text": "4. Chi-square 2 \u03c7 that has been explained in Section 3.2. The experiments in the prior sections were all based on this feature selection method. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Comparison with Different Feature Selection Schemes",
"sec_num": "4.6"
},
{
"text": "We implemented the four feature selection methods in Weka's random forest package, and carried out experiments on the data sets described in Section 4.1. The results are shown in Table 6 . From the results, we can see that the Chi-square method is better than the others. The reason that Information Gain and Gain Ratio scheme did not attain better performance can be explained as follows: since both of them use the same feature evaluation criterion in feature sampling and tree construction (for node splitting), the information of features cannot be fully exploited.",
"cite_spans": [],
"ref_spans": [
{
"start": 179,
"end": 186,
"text": "Table 6",
"ref_id": "TABREF8"
}
],
"eq_spans": [],
"section": "93.88%",
"sec_num": null
},
{
"text": "This paper has proposed IRFA, an improved random forest algorithm, for detecting search interfaces of the hidden Web. We extend the original random forest algorithm with a weighted feature selection method to automatically select a more representative subset of features for building each decision tree. The new method can overcome the problem of classifying high-dimensional and sparse search interface data through the ensemble of decision trees, each learned from a different subset of the original feature space. We have implemented the new algorithm and compared it with SVM, C4.5, Na\u00efve Bayes, and original random forest algorithm (RFA). The experimental results have shown that our method is more accurate and robust. We have also observed that the new method is scalable to high dimensional data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "5."
},
{
"text": "In the future, we plan to investigate more feature weighting methods for construction of random forests. Currently, we just use the features in the search forms. It is expected that using contextual information near the search forms may improve detection performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Weighting Random Forest for Detection of 403 Hidden Web Search Interfaces",
"sec_num": null
},
{
"text": "http://metaquerier.cs.uiuc.edu/repository/ training data set 2 http://www.searchengineguide.com",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "http://www.cs.waikato.ac.nz/ml/weka/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "http://metaquerier.cs.uiuc.edu/repository/training data set",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This research is supported in part by NSFC under grant No.60603066 and China National High-tech Program under grants No.2007AA01Z436 and No.2006AA01A124. Part of Joshua Huang's research was supported by the 863 project matching fund from The University of Hong Kong.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Random forests",
"authors": [
{
"first": "L",
"middle": [],
"last": "Breiman",
"suffix": ""
}
],
"year": 2001,
"venue": "Machine Learning",
"volume": "45",
"issue": "",
"pages": "5--32",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5-32.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Classification and regression trees",
"authors": [
{
"first": "L",
"middle": [],
"last": "Breiman",
"suffix": ""
},
{
"first": "J",
"middle": [
"H"
],
"last": "Friedman",
"suffix": ""
},
{
"first": "R",
"middle": [
"A"
],
"last": "Olshen",
"suffix": ""
}
],
"year": 1984,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Breiman, L., Friedman, J. H., & Olshen, R. A. (1984). Classification and regression trees. New York: Chapman & Hall.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "The deep web: Surfacing hidden value",
"authors": [
{
"first": "",
"middle": [],
"last": "Brightplanet",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Com",
"suffix": ""
}
],
"year": 2000,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "BrightPlanet.com (2000). The deep web: Surfacing hidden value. URL http://www.brightplanet.com",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Cluster and discover: focused extraction of qa-pagelets from the deep web",
"authors": [
{
"first": "J",
"middle": [],
"last": "Caverlee",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "D",
"middle": [
"B"
],
"last": "Probe",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceeding of the 20th International Conference of Data Engineering",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Caverlee, J., Liu, L., & Probe, D. B. (2004). Cluster and discover: focused extraction of qa-pagelets from the deep web. Proceeding of the 20th International Conference of Data Engineering.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Automated discovery of search nterfaces on the web. Fourteenth Australasian Database Conference",
"authors": [
{
"first": "J",
"middle": [],
"last": "Cope",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Craswell",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Hawking",
"suffix": ""
}
],
"year": 2003,
"venue": "Fourteenth Australasian Database Conference",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Cope, J., Craswell, N., & Hawking, D. (2003). Automated discovery of search nterfaces on the web. Fourteenth Australasian Database Conference. Adelaide, Australia: Fourteenth Australasian Database Conference.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Feature selection for classification",
"authors": [
{
"first": "M",
"middle": [],
"last": "Dash",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Liu",
"suffix": ""
}
],
"year": 1997,
"venue": "",
"volume": "1",
"issue": "",
"pages": "131--156",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dash, M., & Liu, H. (1997). Feature selection for classification. Intelligent Data nalysis, 1(3), 131-156.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Database techniques for the world wide web: A survey",
"authors": [
{
"first": "D",
"middle": [],
"last": "Florescu",
"suffix": ""
},
{
"first": "A",
"middle": [
"Y"
],
"last": "Levy",
"suffix": ""
},
{
"first": "A",
"middle": [
"O"
],
"last": "Mendelzon",
"suffix": ""
}
],
"year": 1998,
"venue": "SIGMOD Record",
"volume": "27",
"issue": "3",
"pages": "59--74",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Florescu, D., Levy, A. Y., & Mendelzon, A. O. (1998). Database techniques for the world wide web: A survey. SIGMOD Record, 27(3), 59-74.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Data Mining: Concepts and Techniques",
"authors": [
{
"first": "J",
"middle": [],
"last": "Han",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Kamber",
"suffix": ""
}
],
"year": 2007,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Han, J., & Kamber, M. (2007). Data Mining: Concepts and Techniques(2nd version). China Machine Press.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Accessing the deep web",
"authors": [
{
"first": "B",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Patel",
"suffix": ""
},
{
"first": "Z",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "K",
"middle": [
"C"
],
"last": "Chang",
"suffix": ""
}
],
"year": 2007,
"venue": "Communications Of The ACM",
"volume": "50",
"issue": "5",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "He, B., Patel, M., Zhang, Z., & Chang, K. C. (2007). Accessing the deep web. Communications Of The ACM, 50(5).",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "The random subspace method of constructing decision forests",
"authors": [
{
"first": "T",
"middle": [
"K"
],
"last": "Ho",
"suffix": ""
}
],
"year": 1998,
"venue": "IEEE Trans. on Pattern Analysis and Machine Intelligence",
"volume": "20",
"issue": "8",
"pages": "832--844",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ho, T. K. (1998). The random subspace method of constructing decision forests. IEEE Trans. on Pattern Analysis and Machine Intelligence, 20(8), 832-844.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "On combining classifiers",
"authors": [
{
"first": "J",
"middle": [],
"last": "Kittler",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Hatef",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Duin",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Matas",
"suffix": ""
}
],
"year": 1998,
"venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"volume": "20",
"issue": "",
"pages": "226--239",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kittler, J., Hatef, M., Duin, R., & Matas, J. (1998). On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20, 226-239.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Automatic generation of agents for collecting hidden web pages for data extraction",
"authors": [
{
"first": "J",
"middle": [
"P"
],
"last": "Lage",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Silva",
"suffix": ""
},
{
"first": "P",
"middle": [
"B"
],
"last": "Golgher",
"suffix": ""
},
{
"first": "A",
"middle": [
"H F"
],
"last": "Laender",
"suffix": ""
}
],
"year": 2004,
"venue": "Data and Knowledge Engineering",
"volume": "49",
"issue": "",
"pages": "177--196",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lage, J. P., Silva, D., Golgher, P. B., & Laender, A. H. F. (2004). Automatic generation of agents for collecting hidden web pages for data extraction. Data and Knowledge Engineering, 49, 177-196.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Introduction to probability theory and statistical inference",
"authors": [
{
"first": "H",
"middle": [
"J"
],
"last": "Larson",
"suffix": ""
}
],
"year": 1982,
"venue": "",
"volume": "3",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Larson, H. J. (1982). Introduction to probability theory and statistical inference. New York: Wiley, 3 ed.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Downloading textual hidden web content through keyword queries",
"authors": [
{
"first": "A",
"middle": [],
"last": "Ntoulas",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Zerfos",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Cho",
"suffix": ""
}
],
"year": 2005,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ntoulas, A., Zerfos, P., & Cho, J. (2005). Downloading textual hidden web content through keyword queries.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "C4.5: Programs for Machine LearningMachine Learning",
"authors": [
{
"first": "J",
"middle": [
"R"
],
"last": "Quinlan",
"suffix": ""
}
],
"year": 1993,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Quinlan, J. R. (1993). C4.5: Programs for Machine LearningMachine Learning. Morgan Kaufmann.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Crawling the hidden web",
"authors": [
{
"first": "S",
"middle": [],
"last": "Raghavan",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Garcia-Molina",
"suffix": ""
}
],
"year": 2001,
"venue": "Proceedings of the 27th International Conference on Very Large Data Bases",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Raghavan, S., & Garcia-Molina, H. (2001). Crawling the hidden web. Proceedings of the 27th International Conference on Very Large Data Bases. Roma, Italy.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Data extraction and label assignment for web databases",
"authors": [
{
"first": "J",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Lochovsky",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of the 12th International World Wide Web Conference",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Wang, J., & Lochovsky, F. (2003). Data extraction and label assignment for web databases. Proceedings of the 12th International World Wide Web Conference. Budapest, Hungary.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Understanding web query interfaces: best-effort parsing with hidden syntax",
"authors": [
{
"first": "Z",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "K",
"middle": [
"C"
],
"last": "Chang",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the 2004 ACM SIGMOD International Conference on Management of Data",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhang, Z., He, B., & Chang, K. C. (2004). Understanding web query interfaces: best-effort parsing with hidden syntax. Proceedings of the 2004 ACM SIGMOD International Conference on Management of Data. Paris, France.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"type_str": "figure",
"text": "",
"uris": null,
"num": null
},
"FIGREF1": {
"type_str": "figure",
"text": "The hierarchy of form elements Feature Weighting Random Forest for Detection of 391 Hidden Web Search Interfaces",
"uris": null,
"num": null
},
"FIGREF2": {
"type_str": "figure",
"text": "A search formFigure 3. The HTML codes of the form inFigure 2",
"uris": null,
"num": null
},
"FIGREF3": {
"type_str": "figure",
"text": "shows an example of a non-search form. The corresponding HTML codes are shown in",
"uris": null,
"num": null
},
"FIGREF4": {
"type_str": "figure",
"text": "A non-search form",
"uris": null,
"num": null
},
"FIGREF5": {
"type_str": "figure",
"text": "The HTML codes of the form inFigure 4",
"uris": null,
"num": null
},
"FIGREF6": {
"type_str": "figure",
"text": "of samples whose value of the class feature is j c .An 2\u03c7 statistic weight is calculated for each feature. From the weights, we select only different subsets of features with high weights to build individual decision trees.Given a set of decision trees built from different subsets of features, we use a probability estimation technique to combine the results of individual classifiers. Assume x is a test instance and is given to each classifier (",
"uris": null,
"num": null
},
"FIGREF7": {
"type_str": "figure",
"text": "",
"uris": null,
"num": null
},
"FIGREF8": {
"type_str": "figure",
"text": "Influence of the number of features on precision Influence of the number of features on time cost 400YunmingYe et al.",
"uris": null,
"num": null
},
"FIGREF9": {
"type_str": "figure",
"text": "Influence of the number of trees on precision Figure 10. Influence of the number of trees on time cost Feature Weighting Random Forest for Detection of 401 Hidden Web Search Interfaces",
"uris": null,
"num": null
},
"TABREF1": {
"html": null,
"text": "",
"content": "<table/>",
"type_str": "table",
"num": null
},
"TABREF2": {
"html": null,
"text": "",
"content": "<table><tr><td colspan=\"4\">class form-action-1 form-action-2 form-action-3</td><td colspan=\"2\">input-text-number input-submit-number</td></tr><tr><td colspan=\"2\">yes www.thearda.com</td><td>cgi-bin</td><td>search</td><td>1</td><td>1</td></tr><tr><td>no</td><td>servlet</td><td>login</td><td>?</td><td>3</td><td>0</td></tr></table>",
"type_str": "table",
"num": null
},
"TABREF4": {
"html": null,
"text": ", where k is the number of decision trees, t is the number of attributes, d is the number of training samples. In IRFA, the enumerating number of the feature attribute is constant(Formula (1)). The computational complexity of all feature weights is ( ) O n . Using the bucket sorting method, the weights can be sorted in linear time. Therefore, the computational complexity of the IRFA",
"content": "<table><tr><td>is ( log ) O ktd d is ( log ) O ktd d n \u03b1 + , where 2 log 1 t n = + \u23a2 \u23a5 \u23a3 \u23a6 . Therefore, the computational complexity of</td></tr><tr><td>IRFA is very close to the complexity of RFA.</td></tr></table>",
"type_str": "table",
"num": null
},
"TABREF5": {
"html": null,
"text": "",
"content": "<table><tr><td/><td>search</td><td>non-search</td><td colspan=\"2\">features content of data sets</td></tr><tr><td/><td/><td/><td/><td>Extracted from the website</td></tr><tr><td>Data set 1</td><td>46</td><td>43</td><td>198</td><td>collection crawled from Search</td></tr><tr><td/><td/><td/><td/><td>Engine Guide by crawler</td></tr><tr><td/><td/><td/><td/><td>Extracted from artificial</td></tr><tr><td>Data Set 2</td><td>65</td><td>116</td><td>208</td><td>website collection of</td></tr><tr><td/><td/><td/><td/><td>Metaquerier project</td></tr><tr><td>Data Set 3</td><td>51</td><td>96</td><td>202</td><td/></tr></table>",
"type_str": "table",
"num": null
},
"TABREF6": {
"html": null,
"text": "",
"content": "<table><tr><td/><td/><td colspan=\"3\">Bayes C4.5 Decision Tree SVM</td><td>RFA</td><td>IRFA</td></tr><tr><td>Data Set 1</td><td>Precision Time Cost(s)</td><td>82.02% < 0.005</td><td>80.90% 0.05</td><td>79.78% 0.52</td><td>88.76% 3.16</td><td>91.01% 7.08</td></tr><tr><td>Data Set 2</td><td>Precision Time Cost(s)</td><td>83.43% < 0.005</td><td>88.40% 0.03</td><td>82.87% 0.88</td><td>91.71% 5.22</td><td>92.27% 17.86</td></tr><tr><td>Data Set 3</td><td>Precision Time Cost(s)</td><td>84.35% < 0.005</td><td>89.79% 0.02</td><td>85.71% 0.88</td><td>91.84% 3.58</td><td>93.88% 15.13</td></tr></table>",
"type_str": "table",
"num": null
},
"TABREF7": {
"html": null,
"text": "",
"content": "<table><tr><td/><td colspan=\"2\">Dataset Size Bayes</td><td>C4.5</td><td>SVM</td><td>RFA</td><td>IRFA</td></tr><tr><td>Books</td><td>251</td><td colspan=\"3\">92.83% 96.41% 89.24%</td><td>92.03% 96.81%</td></tr><tr><td>Movies</td><td>126</td><td colspan=\"4\">91.27% 91.27% 92.06% 87.30%</td><td>91.27%</td></tr><tr><td>Airfares</td><td>265</td><td colspan=\"3\">89.06% 90.94% 87.92%</td><td>91.69% 91.70%</td></tr><tr><td>Jobs</td><td>104</td><td colspan=\"4\">88.46% 94.23% 78.85% 96.15% 96.15%</td></tr></table>",
"type_str": "table",
"num": null
},
"TABREF8": {
"html": null,
"text": "",
"content": "<table><tr><td/><td>Random</td><td colspan=\"3\">Information Gain Gain Ratio Chi-square 2 \u03c7</td></tr><tr><td>Data Set 1</td><td>88.76%</td><td>88.76%</td><td>89.89%</td><td>91.01%</td></tr><tr><td>Data Set 2</td><td>91.71%</td><td>91.16%</td><td>92.27%</td><td>92.27%</td></tr><tr><td>Data Set 3</td><td>91.84%</td><td>93.20%</td><td>93.20%</td><td/></tr></table>",
"type_str": "table",
"num": null
}
}
}
} |