Runs - Spam 2005¶

1BASresults¶

Run ID: 1BASresults
Participant: beijingu.guo
Track: Spam
Year: 2005
Submission: 9/1/2005
Task: run

1cefhuj¶

Run ID: 1cefhuj
Participant: breyer.laird
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: This filter tests the cef tokenizer with full standard header analysis and uniform reference measure, case sensitive tokens. Same as pilot submission.

2adphu¶

Run ID: 2adphu
Participant: breyer.laird
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: This filter tests the adp tokenizer with full standard header analysis and uniform reference measure, lowercase tokens. Same as pilot submission.

2V2Bresults¶

Run ID: 2V2Bresults
Participant: beijingu.guo
Track: Spam
Year: 2005
Submission: 9/1/2005
Task: run

3adphd¶

Run ID: 3adphd
Participant: breyer.laird
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: This filter tests the adp tokenizer with full standard header analysis and dirichlet reference measure, lowercase tokens. Same as pilot submission.

3V2Eresults¶

Run ID: 3V2Eresults
Participant: beijingu.guo
Track: Spam
Year: 2005
Submission: 9/1/2005
Task: run

4adp¶

Run ID: 4adp
Participant: breyer.laird
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: This filter tests the adp tokenizer with only Subject analysis, uniform reference measure, lowercase tokens. Same as pilot submission.

4V5Bresults¶

Run ID: 4V5Bresults
Participant: beijingu.guo
Track: Spam
Year: 2005
Submission: 9/1/2005
Task: run

621SPAM1FIN¶

Run ID: 621SPAM1FIN
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: Our submission consists of two core filters and an aggregator. The first is LNB, and extension of naive bayes that partially removes the independence assumption. The second is the SMTP path analysis algorithm presented at CEAS 2005. The aggregator uses the Amoeba optimatization to find the optimal linear weights to combine each classifier. Submission 1 is the aggregate classifier combining both SMTP path analysis and LNB. The second and third submissions are the SMTP path analysis and the LNB algorithm run individually. This will allow us to ascertain both how the individual algorithms perform and how well the aggregator does is combining them. NOTE this submission contains three submissions; details on evaluation precedence are in README.txt

621SPAM1FT1¶

Run ID: 621SPAM1FT1
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 7/5/2005
Type: pilot
Task: filter
Run description: SpamGuru combines mulitple filtering technologies, including bayesian, smtp path analysis, and pattern matching algorithms from the life sciences and combines their scores using an adaptive weighting algorithm. The test filter submitted here only has the bayesian component enabled.

621SPAM1FUL¶

Run ID: 621SPAM1FUL
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

621SPAM1H25¶

Run ID: 621SPAM1H25
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

621SPAM1H50¶

Run ID: 621SPAM1H50
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

621SPAM1S25¶

Run ID: 621SPAM1S25
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

621SPAM1S50¶

Run ID: 621SPAM1S50
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

621SPAM2FUL¶

Run ID: 621SPAM2FUL
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

621SPAM2H25¶

Run ID: 621SPAM2H25
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

621SPAM2H50¶

Run ID: 621SPAM2H50
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

621SPAM2S25¶

Run ID: 621SPAM2S25
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

621SPAM2S50¶

Run ID: 621SPAM2S50
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

621SPAM3FUL¶

Run ID: 621SPAM3FUL
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

621SPAM3H25¶

Run ID: 621SPAM3H25
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

621SPAM3H50¶

Run ID: 621SPAM3H50
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

621SPAM3S25¶

Run ID: 621SPAM3S25
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

621SPAM3S50¶

Run ID: 621SPAM3S50
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

621SPAMFT1¶

Run ID: 621SPAMFT1
Participant: ibm.segal
Track: Spam
Year: 2005
Submission: 7/5/2005
Task: run

azeSPAM1BNS¶

Run ID: azeSPAM1BNS
Participant: uparis-sud.aze
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: Naive Bayes, stop word list, bigrams.

azeSPAM1res¶

Run ID: azeSPAM1res
Participant: uparis-sud.aze
Track: Spam
Year: 2005
Submission: 9/8/2005
Task: run

azeSPAM2CSS¶

Run ID: azeSPAM2CSS
Participant: uparis-sud.aze
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: Naive Bayes, ChiSq, stop word list, bigrams.

azeSPAMpf01¶

Run ID: azeSPAMpf01
Participant: uparis-sud.aze
Track: Spam
Year: 2005
Submission: 7/5/2005
Type: pilot
Task: filter
Run description: Training offline on spamassassin corpus. Statistical algorithm that doesn't work ).

cao1knf¶

Run ID: cao1knf
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 7/29/2005
Task: run

cao2ruk¶

Run ID: cao2ruk
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 7/29/2005
Task: run

cao3knf¶

Run ID: cao3knf
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 7/29/2005
Task: run

cao4hyb¶

Run ID: cao4hyb
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 7/29/2005
Task: run

crmSPAM1full¶

Run ID: crmSPAM1full
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM1ham2¶

Run ID: crmSPAM1ham2
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM1ham5¶

Run ID: crmSPAM1ham5
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM1osf¶

Run ID: crmSPAM1osf
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: CRM114 - variant OSBF The OSBF algorithm is a typical Bayesian classifier, but with the OSB (Orthogonal Sparse Bigrams) feature extraction technique (see http //www.siefkes.net/papers/winnow-spam.pdf) as its front-end and an intuitively derived Confidence Factor, a.k.a. "voodoo", for noise reduction and greater accuracy. This configuration uses no pre-trained info and the messages are not preprocessed in any way, not even mimedecoded.

crmSPAM1spm2¶

Run ID: crmSPAM1spm2
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM1spm5¶

Run ID: crmSPAM1spm5
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM2full¶

Run ID: crmSPAM2full
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM2ham2¶

Run ID: crmSPAM2ham2
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM2ham5¶

Run ID: crmSPAM2ham5
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM2spm2¶

Run ID: crmSPAM2spm2
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM2spm5¶

Run ID: crmSPAM2spm5
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM2win¶

Run ID: crmSPAM2win
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: This filter variation combines the OSB (orthogonal sparse bigrams) feature combination technique with the Winnow algorithm developed by Nick Littlestone. See C. Siefkes, F. Assis, S. Chhabra, and W. Yerazunis "Combining Winnow and Orthogonal Sparse Bigrams for Incremental Spam Filtering", PKDD 2004, http //www.siefkes.net/papers/winnow-spam.pdf, for a detailed description. The classifier has not been pre-trained and input mails are not preprocessed in any way, not even mimedecode.

crmSPAM3full¶

Run ID: crmSPAM3full
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM3ham2¶

Run ID: crmSPAM3ham2
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM3ham5¶

Run ID: crmSPAM3ham5
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM3osu¶

Run ID: crmSPAM3osu
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: CRM114 - variant OSB Unique The OSB Unique algorithm is a typical Bayesian classifier. It is a variant of the OSB option, also available in CRM114, which uses the OSB (Orthogonal Sparse Bigrams) feature extraction technique (see http//www.siefkes.net/papers/winnow-spam.pdf) but with the restriction that features are considered only once, no matter how many times they appear in a document. This configuration uses no pre-trained info and the messages are not preprocessed in any way, not even mimedecoded.

crmSPAM3spm2¶

Run ID: crmSPAM3spm2
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM3spm5¶

Run ID: crmSPAM3spm5
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM4full¶

Run ID: crmSPAM4full
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM4ham2¶

Run ID: crmSPAM4ham2
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM4ham5¶

Run ID: crmSPAM4ham5
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM4osb¶

Run ID: crmSPAM4osb
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: CRM114 - variant OSB The OSB algorithm is a typical Bayesian classifier, but uses OSB (Orthogonal Sparse Bigrams), a feature extraction technique derived from SBPH (Sparse Binary Polynomial Hashing) which reduces the number of features produced, keeping the same accuracy - see http//www.siefkes.net/papers/winnow-spam.pdf. This configuration uses no pre-trained info and the messages are not preprocessed in any way, not even mimedecoded.

crmSPAM4spm2¶

Run ID: crmSPAM4spm2
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAM4spm5¶

Run ID: crmSPAM4spm5
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 9/5/2005
Task: run

crmSPAMp1of¶

Run ID: crmSPAMp1of
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 7/5/2005
Type: pilot
Task: filter
Run description: The OSBF algorithm is a typical Bayesian classifier, but with the OSB (Orthogonal Sparse Bigrams) feature extraction technique (see http //www.siefkes.net/papers/winnow-spam.pdf) as its front-end and an intuitively derived Confidence Factor, a.k.a. "voodoo", for noise reduction and greater accuracy. This configuration uses no pre-trained info and the messages are not preprocessed in any way, not even mimedecoded.

crmSPAMp2wi¶

Run ID: crmSPAMp2wi
Participant: merl.yerazunis
Track: Spam
Year: 2005
Submission: 7/5/2005
Type: pilot
Task: filter
Run description: This filter variation combines the OSB (orthogonal sparse bigrams) feature combination technique with the Winnow algorithm developed by Nick Littlestone. See C. Siefkes, F. Assis, S. Chhabra, and W. Yerazunis "Combining Winnow and Orthogonal Sparse Bigrams for Incremental Spam Filtering", PKDD 2004, http //www.siefkes.net/papers/winnow-spam.pdf, for a detailed description. The classifier has not been pre-trained and input mails are not preprocessing in any way.

dalSPAM1fsw¶

Run ID: dalSPAM1fsw
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: limited word based naive bayes with white list

dalSPAM1sw1¶

Run ID: dalSPAM1sw1
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 9/9/2005
Task: run

dalSPAM1sx2¶

Run ID: dalSPAM1sx2
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 9/13/2005
Task: run

dalSPAM1sxw¶

Run ID: dalSPAM1sxw
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 7/6/2005
Type: pilot
Task: filter
Run description: used partial spamassassin corpus as training

dalSPAM2f¶

Run ID: dalSPAM2f
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 9/7/2005
Task: run

dalSPAM2h25¶

Run ID: dalSPAM2h25
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 9/6/2005
Task: run

dalSPAM2h50¶

Run ID: dalSPAM2h50
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 9/7/2005
Task: run

DalSPAM2n4¶

Run ID: DalSPAM2n4
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: Using CNG method with profile aging; byte n-grams (n=4)

dalSPAM2s25¶

Run ID: dalSPAM2s25
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 9/8/2005
Task: run

dalSPAM2s50¶

Run ID: dalSPAM2s50
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 9/8/2005
Task: run

dalSPAM2vla¶

Run ID: dalSPAM2vla
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 7/6/2005
Type: pilot
Task: filter
Run description: Later... Perl, ngram-based etc.

dalSPAM3f¶

Run ID: dalSPAM3f
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 9/9/2005
Task: run

dalSPAM3h25¶

Run ID: dalSPAM3h25
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 9/9/2005
Task: run

dalSPAM3h50¶

Run ID: dalSPAM3h50
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 9/11/2005
Task: run

DalSPAM3n5¶

Run ID: DalSPAM3n5
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: Using CNG method with profile aging; byte n-grams (n=5)

dalSPAM3s25¶

Run ID: dalSPAM3s25
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 9/11/2005
Task: run

dalSPAM3s50¶

Run ID: dalSPAM3s50
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 9/11/2005
Task: run

dalSPAM3vla¶

Run ID: dalSPAM3vla
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 7/6/2005
Type: pilot
Task: filter
Run description: Later... Perl, ngram-based etc.

dalSPAM4fsw¶

Run ID: dalSPAM4fsw
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: limited ngram based naive bayes with white list

dalSPAM4sw1¶

Run ID: dalSPAM4sw1
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 9/9/2005
Task: run

dalSPAM4sx2¶

Run ID: dalSPAM4sx2
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 9/13/2005
Task: run

dalSPAM4sxw¶

Run ID: dalSPAM4sxw
Participant: dalhousieu.keselj
Track: Spam
Year: 2005
Submission: 7/6/2005
Type: pilot
Task: filter
Run description: used spamassassin corpus as training

ICTSPAM1WNB¶

Run ID: ICTSPAM1WNB
Participant: cas-ict.wang
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: An winnow algorithms,using structure-based hierarchical model.

ICTSPAM2WNH¶

Run ID: ICTSPAM2WNH
Participant: cas-ict.wang
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: An winnow algorithms,using multi-feature model.

ICTSPAM3NBH¶

Run ID: ICTSPAM3NBH
Participant: cas-ict.wang
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: An nBayes algorithms,using structure-based hierarchical model.

ICTSPAM4NBB¶

Run ID: ICTSPAM4NBB
Participant: cas-ict.wang
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: An nBayes algorithms,using multi-feature model.

ICTSPAMpWNB¶

Run ID: ICTSPAMpWNB
Participant: cas-ict.wang
Track: Spam
Year: 2005
Submission: 7/4/2005
Type: pilot
Task: filter
Run description: An hierarchical filter,and use winnow as the classifier.

ijsSPAM1full¶

Run ID: ijsSPAM1full
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

ijsSPAM1h25¶

Run ID: ijsSPAM1h25
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 9/13/2005
Task: run

ijsSPAM1pm1¶

Run ID: ijsSPAM1pm1
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 7/5/2005
Type: pilot
Task: filter
Run description: The filter uses a character-based algorithm, which attempts to evaluate the probability that a message is spam based on the observed sequence of characters. A markov model is used, i.e. the next character is estimated depending on D preceding characters (D is set to 8 in this case). The Prediction by Partial Matching algorithm is used for prediction, with escape method D and exclusion. An adaptive model is used, so that the classifier continues to update the model as it evaluates an unknown message from beginning to end. Preprocessing is minimal. It includes MIME decoding (non-text parts are excluded), whitespace "normalization" (all sequences of whitespace are truncated to a single space) and minimal preprocessing of headers (e.g. dates are removed). The model is not pre-trained.

ijsSPAM1pm8¶

Run ID: ijsSPAM1pm8
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: This filter models email messages with a character-level markov model. The Prediction by Parital Mathing algorithm is used for this purpose. Two models are trained and the model that best approximates the message being classified determines the classification outcome. An order-8 model is used. The model is adaptive, i.e. it is updated as the message is being processed. No particular preprocessing is used, except MIME decoding and exclusion of embedded attachments. The filter is not pre-trained (the model is empty to begin with).

ijsSPAM2cw1¶

Run ID: ijsSPAM2cw1
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 7/5/2005
Type: pilot
Task: filter
Run description: The filter uses a character-based algorithm, which attempts to evaluate the probability that a message is spam based on the observed sequence of characters. A markov model is used, i.e. the next character is estimated depending on D preceding characters (D is set to 8 in this case). The Context Tree Weighting algorithm is used for prediction, with "custom" escape and exclusion mechanisms. An adaptive model is used, so that the classifier continues to update the model as it evaluates an unknown message from beginning to end. Preprocessing is minimal. It includes MIME decoding (non-text parts are excluded), whitespace "normalization" (all sequences of whitespace are truncated to a single space) and minimal preprocessing of headers (e.g. dates are removed). The model is not pre-trained.

ijsSPAM2full¶

Run ID: ijsSPAM2full
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

ijsSPAM2h25¶

Run ID: ijsSPAM2h25
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 9/13/2005
Task: run

ijsSPAM2pm6¶

Run ID: ijsSPAM2pm6
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: This filter models email messages with a character-level markov model. The Prediction by Parital Mathing algorithm is used for this purpose. Two models are trained and the model that best approximates the message being classified determines the classification outcome. An order-6 model is used. The model is adaptive, i.e. it is updated as the message is being processed. No particular preprocessing is used, except MIME decoding and exclusion of embedded attachments. The filter is not pre-trained (the model is empty to begin with).

ijsSPAM3cw2¶

Run ID: ijsSPAM3cw2
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 7/5/2005
Type: pilot
Task: filter
Run description: The filter uses a character-based algorithm, which attempts to evaluate the probability that a message is spam based on the observed sequence of characters. A markov model is used, i.e. the next character is estimated depending on D preceding characters (D is set to 8 in this case). The Context Tree Weighting algorithm is used for prediction, with "custom" escape and exclusion mechanisms. An adaptive model is used, so that the classifier continues to update the model as it evaluates an unknown message from beginning to end. A special mechanism for handling case is employed in this version with additional "control" characters that are encoded to signal the case letters Preprocessing is minimal. It includes MIME decoding (non-text parts are excluded), whitespace "normalization" (all sequences of whitespace are truncated to a single space) and minimal preprocessing of headers (e.g. dates are removed). The model is not pre-trained.

ijsSPAM3full¶

Run ID: ijsSPAM3full
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

ijsSPAM3h25¶

Run ID: ijsSPAM3h25
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 9/13/2005
Task: run

ijsSPAM3pm2¶

Run ID: ijsSPAM3pm2
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 7/5/2005
Type: pilot
Task: filter
Run description: The filter uses a character-based algorithm, which attempts to evaluate the probability that a message is spam based on the observed sequence of characters. A markov model is used, i.e. the next character is estimated depending on D preceding characters (D is set to 8 in this case). The Prediction by Partial Matching algorithm is used for prediction, escape method D and exclusion. An adaptive model is used, so that the classifier continues to update the model as it evaluates an unknown message from beginning to end. A special mechanism for handling case is employed in this version with additional "control" characters that are encoded to signal the case letters Preprocessing is minimal. It includes MIME decoding (non-text parts are excluded), whitespace "normalization" (all sequences of whitespace are truncated to a single space) and minimal preprocessing of headers (e.g. dates are removed). The model is not pre-trained.

ijsSPAM3pm4¶

Run ID: ijsSPAM3pm4
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: This filter models email messages with a character-level markov model. The Prediction by Parital Mathing algorithm is used for this purpose. Two models are trained and the model that best approximates the message being classified determines the classification outcome. An order-4 model is used. The model is adaptive, i.e. it is updated as the message is being processed. No particular preprocessing is used, except MIME decoding and exclusion of embedded attachments. The filter is not pre-trained (the model is empty to begin with).

ijsSPAM4cw2¶

Run ID: ijsSPAM4cw2
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 7/5/2005
Type: pilot
Task: filter
Run description: The filter uses a character-based algorithm, which attempts to evaluate the probability that a message is spam based on the observed sequence of characters. A markov model is used, i.e. the next character is estimated depending on D preceding characters (D is set to 8 in this case). The Context Tree Weighting algorithm is used for prediction, with "custom" escape and exclusion mechanisms. An adaptive model is used, so that the classifier continues to update the model as it evaluates an unknown message from beginning to end. A special mechanism for handling case is employed in this version with additional "control" characters that are encoded to signal the case letters Preprocessing is minimal. It includes MIME decoding (non-text parts are excluded), whitespace "normalization" (all sequences of whitespace are truncated to a single space) and minimal preprocessing of headers (e.g. dates are removed). The model is not pre-trained.

ijsSPAM4cw8¶

Run ID: ijsSPAM4cw8
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: This filter models email messages with a character-level markov model. The Context Tree Weighting algorithm is used for this purpose. Two models are trained and the model that best approximates the message being classified determines the classification outcome. An order-8 model is used. The model is adaptive, i.e. it is updated as the message is being processed. No particular preprocessing is used, except MIME decoding and exclusion of embedded attachments. The filter is not pre-trained (the model is empty to begin with).

ijsSPAM4full¶

Run ID: ijsSPAM4full
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

ijsSPAM4h25¶

Run ID: ijsSPAM4h25
Participant: jozef-stefan-inst.bratko
Track: Spam
Year: 2005
Submission: 9/13/2005
Task: run

indSPAM1f4f¶

Run ID: indSPAM1f4f
Participant: indianau.yang
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: Fusion filter (Naive Bayes, Rule-based, Pattern-based, Blacklist) with large training data

indSPAM2f42¶

Run ID: indSPAM2f42
Participant: indianau.yang
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: Fusion filter (Naive Bayes, Rule-based, Pattern-based, Blacklist) with training after each classification

indSPAM3f40¶

Run ID: indSPAM3f40
Participant: indianau.yang
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: Fusion filter (Naive Bayes, Rule-based, Pattern-based, Blacklist) with training only after misclassification

indSPAM4pf4¶

Run ID: indSPAM4pf4
Participant: indianau.yang
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: Prototype Fusion filter (Naive Bayes, Rule-based, Pattern-based, Blacklist) with large training data

indSPAMpwf4¶

Run ID: indSPAMpwf4
Participant: indianau.yang
Track: Spam
Year: 2005
Submission: 7/4/2005
Type: pilot
Task: filter
Run description: Fusion filter that combines Bayesian filter, rule-based filter, pattern-based filter, and blacklist filter

kidSPAM1BAS¶

Run ID: kidSPAM1BAS
Participant: beijingu.guo
Track: Spam
Year: 2005
Submission: 7/27/2005
Type: final
Task: filter
Run description: KidultPRIS is a command line program which is exploited by the members in the lab of Pattern Recognition and Intelligent System, Beijing University of Posts and Telecommunications.

kidSPAM2V2B¶

Run ID: kidSPAM2V2B
Participant: beijingu.guo
Track: Spam
Year: 2005
Submission: 7/27/2005
Type: final
Task: filter
Run description: KidultPRIS is a command line program which is exploited by the members in the lab of Pattern Recognition and Intelligent System, Beijing University of Posts and Telecommunications.

kidSPAM3V2E¶

Run ID: kidSPAM3V2E
Participant: beijingu.guo
Track: Spam
Year: 2005
Submission: 7/27/2005
Type: final
Task: filter
Run description: KidultPRIS is a command line program which is exploited by the members in the lab of Pattern Recognition and Intelligent System, Beijing University of Posts and Telecommunications.

kidSPAM4V5B¶

Run ID: kidSPAM4V5B
Participant: beijingu.guo
Track: Spam
Year: 2005
Submission: 7/27/2005
Type: final
Task: filter
Run description: KidultPRIS is a command line program which is exploited by the members in the lab of Pattern Recognition and Intelligent System, Beijing University of Posts and Telecommunications.

kidSPAMpBAS¶

Run ID: kidSPAMpBAS
Participant: beijingu.guo
Track: Spam
Year: 2005
Submission: 6/30/2005
Type: pilot
Task: filter
Run description: KidultPRIS is a command line program which is exploited by the members in the lab of Pattern Recognition and Intelligent System, Beijing University of Posts and Telecommunications.

No2¶

Results | Participants | Summary | Appendix

Run ID: No2
Participant: cas.nlpr.jzhao
Track: Spam
Year: 2005
Submission: 7/5/2005
Type: pilot
Task: filter
Run description: This is just a pilot system. Based on ifile system developed by Jason Rennie, we incorporate other machine learning methods in this system. please see the incoming final filter system and its Readme file.

None¶

Results | Participants | Summary | Appendix

Run ID: None
Participant: cas.nlpr.jzhao
Track: Spam
Year: 2005
Submission: 7/5/2005
Task: run

p1cefhuj¶

Run ID: p1cefhuj
Participant: breyer.laird
Track: Spam
Year: 2005
Submission: 7/3/2005
Type: pilot
Task: filter
Run description: This filter tests the cef tokenizer with full standard header analysis and uniform reference measure, case sensitive tokens.

p2adphu¶

Run ID: p2adphu
Participant: breyer.laird
Track: Spam
Year: 2005
Submission: 7/3/2005
Type: pilot
Task: filter
Run description: This filter tests the adp tokenizer with full standard header analysis and uniform reference measure, lowercase tokens.

p3adphd¶

Run ID: p3adphd
Participant: breyer.laird
Track: Spam
Year: 2005
Submission: 7/3/2005
Type: pilot
Task: filter
Run description: This filter tests the adp tokenizer with full standard header analysis and dirichlet reference measure, lowercase tokens.

p4adp¶

Run ID: p4adp
Participant: breyer.laird
Track: Spam
Year: 2005
Submission: 7/3/2005
Type: pilot
Task: filter
Run description: This filter tests the adp tokenizer with only Subject analysis, uniform reference measure, lowercase tokens.

pub1full¶

Run ID: pub1full
Participant: breyer.laird
Track: Spam
Year: 2005
Submission: 9/6/2005
Task: run

pub1ham50¶

Run ID: pub1ham50
Participant: breyer.laird
Track: Spam
Year: 2005
Submission: 9/6/2005
Task: run

pub1spam50¶

Run ID: pub1spam50
Participant: breyer.laird
Track: Spam
Year: 2005
Submission: 9/6/2005
Task: run

pub2full¶

Run ID: pub2full
Participant: breyer.laird
Track: Spam
Year: 2005
Submission: 9/6/2005
Task: run

pub3full¶

Run ID: pub3full
Participant: breyer.laird
Track: Spam
Year: 2005
Submission: 9/7/2005
Task: run

pub4full¶

Run ID: pub4full
Participant: breyer.laird
Track: Spam
Year: 2005
Submission: 9/6/2005
Task: run

PUC¶

Run ID: PUC
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 7/5/2005
Type: pilot
Task: filter
Run description: It uses a unigram language modelling approach with laplace smoothing. It computes the likelihood, given two lnguage models (one for ham, one for spam) and then the odds of generating the message on each, deciding which class based on the odds. It uses as prior training 2500 spam messages (collected from the web) for the spam language model and 1400 aquaint documents as the ham language model. The backgroung model is interpolated with new models created on-the-fly and updated on each new message (train). The current configuration will generate 98 false negatives and 45 false positives in the spamassassing corpus.

pucrs0¶

Run ID: pucrs0
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: It is based on unigram language models, one for ham messages and another for ham messages. It computes the message likelihood given these two models and uses the values to classify the message. Uses spamassassin corpus as training data.

pucrs0full¶

Run ID: pucrs0full
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 8/23/2005
Task: run

pucrs0ham25¶

Run ID: pucrs0ham25
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 8/23/2005
Task: run

pucrs0ham50¶

Run ID: pucrs0ham50
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 8/23/2005
Task: run

pucrs0spam25¶

Run ID: pucrs0spam25
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 8/23/2005
Task: run

pucrs0spam50¶

Run ID: pucrs0spam50
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 8/23/2005
Task: run

pucrs1¶

Run ID: pucrs1
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: It is based on unigram language models, one for ham messages and another for ham messages. It computes the message likelihood given these two models and uses the values to classify the message.

pucrs1full¶

Run ID: pucrs1full
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 8/23/2005
Task: run

pucrs1ham25¶

Run ID: pucrs1ham25
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 8/23/2005
Task: run

pucrs1ham50¶

Run ID: pucrs1ham50
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 8/23/2005
Task: run

pucrs1spam25¶

Run ID: pucrs1spam25
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 8/23/2005
Task: run

pucrs1spam50¶

Run ID: pucrs1spam50
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 8/23/2005
Task: run

pucrs2¶

Run ID: pucrs2
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: It is based on unigram language models, one for ham messages and another for ham messages. It computes the message likelihood given these two models and uses the values to classify the message.

pucrs2full¶

Run ID: pucrs2full
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 8/23/2005
Task: run

pucrs2ham25¶

Run ID: pucrs2ham25
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 8/23/2005
Task: run

pucrs2ham50¶

Run ID: pucrs2ham50
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 8/23/2005
Task: run

pucrs2spam25¶

Run ID: pucrs2spam25
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 8/23/2005
Task: run

pucrs2spam50¶

Run ID: pucrs2spam50
Participant: puc-rs.terra
Track: Spam
Year: 2005
Submission: 8/23/2005
Task: run

tamSPAM1dte¶

Run ID: tamSPAM1dte
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: No prior training. This is SpamBayes 1.1 (see http //spambayes.org for details, including the source), using all default options, and doing train-on-everything (with both cutoffs set to 0.6). SpamBayes is a standard statistic classifier, using the now-common Fishers/chi-squared combining technique, and various tokenization rules.

tamSPAM1full¶

Run ID: tamSPAM1full
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 9/10/2005
Task: run

tamSPAM1h25¶

Run ID: tamSPAM1h25
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 9/10/2005
Task: run

tamSPAM1h50¶

Run ID: tamSPAM1h50
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 9/10/2005
Task: run

tamSPAM1s25¶

Run ID: tamSPAM1s25
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 9/10/2005
Task: run

tamSPAM1s50¶

Run ID: tamSPAM1s50
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 9/10/2005
Task: run

tamSPAM2ber¶

Run ID: tamSPAM2ber
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: This is SpamBayes 1.1 (see http //spambayes.org for full details, including the source), with the use_bigrams option enabled, and doing train-on-error. Both cutoffs are set to 0.6. The use_bigrams option does both unigram and bigram tokenization, and uses a windowing technique to avoid using tokens twice (i.e. once in the unigram and once in the bigram). SpamBayes is a standard statistic classifier, with various tokenization methods, and using the now-common Fisher/chi-squared combination method.

tamSPAM2full¶

Run ID: tamSPAM2full
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 9/10/2005
Task: run

tamSPAM2h25¶

Run ID: tamSPAM2h25
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 9/10/2005
Task: run

tamSPAM2h50¶

Run ID: tamSPAM2h50
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 9/10/2005
Task: run

tamSPAM2s25¶

Run ID: tamSPAM2s25
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 9/10/2005
Task: run

tamSPAM2s50¶

Run ID: tamSPAM2s50
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 9/10/2005
Task: run

tamSPAM3bex¶

Run ID: tamSPAM3bex
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: SpamBayes 1.1 (see http //spambayes.org for full details, including source), with the use_bigrams option enabled (see description for run #2) and using train-to-exhaustion. Train-to-exhaustion (see Gary Robinson's blog for full details) repeatedly trains on the ham/spam corpus until all messages are correctly classified. None of the SpamBayes submissions have any prior training.

tamSPAM3full¶

Run ID: tamSPAM3full
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 9/10/2005
Task: run

tamSPAM3s25¶

Run ID: tamSPAM3s25
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 9/10/2005
Task: run

tamSPAM4aex¶

Run ID: tamSPAM4aex
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 7/28/2005
Type: final
Task: filter
Run description: SpamBayes 1.1 (see http //spambayes.org for full details, including source), with all possible tokenization options turned on and using train-to-exhaustion (se description for run #3). See previous run descriptions for more details.

tamSPAMpplt¶

Run ID: tamSPAMpplt
Participant: masseyu.meyer
Track: Spam
Year: 2005
Submission: 6/30/2005
Type: pilot
Task: filter
Run description: SpamBayes 1.1. Uses chi-squared combining (both thresholds set at 0.6 to eliminate unsure range) statistical method. No prior training.

yorSPAM1full¶

Run ID: yorSPAM1full
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM1ham2¶

Run ID: yorSPAM1ham2
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM1ham5¶

Run ID: yorSPAM1ham5
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM1knf¶

Run ID: yorSPAM1knf
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: Our pilot filter is based on the k-nearest neighbor classification method. We use a collection of public and private emails as an initial set of training data. Based on this training data set, an initial set of features are selected according to an information gain measure. The feature values are computed according to local term frequencies. Data preprocessing is conducted on both the training data and the test emails. We use a decoding program to parse the content of an email, which retains only the text part of an email including the decoded text body. The parsed email is then tokenized by first using white spaces as separators, then remove some non-alphanumerical characters with a few of them left. When tokenizing, we remove stop words and stem the left word if necessary. During classification, the training data set is maintained by removing examples that constantly cause wrong classifications and adding each test example into the training set. The weighted Euclidean distance measure is used to find the nearest neighbors. To be adaptive to the evolving of spam, we retrain our model periodically. The retained instances are represented start all over to reflect the updated features learnt.

yorSPAM1spa2¶

Run ID: yorSPAM1spa2
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM1spa5¶

Run ID: yorSPAM1spa5
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM2full¶

Run ID: yorSPAM2full
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM2ham2¶

Run ID: yorSPAM2ham2
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM2ham5¶

Run ID: yorSPAM2ham5
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM2ruk¶

Run ID: yorSPAM2ruk
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: Our filter reuses SpamAssassins code to do rule analysis, SA returns a preliminary result. The modified SA returns a parsed text to kNN which then do classify again. The final result is determined base on SA and kNN. In the training part, we use auto learning and then train on errors.

yorSPAM2spa2¶

Run ID: yorSPAM2spa2
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM2spa5¶

Run ID: yorSPAM2spa5
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM3full¶

Run ID: yorSPAM3full
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM3ham2¶

Run ID: yorSPAM3ham2
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM3ham5¶

Run ID: yorSPAM3ham5
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM3knf¶

Run ID: yorSPAM3knf
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: Our pilot filter is based on the k-nearest neighbor classification method. We use a collection of public and private emails as an initial set of training data. Based on this training data set, an initial set of features are selected according to an information gain measure. The feature values are computed according to local term frequencies. Data preprocessing is conducted on both the training data and the test emails. We use a decoding program to parse the content of an email, which retains only the text part of an email including the decoded text body. The parsed email is then tokenized by first using white spaces as separators, then remove some non-alphanumerical characters with a few of them left. When tokenizing, we remove stop words and stem the left word if necessary. During classification, the training data set is maintained by removing examples that constantly cause wrong classifications and adding each test example into the training set. The weighted Euclidean distance measure is used to find the nearest neighbors. To be adaptive to the evolving of spam, we retrain our model periodically. The retained instances are represented start all over to reflect the updated features learnt. This one is different from the first filter in pre-processing methods and retrain parameters.

yorSPAM3spa2¶

Run ID: yorSPAM3spa2
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM3spa5¶

Run ID: yorSPAM3spa5
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM4full¶

Run ID: yorSPAM4full
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM4ham2¶

Run ID: yorSPAM4ham2
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM4ham5¶

Run ID: yorSPAM4ham5
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM4hyb¶

Run ID: yorSPAM4hyb
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 7/29/2005
Type: final
Task: filter
Run description: This filter is a combination of kNN, NaveBayes and RandomForest tree learning.

yorSPAM4spa2¶

Run ID: yorSPAM4spa2
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAM4spa5¶

Run ID: yorSPAM4spa5
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 9/12/2005
Task: run

yorSPAMpknn¶

Run ID: yorSPAMpknn
Participant: yorku.huang
Track: Spam
Year: 2005
Submission: 7/5/2005
Type: pilot
Task: filter
Run description: Our pilot filter is based on the k-nearest neighbor classification method. We use a collection of public and private emails as an initial set of training data. Based on this training data set, an initial set of features are selected according to an information gain measure. The feature values are computed according to the TF-IDF measure. Data preprocessing is conducted on both the training data and the test emails. We use a decoding program to parse the content of an email, which retains only the text part of an email including the decoded text body. The parsed email is then tokenized by using non-alphabetical letters as separators. During classification, the training data set is maintained by removing examples that constantly cause wrong classifications and adding each test example into the training set. The Euclidean distance measure is used to find the nearest neighbors.