Conical Classification For Efficient One-Class Topic Determination

Abstract

As the Internet grows in size, so does the amount of text based information that exists. For many application spaces it is paramount to isolate and identify texts that relate to a particular topic. While one-class classification would be ideal for such analysis, there is a relative lack of research regarding efficient approaches with high predictive power. By noting that the range of documents we wish to identify can be represented as positive linear combinations of the Vector Space Model representing our text, we propose Conical classification, an approach that allows us to identify if a document is of a particular topic in a computationally efficient manner. We also propose Normal Exclusion, a modified version of Bi-Normal Separation that makes it more suitable within the one-class classification context. We show in our analysis that our approach not only has higher predictive power on our datasets, but is also faster to compute. Dataset Model Accuracy Balanced Accuracy Precision Recall F1 Score Time (s) ECommerce Linear OCSVM 0.900 ± 0.001 0.755 ± 0.003 0.978 ± 0.002 0.512 ± 0.007 0.672 ± 0.006 183.255 ± 3.821 RBF OCSVM 0.899 ± 0.001 0.753 ± 0.004 0.981 ± 0.003 0.508 ± 0.008 0.669 ± 0.007 201.339 ± 3.792 Sigmoid OCSVM 0.899 ± 0.002 0.752 ± 0.005 0.979 ± 0.003 0.508 ± 0.011 0.668 ± 0.009 193.402 ± 7.382 Poly OCSVM 0.885 ± 0.006 0.712 ± 0.002 0.995 ± 0.001 0.424 ± 0.003 0.595 ± 0.003 183.813 ± 2.639 IsolFor 0.199 ± 0.000 0.500 ± 0.000 0.199 ± 0.000 1.000 ± 0.000 0.333 ± 0.000 280.893 ± 12.543 OCRF 0.953 ± 0.000 0.947 ± 0.0176 0.800 ± 0.000 0.898 ± 0.036 0.889 ± 0.000 189.252 ± 244.041 PNB 0.638 ± 0.202 0.762 ± 0.092 0.786 ± 0.203 0.757 ± 0.305 0.687 ± 0.153 111.839 ± 59.581 GloVe CVDD 0.948 ± 0.017 0.906 ± 0.026 0.951 ± 0.015 0.983 ± 0.007 0.967 ± 0.011 188.462 ± 12.773 BERT CVDD 0.951 ± 0.128 0.910 ± 0.013 0.954 ± 0.017 0.987 ± 0.012 0.973 ± 0.028 233.626 ± 23.796 CC 0.988 ± 0.009 0.988 ± 0.007 0.956 ± 0.005 0.988 ± 0.009 0.971 ± 0.002 10.878 ± 0.036 FakeNews Linear OCSVM 0.818 ± 0.039 0.685 ± 0.082 0.819 ± 0.157 0.430 ± 0.232 0.495 ± 0.175 496.206 ± 5.353 RBF OCSVM 0.813 ± 0.033 0.706 ± 0.065 0.772 ± 0.184 0.500 ± 0.229 0.538 ± 0.086 533.529 ± 11.002 Sigmoid OCSVM 0.760 ± 0.046 0.648 ± 0.091 0.757 ± 0.259 0.436 ± 0.347 0.389 ± 0.166 514.913 ± 19.340 Poly OCSVM 0.850 ± 0.014 0.722 ± 0.053 0.846 ± 0.071 0.476 ± 0.128 0.592 ± 0.091 470.485 ± 2.950 IsolFor 0.238 ± 0.000 0.500 ± 0.000 0.238 ± 0.000 1.000 ± 0.000 0.384 ± 0.000 278.308 ± 5.449 OCRF 0.930 ± 0.000 0.955 ± 0.000 0.761 ± 0.000 1.000 ± 0.000 0.864 ± 0.000 197.232 ± 255.352 PNB 0.624 ± 0.223 0.773 ± 0.097 0.900 ± 0.142 0.697 ± 0.283 0.729 ± 0.152 218.296 ± 38.515 GloVe CVDD 0.906 ± 0.003 0.881 ± 0.011 0.938 ± 0.018 0.936 ± 0.022 0.936 ± 0.002 282.293 ± 31.88 BERT CVDD 0.899 ± 0.121 0.878 ± 0.218 0.923 ± 0.342 0.933 ± 0.231 0.927 ± 0.153 322.513 ± 49.659 CC 0.985 ± 0.005 0.985 ± 0.004 0.955 ± 0.002 0.987 ± 0.006 0.969 ± 0.001 13.952 ± 0.026 Jobs Linear OCSVM 0.913 ± 0.001 0.768 ± 0.004 0.971 ± 0.002 0.540 ± 0.009 0.694 ± 0.007 368.263 ± 6.227 RBF OCSVM 0.91...