1 00:00:00,430 --> 00:00:01,020 All right. 2 00:00:01,410 --> 00:00:07,230 So now that we've established that there is an inverse relationship between the precision school and 3 00:00:07,230 --> 00:00:16,470 recall we'll need a way to reconcile the two because ideally you want both high precision and high recall. 4 00:00:16,470 --> 00:00:22,350 The good news is that both of these metrics can be blended into a single value which is called the f 5 00:00:22,350 --> 00:00:25,590 score also known as the f one score. 6 00:00:26,580 --> 00:00:38,730 Here's the definition the f score is defined as two times precision times recall divided by precision 7 00:00:39,120 --> 00:00:47,590 plus recall so what this formula is telling us is that we want to create a balanced classifier. 8 00:00:47,670 --> 00:00:57,750 We try and maximize the f1 score let's head over to Jupiter notebook and calculate this I'll insert 9 00:00:57,810 --> 00:01:10,040 a quick section hitting him hold f hyphen school or F one school and here I'll calculate the value f 10 00:01:10,070 --> 00:01:10,920 1 underscore. 11 00:01:10,920 --> 00:01:21,840 Score is equal to well what do we say two times and then I'll be precision on a score score times recall 12 00:01:21,960 --> 00:01:33,240 on the score score all divided by precision on a score score plus recall underscore score and if we 13 00:01:33,240 --> 00:01:48,060 want to print this out c f score is and then curly braces colon C I don't know to dot format parentheses 14 00:01:48,570 --> 00:02:00,000 F one and a score score and before I hit shift enter I'll put a decimal point here and what we get is 15 00:02:00,390 --> 00:02:10,520 the f score is equal to zero point nine five which actually is very very good but how do we interpret 16 00:02:10,700 --> 00:02:13,840 the F school what does it mean well. 17 00:02:14,380 --> 00:02:21,280 So technically speaking the f score is the harmonic average the harmonic mean of precision and recall 18 00:02:22,180 --> 00:02:30,900 and because of this it takes both the false positives and the false negatives into account also it always 19 00:02:30,900 --> 00:02:32,480 has a value between 0 and 1. 20 00:02:32,670 --> 00:02:39,420 So 1 is the maximum value which is why zero point nine five is very very high and zero is the minimum 21 00:02:39,420 --> 00:02:48,720 value so in conclusion we've done a lot of evaluation on our need based classifier we've looked at a 22 00:02:48,720 --> 00:02:56,250 number of metrics precision recall f score and accuracy and we're doing actually very very very well 23 00:02:56,490 --> 00:03:04,020 because the number of false positives and false negatives is very low given the amount of data that 24 00:03:04,020 --> 00:03:06,090 we have now. 25 00:03:06,210 --> 00:03:09,270 Of course our spam filter isn't perfect right. 26 00:03:09,360 --> 00:03:14,910 I mean spam filters these days are actually getting really really good right because they don't just 27 00:03:15,000 --> 00:03:21,120 look at the words in the email body they actually look at a lot of other things as well write that a 28 00:03:21,120 --> 00:03:28,410 spammer might do for example a spammer might falsify the header or always send spam from like the same 29 00:03:28,410 --> 00:03:36,360 IP address which is why there are IP blacklists so truth be told I'm actually always really really impressed 30 00:03:36,390 --> 00:03:43,950 when a spammer manages to inbox in my gmail or in my hotmail account I'm almost always tempted to actually 31 00:03:43,950 --> 00:03:50,070 reply like Hey how did you manage to get around the spam filter and get into my inbox tell me your secrets 32 00:03:50,520 --> 00:03:58,470 but that's probably just going to net me a lot more spam but I think we've come very very far and you 33 00:03:58,470 --> 00:04:04,530 really deserve a big big pat on the back if you've completed all of this and made it through all of 34 00:04:04,530 --> 00:04:11,550 this and understood it this was a very very involved and a very very big project but we've touched on 35 00:04:11,760 --> 00:04:17,940 many many areas of machine learning and data science and we were able to learn a lot of these concepts 36 00:04:18,300 --> 00:04:25,200 as they came up as we needed them as part of this project so again well done and I'll see you in the 37 00:04:25,200 --> 00:04:27,870 next lesson as always take her.