Not a member of Pastebin yet?
Sign Up,
it unlocks many cool features!
- Features:
- ---------
- Basically all ml algorithms need to run on a specific set of data.
- The data has to be in the most convinent form for the machine to read.
- Its our job to supply this kind of data.
- These are commonly called as features. Supply values that you think the algo might find useful in classification.
- For example:
- [1] If you were dealing in audio, then amplitudes of selected frequencies can act as features.
- [2] For a movie dataset, the genere, length, bugget, actors, language etc are all features.
- Lets do for some sentiment analysis of audio:
- When Sad, you (probably) speak in a lower tone
- When Angry, you speak in a higher tone. Loudness is higher etc
- When Happy, you speak in a higher tone, loudness is lower
- For such a dataset you need these "features": [1] frequency, [2] rate of change of frequency, [3] Loudness etc
- This above is some dummy data, but bear with it
- -------------------------------------------------------------------------------------
- | Slno. | freq1 | freq2 | fre3 | ... | freq n | d_f1 | d_f2 | ...| df_n | loudness |
- -------------------------------------------------------------------------------------
- | tone1 | 3 | 2 | ...... | 0 | 4 | ...............| 9 |
- -------------------------------------------------------------------------------------
- | tone2 | 4 | 7 | ...... | 1 | 8 | ...............| 4 |
- -------------------------------------------------------------------------------------
- ....
- -------------------------------------------------------------------------------------
- | tonem | 4 | 5 | ...... | 1 | 8 | ...............| 9 |
- -------------------------------------------------------------------------------------
- I hope you get the idea.
- Label:
- ------
- Well if you need to classify a set of things, each possible output is called a label.
- Generally these are represented as vectors. If you had 3 outputs, most commonly you would choose your outputs as:
- Output 1 (Sad) : [1, 0, 0]
- Output 2 (angry): [0, 1, 0]
- Output 3 (happy): [0, 0, 1]
- So features are the inputs and labels are you output.
- The end result of whatever algorithm you use will produce an output like:
- [0.8, 0.1, 0.1] -> So this gets classified as sad
- [0.4, 0.6, 0.0] -> This is angry
- This values are analogous to output probability values, so 0.4 -> 40% chance for sad
- See: https://stackoverflow.com/questions/40898019/what-is-the-difference-between-a-feature-and-a-label
- Classifier:
- -----------
- This is the actual algorithm that converts you input vector space to output vector space.
- Nice big balckbox that does the conversion.
- Actually the math behind each of these classifiers is pretty intense and fun.
- Forest is one such classifier, I dont remember much else. The wikipedia link is extensive.
- raw inputs Feed to
- (audio) -> Features -> classifier -> Output vectors -> Human readable output
- | ... .
- | .... .
- | ... o
- | .. . oo
- |... oo
- | . oo
- |. oo
- ------------------
- Assume your features (unrelated to the sentiment analysis problem) plotted on a graph look like the above
- the "dots" are for one type of data and the "o" are for the other type.
- Depending on the algorithm that's used in the classifier, you a rule (polynomial equation) will be created
- that will determine what each future value is classified as.
- | ...| .
- | .... | .
- | ... | o
- | .. . | oo
- |... | oo
- | . | oo
- |. | oo
- ------------------
- Assume that your classifier comes with this line as the rule to decide if
- any future value is type '.' or type 'o'. Everything to the left of the line
- is '.' and to the right is 'o'.
- The end result of the training is this rule (or more commonly called model).
- You then use your test dataset to see how accurate your predictions were.
- Confusion matrix:
- -----------------
- So the first step is to train your classifier and the next step is to measure how accurate the predictions are.
- A confusion matrix is a convinent way to plot the results. It represents the predictied values vs the actual values.
- For example:
- ------------------
- | | . | o |
- ------------------
- | . | 90 | 10 |
- ------------------
- | o | 05 | 45 |
- ------------------
- Your Y axis represents the actual values and your X axis represents the predicted values.
- There are 100 (.) used in the test, of which your classifier predicted 90 of them correctly and 10 incorrectly.
- There are 50 (o) used in the test, of which your classifier predicted 45 of them correctly and 5 incorrectly.
- PS: Forest I dont remember
Advertisement
Add Comment
Please, Sign In to add comment