This report was created the using the behavseqanalyser , behaviour were grouped using the Berlin categorisation (date: 2018-06-29).
The data is grouped by treatment. Data transformation: data (%age of time spent doing the behavior) transformed using the square root method..
## first second
## 10 10
We grouped the variables following the Berlin argument to get 19 behavior categories. We used the folowing time windows and got 18 x 9 = 162 variables :
time_reference | windowstart | windowend | windowname |
---|---|---|---|
Bin | 0 | 120 | first 2 hours of recording |
Bintodark | -120 | 0 | last 2h before night |
Bintodark | 0 | 180 | first 3h of night |
Bintodark | 180 | 540 | middle night |
Bintodark | 540 | 720 | late night (3h) |
Bintodark | 720 | 864 | early day(3h) |
Bintodark | -120 | 864 | full recording |
lightcondition | DAY | NA | daytime |
lightcondition | NIGHT | NA | nighttime |
Note that the last window might be truncated if not all dataset is achieving 900 min after light on.
We then run a random forest to get the variables in order of importance to distinguish the groups. We then take the best 20 and run the random forest again (such that the Gini scores obtained will not depend on the initial number of variables). We plot here the table of variables ordered by weight:
Let’s take a teshold of importance (Gini > 0.95) and get all variables satisfying the filter, or at least 8 variables:
digforage7, digforage6, digforage1, Groom5, Drink4, ComeDown5, digforage2 and digforage8
First, lets plot the 2 most discriminative variables following the random forest:
Here, we plot the first two or threecomponents obtained after a ICA performed on the reduced data:
The PCA strategy shows that the behavior profile of the two groups of animal are not identical.
We performed a PCA on the data and tested whether the groups show a difference in their first component score using a Mann-Whitney or a Kruskal-Wallis rank sum test (if more than 2 groups exists). We plot here the first component in a boxplot:
NB: This strategy is pretty good against type I errors. On the other hand, it may well oversee existing differences.
## [1] "No machine learning attempt made."