def display_confmat_compare():
"""
This function compares the estimated tasks through assoc rules (and relating post
proc) to the manual annotations and displays that nicely as a confusion
matrix.
In what follows,
auto refers to automatic annoattion (with the assoc rules and relating post-proc)
hand refers to manual annotation
"""
data = ftools.readData( params.ANNOTATION_FILE )
data = post_pro.addDurationFeature( data ) # Just putting the data into a 'good' shape
titles = ftools.load(params.PATH_TITLE[6:]+params.NEW_DAT+params.TITLE_MAT)
data[params.WINDOW_STR] = [ftools.joinTitles(t) for t in titles]
auto = np.array(data[params.AUTO_TASK_STR])
hand = np.array(data[params.HAND_TASK_STR])
unique_tasks = np.unique(hand)
n_tasks = len(unique_tasks)
conf_mat = np.zeros((n_tasks, n_tasks))
for i, task1 in enumerate(unique_tasks):
for j, task2 in enumerate(unique_tasks):
conf_mat[j,i] = np.nansum(np.logical_and(auto == task1, hand == task2))
conf_mat_s = conf_mat #Save unscaled matrix
conf_mat = conf_mat / np.nansum(conf_mat, axis=1)[:,None]
corrects = np.sum([conf_mat_s[i,i] for i in range(len(conf_mat_s))])
print('Correct ', corrects, corrects/np.sum(conf_mat_s))
printTable(unique_tasks, conf_mat)
printTable(unique_tasks, conf_mat_s)
"""
classifier_tools.kNN(data)
classifier_tools.randomForest(data)
classifier_tools.svm(data)
"""
displayDurationDensities(data)
def example_q_quartiles():
"""
This function serves to display an example of computing Q-quartiles.
Specifically, it is for reproducing the graph given in Appendix-I (Pre-processing
of quantitaive variables). The below saves the data points into a file, which is
later used in a gnu-plot script.
"""
data = ftools.readData( params.ANNOTATION_FILE )
data = post_pro.addDurationFeature( data ) # Just putting the data into a 'good' shape
titles = ftools.load(params.PATH_TITLE[6:]+params.NEW_DAT+params.TITLE_MAT)
data[params.WINDOW_STR] = [ftools.joinTitles(t) for t in titles]
array = list(map(int, data[params.DURATION_STR]))
y, x = np.histogram(array, bins=np.arange(np.max(array), step=1)+1)
x = x[:-1]
y = np.cumsum(y)
f = open('cdf_duration_dev.txt','w')
for i in range(len(x)):
f.write(str(x[i])+'\t'+str(y[i])+'\n')
f.close()
params.TASK_MAT)
keystrokes_quan = ftools.load(params.PATH_KSTROKES +
params.DAT_FILE_PREFIX + params.KSTROKE_MAT)
lunch = ftools.load(params.PATH_LUNCH + params.DAT_FILE_PREFIX +
params.LUNCH_MAT)
l_clicks = ftools.load(params.PATH_CLICKS + params.NEW_DAT +
params.LCLICK_MAT)
r_clicks = ftools.load(params.PATH_CLICKS + params.NEW_DAT +
params.RCLICK_MAT)
duration = ftools.load(params.PATH_DURATION + params.NEW_DAT +
params.DURATION_MAT)
(exe_names, window_names, time_names,
level_of_assoc) = dtools.define_names()
title_combinations = np.unique([ftools.joinTitles(t) for t in windows])
title_codes = np.array([ftools.joinTitles(t) for t in windows])
n_actions = len(exes) # or any other matrix
"""
In est_by_rules_direct, I use use 1 column for line number, 4 columns for
estimations. But actually number of maximum estimations is no more than 3.
"""
est_by_rules_direct = []
for i in range(n_actions):
"""
Estimation by applying the rules directly
"""
(exe_only_rules, title_only_rules, exe_and_title_rules,
exe_and_keyst_rules, exe_and_lunch_rules) = rtools.define_rules()
if params.STAGE == 1:
"""
At stage-1, anything that is not Document or Test, needs to be renamed
as Others
"""
for p in [params.HAND_TASK_STR, params.EST_1_STR, \
params.EST_2_STR, params.EST_3_STR]:
data[p] = np.array(data[p])
data[p][np.invert(np.logical_or(data[p] == params.TEST, \
data[p] == params.DOCUMENT))] = params.OTHER
else:
"""
At stage-2, we filter out the rows that are labeled Document or Test
and consider only the others
"""
boolean_matrix = (data[params.HAND_TASK_STR] == np.array(
params.TASKS)[:, None])
query_array = boolean_matrix.any(axis=0) # Logical or between rows
for k in data.keys():
data[k] = np.array(data[k])[query_array]
titles = ftools.load(params.PATH_TITLE + params.DAT_FILE_PREFIX +
params.TITLE_MAT)
data[params.WINDOW_STR] = [ftools.joinTitles(t) for t in titles]
ctools.kNN(data, multi=False)
ctools.randomForest(data, multi=False)
ctools.svm(data, multi=False)
l_clicks = ftools.load(params.PATH_CLICKS + params.NEW_DAT +
params.LCLICK_MAT)
r_clicks = ftools.load(params.PATH_CLICKS + params.NEW_DAT +
params.RCLICK_MAT)
duration = ftools.load(params.PATH_DURATION + params.NEW_DAT +
params.DURATION_MAT)
exe_names, window_titles = dtools.define_names()[:2]
window_titles.insert(0, '') # for alien titles, i.e. not a known title
"""
One hot encoding for exes and window titles
"""
exe_codes = [exe_names.index(e) + 1
for e in exes] # + 1 because I don't want exes with ID code 0
title_combination = np.unique([ftools.joinTitles(t)
for t in titles]).tolist()
title_codes = [
title_combination.index(ftools.joinTitles(t)) for t in titles
]
descriptors = [exe_codes, title_codes, keystrokes_quan, \
l_clicks, r_clicks, duration] # eventually we omit lunch time info
descriptors_names = ['exes', 'titles', 'keystrokes', 'Left clicks', \
'Right clicks', 'Duration'] # Only for printing purposes
jointProbs = []
for i in range(len(descriptors) - 1):
for j in range(i + 1, len(descriptors)):
desc1 = descriptors[i]
desc2 = descriptors[j]
q1 = etools.compute_histogram(desc1)