def create_legend(self, img, x, y, dw, dh, x_start, x_end, y_range):
x_axis_type = 'linear' if self.transfer_function == 'linear' else 'log'
legend_fig = Figure(x_range=(x_start, x_end),
plot_height=max(dh, 50),
plot_width=self.plot_width,
lod_threshold=None,
toolbar_location=None,
y_range=y_range,
x_axis_type=x_axis_type)
legend_fig.min_border_top = 0
legend_fig.min_border_bottom = 10
legend_fig.min_border_left = 15
legend_fig.min_border_right = 15
legend_fig.yaxis.visible = False
legend_fig.grid.grid_line_alpha = 0
legend_fig.image_rgba(image=[img],
x=[x],
y=[y],
dw=[dw],
dh=[dh],
dw_units='screen')
return legend_fig
def create_ramp_legend(agg, cmap, how='linear', width=600):
'''
Helper function to create a Bokeh ``Figure`` object
with a color ramp corresponding to input aggregate and transfer function.
Parameters
----------
agg : xarray
Datashader aggregate object (e.g. result of Canvas.points())
cmap : list of colors or matplotlib.colors.Colormap, optional
The colormap to use. Can be either a list of colors (in any of the
formats described above), or a matplotlib colormap object.
how : str
Datashader transfer function name (either linear or log)
width : int
Width in pixels of resulting legend figure (default=600)
'''
vals_arr, min_val, max_val = summarize_aggregate_values(agg, how=how)
img = tf.shade(vals_arr, cmap=cmap, how=how)
x_axis_type = how
assert x_axis_type == 'linear' or x_axis_type == 'log'
legend_fig = Figure(x_range=(min_val, max_val),
plot_height=50,
plot_width=width,
lod_threshold=None,
toolbar_location=None,
y_range=(0, 18),
x_axis_type=x_axis_type)
legend_fig.min_border_top = 0
legend_fig.min_border_bottom = 10
legend_fig.min_border_left = 15
legend_fig.min_border_right = 15
legend_fig.yaxis.visible = False
legend_fig.grid.grid_line_alpha = 0
legend_fig.image_rgba(image=[img.values],
x=[min_val],
y=[0],
dw=[max_val - min_val],
dh=[18],
dw_units='screen')
return legend_fig
def create_ramp_legend(agg, cmap, how='linear', width=600):
'''
Helper function to create a Bokeh ``Figure`` object
with a color ramp corresponding to input aggregate and transfer function.
Parameters
----------
agg : xarray
Datashader aggregate object (e.g. result of Canvas.points())
cmap : list of colors or matplotlib.colors.Colormap, optional
The colormap to use. Can be either a list of colors (in any of the
formats described above), or a matplotlib colormap object.
how : str
Datashader transfer function name (either linear or log)
width : int
Width in pixels of resulting legend figure (default=600)
'''
vals_arr, min_val, max_val = summarize_aggregate_values(agg, how=how)
img = tf.shade(vals_arr, cmap=cmap, how=how)
x_axis_type = how
assert x_axis_type == 'linear' or x_axis_type == 'log'
legend_fig = Figure(x_range=(min_val, max_val),
plot_height=50,
plot_width=width,
lod_threshold=None,
toolbar_location=None,
y_range=(0, 18),
x_axis_type=x_axis_type)
legend_fig.min_border_top = 0
legend_fig.min_border_bottom = 10
legend_fig.min_border_left = 15
legend_fig.min_border_right = 15
legend_fig.yaxis.visible = False
legend_fig.grid.grid_line_alpha = 0
legend_fig.image_rgba(image=[img.values],
x=[min_val],
y=[0],
dw=[max_val - min_val],
dh=[18],
dw_units='screen')
return legend_fig
def _replacement_plot(self, source):
start = self.now - datetime.timedelta(days=self.RECOATING_PERIOD)
start_timestamp = datetime.datetime(start.year, start.month, start.day, 0, 0, 0, 0).timestamp()
end_timestamp = datetime.datetime(self.now.year, self.now.month, self.now.day, 0, 0, 0, 0).timestamp()
x_range = DataRange1d(start=1000 * start_timestamp, end=1000 * end_timestamp)
y_range = DataRange1d(start=0, end=120)
plot = Figure(x_axis_type='datetime',
x_range=x_range,
y_range=y_range,
tools=['tap'],
toolbar_location=None)
# required recoated segments
plot.line(x=[start, self.now],
y=[0, 91],
legend='required recoated segments',
line_width=2,
line_color='blue')
# recoated segments
plot.line(x='ReplacementDate',
y='RecoatingsSinceYearStart',
source=source,
legend='recoated segments',
line_width=2,
color='orange')
plot.circle(x='ReplacementDate',
y='RecoatingsSinceYearStart',
source=source,
legend='recoated segments',
size=6,
color='orange')
date_format = '%d %b %Y'
formats = dict(hours=[date_format], days=[date_format], months=[date_format], years=[date_format])
plot.xaxis[0].formatter = DatetimeTickFormatter(formats=formats)
plot.legend.location = 'top_left'
plot.min_border_top = 20
plot.min_border_bottom = 30
return plot
def make_plots(linesources, pointsources):
plots = []
i=0
for linesource, pointsource in zip(linesources, pointsources):
fig = Figure(title=None, toolbar_location=None, tools=[],
x_axis_type="datetime",
width=300, height=70)
fig.xaxis.visible = False
if i in [0, 9] :
fig.xaxis.visible = True
fig.height = 90
fig.yaxis.visible = False
fig.xgrid.visible = True
fig.ygrid.visible = False
fig.min_border_left = 10
fig.min_border_right = 10
fig.min_border_top = 5
fig.min_border_bottom = 5
if not i in [0, 9]:
fig.xaxis.major_label_text_font_size = "0pt"
#fig.yaxis.major_label_text_font_size = "0pt"
fig.xaxis.major_tick_line_color = None
fig.yaxis.major_tick_line_color = None
fig.xaxis.minor_tick_line_color = None
fig.yaxis.minor_tick_line_color = None
fig.background_fill_color = "whitesmoke"
fig.line(x='date', y="y", source=linesource)
fig.circle(x='date', y='y', size=5, source=pointsource)
fig.text(x='date', y='y', text='text', x_offset=5, y_offset=10, text_font_size='7pt', source=pointsource)
fig.title.align = 'left'
fig.title.text_font_style = 'normal'
plots.append(fig)
i+=1
return plots
ymax = 4979238.441
path = './data/projected.tif'
fig = Figure(x_range=(xmin, xmax),
y_range=(ymin, ymax),
plot_height=600,
plot_width=900,
tools='pan,wheel_zoom')
fig.background_fill_color = 'black'
fig.add_tile(STAMEN_TONER, alpha=0) # used to set axis ranges
fig.x_range.callback = CustomJS(code=dims_jscode, args=dict(plot=fig, dims=dims))
fig.y_range.callback = CustomJS(code=dims_jscode, args=dict(plot=fig, dims=dims))
fig.axis.visible = False
fig.grid.grid_line_alpha = 0
fig.min_border_left = 0
fig.min_border_right = 0
fig.min_border_top = 0
fig.min_border_bottom = 0
image_source = ColumnDataSource(dict(image=[], x=[], y=[], dw=[], dh=[]))
fig.image_rgba(source=image_source,
image='image',
x='x',
y='y',
dw='dw',
dh='dh',
dilate=False)
curdoc().add_root(fig)
y_range=(ymin, ymax),
plot_height=600,
plot_width=900,
tools='pan,wheel_zoom')
fig.background_fill_color = 'black'
fig.add_tile(get_provider("STAMEN_TONER"), alpha=.3)
fig.x_range.callback = CustomJS(code=dims_jscode,
args=dict(plot=fig, dims=dims))
fig.y_range.callback = CustomJS(code=dims_jscode,
args=dict(plot=fig, dims=dims))
fig.axis.visible = False
fig.grid.grid_line_alpha = 0
fig.min_border_left = 0
fig.min_border_right = 0
fig.min_border_top = 0
fig.min_border_bottom = 0
image_source = ColumnDataSource(dict(image=[], x=[], y=[], dw=[], dh=[]))
fig.image_rgba(source=image_source,
image='image',
x='x',
y='y',
dw='dw',
dh='dh',
dilate=False)
time_text = Paragraph(text='Time Period: 00:00 - 00:00')
controls = HBox(children=[time_text, time_select], width=fig.plot_width)
layout = VBox(children=[fig, controls])
curdoc().add_root(layout)
def main():
print('''Please select the CSV dataset you\'d like to use.
The dataset should contain these columns:
- metric to apply threshold to
- indicator of event to detect (e.g. malicious activity)
- Please label this as 1 or 0 (true or false);
This will not work otherwise!
''')
# Import the dataset
imported_data = None
while isinstance(imported_data, pd.DataFrame) == False:
file_path = input('Enter the path of your dataset: ')
imported_data = file_to_df(file_path)
time.sleep(1)
print(f'''\nGreat! Here is a preview of your data:
Imported fields:''')
# List headers by column index.
cols = list(imported_data.columns)
for index in range(len(cols)):
print(f'{index}: {cols[index]}')
print(f'Number of records: {len(imported_data.index)}\n')
# Preview the DataFrame
time.sleep(1)
print(imported_data.head(), '\n')
# Prompt for the metric and source of truth.
time.sleep(1)
metric_col, indicator_col = columns_picker(cols)
# User self-validation.
col_check = input('Can you confirm if this is correct? (y/n): ').lower()
# If it's wrong, let them try again
while col_check != 'y':
metric_col, indicator_col = columns_picker(cols)
col_check = input(
'Can you confirm if this is correct? (y/n): ').lower()
else:
print(
'''\nGreat! Thanks for your patience. Generating summary stats now..\n'''
)
# Generate summary stats.
time.sleep(1)
malicious, normal = classification_split(imported_data, metric_col,
indicator_col)
mal_mean = malicious.mean()
mal_stddev = malicious.std()
mal_count = malicious.size
mal_median = malicious.median()
norm_mean = normal.mean()
norm_stddev = normal.std()
norm_count = normal.size
norm_median = normal.median()
print(f'''Normal vs Malicious Summary (metric = {metric_col}):
Normal:
-----------------------------
Observations: {round(norm_count, 2)}
Average: {round(norm_mean, 2)}
Median: {round(norm_median, 2)}
Standard Deviation: {round(norm_stddev, 2)}
Malicious:
-----------------------------
Observations: {round(mal_count, 2)}
Average: {round(mal_mean, 2)}
Median: {round(mal_median, 2)}
Standard Deviation: {round(mal_stddev, 2)}
''')
# Insights and advisories
# Provide the accuracy metrics of a generic threshold at avg + 3 std deviations
generic_threshold = confusion_matrix(
malicious, normal, threshold_calc(norm_mean, norm_stddev, 3))
time.sleep(1)
print(
f'''A threshold at (average + 3x standard deviations) {metric_col} would result in:
- True Positives (correctly identified malicious events: {generic_threshold['TP']:,}
- False Positives (wrongly identified normal events: {generic_threshold['FP']:,}
- True Negatives (correctly identified normal events: {generic_threshold['TN']:,}
- False Negatives (wrongly identified malicious events: {generic_threshold['FN']:,}
Accuracy Metrics:
- Precision (what % of events above threshold are actually malicious): {round(generic_threshold['precision'] * 100, 1)}%
- Recall (what % of malicious events did we catch): {round(generic_threshold['recall'] * 100, 1)}%
- F1 Score (blends precision and recall): {round(generic_threshold['f1_score'] * 100, 1)}%'''
)
# Distribution skew check.
if norm_mean >= (norm_median * 1.1):
time.sleep(1)
print(
f'''\nYou may want to be cautious as your normal traffic\'s {metric_col}
has a long tail towards high values. The median is {round(norm_median, 2)}
compared to {round(norm_mean, 2)} for the average.''')
if mal_mean < threshold_calc(norm_mean, norm_stddev, 2):
time.sleep(1)
print(
f'''\nWarning: you may find it difficult to avoid false positives as the average
{metric_col} for malicious traffic is under the 95th percentile of the normal traffic.'''
)
# For fun/anticipation. Actually a nerd joke because of the method we'll be using.
if '-q' not in sys.argv[1:]:
time.sleep(1)
play_a_game.billy()
decision = input('yes/no: ').lower()
while decision != 'yes':
time.sleep(1)
print('...That\'s no fun...')
decision = input('Let\'s try that again: ').lower()
# Let's get to the simulations!
time.sleep(1)
print('''\nInstead of manually experimenting with threshold multipliers,
let\'s simulate a range of options and see what produces the best result.
This is similar to what is known as \"Monte Carlo simulation\".\n''')
# Initialize session name & create app folder if there isn't one.
time.sleep(1)
session_name = input('Please provide a name for this project/session: ')
session_folder = make_folder(session_name)
# Generate list of multipliers to iterate over.
time.sleep(1)
mult_start = float(
input(
'Please provide the minimum multiplier you want to start at. We recommend 2: '
))
# Set the max to how many std deviations away the sample max is.
mult_end = (imported_data[metric_col].max() - norm_mean) / norm_stddev
mult_interval = float(
input('Please provide the desired gap between multiplier options: '))
# range() only allows integers, let's manually populate a list
multipliers = []
mult_counter = mult_start
while mult_counter < mult_end:
multipliers.append(round(mult_counter, 2))
mult_counter += mult_interval
print('Generating simulations..\n')
# Run simulations using our multipliers.
simulations = monte_carlo(malicious, normal, norm_mean, norm_stddev,
multipliers)
print('Done!')
time.sleep(1)
# Save simulations as CSV for later use.
simulation_filepath = os.path.join(
session_folder, f'{session_name}_simulation_results.csv')
simulations.to_csv(simulation_filepath, index=False)
print(f'Saved results to: {simulation_filepath}')
# Find the first threshold with the highest F1 score.
# This provides a balanced approach between precision and recall.
f1_max = simulations[simulations.f1_score ==
simulations.f1_score.max()].head(1)
f1_max_mult = f1_max.squeeze()['multiplier']
time.sleep(1)
print(
f'''\nBased on the F1 score metric, setting a threshold at {round(f1_max_mult,1)} standard deviations
above the average magnitude might provide optimal results.\n''')
time.sleep(1)
print(f'''{f1_max}
We recommend that you skim the CSV and the following visualization outputs
to sanity check results and make your own judgement.
''')
# Now for the fun part..generating the visualizations via Bokeh.
# Header & internal CSS.
title_text = '''
<style>
@font-face {
font-family: RobotoBlack;
src: url(fonts/Roboto-Black.ttf);
font-weight: bold;
}
@font-face {
font-family: RobotoBold;
src: url(fonts/Roboto-Bold.ttf);
font-weight: bold;
}
@font-face {
font-family: RobotoRegular;
src: url(fonts/Roboto-Regular.ttf);
}
body {
background-color: #f2ebe6;
}
title_header {
font-size: 80px;
font-style: bold;
font-family: RobotoBlack, Helvetica;
font-weight: bold;
margin-bottom: -200px;
}
h1, h2, h3 {
font-family: RobotoBlack, Helvetica;
color: #313596;
}
p {
font-size: 12px;
font-family: RobotoRegular
}
b {
color: #58c491;
}
th, td {
text-align:left;
padding: 5px;
}
tr:nth-child(even) {
background-color: white;
opacity: .7;
}
.vertical {
border-left: 1px solid black;
height: 190px;
}
</style>
<title_header style="text-align:left; color: white;">
Cream.
</title_header>
<p style="font-family: RobotoBold, Helvetica;
font-size:18px;
margin-top: 0px;
margin-left: 5px;">
Because time is money, and <b style="font-size=18px;">"Cash Rules Everything Around Me"</b>.
</p>
</div>
'''
title_div = Div(text=title_text,
width=800,
height=160,
margin=(40, 0, 0, 70))
# Summary stats from earlier.
summary_text = f'''
<h1>Results Overview</h1>
<i>metric = magnitude</i>
<table style="width:100%">
<tr>
<th>Metric</th>
<th>Normal Events</th>
<th>Malicious Events</th>
</tr>
<tr>
<td>Observations</td>
<td>{norm_count:,}</td>
<td>{mal_count:,}</td>
</tr>
<tr>
<td>Average</td>
<td>{round(norm_mean, 2):,}</td>
<td>{round(mal_mean, 2):,}</td>
</tr>
<tr>
<td>Median</td>
<td>{round(norm_median, 2):,}</td>
<td>{round(mal_median, 2):,}</td>
</tr>
<tr>
<td>Standard Deviation</td>
<td>{round(norm_stddev, 2):,}</td>
<td>{round(mal_stddev, 2):,}</td>
</tr>
</table>
'''
summary_div = Div(text=summary_text,
width=470,
height=320,
margin=(3, 0, -70, 73))
# Results of the hypothetical threshold.
hypothetical = f'''
<h1>"Rule of thumb" Hypothetical Threshold</h1>
<p>A threshold at <i>(average + 3x standard deviations)</i> {metric_col} would result in:</p>
<ul>
<li>True Positives (correctly identified malicious events:
<b>{generic_threshold['TP']:,}</b></li>
<li>False Positives (wrongly identified normal events:
<b>{generic_threshold['FP']:,}</b></li>
<li>True Negatives (correctly identified normal events:
<b>{generic_threshold['TN']:,}</b></li>
<li>False Negatives (wrongly identified malicious events:
<b>{generic_threshold['FN']:,}</b></li>
</ul>
<h2>Accuracy Metrics</h2>
<ul>
<li>Precision (what % of events above threshold are actually malicious):
<b>{round(generic_threshold['precision'] * 100, 1)}%</b></li>
<li>Recall (what % of malicious events did we catch):
<b>{round(generic_threshold['recall'] * 100, 1)}%</b></li>
<li>F1 Score (blends precision and recall):
<b>{round(generic_threshold['f1_score'] * 100, 1)}%</b></li>
</ul>
'''
hypo_div = Div(text=hypothetical,
width=600,
height=320,
margin=(5, 0, -70, 95))
line = '''
<div class="vertical"></div>
'''
vertical_line = Div(text=line,
width=20,
height=320,
margin=(80, 0, -70, -10))
# Let's get the exploratory charts generated.
malicious_hist, malicious_edge = np.histogram(malicious, bins=100)
mal_hist_df = pd.DataFrame({
'metric': malicious_hist,
'left': malicious_edge[:-1],
'right': malicious_edge[1:]
})
normal_hist, normal_edge = np.histogram(normal, bins=100)
norm_hist_df = pd.DataFrame({
'metric': normal_hist,
'left': normal_edge[:-1],
'right': normal_edge[1:]
})
exploratory = figure(
plot_width=plot_width,
plot_height=plot_height,
sizing_mode='fixed',
title=f'{metric_col.capitalize()} Distribution (σ = std dev)',
x_axis_label=f'{metric_col.capitalize()}',
y_axis_label='Observations')
exploratory.title.text_font_size = title_font_size
exploratory.border_fill_color = cell_bg_color
exploratory.border_fill_alpha = cell_bg_alpha
exploratory.background_fill_color = cell_bg_color
exploratory.background_fill_alpha = plot_bg_alpha
exploratory.min_border_left = left_border
exploratory.min_border_right = right_border
exploratory.min_border_top = top_border
exploratory.min_border_bottom = bottom_border
exploratory.quad(bottom=0,
top=mal_hist_df.metric,
left=mal_hist_df.left,
right=mal_hist_df.right,
legend_label='malicious',
fill_color=malicious_color,
alpha=.85,
line_alpha=.35,
line_width=.5)
exploratory.quad(bottom=0,
top=norm_hist_df.metric,
left=norm_hist_df.left,
right=norm_hist_df.right,
legend_label='normal',
fill_color=normal_color,
alpha=.35,
line_alpha=.35,
line_width=.5)
exploratory.add_layout(
Arrow(end=NormalHead(fill_color=malicious_color, size=10,
line_alpha=0),
line_color=malicious_color,
x_start=mal_mean,
y_start=mal_count,
x_end=mal_mean,
y_end=0))
arrow_label = Label(x=mal_mean,
y=mal_count,
y_offset=5,
text='Malicious Events',
text_font_style='bold',
text_color=malicious_color,
text_font_size='10pt')
exploratory.add_layout(arrow_label)
exploratory.xaxis.formatter = NumeralTickFormatter(format='0,0')
exploratory.yaxis.formatter = NumeralTickFormatter(format='0,0')
# 3 sigma reference line
sigma_ref(exploratory, norm_mean, norm_stddev)
exploratory.legend.location = "top_right"
exploratory.legend.background_fill_alpha = .3
# Zoomed in version
overlap_view = figure(
plot_width=plot_width,
plot_height=plot_height,
sizing_mode='fixed',
title=f'Overlap Highlight',
x_axis_label=f'{metric_col.capitalize()}',
y_axis_label='Observations',
y_range=(0, mal_count * .33),
x_range=(norm_mean + (norm_stddev * 2.5), mal_mean + (mal_stddev * 3)),
)
overlap_view.title.text_font_size = title_font_size
overlap_view.border_fill_color = cell_bg_color
overlap_view.border_fill_alpha = cell_bg_alpha
overlap_view.background_fill_color = cell_bg_color
overlap_view.background_fill_alpha = plot_bg_alpha
overlap_view.min_border_left = left_border
overlap_view.min_border_right = right_border
overlap_view.min_border_top = top_border
overlap_view.min_border_bottom = bottom_border
overlap_view.quad(bottom=0,
top=mal_hist_df.metric,
left=mal_hist_df.left,
right=mal_hist_df.right,
legend_label='malicious',
fill_color=malicious_color,
alpha=.85,
line_alpha=.35,
line_width=.5)
overlap_view.quad(bottom=0,
top=norm_hist_df.metric,
left=norm_hist_df.left,
right=norm_hist_df.right,
legend_label='normal',
fill_color=normal_color,
alpha=.35,
line_alpha=.35,
line_width=.5)
overlap_view.xaxis.formatter = NumeralTickFormatter(format='0,0')
overlap_view.yaxis.formatter = NumeralTickFormatter(format='0,0')
sigma_ref(overlap_view, norm_mean, norm_stddev)
overlap_view.legend.location = "top_right"
overlap_view.legend.background_fill_alpha = .3
# Probability Density - bigger bins for sparser malicous observations
malicious_hist_dense, malicious_edge_dense = np.histogram(malicious,
density=True,
bins=50)
mal_hist_dense_df = pd.DataFrame({
'metric': malicious_hist_dense,
'left': malicious_edge_dense[:-1],
'right': malicious_edge_dense[1:]
})
normal_hist_dense, normal_edge_dense = np.histogram(normal,
density=True,
bins=100)
norm_hist_dense_df = pd.DataFrame({
'metric': normal_hist_dense,
'left': normal_edge_dense[:-1],
'right': normal_edge_dense[1:]
})
density = figure(plot_width=plot_width,
plot_height=plot_height,
sizing_mode='fixed',
title='Probability Density',
x_axis_label=f'{metric_col.capitalize()}',
y_axis_label='% of Group Total')
density.title.text_font_size = title_font_size
density.border_fill_color = cell_bg_color
density.border_fill_alpha = cell_bg_alpha
density.background_fill_color = cell_bg_color
density.background_fill_alpha = plot_bg_alpha
density.min_border_left = left_border
density.min_border_right = right_border
density.min_border_top = top_border
density.min_border_bottom = bottom_border
density.quad(bottom=0,
top=mal_hist_dense_df.metric,
left=mal_hist_dense_df.left,
right=mal_hist_dense_df.right,
legend_label='malicious',
fill_color=malicious_color,
alpha=.85,
line_alpha=.35,
line_width=.5)
density.quad(bottom=0,
top=norm_hist_dense_df.metric,
left=norm_hist_dense_df.left,
right=norm_hist_dense_df.right,
legend_label='normal',
fill_color=normal_color,
alpha=.35,
line_alpha=.35,
line_width=.5)
density.xaxis.formatter = NumeralTickFormatter(format='0,0')
density.yaxis.formatter = NumeralTickFormatter(format='0.000%')
sigma_ref(density, norm_mean, norm_stddev)
density.legend.location = "top_right"
density.legend.background_fill_alpha = .3
# Simulation Series to be used
false_positives = simulations.FP
false_negatives = simulations.FN
multiplier = simulations.multiplier
precision = simulations.precision
recall = simulations.recall
f1_score = simulations.f1_score
f1_max = simulations[simulations.f1_score == simulations.f1_score.max(
)].head(1).squeeze()['multiplier']
# False Positives vs False Negatives
errors = figure(plot_width=plot_width,
plot_height=plot_height,
sizing_mode='fixed',
x_range=(multiplier.min(), multiplier.max()),
y_range=(0, false_positives.max()),
title='False Positives vs False Negatives',
x_axis_label='Multiplier',
y_axis_label='Count')
errors.title.text_font_size = title_font_size
errors.border_fill_color = cell_bg_color
errors.border_fill_alpha = cell_bg_alpha
errors.background_fill_color = cell_bg_color
errors.background_fill_alpha = plot_bg_alpha
errors.min_border_left = left_border
errors.min_border_right = right_border
errors.min_border_top = top_border
errors.min_border_bottom = right_border
errors.line(multiplier,
false_positives,
legend_label='false positives',
line_width=2,
color=fp_color)
errors.line(multiplier,
false_negatives,
legend_label='false negatives',
line_width=2,
color=fn_color)
errors.yaxis.formatter = NumeralTickFormatter(format='0,0')
errors.extra_y_ranges = {"y2": Range1d(start=0, end=1.1)}
errors.add_layout(
LinearAxis(y_range_name="y2",
axis_label="Score",
formatter=NumeralTickFormatter(format='0.00%')), 'right')
errors.line(multiplier,
f1_score,
line_width=2,
color=f1_color,
legend_label='F1 Score',
y_range_name="y2")
# F1 Score Maximization point
f1_thresh = Span(location=f1_max,
dimension='height',
line_color=f1_color,
line_dash='dashed',
line_width=2)
f1_label = Label(x=f1_max + .05,
y=180,
y_units='screen',
text=f'F1 Max: {round(f1_max,2)}',
text_font_size='10pt',
text_font_style='bold',
text_align='left',
text_color=f1_color)
errors.add_layout(f1_thresh)
errors.add_layout(f1_label)
errors.legend.location = "top_right"
errors.legend.background_fill_alpha = .3
# False Negative Weighting.
# Intro.
weighting_intro = f'''
<h3>Error types differ in impact.</h3>
<p>In the case of security incidents, a false negative,
though possibly rarer than false positives, is likely more costly. For example, downtime suffered
from a DDoS attack (lost sales/customers) incurs more loss than time wasted chasing a false positive
(labor hours). </p>
<p>Try playing around with the slider to the right to see how your thresholding strategy might need to change
depending on the relative weight of false negatives to false positives. What does it look like at
1:1, 50:1, etc.?</p>
'''
weighting_div = Div(text=weighting_intro,
width=420,
height=180,
margin=(0, 75, 0, 0))
# Now for the weighted errors viz
default_weighting = 10
initial_fp_cost = 100
simulations['weighted_FN'] = simulations.FN * default_weighting
weighted_fn = simulations.weighted_FN
simulations[
'total_weighted_error'] = simulations.FP + simulations.weighted_FN
total_weighted_error = simulations.total_weighted_error
simulations['fp_cost'] = initial_fp_cost
fp_cost = simulations.fp_cost
simulations[
'total_estimated_cost'] = simulations.total_weighted_error * simulations.fp_cost
total_estimated_cost = simulations.total_estimated_cost
twe_min = simulations[simulations.total_weighted_error ==
simulations.total_weighted_error.min()].head(
1).squeeze()['multiplier']
twe_min_count = simulations[simulations.multiplier == twe_min].head(
1).squeeze()['total_weighted_error']
generic_twe = simulations[simulations.multiplier.apply(
lambda x: round(x, 2)) == 3.00].squeeze()['total_weighted_error']
comparison = f'''
<p>Based on your inputs, the optimal threshold is around <b>{twe_min}</b>.
This would result in an estimated <b>{int(twe_min_count):,}</b> total weighted errors and
<b>${int(twe_min_count * initial_fp_cost):,}</b> in losses.</p>
<p>The generic threshold of 3.0 standard deviations would result in <b>{int(generic_twe):,}</b>
total weighted errors and <b>${int(generic_twe * initial_fp_cost):,}</b> in losses.</p>
<p>Using the optimal threshold would save <b>${int((generic_twe - twe_min_count) * initial_fp_cost):,}</b>,
reducing costs by <b>{(generic_twe - twe_min_count) / generic_twe * 100:.1f}%</b>
(assuming near-future events are distributed similarly to those from the past).</p>
'''
comparison_div = Div(text=comparison,
width=420,
height=230,
margin=(0, 75, 0, 0))
loss_min = ColumnDataSource(data=dict(multiplier=multiplier,
fp=false_positives,
fn=false_negatives,
weighted_fn=weighted_fn,
twe=total_weighted_error,
fpc=fp_cost,
tec=total_estimated_cost,
precision=precision,
recall=recall,
f1=f1_score))
evaluation = Figure(plot_width=900,
plot_height=520,
sizing_mode='fixed',
x_range=(multiplier.min(), multiplier.max()),
title='Evaluation Metrics vs Total Estimated Cost',
x_axis_label='Multiplier',
y_axis_label='Cost')
evaluation.title.text_font_size = title_font_size
evaluation.border_fill_color = cell_bg_color
evaluation.border_fill_alpha = cell_bg_alpha
evaluation.background_fill_color = cell_bg_color
evaluation.background_fill_alpha = plot_bg_alpha
evaluation.min_border_left = left_border
evaluation.min_border_right = right_border
evaluation.min_border_top = top_border
evaluation.min_border_bottom = bottom_border
evaluation.line('multiplier',
'tec',
source=loss_min,
line_width=3,
line_alpha=0.6,
color=total_weighted_color,
legend_label='Total Estimated Cost')
evaluation.yaxis.formatter = NumeralTickFormatter(format='$0,0')
# Evaluation metrics on second right axis.
evaluation.extra_y_ranges = {"y2": Range1d(start=0, end=1.1)}
evaluation.add_layout(
LinearAxis(y_range_name="y2",
axis_label="Score",
formatter=NumeralTickFormatter(format='0.00%')), 'right')
evaluation.line('multiplier',
'precision',
source=loss_min,
line_width=3,
line_alpha=0.6,
color=precision_color,
legend_label='Precision',
y_range_name="y2")
evaluation.line('multiplier',
'recall',
source=loss_min,
line_width=3,
line_alpha=0.6,
color=recall_color,
legend_label='Recall',
y_range_name="y2")
evaluation.line('multiplier',
'f1',
source=loss_min,
line_width=3,
line_alpha=0.6,
color=f1_color,
legend_label='F1 score',
y_range_name="y2")
evaluation.legend.location = "bottom_right"
evaluation.legend.background_fill_alpha = .3
twe_thresh = Span(location=twe_min,
dimension='height',
line_color=total_weighted_color,
line_dash='dashed',
line_width=2)
twe_label = Label(x=twe_min - .05,
y=240,
y_units='screen',
text=f'Cost Min: {round(twe_min,2)}',
text_font_size='10pt',
text_font_style='bold',
text_align='right',
text_color=total_weighted_color)
evaluation.add_layout(twe_thresh)
evaluation.add_layout(twe_label)
# Add in same f1 thresh as previous viz
evaluation.add_layout(f1_thresh)
evaluation.add_layout(f1_label)
handler = CustomJS(args=dict(source=loss_min,
thresh=twe_thresh,
label=twe_label,
comparison=comparison_div),
code="""
var data = source.data
var ratio = cb_obj.value
var multiplier = data['multiplier']
var fp = data['fp']
var fn = data['fn']
var weighted_fn = data['weighted_fn']
var twe = data['twe']
var fpc = data['fpc']
var tec = data['tec']
var generic_twe = 0
function round(value, decimals) {
return Number(Math.round(value+'e'+decimals)+'e-'+decimals);
}
function comma_sep(x) {
return x.toString().replace(/\B(?<!\.\d*)(?=(\d{3})+(?!\d))/g, ",");
}
for (var i = 0; i < multiplier.length; i++) {
weighted_fn[i] = Math.round(fn[i] * ratio)
twe[i] = weighted_fn[i] + fp[i]
tec[i] = twe[i] * fpc[i]
if (round(multiplier[i],2) == 3.00) {
generic_twe = twe[i]
}
}
var min_loss = Math.min.apply(null,twe)
var new_thresh = 0
for (var i = 0; i < multiplier.length; i++) {
if (twe[i] == min_loss) {
new_thresh = multiplier[i]
thresh.location = new_thresh
thresh.change.emit()
label.x = new_thresh
label.text = `Cost Min: ${new_thresh}`
label.change.emit()
comparison.text = `
<p>Based on your inputs, the optimal threshold is around <b>${new_thresh}</b>.
This would result in an estimated <b>${comma_sep(round(min_loss,0))}</b> total weighted errors and
<b>$${comma_sep(round(min_loss * fpc[i],0))}</b> in losses.</p>
<p>The generic threshold of 3.0 standard deviations would result in <b>${comma_sep(round(generic_twe,0))}</b>
total weighted errors and <b>$${comma_sep(round(generic_twe * fpc[i],0))}</b> in losses.</p>
<p>Using the optimal threshold would save <b>$${comma_sep(round((generic_twe - min_loss) * fpc[i],0))}</b>,
reducing costs by <b>${comma_sep(round((generic_twe - min_loss) / generic_twe * 100,0))}%</b>
(assuming near-future events are distributed similarly to those from the past).</p>
`
comparison.change.emit()
}
}
source.change.emit();
""")
slider = Slider(start=1.0,
end=500,
value=default_weighting,
step=.25,
title="FN:FP Ratio",
bar_color='#FFD100',
height=50,
margin=(5, 0, 5, 0))
slider.js_on_change('value', handler)
cost_handler = CustomJS(args=dict(source=loss_min,
comparison=comparison_div),
code="""
var data = source.data
var new_cost = cb_obj.value
var multiplier = data['multiplier']
var fp = data['fp']
var fn = data['fn']
var weighted_fn = data['weighted_fn']
var twe = data['twe']
var fpc = data['fpc']
var tec = data['tec']
var generic_twe = 0
function round(value, decimals) {
return Number(Math.round(value+'e'+decimals)+'e-'+decimals);
}
function comma_sep(x) {
return x.toString().replace(/\B(?<!\.\d*)(?=(\d{3})+(?!\d))/g, ",");
}
for (var i = 0; i < multiplier.length; i++) {
fpc[i] = new_cost
tec[i] = twe[i] * fpc[i]
if (round(multiplier[i],2) == 3.00) {
generic_twe = twe[i]
}
}
var min_loss = Math.min.apply(null,twe)
var new_thresh = 0
for (var i = 0; i < multiplier.length; i++) {
if (twe[i] == min_loss) {
new_thresh = multiplier[i]
comparison.text = `
<p>Based on your inputs, the optimal threshold is around <b>${new_thresh}</b>.
This would result in an estimated <b>${comma_sep(round(min_loss,0))}</b> total weighted errors and
<b>$${comma_sep(round(min_loss * new_cost,0))}</b> in losses.</p>
<p>The generic threshold of 3.0 standard deviations would result in
<b>${comma_sep(round(generic_twe,0))}</b> total weighted errors and
<b>$${comma_sep(round(generic_twe * new_cost,0))}</b> in losses.</p>
<p>Using the optimal threshold would save
<b>$${comma_sep(round((generic_twe - min_loss) * new_cost,0))}</b>,
reducing costs by <b>${comma_sep(round((generic_twe - min_loss)/generic_twe * 100,0))}%</b>
(assuming near-future events are distributed similarly to those from the past).</p>
`
comparison.change.emit()
}
}
source.change.emit();
""")
cost_input = TextInput(value=f"{initial_fp_cost}",
title="How much a false positive costs:",
height=75,
margin=(20, 75, 20, 0))
cost_input.js_on_change('value', cost_handler)
# Include DataTable of simulation results
dt_columns = [
TableColumn(field="multiplier", title="Multiplier"),
TableColumn(field="fp",
title="False Positives",
formatter=NumberFormatter(format='0,0')),
TableColumn(field="fn",
title="False Negatives",
formatter=NumberFormatter(format='0,0')),
TableColumn(field="weighted_fn",
title="Weighted False Negatives",
formatter=NumberFormatter(format='0,0.00')),
TableColumn(field="twe",
title="Total Weighted Errors",
formatter=NumberFormatter(format='0,0.00')),
TableColumn(field="fpc",
title="Estimated FP Cost",
formatter=NumberFormatter(format='$0,0.00')),
TableColumn(field="tec",
title="Estimated Total Cost",
formatter=NumberFormatter(format='$0,0.00')),
TableColumn(field="precision",
title="Precision",
formatter=NumberFormatter(format='0.00%')),
TableColumn(field="recall",
title="Recall",
formatter=NumberFormatter(format='0.00%')),
TableColumn(field="f1",
title="F1 Score",
formatter=NumberFormatter(format='0.00%')),
]
data_table = DataTable(source=loss_min,
columns=dt_columns,
width=1400,
height=700,
sizing_mode='fixed',
fit_columns=True,
reorderable=True,
sortable=True,
margin=(30, 0, 20, 0))
# weighting_layout = column([weighting_div, evaluation, slider, data_table])
weighting_layout = column(
row(column(weighting_div, cost_input, comparison_div),
column(slider, evaluation), Div(text='', height=200, width=60)),
data_table)
# Initialize visualizations in browser
time.sleep(1.5)
layout = grid([
[title_div],
[row(summary_div, vertical_line, hypo_div)],
[
row(Div(text='', height=200, width=60), exploratory,
Div(text='', height=200, width=10), overlap_view,
Div(text='', height=200, width=40))
],
[Div(text='', height=10, width=200)],
[
row(Div(text='', height=200, width=60), density,
Div(text='', height=200, width=10), errors,
Div(text='', height=200, width=40))
],
[Div(text='', height=10, width=200)],
[
row(Div(text='', height=200, width=60), weighting_layout,
Div(text='', height=200, width=40))
],
])
# Generate html resources for dashboard
fonts = os.path.join(os.getcwd(), 'fonts')
if os.path.isdir(os.path.join(session_folder, 'fonts')):
shutil.rmtree(os.path.join(session_folder, 'fonts'))
shutil.copytree(fonts, os.path.join(session_folder, 'fonts'))
else:
shutil.copytree(fonts, os.path.join(session_folder, 'fonts'))
html = file_html(layout, INLINE, "Cream")
with open(os.path.join(session_folder, f'{session_name}.html'),
"w") as file:
file.write(html)
webbrowser.open("file://" +
os.path.join(session_folder, f'{session_name}.html'))
def _segment_plot(self, source):
"""Plot showing the mirror segments.
The plot shows the 91 segment positions with their position number. The colour of the segments indicates how
long ago they have been recoated.
Params:
-------
source: pandas.DataFrame
Data source.
Returns:
--------
bokeh.plotting.Figure
Plot showing the mirror segments.
"""
def segment_color(date):
"""Fill colour of a mirror segment."""
days = (datetime.datetime.now().date() - date).days
days = min(days, self.RECOATING_PERIOD)
days = max(days, 0)
def hex_value(v):
"""Convert integer into a hexadecimal string suitable for specifying an RGB value in a css rule."""
s = format(int(round(255 * v)), 'x')
if len(s) < 2:
s = '0' + s
return s
# The colour ranges from dark green for recently recoated segments to white for segments which need to be
# recoated.
h = 99
s = 100
l = 33 + 67 * days / self.RECOATING_PERIOD
rgb = colorsys.hls_to_rgb(h / 360, l / 100, s / 100)
return '#{r}{g}{b}'.format(r=hex_value(rgb[0]),
g=hex_value(rgb[1]),
b=hex_value(rgb[2]))
plot = Figure(tools='tap', toolbar_location=None)
colors = [segment_color(d) for d in source.data['ReplacementDate']]
plot.patches(xs='SegmentPositionCornerXs',
ys='SegmentPositionCornerYs',
source=source,
color=colors,
line_color='#dddddd')
plot.text(x='SegmentPositionX',
y='SegmentPositionY',
text='SegmentPosition',
source=source,
text_color='black',
text_align='center',
text_baseline='middle')
plot.grid.grid_line_color = None
plot.axis.visible = False
plot.min_border_top = 20
plot.min_border_bottom = 30
plot.outline_line_color = None
return plot