Mass invasions: Temporal patterns for 13 Instagram vantage points¶
â–“â–“â–“â–“â–“â–“â–“ â–“â–“â–“â–“â–“â–“â–“, â–“â–“â–“â–“â–“â–“â–“, â–“â–“â–“â–“â–“â–“â–“ â–“â–“â–“â–“â–“â–“â–“ â–“â–“â–“â–“â–“â–“â–“; â–“â–“â–“â–“â–“â–“â–“ â–“â–“â–“â–“â–“â–“â–“, â–“â–“â–“â–“â–“â–“â–“ â–“â–“â–“â–“â–“â–“â–“ â–“â–“â–“â–“â–“â–“â–“, â–“â–“â–“â–“â–“â–“â–“
•••
Out[1]:
Visualizations of temporal patterns for places affected by "mass invasion". The list of 13 places was prepared by Claudia Tautenhahn for her master's thesis.
Tautenhahn, C. Das Phänomen der Masse: Landschaftliche Wirkungen sozialer Medien. Master's thesis. TU Dresden 2020, Department of Landscape Architecture and Environmental Planning
I was the second advisor of this thesis and performed data collection. Based on Claudia's list, 1.5 Million Instagram Posts were queried. The query period covered was 2010 to 2019, but the lower analysis bound was limited to 2015 due to low data availability in early years. This notebook uses the originally collected data to create visualizations based on privacy-aware HyperLogLog cardinality estimation.
Preparations¶
•••
In [3]:
OUTPUT = Path.cwd().parents[0] / "out" # output directory for figures (etc.)
WORK_DIR = Path.cwd().parents[0] / "tmp" # Working directory
In [4]:
OUTPUT.mkdir(exist_ok=True)
(OUTPUT / "figures").mkdir(exist_ok=True)
(OUTPUT / "svg").mkdir(exist_ok=True)
WORK_DIR.mkdir(exist_ok=True)
In [5]:
%load_ext autoreload
%autoreload 2
Select M for monthly aggregation, Y for yearly aggregation
In [6]:
AGG_BASE = "M"
First, define whether to study usercount or postcount
In [7]:
# METRIC = 'user'
METRIC = 'post'
In [8]:
metric_col = 'post_hll'
if METRIC == 'user':
metric_col = 'user_hll'
Set global font
In [9]:
plt.rcParams['font.family'] = 'serif'
plt.rcParams['font.serif'] = ['Times New Roman'] + plt.rcParams['font.serif']
In [10]:
color_instagram = '#737EBD'
Load HLL aggregate data¶
Load the data from CSV, generated in the previous notebook. Data is stored as aggregate HLL data (postcount, usercount) for each month.
•••
In [12]:
%%time
data_files = {
"MASS_INVASION_ALL":MASS_INVASION_ALL,
}
tools.display_file_stats(data_files)
In [13]:
pd.read_csv(MASS_INVASION_ALL, nrows=10)
Out[13]:
Connect hll worker db¶
In [14]:
db_user = "postgres"
db_pass = os.getenv('POSTGRES_PASSWORD')
db_host = "127.0.0.1"
db_port = "25432"
db_name = "hlldb"
Connect to empty Postgres database running HLL Extension:
In [15]:
DB_CONN = psycopg2.connect(
host=db_host,
port=db_port,
dbname=db_name,
user=db_user,
password=db_pass
)
DB_CONN.set_session(
readonly=True)
DB_CALC = tools.DbConn(
DB_CONN)
CUR_HLL = DB_CONN.cursor()
test
Calculate HLL Cardinality per month and year¶
Define additional functions for reading and formatting CSV as pd.DataFrame
In [16]:
from datetime import datetime
def read_csv_datetime(csv: Path) -> pd.DataFrame:
"""Read CSV with parsing datetime index (months)
First CSV column: Year
Second CSV column: Month
"""
date_cols = ["year", "month"]
df = pd.read_csv(
csv, index_col='datetime',
parse_dates={'datetime':date_cols},
date_format='%Y %m',
keep_date_col='False')
df.drop(columns=date_cols, inplace=True)
return df
def append_cardinality_df(df: pd.DataFrame, hll_col: str = "post_hll", cardinality_col: str = 'postcount_est'):
"""Calculate cardinality from HLL and append to extra column in df"""
df[cardinality_col] = df.apply(
lambda x: hll.cardinality_hll(
x[hll_col], CUR_HLL),
axis=1)
df.drop(columns=[hll_col], inplace=True)
return df
def filter_fill_time(
df: pd.DataFrame, min_year: int,
max_year: int, val_col: str = "postcount_est",
min_month: str = "01", max_month: str = "01", agg_base: str = None,
agg_method = None):
"""Filter time values between min - max year and fill missing values"""
max_day = "01"
if agg_base is None:
agg_base = "M"
elif agg_base == "Y":
max_month = "12"
max_day = "31"
min_date = pd.Timestamp(f'{min_year}-{min_month}-01')
max_date = pd.Timestamp(f'{max_year}-{max_month}-{max_day}')
# clip by start and end date
if not min_date in df.index:
df.loc[min_date, val_col] = 0
if not max_date in df.index:
df.loc[max_date, val_col] = 0
df.sort_index(inplace=True)
# mask min and max time
time_mask = ((df.index >= min_date) & (df.index <= max_date))
resampled = df.loc[time_mask][val_col].resample(agg_base)
if agg_method is None:
series = resampled.sum()
elif agg_method == "count":
series = resampled.count()
elif agg_method == "nunique":
series = resampled.nunique()
# fill missing months with 0
# this will also set the day to max of month
return series.fillna(0).to_frame()
Select dataset to process below
Apply functions to all data sets.
- Read from CSV
- calculate cardinality
- merge year and month to single column
- filter 2007 - 2018 range, fill missing values
In [34]:
def process_dataset(
dataset: Path = None, metric: str = None, df_post: pd.DataFrame = None,
min_year: int = None, max_year: int = None, agg_base: str = None) -> pd.DataFrame:
"""Apply temporal filter/pre-processing to all data sets."""
if metric is None:
metric = 'post_hll'
warn(f"Using default value {metric}")
if metric == 'post_hll':
cardinality_col = 'postcount_est'
else:
cardinality_col = 'usercount_est'
if min_year is None:
min_year = 2015
if max_year is None:
max_year = 2019
if df_post is None:
df_post = read_csv_datetime(dataset)
df_post = append_cardinality_df(df_post, metric, cardinality_col)
return filter_fill_time(df_post, min_year, max_year, cardinality_col, agg_base=agg_base)
In [35]:
%%time
df_post = process_dataset(MASS_INVASION_ALL, agg_base=AGG_BASE, metric='post_hll')
In [36]:
df_post.head(5)
Out[36]:
In [37]:
%%time
df_user = process_dataset(MASS_INVASION_ALL, metric=metric_col, agg_base=AGG_BASE)
In [38]:
df_user.head(5)
Out[38]:
Visualize Cardinality¶
Flickr¶
Define plot function.
In [39]:
def fill_plot_time(
df: pd.DataFrame, ax: Axes, color: str,
label: str, val_col: str = "postcount_est") -> Axes:
"""Matplotlib Barplot with time axis formatting
If "significant" in df columns, applies different colors to fill/edge
of non-significant values.
"""
if color is None:
colors = sns.color_palette("vlag", as_cmap=True, n_colors=2)
color_rgba = colors([1.0])[0]
color = mcolor.rgb2hex((color_rgba), keep_alpha=True)
color_significant = color
color_significant_edge = "white"
if "significant" in df.columns:
colors_bar = {True: color, False: "white"}
color_significant = df['significant'].replace(colors_bar)
colors_edge = {True: "white", False: "black"}
color_significant_edge = df['significant'].replace(colors_edge)
df_plot = df.set_index(
df.index.map(lambda s: s.strftime('%Y')))
ax = df_plot.plot(
ax=ax, y=val_col, color=color_significant,
label=label, linewidth=0.5, alpha=0.6)
ax.fill_between(range(len(df_plot.index)), df_plot[val_col], facecolor=color, alpha=0.6)
return ax
def plot_time(
df: Tuple[pd.DataFrame, pd.DataFrame], title, color = None, filename = None,
output = OUTPUT, legend: str = "Postcount", val_col: str = None,
trend: bool = None, seasonal: bool = None, residual: bool = None,
agg_base: str = None, fig = None, ax = None, return_fig_ax = None):
"""Create dataframe(s) time plot"""
x_ticks_every = 12
fig_x = 10
fig_y = 2
font_mod = True
x_label = "Year"
linewidth = 3
if agg_base and agg_base == "Y":
x_ticks_every = 1
fig_x = 3
fig_y = 1.5
font_mod = True
x_label = "Year"
linewidth = 1
if fig is None or ax is None:
fig, ax = plt.subplots()
fig.set_size_inches(fig_x, fig_y)
ylabel = f'{legend}'
if val_col is None:
val_col = f'{legend.lower()}_est'
ax = fill_plot_time(
df=df, ax=ax, color=color, val_col=val_col, label=legend)
# TODO: below is a bit hacky way to format the x-axis;
tick_loc = mticker.MultipleLocator(x_ticks_every)
ax.xaxis.set_major_locator(tick_loc)
ax.tick_params(axis='x', rotation=45, length=0.5)
ax.yaxis.set_major_formatter(mticker.StrMethodFormatter('{x:,.0f}'))
xrange_min_max = range(2015, 2021)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
ax.set_xticklabels(xrange_min_max, rotation=45)
ax.set(xlabel=x_label, ylabel=ylabel)
ax.spines["left"].set_linewidth(0.25)
ax.spines["bottom"].set_linewidth(0.25)
ax.spines["top"].set_linewidth(0)
ax.spines["right"].set_linewidth(0)
ax.yaxis.set_tick_params(width=0.5)
# remove legend
ax.get_legend().remove()
ax.set_title(title)
ax.set_xlim(-0.5, len(df)-0.5)
ax.set_ylim(bottom=0)
if font_mod:
for item in (
[ax.xaxis.label, ax.title, ax.yaxis.label] +
ax.get_xticklabels() + ax.get_yticklabels()):
item.set_fontsize(8)
# store figure to file
if filename:
fig.savefig(
output / "figures" / f"{filename}.png", dpi=300, format='PNG',
bbox_inches='tight', pad_inches=1, facecolor="white")
# also save as svg
fig.savefig(
output / "svg" / f"{filename}.svg", format='svg',
bbox_inches='tight', pad_inches=1, facecolor="white")
if return_fig_ax:
return fig, ax
In [40]:
def load_and_plot(
dataset: Path = None, metric: str = None, src_ref: str = "flickr", colors: cm.colors.ListedColormap = None,
agg_base: str = None, trend: bool = None, return_df: bool = None, df_post: pd.DataFrame = None, return_fig_ax = None):
"""Load data and plot"""
if metric is None:
metric = 'post_hll'
if metric == 'post_hll':
metric_label = 'postcount'
else:
metric_label = 'usercount'
if colors is None:
colors = sns.color_palette("vlag", as_cmap=True, n_colors=2)
colors = colors([1.0])
df = process_dataset(dataset, metric=metric, agg_base=agg_base, df_post=df_post)
fig, ax = plot_time(
df, legend=metric_label.capitalize(), color=colors,
title=f'{src_ref}',
filename=f"temporal_{metric_label}_{src_ref}_absolute", trend=trend, agg_base=agg_base, return_fig_ax=True)
fig.show()
if return_fig_ax:
return fig, ax
if return_df:
return df
In [41]:
colors = sns.color_palette("vlag", as_cmap=True, n_colors=2)
In [42]:
fig, ax = load_and_plot(MASS_INVASION_ALL, src_ref=f"Mass invasion total posts 2015-2019, all 13 places", agg_base=AGG_BASE, trend=False, metric=metric_col, return_fig_ax=True)
fig.show()
Visualize individual places¶
Monthly¶
In [154]:
df = pd.read_csv(MASS_INVASION_ALL)
df.tail()
Out[154]:
In [155]:
mass_invasion_places = df['topic_group'].unique()
In [156]:
mass_invasion_places
Out[156]:
In [157]:
def plot_bars(
df: pd.DataFrame, ax: matplotlib.axes = None, title: str = None):
"""Plot bars from two DataFrames"""
bar_param = {
"width":1.0,
"label":"Mass invasion total post count aggregated for months",
"edgecolor":"white",
"linewidth":0.5,
"alpha":1.0
}
# create figure
if not ax:
fig, ax = plt.subplots(1, 1, figsize=(3, 1.5))
# plot
df.groupby(df.index.month)["postcount_est"] \
.mean().plot.bar(ax=ax, color=color_instagram, y="postcount_est", **bar_param)
# format
ax.set_xlim(-0.5,11.5)
month_names = ['Jan','Feb','Mar','Apr','May','Jun',
'Jul','Aug','Sep','Oct','Nov','Dec']
ax.set_xticklabels(month_names)
ax.tick_params(axis='x', rotation=45, length=0) # length: of ticks
ax.spines["left"].set_linewidth(0.25)
ax.spines["bottom"].set_linewidth(0.25)
ax.spines["top"].set_linewidth(0)
ax.spines["right"].set_linewidth(0)
ax.yaxis.set_tick_params(width=0.5)
ax.set(xlabel="", ylabel="")
if not title:
title = "Post Count per Month (mean)"
ax.set_title(title, y=-0.2, pad=-14)
for item in (
[ax.xaxis.label, ax.title, ax.yaxis.label] +
ax.get_xticklabels() + ax.get_yticklabels()):
item.set_fontsize(8)
In [158]:
top_50_cnt = 13
top_50 = mass_invasion_places.tolist()
In [159]:
top_50.sort()
In [160]:
top_50
Out[160]:
In [161]:
df_post = read_csv_datetime(MASS_INVASION_ALL)
# create figure object with multiple subplots
fig, axes = plt.subplots(nrows=int(top_50_cnt/3), ncols=4, figsize=(14, 11))
fig.subplots_adjust(hspace=.5) # adjust vertical space, to allow title below plot
# iterate places
for ix, ax in enumerate(axes.reshape(-1)):
if ix >= len(top_50):
ax.set_axis_off()
continue
np_str = top_50[ix]
# filter np_str and calculate cardinality
df_post_filter = df_post[df_post["topic_group"]==np_str].copy()
df_post_filter = append_cardinality_df(df_post_filter, 'post_hll', 'postcount_est')
df_post_filter = filter_fill_time(df_post_filter, 2010, 2023, 'postcount_est')
# plot bars individually
plot_bars(
df=df_post_filter, title=np_str, ax=ax)
In [162]:
tools.save_fig(fig, output=OUTPUT, name="barplot_massinvasion_seasonal")
Yearly¶
In [163]:
def bar_plot_time(
df: pd.DataFrame, ax: Axes, color: str,
label: str, val_col: str = "postcount_est") -> Axes:
"""Matplotlib Barplot with time axis formatting
If "significant" in df columns, applies different colors to fill/edge
of non-significant values.
"""
if color is None:
color = color_instagram
color_significant = color
color_significant_edge = "white"
if "significant" in df.columns:
colors_bar = {True: color, False: "white"}
color_significant = df['significant'].replace(colors_bar)
colors_edge = {True: "white", False: "black"}
color_significant_edge = df['significant'].replace(colors_edge)
bar_param = {
"width":1.0,
"label":label,
"edgecolor":"white",
"linewidth":0.5,
"alpha":1.0
}
ax = df.set_index(
df.index.map(lambda s: s.strftime('%Y'))).plot.bar(
ax=ax, color=color_instagram, y="postcount_est", **bar_param)
return ax
def plot_time(
df: Tuple[pd.DataFrame, pd.DataFrame], title, color = None, filename = None,
output = OUTPUT, legend: str = "Postcount", val_col: str = None,
trend: bool = None, seasonal: bool = None, residual: bool = None,
agg_base: str = None, ax = None):
"""Create dataframe(s) time plot"""
x_ticks_every = 12
fig_x = 15.7
fig_y = 4.27
font_mod = False
x_label = ""
linewidth = 3
if agg_base and agg_base == "Y":
x_ticks_every = 1
fig_x = 3
fig_y = 1.5
font_mod = True
x_label = ""
linewidth = 1
if ax is None:
fig, ax = plt.subplots()
fig.set_size_inches(fig_x, fig_y)
ylabel = ""
if val_col is None:
val_col = f'{legend.lower()}_est'
ax = bar_plot_time(
df=df, ax=ax, color=color, val_col=val_col, label=legend)
# format
tick_loc = mticker.MultipleLocator(x_ticks_every)
ax.xaxis.set_major_locator(tick_loc)
ax.tick_params(axis='x', rotation=45, length=0)
ax.yaxis.set_major_formatter(mticker.StrMethodFormatter('{x:,.0f}'))
ax.set(xlabel=x_label, ylabel=ylabel)
ax.spines["left"].set_linewidth(0.25)
ax.spines["bottom"].set_linewidth(0.25)
ax.spines["top"].set_linewidth(0)
ax.spines["right"].set_linewidth(0)
ax.yaxis.set_tick_params(width=0.5)
# remove legend
ax.get_legend().remove()
ax.set_title(title, y=-0.2, pad=-14)
ax.set_xlim(-0.5,len(df)-0.5)
for item in (
[ax.xaxis.label, ax.title, ax.yaxis.label] +
ax.get_xticklabels() + ax.get_yticklabels()):
item.set_fontsize(8)
In [164]:
df_post = read_csv_datetime(MASS_INVASION_ALL)
# create figure object with multiple subplots
fig, axes = plt.subplots(nrows=int(top_50_cnt/3), ncols=4, figsize=(17, 8))
fig.subplots_adjust(hspace=.5) # adjust vertical space, to allow title below plot
# iterate places
for ix, ax in enumerate(axes.reshape(-1)):
if ix >= len(top_50):
ax.set_axis_off()
continue
np_str = top_50[ix]
# filter np_str and calculate cardinality
df_post_filter = df_post[df_post["topic_group"]==np_str].copy()
df_post_filter = append_cardinality_df(df_post_filter, 'post_hll', 'postcount_est')
df_post_filter = filter_fill_time(df_post_filter, 2015, 2019, 'postcount_est', max_month=8)
# plot bars individually
plot_time(
df=df_post_filter, title=np_str, ax=ax)
In [165]:
tools.save_fig(fig, output=OUTPUT, name="barplot_massinvasion_trend")
Create notebook HTML¶
In [167]:
!jupyter nbconvert --to html_toc \
--output-dir=../resources/html/ ./01_mass_invasion.ipynb \
--output 04_mass_invasion \
--template=../nbconvert.tpl \
--ExtractOutputPreprocessor.enabled=False >&- 2>&-