from __future__ import annotations
import warnings
from typing import Any, Literal, Optional
import matplotlib
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns
from retentioneering.backend.tracker import (
collect_data_performance,
time_performance,
track,
)
from retentioneering.common.constants import DATETIME_UNITS
from retentioneering.eventstream.types import EventstreamType
from retentioneering.tooling.constants import BINS_ESTIMATORS
from retentioneering.tooling.timedelta_hist.constants import (
AGGREGATION_NAMES,
EVENTSTREAM_GLOBAL_EVENTS,
)
[docs]class TimedeltaHist:
"""
Plot the distribution of the time deltas between two events. Support various
distribution types, such as distribution of time for adjacent consecutive events, or
for a pair of pre-defined events, or median transition time from event to event per user/session.
Parameters
----------
eventstream : EventstreamType
See Also
--------
.UserLifetimeHist : Plot the distribution of user lifetimes.
.EventTimestampHist : Plot the distribution of events over time.
.Eventstream.describe : Show general eventstream statistics.
.Eventstream.describe_events : Show general eventstream events statistics.
.AddStartEndEvents : Create new synthetic events ``path_start`` and ``path_end`` to each user trajectory.
.SplitSessions : Create new synthetic events, that divide users’ paths on sessions.
.LabelCroppedPaths : Create new synthetic event(s) for each user based on the timeout threshold.
.DropPaths : Filter user paths based on the path length, removing the paths that are shorter than the
specified number of events or cut_off.
Notes
-----
See :ref:`Eventstream user guide<eventstream_timedelta_hist>` for the details.
"""
EVENTSTREAM_START = "eventstream_start"
EVENTSTREAM_END = "eventstream_end"
__eventstream: EventstreamType
raw_events_only: bool
event_pair: list[str | EVENTSTREAM_GLOBAL_EVENTS] | None
adjacent_events_only: bool
weight_col: str | None
time_agg: AGGREGATION_NAMES | None
timedelta_unit: DATETIME_UNITS
log_scale: bool | tuple[bool, bool] | None
lower_cutoff_quantile: float | None
upper_cutoff_quantile: float | None
bins: int | Literal[BINS_ESTIMATORS]
bins_to_show: np.ndarray
values_to_plot: np.ndarray
@time_performance(
scope="timedelta_hist",
event_name="init",
)
def __init__(self, eventstream: EventstreamType) -> None:
self.__eventstream = eventstream
self.user_col = self.__eventstream.schema.user_id
self.event_col = self.__eventstream.schema.event_name
self.time_col = self.__eventstream.schema.event_timestamp
self.type_col = self.__eventstream.schema.event_type
self.bins_to_show = np.array([])
self.values_to_plot = np.array([])
def _prepare_time_diff(self, data: pd.DataFrame) -> pd.DataFrame:
if not self.adjacent_events_only:
data = data[data[self.event_col].isin(self.event_pair)] # type: ignore
weight_col_group = data.groupby([self.weight_col])
with pd.option_context("mode.chained_assignment", None):
data["time_passed"] = weight_col_group[self.time_col].diff() / np.timedelta64(1, self.timedelta_unit) # type: ignore
if self.event_pair:
data["prev_event"] = weight_col_group[self.event_col].shift()
data = data[(data[self.event_col] == self.event_pair[1]) & (data["prev_event"] == self.event_pair[0])]
return data.dropna(subset="time_passed") # type: ignore
def _aggregate_data(self, data: pd.DataFrame) -> pd.DataFrame:
if self.time_agg is not None:
data = data.groupby(self.weight_col)["time_passed"].agg(self.time_agg).reset_index()
return data
def _remove_cutoff_values(self, series: pd.Series) -> pd.Series:
idx = [True] * len(series)
if self.upper_cutoff_quantile is not None:
idx &= series <= series.quantile(self.upper_cutoff_quantile)
if self.lower_cutoff_quantile is not None:
idx &= series >= series.quantile(self.lower_cutoff_quantile)
return series[idx]
def _prepare_global_events_diff(self, data: pd.DataFrame) -> pd.DataFrame:
if self.EVENTSTREAM_START in self.event_pair: # type: ignore
global_event_time = data[self.time_col].min()
global_event = self.EVENTSTREAM_START
else:
global_event_time = data[self.time_col].max()
global_event = self.EVENTSTREAM_END
global_events = data.groupby([self.weight_col]).first().reset_index().copy()
global_events[self.time_col] = global_event_time
global_events[self.event_col] = global_event
data = data[data[self.event_col].isin(self.event_pair)].copy() # type: ignore
data = pd.concat([data, global_events]).sort_values([self.weight_col, self.time_col]).reset_index(drop=True) # type: ignore
return data
def __validate_input(
self,
log_scale: bool | tuple[bool, bool] | None = None,
lower_cutoff_quantile: float | None = None,
upper_cutoff_quantile: float | None = None,
) -> tuple[tuple[bool, bool], float | None, float | None]:
if lower_cutoff_quantile is not None:
if not 0 < lower_cutoff_quantile < 1:
raise ValueError("lower_cutoff_quantile should be a fraction between 0 and 1.")
if upper_cutoff_quantile is not None:
if not 0 < upper_cutoff_quantile < 1:
raise ValueError("upper_cutoff_quantile should be a fraction between 0 and 1.")
if lower_cutoff_quantile is not None and upper_cutoff_quantile is not None:
if lower_cutoff_quantile > upper_cutoff_quantile:
warnings.warn("lower_cutoff_quantile exceeds upper_cutoff_quantile; no data passed to the histogram")
if log_scale:
if isinstance(log_scale, bool):
log_scale = (log_scale, False)
else:
log_scale = log_scale
else:
log_scale = (False, False)
return log_scale, upper_cutoff_quantile, lower_cutoff_quantile
[docs] @time_performance(
scope="timedelta_hist",
event_name="fit",
)
def fit(
self,
raw_events_only: bool = False,
event_pair: Optional[list[str | EVENTSTREAM_GLOBAL_EVENTS]] = None,
adjacent_events_only: bool = True,
weight_col: str | None = None,
time_agg: Optional[AGGREGATION_NAMES] = None,
timedelta_unit: DATETIME_UNITS = "s",
log_scale: bool | tuple[bool, bool] | None = None,
lower_cutoff_quantile: Optional[float] = None,
upper_cutoff_quantile: Optional[float] = None,
bins: int | Literal[BINS_ESTIMATORS] = 20,
) -> None:
"""
Calculate values and bins for the histplot.
Parameters
----------
raw_events_only : bool, default True
If ``True`` - statistics will be shown only for raw events.
If ``False`` - statistics will be shown for all events presented in your data.
event_pair : tuple of str, optional
Specify an event pair to plot the time distance between. The first
item corresponds to chronologically first event, the second item corresponds to the second event. If
``event_pair=None``, plot distribution of timedelta for all adjacent events.
Examples: ('login', 'purchase'); ['start', 'cabinet']
Besides the generic eventstream events, ``event_pair`` can accept special ``eventstream_start`` and
``eventstream_end`` events which denote the first and the last event in the entire eventstream correspondingly.
Note that the sequence of events and ``weight_col`` is important.
adjacent_events_only : bool, default True
Is used only when ``event_pair`` is not ``None``; specifies whether events need to be
adjacent to be included.
For example, if ``event_pair=("login", "purchase")`` and ``adjacent_events_only=False``,
then the sequence ("login", "main", "trading", "purchase") will contain a valid pair
(which is not the case with ``adjacent_events_only=True``).
weight_col : str, default None
Specify a unit of observation, inside which time differences will be computed.
By default, the values from ``user_id`` column in :py:class:`.EventstreamSchema` is taken.
For example:
- If ``user_id`` - time deltas will be computed only for events inside each user path.
- If ``session_id`` - the same, but inside each session.
time_agg : {None, "mean", "median"}, default None
Specify the aggregation policy for the time distances. Aggregate based on passed ``weight_col``.
- If ``None`` - no aggregation;
- ``mean`` and ``median`` plot distributions of ``weight_col`` unit mean or unit ``median`` timedeltas.
For example, if session id is specified in ``weight_col``, one observation per
session (for example, session median) will be provided for the histogram.
timedelta_unit : :numpy_link:`DATETIME_UNITS<>`, default 's'
Specify units of time differences the histogram should use. Use "s" for seconds, "m" for minutes,
"h" for hours and "D" for days.
log_scale: bool | tuple of bool | None, optional
- If ``True`` - apply log scaling to the ``x`` axis.
- If tuple of bool - apply log scaling to the (``x``,``y``) axes correspondingly.
lower_cutoff_quantile : float, optional
Specify time distance quantile as the lower boundary. The values below the boundary are truncated.
upper_cutoff_quantile : float, optional
Specify time distance quantile as the upper boundary. The values above the boundary are truncated.
bins : int or {"auto", "fd", "doane", "scott", "stone", "rice", "sturges", "sqrt"}, default 20
Generic bin parameter that can be the name of a reference rule or
the number of bins. Passed to :numpy_bins_link:`numpy.histogram_bin_edges<>`.
Returns
-------
None
"""
called_params = {
"raw_events_only": raw_events_only,
"event_pair": event_pair,
"adjacent_events_only": adjacent_events_only,
"weight_col": weight_col,
"time_agg": time_agg,
"timedelta_unit": timedelta_unit,
"log_scale": log_scale,
"lower_cutoff_quantile": lower_cutoff_quantile,
"upper_cutoff_quantile": upper_cutoff_quantile,
"bins": bins,
}
not_hash_values = ["time_agg", "timedelta_unit"]
self.log_scale, self.upper_cutoff_quantile, self.lower_cutoff_quantile = self.__validate_input(
log_scale, lower_cutoff_quantile, upper_cutoff_quantile
)
self.raw_events_only = raw_events_only
self.event_pair = event_pair
self.adjacent_events_only = adjacent_events_only
self.weight_col = weight_col or self.__eventstream.schema.user_id
self.time_agg = time_agg
self.timedelta_unit = timedelta_unit
self.bins = bins
data = self.__eventstream.to_dataframe(copy=True)
if self.raw_events_only:
data = data[data[self.type_col].isin(["raw"])]
data = data.sort_values([self.weight_col, self.time_col])
if self.event_pair is not None and set([self.EVENTSTREAM_START, self.EVENTSTREAM_END]).intersection(
self.event_pair
):
data = self._prepare_global_events_diff(data)
data = self._prepare_time_diff(data)
data = self._aggregate_data(data)
values_to_plot = data["time_passed"].reset_index(drop=True)
if self._remove_cutoff_values: # type: ignore
values_to_plot = self._remove_cutoff_values(values_to_plot).to_numpy()
if self.log_scale[0]:
log_adjustment = np.timedelta64(100, "ms") / np.timedelta64(1, self.timedelta_unit)
values_to_plot = np.where(values_to_plot != 0, values_to_plot, values_to_plot + log_adjustment) # type: ignore
bins_to_show = np.power(10, np.histogram_bin_edges(np.log10(values_to_plot), bins=self.bins))
else:
bins_to_show = np.histogram_bin_edges(values_to_plot, bins=self.bins)
if len(values_to_plot) == 0:
bins_to_show = np.array([])
self.bins_to_show = bins_to_show
self.values_to_plot = values_to_plot # type: ignore
collect_data_performance(
scope="timedelta_hist",
event_name="metadata",
called_params=called_params,
not_hash_values=not_hash_values,
performance_data={},
eventstream_index=self.__eventstream._eventstream_index,
)
@property
@time_performance(
scope="timedelta_hist",
event_name="values",
)
def values(self) -> tuple[np.ndarray, np.ndarray]:
"""
Returns
-------
tuple(np.ndarray, np.ndarray)
1. The first array contains the values for histogram.
2. The first array contains the bin edges.
"""
return self.values_to_plot, self.bins_to_show
[docs] @time_performance(
scope="timedelta_hist",
event_name="plot",
)
def plot(self, width: float = 6.0, height: float = 4.5) -> Any:
"""
Create a sns.histplot based on the calculated values.
Parameters
----------
width : float, default 6.0
Width in inches.
height : float, default 4.5
Height in inches.
Returns
-------
:matplotlib_axes:`matplotlib.axes.Axes<>`
The matplotlib axes containing the plot.
"""
called_params = {
"width": width,
"height": height,
}
figsize = (width, height)
plt.subplots(figsize=figsize)
hist = sns.histplot(self.values_to_plot, bins=self.bins, log_scale=self.log_scale)
hist.set_title(
f"Timedelta histogram, event pair - {self.event_pair}, weight column - {self.weight_col}"
f"{', group - ' + self.time_agg if self.time_agg is not None else ''}"
)
hist.set_xlabel(f"Time units: {self.timedelta_unit}")
collect_data_performance(
scope="timedelta_hist",
event_name="metadata",
called_params=called_params,
performance_data={},
eventstream_index=self.__eventstream._eventstream_index,
)
return hist