Source code for retentioneering.tooling.step_sankey.step_sankey

from __future__ import annotations

from typing import Any, Dict

import numpy as np
import pandas as pd
import plotly.graph_objects as go
import seaborn as sns

from retentioneering.backend.tracker import (
    collect_data_performance,
    time_performance,
    track,
    tracker,
)
from retentioneering.eventstream.types import EventstreamType
from retentioneering.tooling.mixins.ended_events import EndedEventsMixin


[docs]class StepSankey(EndedEventsMixin): """ A class for the visualization of user paths in stepwise manner using Sankey diagram. Parameters ---------- eventstream : EventstreamType See Also -------- .Eventstream.step_sankey : Call StepSankey tool as an eventstream method. .CollapseLoops : Find loops and create new synthetic events in the paths of all users having such sequences. .StepMatrix : This class provides methods for step matrix calculation and visualization. Notes ----- See :doc:`StepSankey user guide</user_guides/step_sankey>` for the details. """ max_steps: int threshold: int | float sorting: list | None targets: list[str] | str | None data_grp_nodes: pd.DataFrame data: pd.DataFrame data_grp_links: pd.DataFrame data_for_plot: dict @time_performance( scope="step_sankey", 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.event_index_col = self.__eventstream.schema.event_index self.data_grp_nodes = pd.DataFrame() self.data = pd.DataFrame() self.data_grp_links = pd.DataFrame() self.data_for_plot = {} @staticmethod def _make_color( event: str, all_events: list, palette: list, ) -> tuple[int, int, int]: """ It is a color picking function Parameters ---------- event : str An event for color setting all_events : list A list of all events palette : list A list of colors Returns ------- str A picked color for certain event """ return palette[list(all_events).index(event)] @staticmethod def _round_up( n: float, dec: float, ) -> float: """ Rounds the value up to the nearest value assuming a grid with ``dec`` step. E.g. ``_round_up(0.51, 0.05) = 0.55``, ``_round_up(0.55, 0.05) = 0.6`` Parameters ---------- n : float A number to round up dec : float A decimal for correct rounding up Returns ------- float Rounded value """ return round(n - n % dec + dec, 2) def _get_nodes_positions(self, df: pd.DataFrame) -> tuple[list[float], list[float]]: """ It is a function for placing nodes at the x and y coordinates of plotly lib plot canvas. Parameters ---------- df : pandas Dataframe A dataframe that contains aggregated information about the nodes. Returns ------- tuple[list[float], list[float]] Two lists with the corresponding coordinates x and y. """ # NOTE get x axis length x_len = len(df["step"].unique()) # NOTE declare positions x_positions = [] y_positions = [] # NOTE get maximum range for placing middle points y_range = 0.95 - 0.05 # NOTE going inside ranked events for step in sorted(df["step"].unique()): # NOTE placing x-axis points as well for _ in df[df["step"] == step][self.event_col]: x_positions.append([round(x, 2) for x in np.linspace(0.05, 0.95, x_len)][step - 1]) # NOTE it always works very well if you have less than 4 values at current rank y_len = len(df[df["step"] == step][self.event_col]) # NOTE at this case using came positions as x-axis because we don't need to calculate something more if y_len < 4: for p in [round(y, 2) for y in np.linspace(0.05, 0.95, y_len)]: y_positions.append(p) # NOTE jumping in to complex part else: # NOTE total sum for understanding do we need extra step size or not total_sum = df[df["step"] == step]["usr_cnt"].sum() # NOTE step size for middle points y_step = round(y_range / total_sum, 2) # NOTE cumulative sum for understanding do we need use default step size or not cumulative_sum = 0 # NOTE ENDED action ended_sum = df[(df["step"] == step) & (df[self.event_col] == "ENDED")]["usr_cnt"].sum() last_point = self._round_up(ended_sum / total_sum, 0.05) iterate_sum = 0 # NOTE going deeper inside each event for n, event in enumerate(df[df["step"] == step][self.event_col]): # NOTE placing first event at first possible position if n == 0: y_positions.append(0.05) # NOTE placing last event at last possible position elif n + 1 == y_len: y_positions.append(0.95) # NOTE placing middle points else: # NOTE we found out that 70% of total sum is the best cap for doing this case if iterate_sum / total_sum > 0.2 and event != "ENDED": # NOTE placing first point after the biggest one at the next position # but inside [.1; .3] range y_positions.append( round(y_positions[-1] + np.minimum(np.maximum(y_step * iterate_sum, 0.1), 0.3), 2) ) # NOTE placing points after the biggest else: # NOTE placing little points at the all available space y_positions.append( round(y_positions[-1] + (0.95 - last_point - y_positions[-1]) / (y_len - n), 2) ) # NOTE set sum for next step iterate_sum = df[(df["step"] == step) & (df[self.event_col] == event)]["usr_cnt"].to_numpy()[0] # NOTE update cumulative sum cumulative_sum += iterate_sum return x_positions, y_positions def _pad_end_events(self, data: pd.DataFrame) -> pd.DataFrame: """ If the number of events in a user's path is less than self.max_steps, then the function pads the path with multiply ENDED events. It is required for correct visualization of the trajectories which are shorter than self.max_steps. """ pad = ( data.groupby(self.user_col, as_index=False)[self.event_col] .count() .loc[lambda df_: df_[self.event_col] < self.max_steps] # type: ignore .assign(repeat_number=lambda df_: self.max_steps - df_[self.event_col]) ) repeats = pd.DataFrame({self.user_col: np.repeat(pad[self.user_col], pad["repeat_number"])}) padded_end_events = pd.merge(repeats, data[data[self.event_col] == "ENDED"], on=self.user_col) result = pd.concat([data, padded_end_events]).sort_values([self.user_col, self.event_index_col]) return result def _prepare_data(self, data: pd.DataFrame) -> pd.DataFrame: data = self._add_ended_events( data=data, schema=self.__eventstream.schema, weight_col=self.__eventstream.schema.user_id ) data = self._pad_end_events(data) # NOTE set new columns using declared functions data[self.time_col] = pd.to_datetime(data[self.time_col]) data["step"] = data.groupby(self.user_col)[self.event_index_col].rank(method="first").astype(int) data = data.sort_values(by=["step", self.time_col]).reset_index(drop=True) data = self._get_next_event_and_timedelta(data) # NOTE threshold data["event_users"] = data.groupby(by=["step", self.event_col])[self.user_col].transform("nunique") data["total_users"] = data.loc[data["step"] == 1, self.user_col].nunique() data["perc"] = data["event_users"] / data["total_users"] if isinstance(self.threshold, float): column_to_compare = "perc" else: # assume that self.threshold must be of int type here column_to_compare = "event_users" events_to_keep = ["ENDED"] if self.targets is not None: events_to_keep += self.targets threshold_events = ( data.loc[data["step"] <= self.max_steps, :] .groupby(by=self.event_col, as_index=False)[column_to_compare] .max() .loc[ lambda df_: (df_[column_to_compare] <= self.threshold) & (~df_[self.event_col].isin(events_to_keep)) ] # type: ignore .loc[:, self.event_col] ) data.loc[data[self.event_col].isin(threshold_events), self.event_col] = f"thresholded_{len(threshold_events)}" # NOTE rearrange the data taking into account recently added thresholded events data["step"] = data.groupby(self.user_col)[self.event_index_col].rank(method="first").astype(int) data = self._get_next_event_and_timedelta(data) # NOTE use max_steps for filtering data data = data.loc[data["step"] <= self.max_steps, :] # TODO: Do we really need to replace NA values? # NOTE skip mean calculating error data["time_to_next"].fillna(data["time_to_next"].min(), inplace=True) return data def _render_plot( self, data_for_plot: dict, data_grp_nodes: pd.DataFrame, autosize: bool = True, width: int | None = None, height: int | None = None, ) -> go.Figure: # NOTE fill lists for plot targets = [] sources = [] values = [] time_to_next = [] for source_key in data_for_plot["links_dict"].keys(): for target_key, target_value in data_for_plot["links_dict"][source_key].items(): sources.append(source_key) targets.append(target_key) values.append(target_value["unique_users"]) time_to_next.append( str(pd.to_timedelta(target_value["avg_time_to_next"] / target_value["unique_users"])).split(".")[0] ) # NOTE fill another lists for plot labels = [] colors = [] percs = [] for key in data_for_plot["nodes_dict"].keys(): labels += list(data_for_plot["nodes_dict"][key]["sources"]) colors += list(data_for_plot["nodes_dict"][key]["color"]) percs += list(data_for_plot["nodes_dict"][key]["percs"]) # NOTE get colors for plot for idx, color in enumerate(colors): colors[idx] = "rgb" + str(color) + "" # NOTE get positions for plot x, y = self._get_nodes_positions(df=data_grp_nodes) # NOTE make plot fig = go.Figure( data=[ go.Sankey( arrangement="snap", node=dict( thickness=15, line=dict(color="black", width=0.5), label=labels, color=colors, customdata=percs, hovertemplate="Total unique users: %{value} (%{customdata}% of total)<extra></extra>", x=x, y=y, pad=20, ), link=dict( source=sources, target=targets, value=values, label=time_to_next, hovertemplate="%{value} unique users went from %{source.label} to %{target.label}.<br />" + "<br />It took them %{label} in average.<extra></extra>", ), ) ] ) fig.update_layout(font=dict(size=15), plot_bgcolor="white", autosize=autosize, width=width, height=height) return fig def _get_links( self, data: pd.DataFrame, data_for_plot: dict, data_grp_nodes: pd.DataFrame ) -> tuple[dict, pd.DataFrame]: # NOTE create links aggregated dataframe data_grp_links = ( data[data["step"] <= self.max_steps - 1] .groupby(by=["step", self.event_col, "next_event"])[[self.user_col, "time_to_next"]] .agg({self.user_col: ["count"], "time_to_next": ["sum"]}) .reset_index() .rename(columns={self.user_col: "usr_cnt", "time_to_next": "time_to_next_sum"}) ) data_grp_links.columns = data_grp_links.columns.droplevel(1) data_grp_links = data_grp_links.merge( data_grp_nodes[["step", self.event_col, "index"]], how="inner", on=["step", self.event_col], ) data_grp_links.loc[:, "next_step"] = data_grp_links["step"] + 1 data_grp_links = data_grp_links.merge( data_grp_nodes[["step", self.event_col, "index"]].rename( columns={"step": "next_step", self.event_col: "next_event", "index": "next_index"} ), how="inner", on=["next_step", "next_event"], ) data_grp_links.sort_values(by=["index", "usr_cnt"], ascending=[True, False], inplace=True) data_grp_links.reset_index(drop=True, inplace=True) # NOTE generating links plot dict data_for_plot.update({"links_dict": dict()}) for index in data_grp_links["index"].unique(): for next_index in data_grp_links[data_grp_links["index"] == index]["next_index"].unique(): _unique_users, _avg_time_to_next = ( data_grp_links.loc[ (data_grp_links["index"] == index) & (data_grp_links["next_index"] == next_index), ["usr_cnt", "time_to_next_sum"], ] .to_numpy() .T ) if index in data_for_plot["links_dict"]: if next_index in data_for_plot["links_dict"][index]: data_for_plot["links_dict"][index][next_index]["unique_users"] = _unique_users[0] data_for_plot["links_dict"][index][next_index]["avg_time_to_next"] = np.timedelta64( _avg_time_to_next[0] ) else: data_for_plot["links_dict"][index].update( { next_index: { "unique_users": _unique_users[0], "avg_time_to_next": np.timedelta64(_avg_time_to_next[0]), } } ) else: data_for_plot["links_dict"].update( { index: { next_index: { "unique_users": _unique_users[0], "avg_time_to_next": np.timedelta64(_avg_time_to_next[0]), } } } ) return data_for_plot, data_grp_links def _get_nodes(self, data: pd.DataFrame) -> tuple[dict, pd.DataFrame]: all_events = list(data[self.event_col].unique()) palette = self._prepare_palette(all_events) # NOTE create nodes aggregate dataframe data_grp_nodes = ( data.groupby(by=["step", self.event_col])[self.user_col] .nunique() .reset_index() .rename(columns={self.user_col: "usr_cnt"}) ) data_grp_nodes.loc[:, "usr_cnt_total"] = data_grp_nodes.groupby(by=["step"])["usr_cnt"].transform("sum") data_grp_nodes.loc[:, "perc"] = np.round( (data_grp_nodes.loc[:, "usr_cnt"] / data_grp_nodes.loc[:, "usr_cnt_total"]) * 100, 2 ) data_grp_nodes.sort_values( by=["step", "usr_cnt", self.event_col], ascending=[True, False, True], inplace=True, ) data_grp_nodes.reset_index( drop=True, inplace=True, ) data_grp_nodes.loc[:, "color"] = data_grp_nodes[self.event_col].apply( lambda x: self._make_color(x, all_events, palette) ) data_grp_nodes.loc[:, "index"] = data_grp_nodes.index # type: ignore # NOTE doing right ranking if self.sorting is None: data_grp_nodes.loc[:, "sorting"] = 100 else: for n, s in enumerate(self.sorting): data_grp_nodes.loc[data_grp_nodes[self.event_col] == s, "sorting"] = n data_grp_nodes.loc[:, "sorting"].fillna(100, inplace=True) # NOTE placing ENDED at the end data_grp_nodes.loc[data_grp_nodes[self.event_col] == "ENDED", "sorting"] = 101 # NOTE using custom ordering data_grp_nodes.loc[:, "sorting"] = data_grp_nodes.loc[:, "sorting"].astype(int) # @TODO: step variable is not used inside the loop. The loop might be invalid. Vladimir Kukushkin # NOTE doing loop for valid ranking for step in data_grp_nodes["step"].unique(): # NOTE saving last level order data_grp_nodes.loc[:, "order_by"] = ( data_grp_nodes.groupby(by=[self.event_col])["index"].transform("shift").fillna(100).astype(int) ) # NOTE placing ENDED events at the end data_grp_nodes.loc[data_grp_nodes[self.event_col] == "ENDED", "sorting"] = 101 # NOTE creating new indexes data_grp_nodes.sort_values( by=["step", "sorting", "order_by", "usr_cnt", self.event_col], ascending=[True, True, True, False, True], inplace=True, ) data_grp_nodes.reset_index( drop=True, inplace=True, ) data_grp_nodes.loc[:, "index"] = data_grp_nodes.index # type: ignore # NOTE generating nodes plot dict data_for_plot: Dict[str, Any] = dict() data_for_plot.update({"nodes_dict": dict()}) for step in data_grp_nodes["step"].unique(): data_for_plot["nodes_dict"].update({step: dict()}) _sources, _color, _sources_index, _percs = ( data_grp_nodes.loc[data_grp_nodes["step"] == step, [self.event_col, "color", "index", "perc"]] .to_numpy() .T ) data_for_plot["nodes_dict"][step].update( { "sources": list(_sources), "color": list(_color), "sources_index": list(_sources_index), "percs": list(_percs), } ) return data_for_plot, data_grp_nodes @staticmethod def _prepare_palette(all_events: list) -> list[tuple]: # NOTE default color palette palette_hex = ["50BE97", "E4655C", "FCC865", "BFD6DE", "3E5066", "353A3E", "E6E6E6"] # NOTE convert HEX to RGB palette = [] for color in palette_hex: rgb_color = tuple(int(color[i : i + 2], 16) for i in (0, 2, 4)) palette.append(rgb_color) # NOTE extend color palette if number of events more than default colors list complementary_palette = sns.color_palette("deep", len(all_events) - len(palette)) if len(complementary_palette) > 0: colors = complementary_palette.as_hex() for c in colors: col = c[1:] palette.append(tuple(int(col[i : i + 2], 16) for i in (0, 2, 4))) return palette def _get_next_event_and_timedelta(self, data: pd.DataFrame) -> pd.DataFrame: grouped = data.groupby(self.user_col) data["next_event"] = grouped[self.event_col].shift(-1) data["next_timestamp"] = grouped[self.time_col].shift(-1) data["time_to_next"] = data["next_timestamp"] - data[self.time_col] data = data.drop("next_timestamp", axis=1) return data
[docs] @time_performance(scope="step_sankey", event_name="fit") def fit( self, max_steps: int = 10, threshold: int | float = 0.05, sorting: list | None = None, targets: list[str] | str | None = None, ) -> None: """ Calculate the sankey diagram internal values with the defined parameters. Applying ``fit`` method is necessary for the following usage of any visualization or descriptive ``StepSankey`` methods. Parameters ---------- max_steps : int, default 10 Maximum number of steps in trajectories to include. Should be > 1. threshold : float | int, default 0.05 Used to remove rare events from the plot. An event is collapsed to ``thresholded_N`` artificial event if its maximum frequency across all the steps is less than or equal to ``threshold``. The frequency is set with respect to ``threshold`` type: - If ``int`` - the frequency is the number of unique users who had given event at given step. - If ``float`` - percentage of users: the same as for ``int``, but divided by the number of unique users. The events which are prohibited for collapsing could be enlisted in ``target`` parameter. sorting : list of str, optional Define the order of the events visualized at each step. The events that are not represented in the list will follow after the events from the list. targets : str or list of str, optional Contain events that are prohibited for collapsing with ``threshold`` parameter. Raises ------ ValueError If ``max_steps`` parameter is <= 1. """ data = self.__eventstream.to_dataframe(copy=True)[ [self.user_col, self.event_col, self.time_col, self.event_index_col] ] if max_steps <= 1: raise ValueError("max_steps parameter must be > 1!") called_params = { "max_steps": max_steps, "threshold": threshold, "sorting": sorting, "targets": targets, } self.max_steps = max_steps self.threshold = threshold self.sorting = sorting self.targets = targets self.data = self._prepare_data(data) data_for_plot, self.data_grp_nodes = self._get_nodes(self.data) self.data_for_plot, self.data_grp_links = self._get_links(self.data, data_for_plot, self.data_grp_nodes) performance_info = { "nodes_count": len(self.data_grp_nodes), "links_count": len(self.data_grp_links), } collect_data_performance( scope="step_sankey", event_name="metadata", called_params=called_params, performance_data=performance_info, eventstream_index=self.__eventstream._eventstream_index, )
[docs] @time_performance( scope="step_sankey", event_name="plot", ) def plot(self, autosize: bool = True, width: int | None = None, height: int | None = None) -> go.Figure: """ Create a Sankey interactive plot based on the calculated values. Should be used after :py:func:`fit`. Parameters ---------- autosize : bool, default True Plotly autosize parameter. See :plotly_autosize:`plotly documentation<>`. width : int, optional Plot's width (in px). See :plotly_width:`plotly documentation<>`. height : int, optional Plot's height (in px). See :plotly_height:`plotly documentation<>`. Returns ------- plotly.graph_objects.Figure """ called_params = { "autosize": autosize, "width": width, "height": height, } figure = self._render_plot( data_for_plot=self.data_for_plot, data_grp_nodes=self.data_grp_nodes, autosize=autosize, width=width, height=height, ) performance_info = { "nodes_count": len(self.data_grp_nodes), "links_count": len(self.data_grp_links), } collect_data_performance( scope="step_sankey", event_name="metadata", called_params=called_params, performance_data=performance_info, eventstream_index=self.__eventstream._eventstream_index, ) return figure
@property @time_performance( scope="step_sankey", event_name="values", ) def values(self) -> tuple[pd.DataFrame, pd.DataFrame]: """ Returns two pd.DataFrames which the Sankey diagram is based on. Should be used after :py:func:`fit`. Returns ------- tuple[pd.DataFrame, pd.DataFrame] 1. Contains the nodes of the diagram. 2. Contains the edges of the diagram. """ return self.data_grp_nodes, self.data_grp_links @property @time_performance( scope="step_sankey", event_name="params", ) def params(self) -> dict: """ Returns the parameters used for the last fitting. Should be used after :py:func:`fit`. """ return { "max_steps": self.max_steps, "threshold": self.threshold, "sorting": self.sorting, "targets": self.targets, }