Add sleep intraday features with PRICE provider
Meng Li
parent
2d5e966530
commit
d74196cab4
14
config.yaml
14
config.yaml
|
|
@ -400,6 +400,20 @@ FITBIT_SLEEP_INTRADAY:
|
|||
REFERENCE_TIME: MIDNIGHT # chosen from "MIDNIGHT" and "START_OF_THE_SEGMENT"
|
||||
SRC_FOLDER: "rapids" # inside src/features/fitbit_sleep_intraday
|
||||
SRC_LANGUAGE: "python"
|
||||
|
||||
PRICE:
|
||||
COMPUTE: False
|
||||
FEATURES: [avgduration, avgratioduration, avgstarttimeofepisodemain, avgendtimeofepisodemain, avgmidpointofepisodemain, "stdstarttimeofepisodemain", "stdendtimeofepisodemain", "stdmidpointofepisodemain", socialjetlag, meanssdstarttimeofepisodemain, meanssdendtimeofepisodemain, meanssdmidpointofepisodemain, medianssdstarttimeofepisodemain, medianssdendtimeofepisodemain, medianssdmidpointofepisodemain]
|
||||
SLEEP_LEVELS:
|
||||
CLASSIC: [awake, restless, asleep]
|
||||
STAGES: [wake, deep, light, rem]
|
||||
UNIFIED: [awake, asleep]
|
||||
DAY_TYPES: [WEEKEND, WEEK, ALL]
|
||||
GROUP_EPISODES_WITHIN: # by default: today's 6pm to tomorrow's noon
|
||||
START_TIME: 1080 # number of minutes after the midnight (18:00) 18*60
|
||||
LENGTH: 1080 # in minutes (18 hours) 18*60
|
||||
SRC_FOLDER: "price" # inside src/features/fitbit_sleep_intraday
|
||||
SRC_LANGUAGE: "python"
|
||||
|
||||
# See https://www.rapids.science/latest/features/fitbit-steps-summary/
|
||||
FITBIT_STEPS_SUMMARY:
|
||||
|
|
|
|||
|
|
@ -0,0 +1,242 @@
|
|||
import pandas as pd
|
||||
import itertools
|
||||
|
||||
|
||||
|
||||
def featuresFullNames(intraday_features_to_compute, sleep_levels_to_compute, day_types_to_compute):
|
||||
|
||||
features_fullnames = ["local_segment"]
|
||||
|
||||
sleep_level_with_group = []
|
||||
for sleep_level_group in sleep_levels_to_compute:
|
||||
for sleep_level in sleep_levels_to_compute[sleep_level_group]:
|
||||
sleep_level_with_group.append(sleep_level + sleep_level_group.lower())
|
||||
|
||||
for feature in intraday_features_to_compute:
|
||||
if feature == "avgduration":
|
||||
features_fullnames.extend(["avgduration" + x[0] + "main" + x[1].lower() for x in itertools.product(sleep_level_with_group, day_types_to_compute)])
|
||||
elif feature == "avgratioduration":
|
||||
features_fullnames.extend(["avgratioduration" + x[0] + "withinmain" + x[1].lower() for x in itertools.product(sleep_level_with_group, day_types_to_compute)])
|
||||
elif feature in ["avgstarttimeofepisodemain", "avgendtimeofepisodemain", "avgmidpointofepisodemain", "stdstarttimeofepisodemain", "stdendtimeofepisodemain", "stdmidpointofepisodemain"]:
|
||||
features_fullnames.extend([feature + x.lower() for x in day_types_to_compute])
|
||||
else:
|
||||
features_fullnames.append(feature)
|
||||
|
||||
return features_fullnames
|
||||
|
||||
def mergeSleepEpisodes(sleep_data, cols_for_groupby, base_sleep_levels):
|
||||
|
||||
sleep_level_with_group = []
|
||||
for sleep_level_group in base_sleep_levels:
|
||||
for sleep_level in base_sleep_levels[sleep_level_group]:
|
||||
sleep_level_with_group.append(sleep_level + sleep_level_group.lower())
|
||||
|
||||
sleep_episodes = pd.DataFrame(columns=["local_segment", "durationinbed", "start_timestamp", "end_timestamp", "local_start_date_time", "local_end_date_time"] + ["duration" + x for x in sleep_level_with_group])
|
||||
|
||||
if cols_for_groupby and (not sleep_data.empty):
|
||||
sleep_data = sleep_data.groupby(by=cols_for_groupby)
|
||||
sleep_episodes = sleep_data[["duration"]].sum().rename(columns={"duration": "durationinbed"})
|
||||
|
||||
sleep_episodes["start_timestamp"] = sleep_data["start_timestamp"].first()
|
||||
sleep_episodes["end_timestamp"] = sleep_data["end_timestamp"].last()
|
||||
|
||||
sleep_episodes["local_start_date_time"] = sleep_data["local_start_date_time"].first()
|
||||
sleep_episodes["local_end_date_time"] = sleep_data["local_end_date_time"].last()
|
||||
|
||||
for sleep_level in sleep_level_with_group:
|
||||
sleep_episodes["duration" + sleep_level] = sleep_data.apply(lambda group: group[group["level"] == sleep_level.replace("classic", "").replace("stages", "").replace("unified", "")]["duration"].sum())
|
||||
|
||||
sleep_episodes.reset_index(inplace=True, drop=False)
|
||||
del sleep_episodes["type_episode_id"]
|
||||
|
||||
return sleep_episodes
|
||||
|
||||
def extractDailyFeatures(sleep_data):
|
||||
daily_grouped = sleep_data.groupby(["local_segment", "fake_date"])
|
||||
daily_features = daily_grouped[["start_minutes"]].first().rename(columns={"start_minutes": "starttimeofepisodemain"})
|
||||
daily_features["endtimeofepisodemain"] = daily_grouped["end_minutes"].last()
|
||||
daily_features["midpointofepisodemain"] = (daily_features["starttimeofepisodemain"] + daily_features["endtimeofepisodemain"]) / 2
|
||||
daily_features["durationinbedmain"] = daily_grouped["durationinbed"].sum()
|
||||
|
||||
for col in sleep_data.columns:
|
||||
if col.startswith("duration") and col != "durationinbed":
|
||||
daily_features[col + "main"] = daily_grouped[col].sum().fillna(0)
|
||||
daily_features["ratio" + col + "withinmain"] = daily_features[col + "main"] / daily_features["durationinbedmain"]
|
||||
daily_features.reset_index(inplace=True)
|
||||
|
||||
# The day of the week with Monday=0, Sunday=6. Set Friday and Saturday as Weekend, others as Weekday.
|
||||
daily_features["is_weekend"] = pd.to_datetime(daily_features["fake_date"]).dt.dayofweek.apply(lambda x: 1 if (x == 4 or x == 5) else 0)
|
||||
|
||||
return daily_features
|
||||
|
||||
def statsOfDailyFeatures(daily_features, day_type, sleep_levels, intraday_features_to_compute, sleep_intraday_features):
|
||||
if day_type == "WEEKEND":
|
||||
daily_features = daily_features[daily_features["is_weekend"] == 0]
|
||||
elif day_type == "WEEK":
|
||||
daily_features = daily_features[daily_features["is_weekend"] == 1]
|
||||
elif day_type == "ALL":
|
||||
pass
|
||||
else:
|
||||
raise ValueError("Please make sure the [FITBIT_SLEEP_INTRADAY][PROVIDERS][PRICE][DAY_TYPES] parameter in config.yaml file only contains the subset of [WEEKEND, WEEK, ALL].")
|
||||
|
||||
if daily_features.empty:
|
||||
return sleep_intraday_features
|
||||
|
||||
if sleep_intraday_features.empty:
|
||||
sleep_intraday_features = pd.DataFrame()
|
||||
|
||||
# Average of time related features
|
||||
if "avgstarttimeofepisodemain" in intraday_features_to_compute:
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features[["local_segment","starttimeofepisodemain"]].groupby("local_segment")["starttimeofepisodemain"].mean().to_frame().rename(columns={"starttimeofepisodemain": "avgstarttimeofepisodemain" + day_type.lower()})], axis=1)
|
||||
if "avgendtimeofepisodemain" in intraday_features_to_compute:
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features[["local_segment","endtimeofepisodemain"]].groupby("local_segment")["endtimeofepisodemain"].mean().to_frame().rename(columns={"endtimeofepisodemain": "avgendtimeofepisodemain" + day_type.lower()})], axis=1)
|
||||
if "avgmidpointofepisodemain" in intraday_features_to_compute:
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features[["local_segment","midpointofepisodemain"]].groupby("local_segment")["midpointofepisodemain"].mean().to_frame().rename(columns={"midpointofepisodemain": "avgmidpointofepisodemain" + day_type.lower()})], axis=1)
|
||||
|
||||
# Std of time related features
|
||||
if "stdstarttimeofepisodemain" in intraday_features_to_compute:
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features[["local_segment","starttimeofepisodemain"]].groupby("local_segment")["starttimeofepisodemain"].std().to_frame().rename(columns={"starttimeofepisodemain": "stdstarttimeofepisodemain" + day_type.lower()})], axis=1)
|
||||
if "stdendtimeofepisodemain" in intraday_features_to_compute:
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features[["local_segment","endtimeofepisodemain"]].groupby("local_segment")["endtimeofepisodemain"].std().to_frame().rename(columns={"endtimeofepisodemain": "stdendtimeofepisodemain" + day_type.lower()})], axis=1)
|
||||
if "stdmidpointofepisodemain" in intraday_features_to_compute:
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features[["local_segment","midpointofepisodemain"]].groupby("local_segment")["midpointofepisodemain"].std().to_frame().rename(columns={"midpointofepisodemain": "stdmidpointofepisodemain" + day_type.lower()})], axis=1)
|
||||
|
||||
# Duration & Ratio features
|
||||
for sleep_level_group in sleep_levels:
|
||||
for sleep_level in sleep_levels[sleep_level_group]:
|
||||
if "avgduration" in intraday_features_to_compute:
|
||||
col = "duration" + sleep_level + sleep_level_group.lower() + "main"
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features[["local_segment", col]].groupby("local_segment")[col].mean().to_frame().rename(columns={col: "avg" + col + day_type.lower()})], axis=1)
|
||||
if "avgratioduration" in intraday_features_to_compute:
|
||||
col = "ratioduration" + sleep_level + sleep_level_group.lower() + "withinmain"
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features[["local_segment", col]].groupby("local_segment")[col].mean().to_frame().rename(columns={col: "avg" + col + day_type.lower()})], axis=1)
|
||||
|
||||
return sleep_intraday_features
|
||||
|
||||
def socialJetLagFeature(daily_features, sleep_intraday_features):
|
||||
daily_features_weekend = daily_features[daily_features["is_weekend"] == 1]
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features_weekend[["local_segment","starttimeofepisodemain"]].groupby("local_segment")["starttimeofepisodemain"].mean().to_frame().rename(columns={"starttimeofepisodemain": "helper_weekend"})], axis=1)
|
||||
|
||||
daily_features_weekday = daily_features[daily_features["is_weekend"] == 0]
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features_weekday[["local_segment","starttimeofepisodemain"]].groupby("local_segment")["starttimeofepisodemain"].mean().to_frame().rename(columns={"starttimeofepisodemain": "helper_weekday"})], axis=1)
|
||||
|
||||
sleep_intraday_features["socialjetlag"] = sleep_intraday_features["helper_weekend"] - sleep_intraday_features["helper_weekday"]
|
||||
|
||||
for col in ["helper_weekend", "helper_weekday"]:
|
||||
del sleep_intraday_features[col]
|
||||
|
||||
return sleep_intraday_features
|
||||
|
||||
def MSSDFeatures(daily_features, intraday_features_to_compute, sleep_intraday_features):
|
||||
|
||||
date_idx = pd.DataFrame(pd.date_range(start=daily_features["fake_date"].min(), end=daily_features["fake_date"].max(), freq="D"), columns=["fake_date"])
|
||||
date_idx["fake_date"] = date_idx["fake_date"].dt.date
|
||||
daily_features = daily_features.merge(date_idx, on="fake_date", how="right")
|
||||
|
||||
for col in ["starttimeofepisodemain", "endtimeofepisodemain", "midpointofepisodemain"]:
|
||||
daily_features[col + "_diff"] = daily_features[col].diff()
|
||||
|
||||
if "meanssdstarttimeofepisodemain" in intraday_features_to_compute:
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features[["local_segment","starttimeofepisodemain_diff"]].groupby("local_segment")["starttimeofepisodemain_diff"].mean().to_frame().rename(columns={"starttimeofepisodemain_diff": "meanssdstarttimeofepisodemain"})], axis=1)
|
||||
if "meanssdendtimeofepisodemain" in intraday_features_to_compute:
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features[["local_segment","endtimeofepisodemain_diff"]].groupby("local_segment")["endtimeofepisodemain_diff"].mean().to_frame().rename(columns={"endtimeofepisodemain_diff": "meanssdendtimeofepisodemain"})], axis=1)
|
||||
if "meanssdmidpointofepisodemain" in intraday_features_to_compute:
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features[["local_segment","midpointofepisodemain_diff"]].groupby("local_segment")["midpointofepisodemain_diff"].mean().to_frame().rename(columns={"midpointofepisodemain_diff": "meanssdmidpointofepisodemain"})], axis=1)
|
||||
|
||||
if "medianssdstarttimeofepisodemain" in intraday_features_to_compute:
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features[["local_segment","starttimeofepisodemain_diff"]].groupby("local_segment")["starttimeofepisodemain_diff"].median().to_frame().rename(columns={"starttimeofepisodemain_diff": "medianssdstarttimeofepisodemain"})], axis=1)
|
||||
if "medianssdendtimeofepisodemain" in intraday_features_to_compute:
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features[["local_segment","endtimeofepisodemain_diff"]].groupby("local_segment")["endtimeofepisodemain_diff"].median().to_frame().rename(columns={"endtimeofepisodemain_diff": "medianssdendtimeofepisodemain"})], axis=1)
|
||||
if "medianssdmidpointofepisodemain" in intraday_features_to_compute:
|
||||
sleep_intraday_features = pd.concat([sleep_intraday_features, daily_features[["local_segment","midpointofepisodemain_diff"]].groupby("local_segment")["midpointofepisodemain_diff"].median().to_frame().rename(columns={"midpointofepisodemain_diff": "medianssdmidpointofepisodemain"})], axis=1)
|
||||
|
||||
return sleep_intraday_features
|
||||
|
||||
|
||||
|
||||
|
||||
def price_features(sensor_data_files, time_segment, provider, filter_data_by_segment, *args, **kwargs):
|
||||
|
||||
daily_start_time = provider["GROUP_EPISODES_WITHIN"]["START_TIME"]
|
||||
daily_end_time = daily_start_time + provider["GROUP_EPISODES_WITHIN"]["LENGTH"]
|
||||
|
||||
sleep_intraday_data = pd.read_csv(sensor_data_files["sensor_data"])
|
||||
requested_intraday_features = provider["FEATURES"]
|
||||
requested_sleep_levels = provider["SLEEP_LEVELS"]
|
||||
requested_day_types = provider["DAY_TYPES"]
|
||||
|
||||
# Name of the features this function can compute
|
||||
base_intraday_features = ["avgduration", "avgratioduration", "avgstarttimeofepisodemain", "avgendtimeofepisodemain", "avgmidpointofepisodemain", "stdstarttimeofepisodemain", "stdendtimeofepisodemain", "stdmidpointofepisodemain", "socialjetlag", "meanssdstarttimeofepisodemain", "meanssdendtimeofepisodemain", "meanssdmidpointofepisodemain", "medianssdstarttimeofepisodemain", "medianssdendtimeofepisodemain", "medianssdmidpointofepisodemain"]
|
||||
base_sleep_levels = {"CLASSIC": ["awake", "restless", "asleep"],
|
||||
"STAGES": ["wake", "deep", "light", "rem"],
|
||||
"UNIFIED": ["awake", "asleep"]}
|
||||
base_day_types = ["WEEKEND", "WEEK", "ALL"]
|
||||
|
||||
# The subset of requested features this function can compute
|
||||
intraday_features_to_compute = list(set(requested_intraday_features) & set(base_intraday_features))
|
||||
sleep_levels_to_compute = {key: list(set(requested_sleep_levels[key]) & set(base_sleep_levels[key])) for key in requested_sleep_levels if key in base_sleep_levels}
|
||||
day_types_to_compute = list(set(requested_day_types) & set(base_day_types))
|
||||
|
||||
# Full names
|
||||
features_fullnames = featuresFullNames(intraday_features_to_compute, sleep_levels_to_compute, day_types_to_compute)
|
||||
sleep_intraday_features = pd.DataFrame(columns=features_fullnames)
|
||||
|
||||
# Filter by segemnts and chunk episodes
|
||||
sleep_intraday_data = filter_data_by_segment(sleep_intraday_data, time_segment)
|
||||
|
||||
# Discard segments shorter than one day
|
||||
sleep_intraday_data["segment_length"] = (sleep_intraday_data["segment_end_timestamp"] - sleep_intraday_data["segment_start_timestamp"]) / 1000 # in seconds
|
||||
sleep_intraday_data = sleep_intraday_data[sleep_intraday_data["segment_length"] >= 24 * 60 * 60 - 1]
|
||||
del sleep_intraday_data["segment_length"]
|
||||
|
||||
# Select main sleep records
|
||||
sleep_intraday_data = sleep_intraday_data[sleep_intraday_data["is_main_sleep"] == 1]
|
||||
|
||||
if sleep_intraday_data.empty:
|
||||
return sleep_intraday_features
|
||||
|
||||
# Merge rows to get sleep episodes
|
||||
main_sleep_episodes = mergeSleepEpisodes(sleep_intraday_data, ["local_segment", "type_episode_id"], base_sleep_levels)
|
||||
|
||||
# Extract number of minutes after midnight as start time; add duration to get the end time
|
||||
main_sleep_episodes["start_minutes"] = main_sleep_episodes["local_start_date_time"].apply(lambda x: x.hour * 60 + x.minute + x.second / 60)
|
||||
main_sleep_episodes["end_minutes"] = main_sleep_episodes["start_minutes"] + main_sleep_episodes["durationinbed"]
|
||||
# Extract fake date
|
||||
""" The rule used for fake date extraction
|
||||
set DS = daily_start_time, DE = daily_end_time
|
||||
set start = start_minutes, end = end_minutes
|
||||
|
||||
if (DS <= start < DE) or (DS < end <= DE) or (start <= DS and end >= DE):
|
||||
assign today
|
||||
elif if end <= DS:
|
||||
assign yesterday
|
||||
else: (same as start >=DE)
|
||||
assign tomorrow
|
||||
"""
|
||||
main_sleep_episodes["fake_date_delta"] = main_sleep_episodes[["start_minutes", "end_minutes"]].apply(lambda row: 0 if ((row["start_minutes"] >= daily_start_time and row["start_minutes"] < daily_end_time) or (row["end_minutes"] > daily_start_time and row["end_minutes"] <= daily_end_time) or (row["start_minutes"] <= daily_start_time and row["end_minutes"] >= daily_end_time)) else -1 if (row["end_minutes"] <= daily_start_time) else 1, axis=1)
|
||||
main_sleep_episodes["fake_date"] = (main_sleep_episodes["local_start_date_time"] + pd.to_timedelta(main_sleep_episodes["fake_date_delta"], unit="d")).dt.date
|
||||
|
||||
# Update "start_minutes" column based on START_TIME
|
||||
main_sleep_episodes["start_minutes"] = main_sleep_episodes[["start_minutes", "fake_date_delta"]].apply(lambda row: row["start_minutes"] - 24 * 60 * row["fake_date_delta"], axis=1)
|
||||
main_sleep_episodes["end_minutes"] = main_sleep_episodes["start_minutes"] + main_sleep_episodes["durationinbed"]
|
||||
|
||||
# We keep a sleep episode that intersets or contains the period between [START_TIME, START_TIME + LENGTH], aka [daily_start_time, daily_end_time].
|
||||
main_sleep_episodes = main_sleep_episodes.query("(start_minutes >= @daily_start_time and start_minutes < @daily_end_time) or (end_minutes > @daily_start_time and end_minutes <= @daily_end_time) or (start_minutes <= @daily_start_time and end_minutes >= @daily_end_time)")
|
||||
|
||||
# Sort main sleep episodes based on fake_date and start_minutes
|
||||
main_sleep_episodes = main_sleep_episodes.sort_values(["fake_date", "start_minutes"])
|
||||
# Extract daily features
|
||||
daily_features = extractDailyFeatures(main_sleep_episodes)
|
||||
|
||||
# Extract features per segment based on daily features
|
||||
for day_type in day_types_to_compute:
|
||||
sleep_intraday_features = statsOfDailyFeatures(daily_features, day_type, sleep_levels_to_compute, intraday_features_to_compute, sleep_intraday_features)
|
||||
if "socialjetlag" in intraday_features_to_compute:
|
||||
sleep_intraday_features = socialJetLagFeature(daily_features, sleep_intraday_features)
|
||||
sleep_intraday_features = MSSDFeatures(daily_features, intraday_features_to_compute, sleep_intraday_features)
|
||||
|
||||
sleep_intraday_features.index.name = "local_segment"
|
||||
sleep_intraday_features.reset_index(inplace=True)
|
||||
|
||||
|
||||
return sleep_intraday_features
|
||||
Loading…
Reference in New Issue