GWなので、妻向けにAI diet Botを作ってみた
最近あまり個人開発もしていなかったのですが、
妻が毎日トレーナーさんへ食事記録や体重データを送っているのを見て、
「これ、LINE Botで自動化できるのでは…？」
と思い、勢いで作ってみました💪
やったことはシンプルで、
- LINEに自然文で食事内容を送信
- GASでデータ解析
- スプレッドシートへ自動蓄積
- GitHub ActionsでML推論
- LightGBMで翌日の体重増減を予測
- 結果をLINEへPush通知
という構成です。
思った以上に、
「家庭内MLOps」みたいな構成になりました😂
システムアーキテクチャ
構成はこんな感じ LINE DeveloperでLINE Message APIを活用し、
LINE→Google Apps Script(GAS)→スプレッドシート→GitHub Action→LINE
構造化
LINEから送られてくる自然文を、
GAS側で構造化(JSON化)し、
そのままスプレッドシートへ蓄積しています。
モデル選択
今回は推論速度と実装コストを優先し、
LightGBMを採用しました。
LSTMなどの時系列モデルも候補でしたが、
今回の用途では、
「即レスポンスが返ること」を重視しています。
LINE側
LINE Developer > プロバイダ > Messaging API を作成していく
必要なMessaging APIのチャネルアクセストークンを用意する
Google Apps Scriptでデプロイしてからにはなるが
Webhook URLも必要なので後程用意する!
Google Apps Script
https://script.google.com/home
Google Apps ScriptでJavaScriptを作成していく。
下記はサンプルだが大体こんな感じで作りこんでいる。
必要な特徴量、因子などは各自で構造化してください
GAS用のjavascriptサンプルコード
// =========================
// Script Properties から取得
// =========================
const props = PropertiesService.getScriptProperties();
const ACCESS_TOKEN = props.getProperty("LINE_ACCESS_TOKEN");
const CHANNEL_SECRET = props.getProperty("LINE_CHANNEL_SECRET");
const SPREADSHEET_ID = props.getProperty("SPREADSHEET_ID");
const SHEET_NAME = props.getProperty("SHEET_NAME");
const GITHUB_TOKEN = props.getProperty("GITHUB_TOKEN");
const GITHUB_OWNER = props.getProperty("GITHUB_OWNER");
const GITHUB_REPO = props.getProperty("GITHUB_REPO");
const GITHUB_WORKFLOW_FILE = props.getProperty("GITHUB_WORKFLOW_FILE");
const GITHUB_BRANCH = props.getProperty("GITHUB_BRANCH");
// =========================
// LINE Webhook
// =========================
function doPost(e) {
// 本番では署名検証推奨
// verifyLineSignature(e);
const lineJson = JSON.parse(e.postData.contents);
const event = lineJson.events[0];
const replyToken = event.replyToken;
const userId = event.source.userId;
const userMessage = event.message.text;
try {
// 自然言語解析
const data = parseDietText(userMessage);
// スプレッドシート保存
saveDietData(data);
// GitHub Actions起動
let githubTriggered = false;
try {
triggerGitHubActions(userId);
githubTriggered = true;
} catch (githubError) {
Logger.log(
"GitHub Actions Trigger Error: " +
githubError.message
);
}
// LINE返信
const replyText =
"✅ データを保存しました\n\n" +
`日付：${data["日付"]}\n` +
`朝の体重：${data["朝の体重"]} kg\n` +
`合計カロリー：${calcTotalKcal(data)} kcal\n\n` +
(
githubTriggered
? "📊 AI分析を実行中です。"
: "⚠️ 保存は完了しましたが、分析起動に失敗しました。"
);
reply(replyToken, replyText);
} catch (error) {
Logger.log("Parse Error: " + error.message);
reply(
replyToken,
"⚠️ データを読み取れませんでした。"
);
}
return ContentService.createTextOutput("OK");
}
// =========================
// GitHub Actions起動
// =========================
function triggerGitHubActions(userId) {
const url =
`https://api.github.com/repos/${GITHUB_OWNER}/${GITHUB_REPO}` +
`/actions/workflows/${GITHUB_WORKFLOW_FILE}/dispatches`;
const payload = {
ref: GITHUB_BRANCH,
inputs: {
user_id: userId
}
};
const options = {
method: "post",
headers: {
"Accept": "application/vnd.github+json",
"Authorization": "Bearer " + GITHUB_TOKEN,
"X-GitHub-Api-Version": "2022-11-28"
},
contentType: "application/json",
payload: JSON.stringify(payload),
muteHttpExceptions: true
};
const response = UrlFetchApp.fetch(url, options);
// ステータスコードのみログ
Logger.log(
"GitHub Status: " +
response.getResponseCode()
);
if (response.getResponseCode() !== 204) {
throw new Error(
"GitHub Actions起動失敗"
);
}
}
// =========================
// LINE返信
// =========================
function reply(replyToken, text) {
const url =
"https://api.line.me/v2/bot/message/reply";
const payload = {
replyToken: replyToken,
messages: [{
type: "text",
text: text
}]
};
const options = {
method: "post",
headers: {
"Content-Type": "application/json",
"Authorization": "Bearer " + ACCESS_TOKEN
},
payload: JSON.stringify(payload),
muteHttpExceptions: true
};
const response =
UrlFetchApp.fetch(url, options);
Logger.log(
"LINE Reply Status: " +
response.getResponseCode()
);
}
このときSettings/ Developer Settings > Personal access tokens > Fine-grained personal access tokensが必要なので用意してくださいね
Permissions ではActionsはRead and write, ContentsはRead-onlyで！
そしてデプロイ
ウェブアプリのURLをコピーし、LINEのWebhook URLにペーストする
Github Actions
次にGithub側の準備
repositoryを作って、
①.github/workflows/analyze.yaml
②src/analyze.py
③requirements.txt
を用意する
analyze.yaml
name: Analyze Diet Data
on:
workflow_dispatch:
inputs:
user_id:
description: "LINE user ID"
required: true
type: string
jobs:
analyze:
runs-on: ubuntu-latest
steps:
- name: Checkout repository
uses: actions/checkout@v4
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: "3.11"
- name: Install dependencies
run: pip install -r requirements.txt
- name: Run analysis and push LINE message
env:
SHEET_CSV_URL: ${{ secrets.SHEET_CSV_URL }}
LINE_ACCESS_TOKEN: ${{ secrets.LINE_ACCESS_TOKEN }}
LINE_USER_ID: ${{ github.event.inputs.user_id }}
run: python src/analyze.py
analyze.py
import os
from io import BytesIO, StringIO
from urllib.parse import parse_qs, urlparse
import numpy as np
import pandas as pd
import requests
from lightgbm import LGBMClassifier, LGBMRegressor
SHEET_CSV_URL = os.environ["SHEET_CSV_URL"]
LINE_ACCESS_TOKEN = os.environ["LINE_ACCESS_TOKEN"]
LINE_USER_ID = os.environ["LINE_USER_ID"]
FEATURE_COLUMNS = [
"morning_weight",
"evening_weight",
"breakfast_cal",
"lunch_cal",
"dinner_cal",
"activity_cal",
"water_ml",
"stool_count",
"exercise_cal",
"exercise_flag",
"menstruation_flag",
"morning_weight_diff",
"day_weight_range",
"energy_balance",
"meal_cal_3day_avg",
"weight_3day_avg",
]
FACTOR_LABELS = {
"morning_weight": "朝体重",
"evening_weight": "夜体重",
"breakfast_cal": "朝食カロリー",
"lunch_cal": "昼食カロリー",
"dinner_cal": "夕食カロリー",
"activity_cal": "活動カロリー",
"water_ml": "水分量",
"stool_count": "排便回数",
"exercise_cal": "運動カロリー",
"exercise_flag": "運動の有無",
"menstruation_flag": "生理の有無",
"morning_weight_diff": "前日からの朝体重差",
"day_weight_range": "日内の体重差",
"energy_balance": "摂取カロリーと活動の差",
"meal_cal_3day_avg": "直近3日の食事カロリー平均",
"weight_3day_avg": "直近3日の朝体重平均",
}
USER_FACING_EXCLUDED_FACTORS = {
"morning_weight",
"evening_weight",
"breakfast_cal",
"lunch_cal",
"dinner_cal",
"activity_cal",
"exercise_cal",
"morning_weight_diff",
"day_weight_range",
"energy_balance",
"meal_cal_3day_avg",
"weight_3day_avg",
}
def _to_google_csv_export_url(url: str) -> str:
if "docs.google.com/spreadsheets" not in url:
return url
parsed = urlparse(url)
query = parse_qs(parsed.query)
gid = query.get("gid", ["0"])[0]
# /d/<sheet_id>/... 形式から sheet_id を抽出
parts = [p for p in parsed.path.split("/") if p]
sheet_id = ""
if "d" in parts:
i = parts.index("d")
if i + 1 < len(parts):
sheet_id = parts[i + 1]
if sheet_id:
return f"https://docs.google.com/spreadsheets/d/{sheet_id}/export?format=csv&gid={gid}"
return url
def _parse_date_series(series: pd.Series) -> pd.Series:
text = series.astype(str).str.strip()
# 1) 通常の日付文字列
dt = pd.to_datetime(text, errors="coerce")
# 2) Google Sheets / Excel シリアル日付
missing = dt.isna()
if missing.any():
serial = pd.to_numeric(text[missing], errors="coerce")
serial_dt = pd.to_datetime(serial, unit="D", origin="1899-12-30", errors="coerce")
dt.loc[missing] = serial_dt
return dt
def _coerce_numeric(series: pd.Series) -> pd.Series:
cleaned = (
series.astype(str)
.str.replace(",", "", regex=False)
.str.replace(r"[^0-9+\-.]", "", regex=True)
.replace("", np.nan)
)
return pd.to_numeric(cleaned, errors="coerce")
def _coerce_bool(series: pd.Series) -> pd.Series:
text = series.astype(str).str.strip().str.lower()
true_values = {"true", "1", "yes", "y", "on", "あり", "有", "有り"}
false_values = {"false", "0", "no", "n", "off", "なし", "無"}
mapped = pd.Series(np.nan, index=series.index, dtype="float64")
mapped[text.isin(true_values)] = 1.0
mapped[text.isin(false_values)] = 0.0
numeric_fallback = pd.to_numeric(text, errors="coerce")
mapped = mapped.fillna(numeric_fallback)
return mapped.fillna(0).clip(lower=0, upper=1)
def _download_sheet(url: str) -> requests.Response:
res = requests.get(url, timeout=30)
res.raise_for_status()
return res
def _read_sheet_like(response: requests.Response) -> pd.DataFrame:
# まず区切り自動判定（CSV/TSV 等）
text = response.text
try:
df = pd.read_csv(StringIO(text), sep=None, engine="python")
if len(df.columns) > 1:
return df
except Exception:
pass
# fallback: 明示CSV
try:
return pd.read_csv(StringIO(text))
except Exception:
pass
# fallback: Excel
return pd.read_excel(BytesIO(response.content))
def load_data() -> pd.DataFrame:
url = _to_google_csv_export_url(SHEET_CSV_URL)
response = _download_sheet(url)
raw_df = _read_sheet_like(response)
if raw_df.empty:
raise ValueError("スプレッドシートのデータが空です。")
# 既存実装の列順を前提に安全に内部名へマッピング
index_map = {
0: "date",
1: "morning_weight",
2: "evening_weight",
3: "breakfast_cal",
5: "lunch_cal",
7: "dinner_cal",
9: "target_cal",
10: "activity_cal",
11: "water_ml",
12: "stool_count",
13: "sleep_hours",
14: "exercise_cal",
15: "exercise_flag",
16: "menstruation_flag",
}
cols = {}
for i, name in index_map.items():
if i < len(raw_df.columns):
cols[raw_df.columns[i]] = name
df = raw_df.rename(columns=cols).copy()
if "date" not in df.columns:
raise ValueError("日付列を見つけられません。シートの1列目に日付を入れてください。")
df["date"] = _parse_date_series(df["date"])
df = df.dropna(subset=["date"]).sort_values("date").reset_index(drop=True)
numeric_cols = [
"morning_weight",
"evening_weight",
"breakfast_cal",
"lunch_cal",
"dinner_cal",
"target_cal",
"activity_cal",
"water_ml",
"stool_count",
"exercise_cal",
]
bool_cols = ["exercise_flag", "menstruation_flag"]
for col in numeric_cols:
if col not in df.columns:
df[col] = 0
df[col] = _coerce_numeric(df[col]).fillna(0)
for col in bool_cols:
if col not in df.columns:
df[col] = 0
df[col] = _coerce_bool(df[col])
# 運動カロリー列が空なら食事カロリー合計を採用
if df["exercise_cal"].sum() == 0:
df["exercise_cal"] = df["breakfast_cal"] + df["lunch_cal"] + df["dinner_cal"]
return df
def add_features(df: pd.DataFrame) -> pd.DataFrame:
df = df.copy()
df["morning_weight_diff"] = df["morning_weight"].diff()
df["day_weight_range"] = df["evening_weight"] - df["morning_weight"]
df["energy_balance"] = df["exercise_cal"] - df["activity_cal"]
df["meal_cal_3day_avg"] = df["exercise_cal"].rolling(window=3, min_periods=1).mean()
df["weight_3day_avg"] = df["morning_weight"].rolling(window=3, min_periods=1).mean()
df["next_morning_weight"] = df["morning_weight"].shift(-1)
df["next_weight_change"] = df["next_morning_weight"] - df["morning_weight"]
df["next_increase_flag"] = (df["next_weight_change"] > 0).astype(int)
df[FEATURE_COLUMNS] = df[FEATURE_COLUMNS].replace([np.inf, -np.inf], np.nan).fillna(0)
return df
def make_comment(weight_change: float, prob: float) -> str:
if prob >= 0.70:
return "明日は体重が増えやすい見込みです。食事量と水分・睡眠のバランスを意識してください。焦らず整えていきましょう💪🙂"
if prob >= 0.55:
return "やや増加傾向です。塩分と夜食、活動量を少し調整するのがおすすめです。小さな改善の積み重ねが効きます🌱😊"
if prob <= 0.35:
return "明日は体重が下がりやすい見込みです。現在の習慣を維持してください。この調子で続けましょう✨👍"
return "明日の体重変化は小さめの見込みです。一歩ずつ着実に進めています🙌🙂"
def train_and_predict(df: pd.DataFrame) -> dict:
train_df = df.dropna(subset=["next_morning_weight"]).copy()
latest = df.iloc[[-1]][FEATURE_COLUMNS]
current_weight = float(df.iloc[-1]["morning_weight"])
if len(train_df) < 10:
return {
"predicted_weight": round(current_weight, 2),
"weight_change": 0.0,
"increase_probability": 0.5,
"top_factors": ["データ不足"] * 5,
"top_factor_importances": [0.0] * 5,
"comment": "学習データが少ないため、現在体重を基準に保守的に予測しています。記録を続けるほど精度が上がります📈😊",
}
X = train_df[FEATURE_COLUMNS]
y_reg = train_df["next_weight_change"]
y_cls = train_df["next_increase_flag"]
reg_model = LGBMRegressor(
n_estimators=300,
max_depth=4,
min_child_samples=2,
learning_rate=0.05,
random_state=42,
verbose=-1,
)
reg_model.fit(X, y_reg)
pred_change = float(reg_model.predict(latest)[0])
predicted_weight = current_weight + pred_change
if y_cls.nunique() >= 2:
cls_model = LGBMClassifier(
n_estimators=300,
max_depth=4,
min_child_samples=2,
learning_rate=0.05,
random_state=42,
verbose=-1,
)
cls_model.fit(X, y_cls)
increase_probability = float(cls_model.predict_proba(latest)[0][1])
else:
increase_probability = 0.5
importances = reg_model.feature_importances_
ranked_features = [FEATURE_COLUMNS[i] for i in np.argsort(importances)[::-1]]
user_facing_candidates = [
f for f in ranked_features if f not in USER_FACING_EXCLUDED_FACTORS
]
selected = user_facing_candidates[:5]
if len(selected) < 5:
fallback = [f for f in ranked_features if f not in selected]
selected.extend(fallback[: 5 - len(selected)])
top_factors = [FACTOR_LABELS.get(f, f) for f in selected]
total_importance = float(np.sum(importances))
importance_by_feature = {FEATURE_COLUMNS[i]: float(importances[i]) for i in range(len(FEATURE_COLUMNS))}
if total_importance > 0:
top_factor_importances = [
round((importance_by_feature.get(f, 0.0) / total_importance) * 100, 1) for f in selected
]
else:
top_factor_importances = [0.0] * len(selected)
return {
"predicted_weight": round(predicted_weight, 2),
"weight_change": round(pred_change, 2),
"increase_probability": round(increase_probability, 2),
"top_factors": top_factors,
"top_factor_importances": top_factor_importances,
"comment": make_comment(pred_change, increase_probability),
}
def create_line_message(result: dict) -> str:
change = result["weight_change"]
sign = "+" if change > 0 else ""
factors = result["top_factors"]
factor_importances = result.get("top_factor_importances", [0.0] * len(factors))
factors_text = "\n".join(
[
f"{i}. {factor}（重要度 {importance:.1f}%）"
for i, (factor, importance) in enumerate(zip(factors, factor_importances), start=1)
]
)
return (
"明日の体重予測\n\n"
f"予測体重: {result['predicted_weight']} kg\n"
f"予測増減: {sign}{change} kg\n"
f"増加確率: {int(result['increase_probability'] * 100)}%\n\n"
"主な影響因子\n"
f"{factors_text}\n\n"
f"コメント\n{result['comment']}"
)
def push_line_message(text: str) -> None:
url = "https://api.line.me/v2/bot/message/push"
headers = {
"Content-Type": "application/json",
"Authorization": f"Bearer {LINE_ACCESS_TOKEN}",
}
payload = {
"to": LINE_USER_ID,
"messages": [{"type": "text", "text": text}],
}
response = requests.post(url, headers=headers, json=payload, timeout=30)
print("LINE status:", response.status_code)
print("LINE body:", response.text)
response.raise_for_status()
def main() -> None:
df = load_data()
df = add_features(df)
result = train_and_predict(df)
message = create_line_message(result)
print(message)
push_line_message(message)
if __name__ == "__main__":
main()
requirements.txt
pandas
numpy
scikit-learn
lightgbm
requests
このときrepositoryのSecrets and variablesのsecretでLINE_ACCESS_TOKENでラインのチャネルアクセストークンとSHEET_CSV_URLでスプレッドシートのURLを設定するのを忘れずに！
LINEトーク送信
妻が毎日トレーナーさんに送っている送信内容をこちらのボットに送ると
スプレッドシートにデータが自動的に蓄積され、LightGBMによる体重増減などを予測し返答することができるようになった！
LINEを入口にすると、
「毎日使うUI」になるので継続しやすく、
個人開発との相性もかなり良いと感じました。
今回はLightGBMを使いましたが、
将来的には時系列モデルやLLMによる食事分析なども試してみたいです！
同じように、
「家族向けにちょっと便利なものを作りたい」
という方の参考になれば嬉しいです🙌
