@Blog

import pickle
import os

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity
import pandas as pd

# モデル保存ファイル名
MODEL_PATH = "tfidf_vectorizer.pkl"  # TfidfVectorizer
DATA_PATH = "article_tags.pkl"       # 記事IDとタグ

# 類似度モデル構築（再学習対応）
def build_similarity_model(article_tags_input, save_model=True, retrain=False):
    # 再学習時は既存データを読み込んで結合する
    if retrain and os.path.exists(DATA_PATH):
        with open(DATA_PATH, "rb") as f:
            existing_data = pickle.load(f)
        existing_data.update(article_tags_input)  # 新しいデータを追加
        article_tags_input = existing_data

    article_ids = list(article_tags_input.keys())
    processed_corpus = [" ".join(tags) for tags in article_tags_input.values()]

    # 再学習またはモデル未保存時に新たにモデルを学習・保存
    if retrain or not os.path.exists(MODEL_PATH):
        vectorizer = TfidfVectorizer()
        tfidf_matrix = vectorizer.fit_transform(processed_corpus)

        if save_model:
            with open(MODEL_PATH, "wb") as f:
                pickle.dump(vectorizer, f)  # ← ここでモデルを保存
            with open(DATA_PATH, "wb") as f:
                pickle.dump(article_tags_input, f)  # ← ここで元データを保存

    else:
        # 保存済みモデルを使って変換する
        with open(MODEL_PATH, "rb") as f:
            vectorizer = pickle.load(f)
        tfidf_matrix = vectorizer.transform(processed_corpus)

    cosine_sim_matrix = cosine_similarity(tfidf_matrix)
    cosine_sim_df = pd.DataFrame(cosine_sim_matrix, index=article_ids, columns=article_ids)

    return cosine_sim_df, article_ids

# 類似記事を取得する関数
def get_recommendations(article_title, similarity_matrix, articles_map, top_n=3):
    if article_title not in articles_map:
        print(f"エラー: 記事 '{article_title}' が見つかりません。")
        return []

    sim_scores = list(enumerate(similarity_matrix[article_title]))
    sim_scores = sorted(sim_scores, key=lambda x: x[1], reverse=True)

    recommended_articles = []
    for i, score in sim_scores:
        if articles_map[i] != article_title and len(recommended_articles) < top_n:
            recommended_articles.append((articles_map[i], score))
        if len(recommended_articles) >= top_n:
            break

    return recommended_articles

# 入力記事データ ( 例 )
article_tags_input = {
    "記事1": ["Python", "機械学習", "データサイエンス"],
    "記事2": ["Python", "Web開発", "Django"],
    "記事3": ["機械学習", "自然言語処理"],
    "記事4": ["データサイエンス", "統計学"],
    "記事5": ["Python", "データサイエンス", "可視化"]
}

# 類似度モデル構築 + モデル保存 ( 初回学習 )
cosine_sim_df, article_ids = build_similarity_model(article_tags_input)

# 使用例
target_article = "記事1"
recommendations = get_recommendations(target_article, cosine_sim_df, article_ids, top_n=2)
print(f"\n「{target_article}」へのおすすめ記事 ( 上位2件 ):")
for article, score in recommendations:
    print(f"- {article} (類似度: {score:.4f})")

# 新しい記事を追加して再学習
new_article_id = "記事6"
new_article_tags = ["Python", "統計学"]
article_tags_input = {new_article_id: new_article_tags}

# 再構築＋再学習
cosine_sim_df, article_ids = build_similarity_model(article_tags_input, retrain=True)
target_article = new_article_id
recommendations = get_recommendations(target_article, cosine_sim_df, article_ids, top_n=2)
print(f"\n「{target_article}」へのおすすめ記事 ( 上位2件 ):")
for article, score in recommendations:
    print(f"- {article} (類似度: {score:.4f})")