David

hybrid/baseline_script.py

@@ -3,7 +3,7 @@
 from datetime import timedelta
 
 
-def popularity_split_baseline(behaviors_df, articles_df, train_slice=100000, window_size=10000):
+def popularity_split_baseline(behaviors_df, articles_df, train_slice=100_000, window_size=100_00):
     behaviors_df = behaviors_df.sort_values('impression_time').reset_index(drop=True)
 
     train_df = behaviors_df.iloc[:train_slice]
@@ -37,10 +37,9 @@ def popularity_split_baseline(behaviors_df, articles_df, train_slice=100000, win
 def popularity_baseline(behaviors_df, articles_df):
     total_hr20 = 0
     num_clicks = 0
-
     behaviors_df = behaviors_df.sort_values('impression_time')
 
-    batch_size = 10000
+    batch_size = 10_000
     for i in range(0, len(behaviors_df), batch_size):
         chunk = behaviors_df.iloc[i : i + batch_size]
         current_time = chunk['impression_time'].max()
@@ -185,3 +184,50 @@ def to_list(x):
     evaluate_ever_clicked_hr(baseline_recs, behaviors.iloc[100_000 : ])
     check_article_stats(baseline_recs, articles)
     evaluate_detailed_hits(baseline_recs, behaviors.iloc[100_000 : ])
+
+
+
+def popularity_baseline_hybrid(
+    behaviors_df: pd.DataFrame,
+    articles_df: pd.DataFrame,
+    top_k: int = 20,
+) -> dict:
+    """
+    Popularity baseline using the inview candidate set per impression.
+    
+    Ranks inview articles by their total_pageviews, recommends the top-K,
+    and measures HR@K against clicked articles.
+    Falls back to random ranking for articles with null pageviews.
+    """
+    if behaviors_df.empty:
+        return {"hr_at_k": 0.0, "hits": 0, "impressions": 0, "top_k": top_k}
+
+    # Build a pageview lookup; fill nulls with 0 so they rank last
+    pageview_lookup = (
+        articles_df["total_pageviews"]
+        .fillna(0)
+        .astype(int)
+        .to_dict()
+    )
+
+    total_hits = 0
+    total_impressions = 0
+
+    for _, row in behaviors_df.iterrows():
+        inview = row["article_ids_inview"]
+        clicked = row["article_ids_clicked"]
+
+        if not inview.any() or not clicked.any():
+            continue
+
+        # Rank inview articles by global popularity
+        ranked = sorted(inview, key=lambda aid: pageview_lookup.get(aid, 0), reverse=True)
+        recommended = ranked[:top_k]
+
+        if any(aid in recommended for aid in clicked):
+            total_hits += 1
+        total_impressions += 1
+
+    hr = total_hits / total_impressions if total_impressions > 0 else 0.0
+    print(f"HR@{top_k}: {hr:.4%}  ({total_hits}/{total_impressions} impressions)")
+    return {"hr_at_k": hr, "hits": total_hits, "impressions": total_impressions, "top_k": top_k}

hybrid/hybrid.ipynb

@@ -17,15 +17,15 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 2,
+   "execution_count": null,
    "id": "0051c91c",
    "metadata": {},
    "outputs": [],
    "source": [
     "def read_data():\n",
-    "    basedir = \"../data/raw/train/\"\n",
+    "    basedir = \"../data/raw/\"\n",
     "    print(\"Loading data...\")\n",
-    "    behaviors = pd.read_parquet(basedir+'behaviors.parquet')\n",
+    "    behaviors = pd.read_parquet(basedir+'train/behaviors.parquet')\n",
     "    articles = pd.read_parquet(basedir+'articles.parquet').set_index('article_id')\n",
     "    behaviors['impression_time'] = pd.to_datetime(behaviors['impression_time'])\n",
     "    articles['published_time'] = pd.to_datetime(articles['published_time'])\n",

hybrid/hybrid.py

@@ -0,0 +1,24 @@
+import polars as pl
+import pandas as pd
+
+
+from baseline_script import popularity_baseline_hybrid
+from content import *
+from collaborative import *
+
+
+# Read the behaviors and articles data
+def read_data_hybrid(start=None,length=None) -> tuple[pd.DataFrame,pd.DataFrame]:
+    basedir = "./data/raw/"
+    print("Loading data...")
+    if start is None or length is None:
+        raise ValueError
+    behaviors = pl.scan_parquet(basedir+'train/behaviors.parquet').slice(start,length).collect().to_pandas()
+    articles = pl.read_parquet(basedir+'articles.parquet').to_pandas()
+    return behaviors, articles
+
+
+if __name__ == "__main__":
+    behaviors, articles = read_data_hybrid(2_000_000,1_000_000)
+    # Creating the baseline
+    popularity_baseline_hybrid(behaviors,articles, top_k=5)

pyproject.toml

@@ -15,4 +15,5 @@ dependencies = [
     "stop-words>=2025.11.4",
     "jupyter>=1.1.1",
     "ipykernel>=6.29.0",
+    "tqdm>=4.66.0",
 ]