Co-Visitation Matrisleri ve RAPIDS cuDF ile Verimli Öneri Sistemleri Oluşturun

covisitation_matrix = []
for part in range(PARTS):
      print(f"- Part {part + 1}/{PARTS}")
      matrix = None
      for file_idx, file in enumerate(tqdm(train_files)):

for part in range(PARTS):
     for file_idx, file in enumerate(tqdm(train_files)):
	      [...]
          # Load sessions & convert columns to save memory
          df = pd.read_parquet(file, columns=["session", "aid", "ts"])
          df["ts"] = (df["ts"] / 1000).astype("int32")
          df[["session", "aid"]] = df[["session", "aid"]].astype("int32")

          # Restrict to first 30 interactions
          df = df.sort_values(
               ["session", "ts"],
               ascending=[True, False],
               ignore_index=True
          )
          df["n"] = df.groupby("session").cumcount()
          df = df.loc[df.n < 30].drop("n", axis=1)

     # Compute pairs 
     df = df.merge(df, on="session")

     # Split in parts to reduce memory usage
     df = df.loc[
          (df["aid_x"] >= part * SIZE) &
          (df["aid_x"] < (part + 1) * SIZE)
     ]

     # Restrict to same day and remove self-matches
     df = df.loc[
          ((df["ts_x"] - df["ts_y"]).abs() < 60 * 60) & 
          (df.aid_x != df.aid_y)
     ]
	# No duplicated interactions within sessions
    df = df.drop_duplicates(
          subset=["session", "aid_x", "aid_y"],
          keep="first",
     ).reset_index(drop=True)

     # Compute weights of pairs
     df["wgt"] = 1
     df["wgt"] = df["wgt"].astype("float32")

     df.drop(["session", "ts_x", "ts_y"], axis=1, inplace=True)
     df = df.groupby(["aid_x", "aid_y"]).sum()

     # Update covisitation matrix with new weights
     if matrix is None:
          matrix = df
     else:  # this is the bottleneck operation
          matrix = matrix.add(df, fill_value=0)  

     # Clear memory
     del df
     gc.collect()

for part in range(PARTS):
     [...]
     # Final matrix : Sort values
     matrix = matrix.reset_index().rename(
          columns={"aid_x": "aid", "aid_y": "candidate"}
     )	
     matrix = matrix.sort_values(
          ["aid", "wgt"], ascending=[True, False], ignore_index=True
     )

     # Restrict to n candids
     matrix["rank"] = matrix.groupby("aid").candidate.cumcount()
     matrix = matrix[matrix["rank"] < N_CANDIDS].reset_index(drop=True)
     covisitation_matrix.append(matrix)

covisitation_matrix = pd.concat(covisitation_matrix, ignore_index=True)