Apply a function to each cell of a pandas dataframe using information from a particular column

You can try to extend the lists inplace:

for col in df.loc[:, "s_1":]:
    m = df[col].notna()

    for i, v in zip(df.loc[m, "i"], df.loc[m, col]):
        v.extend([v[-1]] * i)

    df.loc[~m, col] = 0

Benchmark:

from timeit import timeit
from ast import literal_eval


def get_df():
    dfs = []

    # create some big dataframe
    for i in range(5000):
        txt = """
        i   s_1         s_1        s_3
        2   [1, 2, 3]   [3, 4, 5]  NaN
        1   NaN         [0, 0, 0]  [2]  """

        df = pd.read_csv(StringIO(txt), sep=r"\s{2,}", engine="python")

        df.loc[:, "s_1":] = df.loc[:, "s_1":].apply(
            lambda x: [v if pd.isna(v) else literal_eval(v) for v in x]
        )
        dfs.append(df)

    return pd.concat(dfs)


def f1(df):
    for col in df.loc[:, "s_1":]:
        m = df[col].notna()

        for i, v in zip(df.loc[m, "i"], df.loc[m, col]):
            v.extend([v[-1]] * i)

        df.loc[~m, col] = 0
    return df


def f2(df):
    df = df.apply(
        lambda x: x.apply(
            lambda y: np.pad(y, (0, x.i), "constant", constant_values=y[-1])
            if type(y) == list
            else 0
        ),
        axis=1,
    )
    return df


df1 = get_df()
df2 = get_df()

t1 = timeit(lambda: f1(df1), number=1)
t2 = timeit(lambda: f2(df2), number=1)

print(t1)
print(t2)

Prints:

0.01171580795198679
2.3192087680799887

So improvement ~200x

We can stack the datafarme and use list comprehension to pad the values, then unstack back to reshape

s = df.set_index('i', append=True).stack()
s[:] = [v + v[-1:] * r for r, v in zip(s.index.get_level_values(1), s)]
out = s.unstack()

   i              s_1            s_1.1     s_3
0  2  [1, 2, 3, 3, 3]  [3, 4, 5, 5, 5]     NaN
1  1              NaN     [0, 0, 0, 0]  [2, 2]

Performance checks

# Prepare a sample dataframe with 10,000 rows
df = pd.concat([df] * 5000, ignore_index=True)


# Solution with stack and unstack
%%timeit
s = df.set_index('i', append=True).stack()
s[:] = [v + v[-1:] * r for r, v in zip(s.index.get_level_values(1), s)]
_ = s.unstack()
# 38.9 ms ± 700 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)


# OP's solution with apply and np.pad
%%timeit 
df.apply(lambda x: x.apply(
    lambda y: np.pad(y, (0, x.i), 'constant', constant_values=y[-1]) if type(y) == list else 0), axis=1)
# 7.92 s ± 102 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

To avoid itering over rows you might try the following

df = pd.DataFrame( data=[[2, [1, 2, 3],[3, 4,5],None],[1,None,[0, 0, 0] ,[2]]],columns = ['i','s_1','s_2','s_3'])

for col in ['s_1','s_2','s_3']:
    df[col] = df[col] + df['i']*df[col].apply(lambda x : [x[-1]] if type(x)==list else x)

Output