For loop pandas 和 numpy：性能

Question

I have coded the following for loop. The main idea is that in each occurrence of 'D' in the column 'A_D', it looks for all the possible cases where some specific conditions should happen. When all the conditions are verified, a value is added to a list.

a = []
for i in df.index:
    if df['A_D'][i] == 'D':
         if df['TROUND_ID'][i] == '        ':
             vb = df[(df['O_D'] == df['O_D'][i])
             & (df['A_D'] == 'A' )
             & (df['Terminal'] == df['Terminal'][i])
             & (df['Operator'] == df['Operator'][i])]

            number = df['number_ac'][i]
            try: ## if all the conditions above are verified a value is added to a list
                x = df.START[i] - pd.Timedelta(int(number), unit='m')
                value = vb.loc[(vb.START-x).abs().idxmin()].FlightID
            except: ## if are not verified, several strings are added to the list
                value = 'No_link_found'
        else:
            value = 'Has_link'
    else:
        value = 'IsArrival'
a.append(value)

My main problem is that df has millions of rows, therefore this for loop is way too time consuming. Is there any vectorized solution where I do not need to use a for loop?

Answer 1

An initial set of improvements: use apply rather than a loop; create a second dataframe at the start of the rows where df["A_D"] == "A" ; and vectorise the value x .

arr = df[df["A_D"] == "A"]
# if the next line is slow, apply it only to those rows where x is needed
df["x"] = df.START - pd.Timedelta(int(df["number_ac"]), unit='m')

def link_func(row):
    if row["A_D"] != "D":
        return "IsArrival"
    if row["TROUND_ID"] != "        ":
        return "Has_link"
    vb = arr[arr["O_D"] == row["O_D"]
             & arr["Terminal"] == row["Terminal"]
             & arr["Operator"] == row["Operator"]]
    try:
        return vb.loc[(vb.START - row["x"]).abs().idxmin()].FlightID
    except:
        return "No_link_found"            

df["a"] = df.apply(link_func, axis=1)

Using apply is apparently more efficient but does not automatically vectorise the calculation. But finding a value in arr based on each row of df is inherently time consuming, however efficiently it is implemented. Consider whether the two parts of the original dataframe (where df["A_D"] == "A" and df["A_D"] == "D" , respectively) can be reshaped into a wide format somehow.

EDIT: You might be able to speed up the querying of arr by storing query strings in df , like this:

df["query_string"] = ('O_D == "' + df["O_D"] 
                    + '" & Terminal == "' + df["Terminal"] 
                    + '" & Operator == "' + df["Operator"] + '"')
def link_func(row):
    vb = arr.query(row["query_string"])
    try:
        row["a"] = vb.loc[(vb.START - row["x"]).abs().idxmin()].FlightID
    except:
        row["a"] = "No_link_found"

df.query('(A_D == "D") & (TROUND_ID == "        ")').apply(link_func, axis=1)