Imputing Missing Values

While doing supoervised learning-based matching, you would need to create labeled sample, convert the sample into table of feature vectors, fill in the missing values, select a machine learning (ML) model and use it to produce matches.

The step of filling in the missing values (also called imputing missing values) is important and necessary. If there are missing values in the input tables A and B, then they would be passed on to candidate set and most likely to the feature vectors. In py_entitymatching, if the feature vectors contain missing values, then most of the ML algorithms would not work as they rely on scikit-learn package to provide ML-algorithm implementations (and their implementations would not work if the feature vectors contain NaN’s).

To avoid missing value problem in the feature vectors, you must impute the values of the NaN’s. There are many different ways to impute missing values such as filling the NaN’s (in the whole table or just some columns) with a constant value, or fill the NaN’s with an aggregate value (mean, median, etc.).

Since the table is represented as a pandas Dataframe, there are two common ways to impute missing values: (1) use fillna method from pandas Dataframe, and (2) impute missing values using Imputer from Scikit-learn package.

But there are two problems that we have to tackle if we have to using the above commands or objects directly:

• They are not metadata aware, so the user has to explicitly take care of it.
• The Dataframe type that gets imputed typically contains attributes such as key, foreign keys to A and B. The user must have to rightly project them out to impute missing values using aggregates.

In py_entitymatching, we propose a hybrid method to impute missing values. To fill NaN’s with a constant value use fillna command from pandas Dataframe. Please look at the API reference of fillna for more details. An example of using fillna to the whole table is shown below:

>>> H.fillna(value=0, inplace=True)

In the above, H is a Dataframe containing feature vectors, 0 is the constant value that to be filled in, and inplace=True means that the updation should be done in place (i.e., without creating a copy). It is important to set inplace=True as we do not want the metadata for H in Catalog to be corrupted.

Another example of using fillna on a column is shown below:

>>> H['name_name_lev'] = H['name_name_lev'].fillna(value=0, inplace=False)

Note that, in the above inplace should be specified as False, this is because the output is getting assigned to a column in the old Dataframe H and the metadata of H does not get affected.

To fill NaN’s with an aggregate value, in py_entitymatching you can use impute_table command. It is a wrapper around scikit-learn’s Imputer object (to make it metadata aware). An example of using impute_table is shown below:

>>> H = em.impute_table(H, exclude_attrs=['_id', 'ltable_id', 'rtable_id'], strategy='mean')

Note

If all the values in a column or a row are NaN’s, then the above aggregation strategy will not work (i.e. we cannot compute the mean and use it to fill the missing values). In such cases, you need to specify a value in val_all_nans parameter and the command will use this value to fill in all the missing values.

Please refer to the API reference of impute_table() for more details.