Combining Predictions from Multiple Matchers

In the matching step, if you use multiple matchers then you will have to combine the predictions from them to get a consolidated prediction. There are many different ways to combine these predictions such as weighted vote, majority vote, stacking, etc. Currently, py_entitymatching supports majority and weighted voting-based combining. These combiners are experimental and not tested.

An example of using majority voting-based combining is shown below.

>>> dt = DTMatcher()
>>> rf = RFMatcher()
>>> nb = NBMatcher()
>>> dt.fit(table=H, exclude_attrs=['_id', 'l_id', 'r_id'], target_attr='label') # H is training set containing feature vectors
>>> dt.predict(table=L, exclude_attrs=['id', 'l_id', 'r_id'], append=True, inplace=True, target_attr='dt_predictions') # L is the test set for which we should get predictions.
>>> rf.fit(table=H, exclude_attrs=['_id', 'l_id', 'r_id'], target_attr='label')
>>> rf.predict(table=L, exclude_attrs=['id', 'l_id', 'r_id'], append=True, inplace=True, target_attr='rf_predictions')
>>> nb.fit(table=H, exclude_attrs=['_id', 'l_id', 'r_id'], target_attr='label')
>>> nb.predict(table=L, exclude_attrs=['id', 'l_id', 'r_id'], append=True, inplace=True, target_attr='nb_predictions')
>>> mv_combiner = MajorityVote()
>>> L['consol_predictions'] = mv_combiner.combine(L[['dt_predictions', 'rf_predictions', 'nb_predictions']])

Conceptually, given a list of predictions (from different matchers) the prediction that occurs most is returned as the consolidated prediction. If there is no clear winning prediction (for example, 0 and 1 occuring equal number of times) then 0 is returned.

An example of using weighted voting-based combining is shown below.

>>> dt = DTMatcher()
>>> rf = RFMatcher()
>>> nb = NBMatcher()
>>> dt.fit(table=H, exclude_attrs=['_id', 'l_id', 'r_id'], target_attr='label') # H is training set containing feature vectors
>>> dt.predict(table=L, exclude_attrs=['id', 'l_id', 'r_id'], append=True, inplace=True, target_attr='dt_predictions') # L is the test set for which we should get predictions.
>>> rf.fit(table=H, exclude_attrs=['_id', 'l_id', 'r_id'], target_attr='label')
>>> rf.predict(table=L, exclude_attrs=['id', 'l_id', 'r_id'], append=True, inplace=True, target_attr='rf_predictions')
>>> nb.fit(table=H, exclude_attrs=['_id', 'l_id', 'r_id'], target_attr='label')
>>> nb.predict(table=L, exclude_attrs=['id', 'l_id', 'r_id'], append=True, inplace=True, target_attr='nb_predictions')
>>> wv_combiner = WeightedVote(weights=[0.3, 0.2, 0.1], threshold=0.4)
>>> L['consol_predictions'] = wv_combiner.combine(L[['dt_predictions',
'rf_predictions', 'nb_predictions']])

Conceptually, given a list of predictions, each prediction is given a weight, we compute a weighted sum of these predictions and compare the result to a threshold. If the result is greater than or equal to the threshold then the consolidated prediction is returned as 1 (i.e., a match) else returned as 0 (no-match).