Skip to content

Custom Models

Although PolyFuzz has several models implemented, what if you have developed your own? What if you want a different similarity/distance measure that is not defined in PolyFuzz? That is where custom models come in. If you follow the structure of PolyFuzz's BaseMatcher you can quickly implement any model you would like.

You simply create a class using BaseMatcher, make sure it has a function match that inputs two lists and outputs a pandas dataframe. That's it!

We start by creating our own model that implements the ratio similarity measure from RapidFuzz:

import numpy as np
import pandas as pd
from rapidfuzz import fuzz

from polyfuzz import PolyFuzz
from polyfuzz.models import BaseMatcher


class MyModel(BaseMatcher):
    def match(self, from_list, to_list, **kwargs):
        # Calculate distances
        matches = [[fuzz.ratio(from_string, to_string) / 100 
                   for to_string in to_list] for from_string in from_list]

        # Get best matches
        mappings = [to_list[index] for index in np.argmax(matches, axis=1)]
        scores = np.max(matches, axis=1)

        # Prepare dataframe
        matches = pd.DataFrame({'From': from_list,
                                'To': mappings, 
                                'Similarity': scores})
        return matches

MyModel can now be used within PolyFuzz and runs like every other model:

from_list = ["apple", "apples", "appl", "recal", "house", "similarity"]
to_list = ["apple", "apples", "mouse"]

custom_matcher = MyModel()

model = PolyFuzz(custom_matcher).match(from_list, to_list)

Now we can visualize the results:

model.visualize_precision_recall(kde=True)

fit, transform, fit_transform

Although the above model can be used in production using fit, it does not track its state between fit and transform. This is not necessary here, since edit distances should be recalculated but if you have embeddings that you do not want to re-calculate, then it is helpful to track the states between fit and transform so that embeddings do not need to be re-calculated. To do so, we can use the re_train parameter to define what happens if we re-train a model (for example when using fit) and what happens when we do not re-train a model (for example when using transform).

In the example below, when we set re_train=True we calculate the embeddings from both the from_list and to_list if they are defined and save the embeddings to the self.embeddings_to variable. Then, when we set re_train=True, we can prevent redoing the fit by leveraging the pre-calculated self.embeddings_to variable.

import numpy as np
from sentence_transformers import SentenceTransformer

from ._utils import cosine_similarity
from ._base import BaseMatcher


class SentenceEmbeddings(BaseMatcher):
    def __init__(self, model_id):
        super().__init__(model_id)
        self.type = "Embeddings"

        self.embedding_model = SentenceTransformer("all-MiniLM-L6-v2")
        self.embeddings_to = None

    def match(self, from_list, to_list, re_train=True) -> pd.DataFrame:
        # Extract embeddings from the `from_list`
        embeddings_from = self.embedding_model.encode(from_list, show_progress_bar=False)

        # Extract embeddings from the `to_list` if it exists
        if not isinstance(embeddings_to, np.ndarray):
            if not re_train:
                embeddings_to = self.embeddings_to
            elif to_list is None:
                embeddings_to = self.embedding_model.encode(from_list, show_progress_bar=False)
            else:
                embeddings_to = self.embedding_model.encode(to_list, show_progress_bar=False)

        # Extract matches
        matches = cosine_similarity(embeddings_from, embeddings_to, from_list, to_list)

        self.embeddings_to = embeddings_to

        return matches

Then, we can use it as follows:

from_list = ["apple", "apples", "appl", "recal", "house", "similarity"]
to_list = ["apple", "apples", "mouse"]

custom_matcher = MyModel()

model = PolyFuzz(custom_matcher).fit(from_list)

By using the .fit function, embeddings are created from the from_list variable and saved. Then, when we run model.transform(to_list), the embeddings created from the from_list variable do not need to be recalculated.