Explore entity resolution in Python, including use cases, semantic preprocessing, graph clustering, and weak supervision. Boost business value with hands-on coding and strategic decisions.

Hands-on Entity Resolution with Python-v3.png

dockerfile.tar.gz

conda-main

A typical business stores data across multiple systems, including ERPs for operations, a CRM for marketing, files, notebooks, and BI apps for other purposes. Records of the same customer (entity) exist in multiple places, likely not in sync across nor unique within sources. This inconsistent situation generates an opportunity for us to drive business value by cross-referencing and deduplicating records with entity resolution.

This course covers business acumen and hands-on coding. It starts with several business cases and a quick introduction to entity resolution in Python. Then, it explores semantic-preserving preprocessing, similarity feature engineering, graph clustering, weak supervision, confident learning, and integration.

As a developer, you’ll increase your company’s business value by developing and deploying entity resolution pipelines. As a decision-maker, you’ll know which solution best suits your business cases and how to negotiate the best value for your money.

An Introduction to Entity Resolution in Python

The $F_1$ score is the arithmetic average of precision and recall.

We evaluate the performance of a binary classification model on a test set of $10$ labeled examples. We observe $tp=2$ true positives, $tn=5$ true negatives, $fn=4$ false negatives, and $fp=1$ false positives. Which statement is correct?

We evaluate the performance of a binary classification model on a test set of <math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mn>10</mn></mrow><annotation encoding="application/x-tex">10</annotation></semantics></math>10 labeled examples. We observe <math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>t</mi><mi>p</mi><mo>=</mo><mn>2</mn></mrow><annotation encoding="application/x-tex">tp=2</annotation></semantics></math>tp=2 true positives, <math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>t</mi><mi>n</mi><mo>=</mo><mn>5</mn></mrow><annotation encoding="application/x-tex">tn=5</annotation></semantics></math>tn=5 true negatives, <math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>f</mi><mi>n</mi><mo>=</mo><mn>4</mn></mrow><annotation encoding="application/x-tex">fn=4</annotation></semantics></math>fn=4 false negatives, and <math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mi>f</mi><mi>p</mi><mo>=</mo><mn>1</mn></mrow><annotation encoding="application/x-tex">fp=1</annotation></semantics></math>fp=1 false positives. Which statement is correct?

Absurdly imbalanced classes, usually with less than 0.1% matching pairs.

The likelihood of a match should be nondecreasing in properly engineered similarity features.

Training labels are not readily available and must be prepared first with the available sources of supervision, for example, manual labels.

What is not a typical characteristic of binary classification in real-world entity resolution scenarios?

What is not a typical characteristic of binary classification in real-world entity resolution scenarios?

By suggesting the next batch of pairs for manual labeling.

By activating the most robust match rules with adapted thresholds.

By data-driven activation and deactivation of model features.

How does active machine learning contribute to the quality of entity resolution?

How does active machine learning contribute to the quality of entity resolution?

Training labels are human-made and therefore prone to mistakes. cleanlab helps detect such instances.

cleanlab's `CleanLearning` algorithm detects and automatically reduces the impact of out-of-distribution outliers on the model training.

With cleanlab, we can assess the accuracy of different data annotators.

cleanlab's `CleanLearning` algorithm detects and automatically reduces the impact of label errors on the model training.

What is not a reason to integrate cleanlab into an entity resolution framework?

What is not a reason to integrate cleanlab into an entity resolution framework?

We must decide which labeling function overrules another if they set conflicting labels for a record pair.

Snorkel combines multiple sources of supervision and outputs a single label per record pair.

Weak supervision generalizes rule-based models by defining them as an intermediate labeling function, contributing to the learned labels as one of many supervision sources.

Snorkel's `LabelModel` can predict match vs. no-match for a subset of the feature space.

With weak supervision, we can quickly build large training datasets. What is not a feature of Snorkel?

With weak supervision, we can quickly build large training datasets. What is not a feature of Snorkel?

Oversampling is used to add more examples to an underrepresented scenario. This can also include exotic examples labeled with the majority class.

Oversampling is always superior to undersampling because we add information instead of throwing it away.

Undersampling always means selecting a subset from the majority class while minimizing relevant information loss.

Oversampling the training data always improves the performance on the test set because the model learns from more data.

Which statement about over/undersampling is correct? 

Which statement about over/undersampling is correct?

When we add relevant augmented examples to the training dataset, a refitted model will perform better on the original test set than without learning from augmented examples.

A transformation of a record is called entity-invariant if humans recognize both the original and the new record as descriptions of the same entity.

Any chaining of entity-invariant transformation function is a text-invariant transformation.

The nlpaug package supports text, audio, and image augmentation.

We start with a labeled dataset of pairs and split them into training and test subsets. With text augmentation, we create artificial pairs of matching records by applying entity-invariant transformations on a record. We pair the original record with the augmented record. Which statement is true about this approach?

We start with a labeled dataset of pairs and split them into training and test subsets. With text augmentation, we create artificial pairs of matching records by applying entity-invariant transformations on a record. We pair the original record with the augmented record. Which statement is true about this approach?

In the feature-based approach, we must use the final embedding of the [CLS] token as input for a classification model learning from examples.

The fine-tuning approach means we start from pretrained parameters in BERT, add a classification header with another set of parameters, and relearn all parameters end-to-end, running a few training epochs.

Multiple versions of BERT vary in embedding dimensions, the number of transformer layers, and how parameters have been pretrained. Some aspects are a trade-off of quality and computational costs. The pretraining tasks ideally are close to our specific task.

BERT, SBERT, and RoBERTa (all base versions) differ significantly in their architecture.

We want to use BERT as a building block for our deep matching model. According to the original authors of BERT, we can follow the fine-tuning or the feature-based approach. Which statement is true?

We want to use BERT as a building block for our deep matching model. According to the original authors of BERT, we can follow the fine-tuning or the feature-based approach. Which statement is true?

The `TreeExplainer` tool is an efficient alternative to `KernelExplainer` using decision trees to approximate a model in a local neighborhood.

The `KernelExplainer` method is model-agnostic. Its downside is the high computation cost compared to `TreeExplainer`.

The individual SHAP contributions are additive in the original probability space. This means the sum of all SHAP values of any given example gives us the predicted probability.

The SHAP dependence plot visualizes SHAP contributions of up to two variables.

We use SHAP to explain the predictions of our matching model. Which statement is true?

We use SHAP to explain the predictions of our matching model. Which statement is true?

Evaluate your learnings about binary classification and pairwise matching.

Introduction to Entity Resolution and Applications

A Quickstart Guide Using the RecordLinkage Package

Preprocessing

Indexing

Feature Engineering

Pairwise Matching

Clustering

Integration

Entity Resolution Fundamentals

Matching Products Across Two Online Shops

Conclusion

Appendix

Auto-Tagging System for Content Categorization

Quiz on Pairwise Matching