Leveling Up Identity Resolution: Best Practices for Data Scientists

Identity resolution is essential for creating personalized customer journeys. However, building a scalable model for identity resolution is challenging and requires a consistent approach, a true universal identifier, and simplified data pipelines.

By Segment

Understanding customers at a personal level, and treating them like individuals, is a requisite skill for becoming a customer-first organization. In fact, according to our 2022 “State of Personalization Report”, 62% of consumers say a brand will lose their loyalty if they deliver an un-personalized experience.

The ability to provide unique customer journeys is essential, but as a business grows, so does the complexity of identifying customers as individuals. This is where data science teams can deliver an outsized impact on business outcomes by solving the challenge of identity resolution.  

What is Identity Resolution? 

Identity Resolution, also known as ID Resolution, entity resolution, identity mapping, or record linkage, is the practice of creating a single customer profile for every customer by unifying different data sets pulled from a variety of locations, including CRM, marketing and support tools, SMS records, third party databases, and more. Data teams use Identity Resolution to connect real-world data concerning a single person from a variety of sources so that all customer data and behavior is in the same record. 

Building a scalable, sustainable model for identity resolution is not a trivial task and requires extensive work from the data team. Not only do teams need a consistent approach for creating and resolving customer profiles, but those profiles need to be maintained, which is a continuous process of cleaning, merging, enriching, and porting data back and forth between system application layers, the warehouse, and a BI solution of choice. And since the average data scientist is still spending almost 40% of their time on data prep and cleaning. According to a 2021 Anaconda study, introducing another model that requires constant heavy-lifting to keep it on track is a recipe for disaster. 

Many intrepid data science teams have embarked on an identity resolution journey, only to find it’s tougher to build than it looks. However, it is possible, if you know the challenges and best practices. 

What are some common Identity Resolution challenges?

Identity graphs and data pipelines are complex 

Identity resolution is tough, and it sometimes takes years to create a solution that you can fully trust. Sadly, many optimistic data science teams get to the final stage of launching their identity resolution model, only to discover that it is immediately broken. The challenges of maintenance are compounded as business teams continue to onboard new engagement solutions with completely different data structures, which render the approach to merging customer data into the profile out of date.

Customer data is often incomplete and inaccurate

For any given customer profile, the following inaccuracies and redundancies are common: 

  • The same person may submit a form with multiple email addresses and therefore have multiple profiles  

  • Due to typos, the same person may be in the system a variety of times using a different spelling of their name or company names

  • Third party data, which is less trust-worthy, may be used to create or append a customer profile 

  • Abbreviations, varying uses of punctuation, and omitted fields create errors 

These data gaps require data science teams to model a solution for handling, merging, and rejecting data as it is created. This is not a straight-forward task since there are infinite error types that can be introduced to the data. We will expand on best practices for handling this challenge later.

Changing rules and regulations threaten compliance

Both internet providers and governments are making moves to ensure personalization doesn’t happen at the expense of customer privacy. So your solution for identity resolution must strike the right balance of personalization and privacy, even as compliance rules continue to change. For example, Google Chrome, the world’s most popular browser, has promised all third-party cookies will be deprecated by 2024, while fellow industry giants Apple and Mozilla have taken similar steps. Meanwhile, governments have introduced GDPR, CCPA, and industry-specific privacy regulations like HIPAA for healthcare, all making customer data collection, and subsequently personalization, more difficult. 

Identity graphs are difficult to maintain over time

In 2022, Chief Martech identified over 9,900 marketing technology solutions available to businesses. Your business teams may not use all 9,900, but they probably use more than they can count on two hands. With potentially millions of data points coming into the business on a weekly – or even daily – basis, there are bound to be challenges capturing and recognizing data from all 3rd party sources, especially when there are conflicting inputs, or difficulties tracking the historical/changes to profiles. Without a continuous process for capturing, cleaning, and merging new data inputs, it’s impossible to capture a real and true snapshot, let alone maintain an accurate profile over time. 

Resolved customer data is stuck in the warehouse 

While there is growing interest in using the warehouse as the source of truth for identity resolution, there are challenges to this approach. 

Data comes into the warehouse from a variety of sources and a variety of formats, often with different schemas, and manual-based solutions for turning raw data into clean customer profiles can be incredibly complex and time consuming to manage. Additionally, creating pipelines for the multitude of data inputs and assigning meaning and relevance to data that is often incomplete, inaccurate, and in a variety of formats, is a huge challenge, even for the most qualified SQL expert. 

The goal of identity resolution is to help the business understand the full picture of the customer journey so it can become accessible and actionable for business users through analytics. Yet, many parts of the business will often never see the results of these models outside of an analytics dashboard within a BI tool - if they ever see it at all.  This makes running self-service queries, creating audience segments, or actioning complex use cases like marketing attribution difficult if this data isn’t available in the tools teams already use today. 

Getting started with identity resolution: Best practices and considerations 

With changing guidelines around customer privacy, exploding customer touchpoints, and a myriad of data challenges to solve, it’s no surprise that identity resolution, the touchstone of personalized marketing, remains a critical challenge. But it is possible to develop an effective solution for Identity Resolution with best practices at the core. 

Customer Data Platforms, like Segment, help businesses collect, clean, and control their customer data. Segment includes identity resolution, the ability to merge customer activity into a single profile in real time. Our solution – and approach – to identity resolution was developed through our experiences with over 25,000 customers, across a wide variety of industries. Whether you work with a CDP provider, or attempt to develop your own identity resolution pipelines, here are some considerations and best practices we’ve learned that can simplify the complex and challenging world of identity management. 

Start with a true universal identifier

One of the first challenges most data science teams face when approaching identity resolution is that (some) of the business teams may believe the problem has already been solved. This is because many off-the-shelf customer communication tools treat “identity” as matching on an email address or a hashed email address. This approach works for vendors, but falls flat for businesses as email is not a true universal identifier. If a team leveraging “email” as an identifier attempted to sync every single one of your sources – server, mobile, cloud, web, etc. – on behalf of a user, it would result in a bunch of disconnected tables that would need to be stitched together. 

Consider a user with an anonymous ID AABBCC who is then de-anonymized to become user 1243. In this scenario, every event generated before that customer logged in will be stored in your warehouse, forever, with just that anonymous ID. If you want to activate that data and understand user 1243’s whole history with you, you’ll have to look that anonymous ID up. Your data will certainly be fragmented with just an email, just an anonymous ID, or any one of a number of identifiers, leaving you with the work of resolving, each and every time, who that identifier actually represents.

True identity solutions should connect data to consumers in a privacy-compliant way by establishing brand-unique identity resolution rules that create a single view of the consumer across all channels.

This is where CDP solutions like Segment can offload the heavy lifting by providing a canonical identifier that can be the point of consistency that all other identifiers are merged with. Starting with externalIDs, which are identifiers pulled into your CDP from an external data source, like user_id, Android and iOS IDs, Google Analytics, anonymous_id, and group_id.,  you can match external identifiers to a canonical_id, which is then the source of truth as the profile is built over time.

Simplify the approach of building your data pipelines

Building and maintaining the data pipelines required for effective identity resolution typically consists of following steps:

  1. Collecting data from sources

  2. Processing/transforming data in a usable format 

  3. Activating the data in destinations used by partners teams

  4. Orchestrating steps 1-3 

As more business applications, customer identifiers, and dependent teams are added, these steps become increasingly difficult, and often have to be reimagined as the business changes. This is where solutions like Segment can provide a lot of value by streamlining these steps, even as the business goes through change. 

With Segment, data teams can easily collect data from a website, server library, mobile SDK, or cloud application, and then process the raw data to build an identity graph that allows data teams to understand user interactions across web, mobile, server, and third-party partner touch-points in real time. 

Customers can access the rich identity graph tables in their warehouse using Profiles Sync or build personalized experiences by calling Profile API. Additionally customers can use rETL to activate data in destinations used by their marketing and sales partners. With Protocols, customers can ensure that data adheres to data collection best practices, thereby making it simple for data teams to focus on solving business use cases rather than fixing data problems. 

Take a more rigid approach to data merge rules

Identity resolution is a complex process that’s tied to a businesses unique approach to customer engagement. So it makes sense that every identity resolution solution is just a little bit different, and it makes even more sense that every data scientist worth their degree wants to put their stamp on a solution that is tailored for your individual business. While it’s true that flexibility is needed to handle the complex realities of identity resolution, there is such a thing as too much flexibility. Because the behavior of overly flexible systems inherently can't be predicted or controlled, it is prone to risk. It happens all the time – a marketer notices a discrepancy in the ID graph, and the data scientist who built the thing can't satisfactorily explain how it emerged or how to prevent it. When that happens enough, the marketer wants to move on from the solution. 

The best approach builds rigidity around what should be fixed, so data science teams can focus on the highest impact areas for the business, and a deterministic approach (more on that later) to identity resolution solves this challenge neatly for data science teams. 

Segment’s approach to identity resolution enables you to provide many identifiers for the same person (such as user_id, email, phone, device_id, anonymous_id), and then set the priority of matching to control how profiles are stitched together. It’s a powerful model which has helped us create over 5B unified customer profiles from billions of stitched together events. But this model requires alignment and consistency in the way you stamp values for each identifier onto your events.  Otherwise identity resolution won’t work. In order to achieve an aligned and consistent approach to identity resolution, Segment relies on a 100% deterministic model, based on first-party data. Our approach is called “deterministic” because it requires exact matches (instead of “fuzzy” or “probabilistic” matches) on identifier values to unify events into a single profile. This means identities are resolved based on what you know to be true as opposed to resolving identities based on what you predict to be true (probabilistic identity resolution.)  At Segment, we believe that deterministic is the best approach for identity resolution because it’s based on first-party data your customers actually produce.

While Segment’s deterministic identity resolution might seem overly rigorous, it’s actually highly beneficial. It enables 100% reliable profile unification, and it honors the exact first-party data a user provides to you, so your rationale for merging profiles (or keeping them separate) is completely transparent.   

While Segment has always taken the approach that it’s better to be rigid, so you have a consistent process for categorizing and processing data inputs across all your applications, increasingly rigid privacy laws have further supported the deterministic approach. Global organizations are most likely aware that the EUs General Data Protection Regulation (GDPR) already prohibits companies from taking advantage of probabilistic identity resolution. While the GDPR may seem like the most strict form of privacy regulations today, countries and technology companies around the world are taking action to be more respectful of consumer privacy. For example, Apple’s latest iOS 14 release includes a new privacy feature that gives users more control over which apps can track them across sites and apps for advertising. As operating systems and browsers release more privacy features, and more users choose to opt out of tracking, the value of a probabilistic strategy will continue to decrease. 

Understand the role of the warehouse 

Innovative warehouse providers like Snowflake, BigQuery, and Redshift have enabled data teams to execute complex queries and transformations more quickly with less fear of resource constraints. Consequently, some teams have attempted to use the warehouse as the business source of truth, and build SQL models for computing key business metrics such as LTV or health score with data straight from the warehouse.

While the separation of storage and compute model has made the data warehouse a much more resource-friendly place to store all customer data, attempting to perform identity resolution in the warehouse still has its drawbacks. Ideally, the warehouse would come equipped with a sophisticated ability to understand how to match various records to represent one customer against a set of criteria, but this is not the case. Most data teams that try to manage an ID graph with more than two or three customer identifiers using SQL find very quickly that it is intractably time consuming, expensive, and complex. 

Solutions have existed on the market to sync data from their data warehouse to their customer-facing teams’ tools, but this is a simplistic look at what is actually required to create a unified profile with a canonical identifier that can be leveraged across all your business applications to execute across the complex use cases that move business forward. 

In short, while the data warehouse is your source of truth-  it is a hammer that can make every problem look like a nail if we aren’t careful.

Make portability a core component of your identity resolution strategy 

The beauty of identity resolution is that your whole business is in sync, you aren’t having to provide the ID graph to everyone and make it work for everyone. You have a solution for keeping profiles up-to-date via APIs, and can allow teams like marketing to build traits and campaigns in the solutions they use, while data science teams can build models in the back end to answer the businesses’ toughest questions. You can even use APIs to move, clean, to pull in demographic, behavioral, financial, lifestyle, purchase, and other data compiled or licensed from third-party sources, such as online news sites, purchase transactions, surveys, email service providers (ESPs), motor vehicle records, voter registration, and other public records.

Conclusion

Identity resolution is not only essential for companies looking to provide a connected experience across websites, devices, applications, and more, but it is also the foundation of all data-driven decisions, and customer engagement strategies. 

This blogpost outlined the challenges and best practices of Identity Resolution. Our next post in this series will highlight the pitfalls of DIY Identity Resolution, and is a must-read for any data science team interested in creating and managing their own solution for building data pipelines.

The State of Personalization 2023

Our annual look at how attitudes, preferences, and experiences with personalization have evolved over the past year.

Recommended articles

Loading

Want to keep updated on Segment launches, events, and updates?