Predictions are complex, and consumer predictions for marketing aren’t an exception. In many cases, you might find yourself given the green or red light to go after a lead, prevent a churn, or approach a customer in another way that’s specific to a business goal your org has. Having access to that knowledge is great, but it’s a black box–simply knowing whether or not a lead will bite isn’t enough if you want to truly understand your customers and the actions they take.

In Faraday, every predictive model you create for your business is scored based on how well you can expect the model to perform, and includes a full technical report describing how the model was built. While the full technical report includes a swathe of info for data science-minded users, for the sake of this guide we’re going to take a look at how **features of importance** and **directionality** impact a sample **lead conversion** model, which you can find in every outcome and recommender.

## Features of importance

![Image of features of importance in a Faraday outcome](/images/blog/product/feature-importance.png)

In each outcome and recommender, you’ll find a **data** section describing the top features of importance for that model broken down into tenures, or how long an individual existed within the eligibility cohort before entering the attainment cohort. The features that had the most impact in creating the model are listed in the chart, sorted by the level of impact. In the above image, you can see that features such as high ticket home decor quintile, activity, total online dollars quintile, and household income had a major impact on the model. This means that when thinking about who you’re targeting with these predictions, you can tailor creative and messaging appropriately–but that still feels a little broad. Let’s take a look at directionality to see exactly how these features of importance break down.

## Directionality

![Image of directionality in model report](/images/blog/product/feature-directionality.png)

When you take a deeper look at a feature of importance in feature directionality, you’ll find it broken down into ranges if the feature is numeric, or selections if not. In the above example, green indicates positive, and red negative. For **household income** (HHI), it’s clear that consumers with an HHI of over 75k rank higher on average (green), whereas all HHI ranges lower than 75k rank lower (red). For **shopping styles**, _luxury offline_ and _prefers online_ shoppers rank higher, while _retail therapy_ shoppers rank lower.

That’s cool and all, but what’s the takeaway?

In short, these breakdowns tell you that in this lead conversion model, the people that are most likely to convert are largely making more than 75k/yr, and prefer to shop online unless it's for luxury items. Using this info, your marketing to them can confidently personalize toward the positive features so that your content is relevant.

Learn how to analyze your predictive models to deeper understand your customer base.

Faraday model reports: Understanding feature importance and directionality

While data security has always been an important part of doing business, in recent years it’s become an increasingly urgent consideration. Public scandals, like those of [Facebook and Cambridge Analytica](https://www.wired.com/story/cambridge-analytica-facebook-privacy-awakening/), and growing government regulations, to the tune of GDPR and CCPA, have consumers on high alert about how their data is being used.

If you find yourself among the many brands looking for ways to build trust with your audience, you’ll need to consider not only how you manage data security at your organization, but also how you communicate with your clients. Transparency is an important part of security—clients and collaborators alike should be able to quickly and easily understand what you’re doing to protect their data. The more effectively you achieve that, the more trust you’ll be able to build.

Read on to explore the three most common questions our customers ask us (and might ask you!) about data security, and how we answer them. Feel free to use these as a blueprint to guide your own practices, though the particulars will vary from industry to industry.

## “How is your company audited for data security?”

_Our answer: Check out our [SOC 2 audit](/security)!_

If you offer a cloud-based product or service, and want potential customers and collaborators to see you as a trusted partner, a clean SOC 2 audit should be your gold standard. SOC 2 audits are exhaustive reviews performed by certified public accountants to ensure that SaaS companies meet strict security requirements and provide a safe operating environment for managing sensitive data. In order to retain a clean SOC 2 audit, companies must renew their audit annually.

The existence of SOC 2 as a process reassures clients, but it also benefits the company. As Seamus Abshere, Faraday co-founder and Chief Technology Officer, puts it, “SOC 2 audits provide a chain of trust going from the client all the way to Faraday and to our vendors (like cloud hosting providers). At every level, auditors are involved to ensure that policies and procedures are followed.”

## “How is my data transferred and stored?”

_Our answer: All data is stored with trusted cloud vendors (Amazon, Google Cloud), encrypted in transit and at rest, and access is controlled using multifactor authentication._

Simply put, encryption is a no-brainer. When customers share data with brands, they’re trusting them to protect it from bad actors, and encryption is the standard practice for doing so. “At this point, using unencrypted data is like a surgeon not washing their hands before surgery,” asserts Abshere. “If you are not encrypting everything, it's just malpractice.”

It's standard practice for clients and vendors to ask about encryption, but at this point, the answer should always be "yes."

## “What data do you actually use, and how?”

_Our answer: We live by the *[*“If we don’t need it, we don’t collect, store, or transmit it”*](https://faraday.ai/blog/using-ethical-data/)* mentality._

It’s essential for brands to understand what data they have access to in order for them to put the proper safeguards in place. Here at Faraday, where we handle the data of hundreds of leading brands, we’re extremely clear with clients about what we use—and what we don’t.

**“We don’t want your account numbers, social security numbers, health records — any of that,” Abshere says.  “All we need are known identifiers (like billing information) and transaction information. We do not require or accept account numbers, so we don't affect our client's PCI DSS scope.”** In our case, we leverage this personally identifiable information (PII) to match consumers with our identity graph and train machine learning algorithms to make predictions about a brand’s current or potential customers.

Another important step is to prevent any co-mingling of your clients’ data, particularly if your clients are competitors. At Faraday, we understand this is a serious concern, and we’re careful to prevent it. “Data is logically separated, and everyone’s predictions are based only on their own data plus Faraday's identity graph,” explains Abshere. "Mixing client data would be a huge risk."

## Security is always a priority

No matter your role at a given company, you’re always a steward of your customers’ data. Various teams, from security to engineering to marketing, may operate on very different datasets and roadmaps, but it’s crucial that everyone contributes to a trusted relationship with clients and collaborators, working together to efficiently and transparently implement best practices.

Transparency is an important part of your security practices—clients and collaborators alike should be able to quickly and easily understand what you’re doing to protect their data.

Building trust: How brands can answer common data security questions

Headlines abound on the [opportunities](https://www.techrepublic.com/article/companies-are-doubling-down-on-artificial-intelligence-and-machine-learning-due-to-pandemic/) and [dangers](https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing) of machine learning. Now armed with [plenty](https://www.propublica.org/article/breaking-the-black-box-when-algorithms-decide-what-you-pay) of [examples](https://www.nytimes.com/2019/08/20/upshot/housing-discrimination-algorithms-hud.html), the Data Science community is at least aware that our data lakes and algorithms can make life more difficult for real people if not properly used. But the root problem can be summed up in pretty simple terms:

> An algorithm contains the biases of its builder.

Examples of bias include:

- **Reporting bias** _(at the source)_ — Maybe the folks who filled out your online survey were mostly young, male, and computer-literate.
- **Implicit bias** _(the call is coming from inside the building)_ — Perhaps your image of an ideal job candidate fits into a mental archetype of which you're only semi-aware, and that image excludes those who don't match the archetype.

At Faraday, we have a handful of approaches we use to minimize these effects.

## Data sourcing

It's possible to get your hands on a pretty broad range of data points these days. Everything from the basics of home value, age, and income to the more esoteric pet ownership and liklihood to buy high-end men's apparel. **Faraday actively excludes data points relating to protected classes**. Race, ethnicity, primary language, and religion are excised from our database entirely and therefore excluded from our predictive models and audience definitions.

Of course, seemingly-harmless demographic variables can often act as proxies for these features, but our stance begins with the most effective levers we have available to pull. And further on in the pipeline we do check _results_ against features like race, but we do it at an aggregated level using standardized census data (more on that below).

## Model priors

We investigate the examples in our training data before they become machine-actionable profiles. In particular, we look at how well they're balanced by age, gender, and geography against the population at large. If there's a significant skew versus the baseline, we know there's a chance the resulting predictive model will show bias, so we rework the inputs.

## Model predictions

When a predictive model is fully-built and ready to be put into action, we typically assign a predictive score to everyone in the country as a matter of efficiency. But this also allows us to check whole populations against statistics that reflect the inclusion or exclusion of protected groups at a census block group level.

If we see - for example - that the lowest scores are concentrated in the block groups with the highest proportions of particular racial or ethnic groups, we'll know that _some_ bias has made it into the model, and we'll revise it.

The bottom line is that these strategies are only a start, and **the best defense against unfair algorithms is multi-pronged and in for the long haul.** If you have questions about how your own applications might be veering away from fairness, I suggest checking out:

- Google's [Marketers' guide to ML fairness](https://www.thinkwithgoogle.com/feature/ml-fairness-for-marketers) or [ML bias crash course](https://developers.google.com/machine-learning/crash-course/fairness/video-lecture)
- ["Abolish big data"](https://d4bl.org/videos/19-abolish-big-data) from [Data for Black Lives](https://d4bl.org/)
- The machine bias case studies at [diversity.ai](http://diversity.ai/#case)

How to reduce the effects of implicit bias in machine learning, from data sourcing to model predictions.

Product

Faraday model reports: Understanding feature importance and directionality

Building trust: How brands can answer common data security questions

Mitigating implicit bias in machine learning