Where is your target audience? Use location personas to find out

Manogna Nadella, Team lead, data science, MiQ

Digital advertising is changing how marketers connect with their audiences. With cookies under scrutiny, cookieless approaches to advertising like location-based marketing are gaining importance. Almost 90% of marketers find that location-based marketing boosts sales and engagement. As advertisers prioritise privacy, we believe location-based marketing will only increase in popularity.

Here’s how it works:

What are location personas

Location personas categorize zip code data into abstract groups, revealing distinctive traits, interests, behaviors, and needs. Analyzing these groups enhances our grasp of audiences in specific areas.

For instance, a brand targeting young, urban individuals near shopping malls with higher incomes can use location personas to pinpoint relevant zip codes for effective marketing.

Why location?

Residence reveals much about individuals. Factors like cost, housing, location, and amenities impact interactions with brands.

At MiQ, we decode this data to:

• Evolve multiple univariate insights (single variable) into personas (multi-faceted insights) for engaging storytelling.
• Create highly-targeted and personalized campaigns in cookieless environments and drive better results.

So, how are location personas created?

Step one: location-dominant characteristics

To profile zip codes, we use the following location-dominant characteristics to identify commonalities and differences between areas:

1. Infrastructure: points of interest e.g. commercial areas, colleges, hospitals

2. Demographics: e.g. age, population distribution, and economic, social, political and cultural traits

3. Audience Interests: consumer lifestyles, hobbies, behaviors, and culture

4. Stores: e.g. store location, road conditions, traffic and spending patterns

5. Physical Properties: weather, terrain etc.

Datasets

Here are just a few of the future-proofed, cookie-less datasets that we use to understand these characteristics:

Step two: feature engineering

Creating location personas involves merging datasets, aggregating features at zip code level. Exploratory analysis handles outliers and missing data. Then, we remove redundant features through correlation analysis. Categorical features were one-hot encoded where the value reflects the quantitative measure of the feature rather than just 0 or 1. Finally, aggregated data at zip code level is normalised using global averages of respective features to mitigate skewness.

Combining datasets yields ~10K features but such high cardinality impacts model performance and hinders interpretability of personas. To address this, we explored dimensionality reduction techniques like PCA and feature categorisation (e.g. bicycle_parking, mototcycle_parking, parking_space, parking_entrance etc., are categorized under parking) to arrive at 156 final features. The pre-processed dataset is 32K zip codes * 156 features, a lower-dimensional representation of our original dataset.

Step three: modeling

The next step is to model the final preprocessed dataset to define the clusters, essentially the location personas for which we have compared various unsupervised clustering algorithms such as K-Means, DBSCAN and Gaussian Mixture Models. To finalize the number of clusters, we used:

• t-distributed Stochastic Neighbour Embedding (t-SNE) to visualize the cluster distribution
• Davies Bouldin index to measure whether the cluster centers are clear apart
• Zip code and population distribution across clusters to avoid some clusters ending up with many zip codes (or high populations) and some with fewer zip codes

The above process is iterated for clusters ranging from 10 to 30 for all the three algorithms and finalized K-Means with 25 clusters.

Davies Bouldin index and t-SNE plot for cluster distribution

Step four: profiling personas

Persona traits are understood by assessing dominant features via an index score, indicating the degree of certainty of each feature within the cluster. We used 60% of features with indices greater than 1 to determine the uniqueness of each cluster.

For instance, consider the ‘US premium coffee consumers’ persona present across 653 zip codes, each with an average population density of 2.70%. Advertisers can customise ads to the target audience using characteristics observed in this profile:

• Physical properties — Spring is hot and humid, while winter brings strong winds, offering chances for seasonal promotions.
• Infrastructural Properties and Nature of Stores — Abundant hardware, services, recreation, and highways. Billboards along highways and coffee kiosks in partnership with local businesses can be used for brand promotion and attracting footfall.
• Demographics — Mainly males aged 19–40, middle-upper income, diverse education level. Customising ads to resonate with these groups enables brands to address unique needs and preferences for better performing creative.
• Audience Interests — Gourmet coffee lovers, health-conscious folk, outdoor activity enthusiasts, and charity volunteers. They engage with grocery store promotions and loyalty cards. Advertisers should use in-store promotional campaigns to capture their attention and loyalty.

Geo insight for the premium coffee brand in the US

Conclusion

Location personas are a powerful tool for marketers looking to better understand their target audience. By leveraging the power of machine learning, you can quickly and effectively analyse large amounts of data to gain meaningful insights into your audience’s behavioural patterns and preferences. So why wait? Let’s move onto future-proofed and robust insights now to deliver engaging and relevant ads to our target audiences.

Manogna is a data science team lead in MiQ’s Bengaluru office. Outside of work, you’ll find her reading about psychology and leadership, or enjoying a long drive.