How AI-Powered Audience Selection Boosted Brand Conversion by 47%

Background

Uni‑Marketing is a full‑chain, full‑media, full‑data, full‑channel brand big‑data marketing strategy built on the Alibaba ecosystem, aiming to solve the quantification and traceability problems of traditional brand marketing.

Terminology

Brand‑Consumer Relationship: Opportunity, Awareness, Interest, Purchase, Loyalty.

Audience Deepening Rate (O→I): Conversion rate from opportunity audience to interest audience.

Brand Data Bank: Manages and activates brand consumer data assets through integration, analysis, and activation.

Brand Strategy Center: Provides market overview, segmentation, competition analysis, consumer insights, and audience expansion for scenarios such as new product launch and brand upgrade.

Project Goal

Generate a specific‑scale target audience for brand A’s New‑Year promotion, then convert the identified audience into interest and purchase stages to enhance the brand’s consumer asset.

Industry Solutions

Typical programmatic advertising audience targeting methods include:

Tag diffusion

Tag‑based collaborative filtering

Social‑relationship diffusion

Clustering diffusion (e.g., BIRCH, CURE)

Audience classification using PU learning

Technical Solution

Two schemes were considered: “seed audience clustering + clustering diffusion” and “multi‑direction diffusion + audience classification”. The latter was implemented first.

Overall workflow (illustrated below):

5.1 Multi‑Direction Audience Diffusion

Six diffusion directions were explored, each extracting effective features and applying white‑box filtering or black‑box prediction.

Interest Preference Direction

Using TGI and TA concentration metrics, four brand‑related features were selected and thresholds set for white‑box diffusion.

Related Category Direction

1) Main category analysis based on product count and sales. 2) Related brand analysis using Jaccard similarity (formula shown below). 3) Related category analysis via Association Rule Mining with confidence filtering.

Competitor Audience Direction

Analyzed top‑10 competitor brands, identified that a large proportion of new brand audience came from competitors, and built a conversion model using competitor‑related features.

Search Audience Direction

Identified high‑value search keywords and calculated entropy‑based scores to select relevant search terms.

Churn Audience Direction

Recaptured users who had left the brand within the past six months (IPL) or one month (A).

Similar‑Interest Audience Direction

Computed user similarity via vector representations (category vectors, brand vectors, graph embeddings) to find top‑N similar users.

5.2 Target Audience Selection Model

Due to lack of historical campaign data, positive samples were drawn from existing brand purchasers, while negative samples were randomly sampled from other brands.

Evaluation Metrics

Traditional metrics (AUC, Precision) were insufficient for diffusion audiences; a new metric PredictTA TopN Precision was introduced to measure the proportion of true target audience within the top‑N predictions. NewTA TopN Recall measures the proportion of newly acquired target audience captured by the model.

Model Training

Features included numeric discretization, categorical value selection, multi‑value handling, one‑hot encoding, and sparse feature embeddings (category/brand embeddings trained via word‑embedding‑like methods).

Three model families were evaluated:

Logistic Regression (baseline, interpretable)

Random Forest (lower performance, discarded)

PS‑SMART (GBDT‑based, best performance after hyper‑parameter tuning)

Model Prediction

The trained model scored the diffusion audiences from Section 5.1; audiences with scores below 0.5 were filtered out, and remaining users were cross‑checked against existing IPL audiences.

5.3 New‑Year (Year‑Goods) Audience Model

Because the campaign coincided with the Chinese New Year, a dedicated model was built to capture year‑goods purchase intent, using last year’s behavior data as samples.

Sample Selection

Positive samples: users who converted to PL status during the New‑Year period last year. Negative samples: randomly selected users with similar activity.

Model Training & Prediction

Same PS‑SMART pipeline as the daily model; predictions filtered by score > 0.5 and by excluding existing IPL users.

5.4 Model Fusion

Combined daily and year‑goods model outputs based on their PredictTA TopN Precision to produce the final audience set for DMP upload.

5.5 Campaign Effect Tracking

Brand A’s campaign using the algorithm‑selected audience achieved a 47% higher O→IPL deepening rate compared with rule‑based selection. The mixed model (daily + year‑goods) yielded the highest conversion during the New‑Year period.

6 Challenges & Mitigations

Short project timeline – prioritized model‑business alignment.

No historical feedback – relied on random negatives and engineered robust features.

Missing historical attribute features – used recent attributes as proxies.

Sparse high‑dimensional features – applied TGI and TA concentration filtering.

Evaluation – introduced PredictTA TopN Precision to better reflect diffusion performance.

7 Summary & Outlook

The audience diffusion + selection pipeline, custom models, and PredictTA TopN Precision metric proved effective for brand target audience targeting. Future work includes leveraging feedback data for sample optimization, storing richer historical attributes, applying deep‑learning embeddings for sparse features, and exploring multi‑task learning for lifestyle and consumption‑psychology embeddings.