Marketing ops · Production

Zalando AdTech Lab uses recurrent neural networks to predict consumer order propensity from interaction histories

The problem

Zalando's ad-tech team needed to predict consumer order behavior from sequential interaction histories, but traditional machine learning required labor-intensive manual feature engineering and produced models that were difficult to interpret for monitoring and collaborator sign-off.

Workflow diagram · grounded in source

Consumer event history ingested

trigger

“Customer histories are sequences of events. For each event, we know its type (product view, cart addition, order, etc.), its timestamp, and further information such as the viewed product or the fashion category the consumer is currently …”

RNN processes raw sequence

ai_action

“We feed consumer histories directly into RNNs. RNNs are made up of a sequence of computational cells.”

LSTM cells compute consumer state

ai_action

“We use long short-term memory cells (LSTMs)... The latent cell state at the last time-step (shown in orange) is used for the final prediction. This prediction can use a simple linear logistic regression layer or a sequence of non-linear …”

Order probability score output

output

“we take this as our running example and focus on predicting order probability”

Latent state visualization for interpretability

feedback_loop

“These visualizations allow us to assess in a quantitative way how consumer actions affect model predictions. They deepen our understanding of how consumers interact with Zalando and shift vague assumptions about this process on a firm em…”

Reported outcome

RNNs achieved comparable or better predictive performance to the handcrafted-feature baseline while eliminating manual feature engineering, and latent-state visualizations provided interpretable insight into consumer behavior.

Reported metrics

RNN vs handcrafted-feature model accuracyabout the same or better performance

Feature engineering effort eliminatedwithout requiring tedious feature engineering efforts

Consumer order probability (exemplary session)jumps from 29% to 51%

Reported stack

recurrent neural networksLSTMs

Source

https://engineering.zalando.com/posts/2016/10/deep-learning-for-understanding-consumer-histories.html

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

RNNs achieved comparable or better predictive performance to the handcrafted-feature baseline while eliminating manual feature engineering, and latent-state visualizations provided interpretable insight into consumer…

What tools did this team use?

recurrent neural networks, LSTMs.

What results were reported?

RNN vs handcrafted-feature model accuracy: about the same or better performance; Feature engineering effort eliminated: without requiring tedious feature engineering efforts; Consumer order probability (exemplary session): jumps from 29% to 51% (source-reported, not independently verified).

How is this marketing ops AI workflow structured?

Consumer event history ingested → RNN processes raw sequence → LSTM cells compute consumer state → Order probability score output → Latent state visualization for interpretability.