Deep Interest Evolution Network (DIEN): Modeling User Interest Evolution for Click‑Through Rate Prediction

Overview

This talk shares Alibaba's 2018 innovation on click‑through rate (CTR) estimation – the Deep Interest Evolution Network (DIEN). The presentation covers background, model structure, and final results.

Background

1. Business Scenario

In Taobao, ads appear as banners, item‑level placements, or within recommendation streams. Seamlessly integrating ads without harming user experience is crucial, requiring models that capture user interests beyond explicit intent.

2. Simple Models

Early CTR models combined a simple linear model (LR) with extensive handcrafted features. As computational resources grew, deep learning replaced manual feature engineering, leading to models such as DeepFM, PNN, and Deep&Wide.

3. Neural Networks

Basic DNN‑based CTR models embed user, item, and context features, concatenate them, and feed them to multi‑layer perceptrons with a final softmax. However, these generic designs struggle with the highly personalized, noisy behavior sequences in e‑commerce.

4. Deep Interest Network (DIN)

DIN introduces a weighted‑sum pooling mechanism that uses the candidate ad to activate historically clicked items, assigning higher weights to more relevant behaviors, thus producing a more expressive user‑interest vector.

Deep Interest Evolution Network (DIEN)

DIEN extends DIN by modeling the temporal evolution of user interests. Simple sequence modeling (e.g., vanilla RNN) proved insufficient due to irregular, noisy behavior streams.

1. Interest Extraction Module

After embedding user, context, and ad features, the behavior sequence is processed by a GRU to capture hidden interest states. An auxiliary loss, built from click‑through patterns of all web traffic (not just ads), is added to mitigate gradient vanishing and encourage the GRU to retain long‑term interest signals.

2. Interest Evolution Module

An attention mechanism computes relevance scores between each historical behavior and the candidate ad. These scores control the update gate of an AUGRU (Attention‑augmented GRU), allowing the model to selectively update hidden states when behaviors are highly related to the ad, and to preserve states otherwise.

Results

1. Offline Evaluation

Experiments on public Amazon datasets (electronics and books) and on Alibaba's production data show DIEN achieving the highest AUC, improving over the base model by 1.9 percentage points.

2. Online Evaluation

In a month‑long A/B test, DIEN increased eCPM by 17%, delivering significant commercial value.

Conclusion

The DIEN framework effectively extracts and evolves user interests, leading to state‑of‑the‑art CTR prediction performance both offline and in live traffic.