Source-linked AI summary
Deep Interest Evolution Network for Click-Through Rate Prediction
Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, Kun Gai
TL;DR
CTR models need to represent latent user interests and their dynamic evolution, but many existing methods treat explicit behavior as interest directly and overlook changing interest trends. DIEN extracts temporally supervised interests and evolves them with target-item-aware attention, significantly outperforming state-of-the-art solutions on public and industrial datasets and improving Taobao display-ad CTR by 20.7%.
Problem
Many CTR interest models treat explicit behavior as the interest directly and provide limited modeling of latent interests and their evolving trends.
Method
DIEN uses an auxiliary-loss-supervised interest extractor and an interest-evolving layer with an attentional update gate related to the target item.
Results
DIEN significantly outperforms state-of-the-art solutions on public and industrial datasets and obtains 20.7% CTR improvement in Taobao’s display advertisement system.
Takeaways & Limitations
Modeling latent interest extraction and target-relative interest evolution yields more expressive interest representations and more precise CTR prediction.
Takeaways & Limitations
Traditional RNN-based models addressed by DIEN can be disturbed by interest drifting because they use one fixed evolution track for every target item.
Abstract
from arXiv · showhide
Click-through rate~(CTR) prediction, whose goal is to estimate the probability of the user clicks, has become one of the core tasks in advertising systems. For CTR prediction model, it is necessary to capture the latent user interest behind the user behavior data. Besides, considering the changing of the external environment and the internal cognition, user interest evolves over time dynamically. There are several CTR prediction methods for interest modeling, while most of them regard the representation of behavior as the interest directly, and lack specially modeling for latent interest behind the concrete behavior. Moreover, few work consider the changing trend of interest. In this paper, we propose a novel model, named Deep Interest Evolution Network~(DIEN), for CTR prediction. Specifically, we design interest extractor layer to capture temporal interests from history behavior sequence. At this layer, we introduce an auxiliary loss to supervise interest extracting at each step. As user interests are diverse, especially in the e-commerce system, we propose interest evolving layer to capture interest evolving process that is relative to the target item. At interest evolving layer, attention mechanism is embedded into the sequential structure novelly, and the effects of relative interests are strengthened during interest evolution. In the experiments on both public and industrial datasets, DIEN significantly outperforms the state-of-the-art solutions. Notably, DIEN has been deployed in the display advertisement system of Taobao, and obtained 20.7\% improvement on CTR.
Introduction
DIEN addresses CTR prediction’s need to represent latent, evolving user interests rather than treating explicit behaviors as interests directly. It extracts supervised temporal interests and models target-item-related evolution, achieving strong results across public and industrial datasets.
- CTR prediction must capture user interests and their dynamics because users often do not explicitly express current intention in e-commerce settings.
- Existing interest models, including DIN, often treat observed behavior as interest directly, leaving latent interest insufficiently represented.
- DIEN uses an interest extractor layer with GRU-based sequential modeling and auxiliary loss from the next behavior to supervise each hidden state.
- DIEN’s interest-evolution model produces more expressive interest representations and more precise CTR prediction.
- Its interest evolving layer uses an attentional update gate to strengthen target-item-relevant interests and weaken effects from interest drifting.
- DIEN significantly outperforms state-of-the-art solutions on public and industrial datasets and has been deployed in Taobao’s display advertisement system.
Related Work
Prior sequential CTR and recommendation models use temporal behavior information but commonly lack specialized supervision for interest representation and flexibility for diverse interest-evolution tracks. DIEN addresses these limitations with target-aware attentional sequential learning.
- Deep CTR models increasingly model feature interactions, while sequential recommendation methods use temporal interactions to build richer user representations.
- Traditional RNN-based models often treat hidden states as latent interests without special supervision for interest representation.
- These models also process adjacent-behavior dependencies equally, although users have diverse interests with separate evolution tracks.
- A fixed interest-evolution track for every target item can expose traditional sequential models to disturbance from interest drifting.
- Prior work introduces extra supervision such as click-level sequential prediction, while attention-based GRU structures provide more flexible sequential learning.
Deep Interest Evolution Network
DIEN models CTR by first extracting latent temporal interests from user behavior sequences, then modeling their target-item-relative evolution. Its pipeline combines embeddings, GRU-based interest extraction with auxiliary supervision, AUGRU-based evolution, and final CTR prediction.
- Feature Representation: DIEN transforms categorical features and time-ordered behaviors into embedding vectors before modeling user interests.The base structure embeds sparse features, while behavior embeddings preserve the ordered history sequence.
- Interest Extractor Layer: GRU extracts an interest state at each behavior step, but its hidden states require auxiliary supervision to represent latent interests effectively.The auxiliary loss uses the next behavior to supervise each intermediate state, rather than supervising only the final CTR prediction.
- Deep Interest Evolution Network: The final evolved interest state is concatenated with other feature embeddings and passed to MLR for final CTR prediction.Figure 1 describes the end-to-end flow from behavior embeddings through interest evolution to prediction.
- Interest Extractor Layer: The global objective combines target-item CTR loss with auxiliary loss, balancing final prediction and intermediate interest representation through α.The loss is L = Ltarget + α ∗ Laux.
- Interest Evolving Layer: DIEN’s interest evolving layer models how extracted interests change over time, focusing evolution on interests relevant to the target item.The model represents the extracted states as an interest sequence and processes them with a second recurrent module.
- Interest Evolving Layer: AUGRU combines attention with GRU updates so relevant interests exert stronger influence while interest drifting has weaker effects.The design addresses AGRU’s replacement of a vector update gate with a scalar attention score by integrating attention into the update gate.
Experiments
DIEN is evaluated on public and industrial datasets using user behavior histories to predict subsequent user actions or clicks. The public-data results are reported with AUC.
- Datasets: DIEN is evaluated on both public and industrial datasets.The experiments use Amazon Books and Electronics subsets plus an industrial display-advertising dataset.
- Public datasets: On Amazon, reviews are treated as time-ordered behaviors, with the first T−1 reviews used to predict the T-th review.The prediction target is whether the user writes the review shown in the T-th position.
- Industrial dataset: The industrial dataset uses impression and click logs, with 14 days of prior behaviors paired with target advertisements.Training targets are ads clicked during the last 49 days, while test targets come from the following day.
- Evaluation: Table 2 reports results using AUC on public datasets.
Compared Methods
DIEN is compared with mainstream CTR prediction methods spanning pooling, linear-plus-deep modeling, feature products, attention, and recurrent attention.
- Baselines: BaseModel integrates behavior embeddings with sum pooling while using the same embedding and MLR settings as DIEN.
- Baselines: Wide & Deep combines a deep model matching BaseModel with a linear wide model.
- Baselines: PNN captures interactions between interfield categories through a product layer.
- Interest models: DIN activates user behaviors related to the target item through attention, while Two layer GRU Attention applies recurrent modeling followed by attention.
Results on Public Datasets
On public datasets, interest-focused models improve AUC over simpler baselines, and DIEN further models target-relative interest evolution. DIEN’s sequential interest modeling yields the strongest reported performance.
- Results: AUC improves for models that capture user interests, including DIN and Two layer GRU Attention.DIN activates interests relative to the ad, while Two layer GRU Attention further activates relevant interests in the sequence.
- Baselines: PNN improves BaseModel performance through automatic feature interaction, whereas Wide & Deep with manually designed features performs poorly.
- DIEN: DIEN captures sequential interests and models interest evolution relative to the target item.The paper reports that this produces better interest representation and more precise modeling of dynamic interests.
- Results: DIEN obtains better performance than the compared interest-modeling approaches on the public datasets.
Results on Industrial Dataset
On the industrial display-advertising dataset, attention-based methods are especially effective because the data contains many product categories. DIEN achieves the best reported performance by modeling target-relative interest evolution.
- Industrial setup: The industrial model uses six FCN layers with dimensions 600, 400, 300, 200, 80, and 2.The maximum history length is 50 behaviors.
- Baselines: Wide & Deep and PNN outperform BaseModel on the industrial dataset.
- Results: Attention-based methods improve performance substantially because online advertising contains many product categories simultaneously.The paper contrasts this setting with Amazon, where the dataset covers one goods category.
- DIEN: DIEN obtains the best performance on the industrial dataset by capturing interest evolution relative to the target item.
Application Study
DIEN’s auxiliary loss and AUGRU improve interest modeling in the application study, with auxiliary supervision helping sequential-interest learning and AUGRU effectively capturing target-relative evolution.
- AUGRU: AUGRU obtains obvious improvements over alternative interest-evolving methods, effectively fusing attention with sequential learning.The comparison attributes its advantage to better use of the update gate and stronger modeling of relative interests.
- Learning curves: Both global CTR loss and auxiliary loss show similar descending trends during training.This indicates that both objectives contribute during optimization.
- Auxiliary loss: Auxiliary loss brings great improvements on both public datasets, highlighting the value of supervision for sequential interests and embedding representations.On the industrial dataset, the improvement is smaller than on public datasets.
- Auxiliary loss: The industrial auxiliary-loss gain is less obvious because the dataset is large and its cross-category, cross-scene supervision can be heterogeneous with the target item.The passage states that this setting magnifies the effect of AUGRU relative to auxiliary loss.
Visualization of Interest Evolution
The visualization projects AUGRU hidden states to show that interest evolution changes with the target item and follows target-dependent activation patterns across behavior histories.
- Interest evolution: PCA-projected AUGRU hidden states trace different moving routes for the same history behaviors when activated by different target items.The histories include products from computer, audio, vehicle-navigation, storage, and accessory categories.
- Interest evolution: The None-target curve represents equal attention scores, corresponding to interest evolution unaffected by the target item.The projected hidden states are linked in behavior order to visualize the evolution trajectory.
Online Serving & A/B testing
DIEN achieved strong online A/B-test gains in Taobao display advertising and was deployed for main traffic, while online serving required latency- and throughput-oriented optimizations.
- Online A/B testing: 20.7% CTR improvement and 17.1% eCPM improvement were obtained against BaseModel in Taobao online A/B testing.The test ran from 2018-06-07 to 2018-07-12 in the display advertising system.
- Online A/B testing: 3.0% PPC decay accompanied the online gains, and DIEN was deployed to serve the main traffic.The passage also reports significant business revenue growth after deployment.
- Online serving: Serving DIEN at traffic peaks above 1 million users per second required element-parallel GRU kernel fusion and batching techniques.These techniques were used to maintain low latency and high throughput.
Conclusion
DIEN models interest evolution through supervised interest extraction and target-relative AUGRU dynamics, improving CTR prediction while motivating more personalized future interest models.
- Conclusion: DIEN uses an auxiliary-loss interest extractor and an attentional update gate to model target-relative interest evolution for CTR prediction.The paper states that AUGRU helps overcome disturbance from interest drifting.
- Conclusion: Modeling interest evolution helps capture user interest effectively and further improves CTR prediction performance.The paper identifies constructing a more personalized interest model as future work.