The winter holiday shopping season has begun, and just as department stores are dressing up their window displays, so are e-commerce companies working on the product recommendation systems they use to identify and target online shoppers’ wants and needs.
Building an effective product recommendation system usually starts with a click-through rate (CTR) prediction model, which is designed to estimate the probability of users clicking on recommended item candidates. Deployment of a CTR model is considered one of the core tasks in non-searching e-commerce, as its performance not only affects platform revenue but also influences customers’ online shopping experience.
With the development of deep learning, CTR prediction methods have gradually transitioned from traditional solutions to deep learning models such as Wide & Deep, DeepFM and Deep Interest Network (DIN). Although these powerful models can effectively leverage big data to better understand shoppers, they lack the capability to dynamically capture user interests or find latent interests behind consumer behaviors.
To improve the effectiveness of advertisement display, Alibaba group researchers have proposed a new Deep Interest Evolution Network (DIEN) to better predict customers’ CTR. DIEN has two core modules: 1. temporally captures and extracts latent interests based on customer history behaviors; 2. models an evolving process of user interests.
Like most deep CTR models, DIEN is built on an embedding & multilayer perceptron (MLR) paradigm: large-scale sparse features (user interest, ad, user profile, context) are first embedded and compressed into vectors for each category; the embedding vectors from different categories are then sent into sum pooling operation, concatenated together and finally fed into an MLP to learn the nonlinear relationship among all features.
DIEN’s advantage over existing CTR prediction models is its ability to capture user interests and interest dynamics with an interest extractor layer and an interest evolving layer.
Behaviors are first sorted by time and transformed to embedding vectors. Then, at the interest extractor layer, researchers take the gated recurrent unit (GRU) supervised by the auxiliary loss to effectively capture sequential user interests. Because a customer’s taste and preferences change over time, a GRU attentional update gate (AUGRU) that can overcome the inference from interest drifting is implemented at the interest evolving layer to model the interest evolving process related to target items.
Researchers used a dataset built from customer reviews in the Books and Electronics subsets of the Amazon public dataset, and an industrial dataset of ad click logs from Alibaba’s online display advertising system to test DIEN’s performance.
The Statistics of Datasets
In comparisons between DIEN and several state of the art CTR prediction methods, DIEN outperformed the others on both the public and industrial datasets.
Results (AUC) on Public Datasets
Results (AUC) on Industrial Datasets
In an online test on the Taobao.com advertisement system this June, DIEN’s deployment resulted in a 20.7 percent CTR increase compared with the BaseModel proposed by the same research group.
In practice, DIEN’s online deployment and effectiveness also includes challenges for a commercial system, especially if the system encounters extremely high traffic for example during holiday shopping events when millions of users need to be served at the same time. To achieve better serving performance in such scenarios, techniques such as kernel fusion, adjacent requests batching and model compressing with rocket launching were utilized to reduce latency (from 38.2 ms to 6.6 ms) and keep throughput high (up to 360 queries per seconds).
Deep Interest Evolution Network for Click-Through Rate Prediction has been accepted by 2019 Advancement of Artificial Intelligence (AAAI) conference. The paper is available on arXiv.
Source: Synced China
Localization: Tingting Cao | Editor: Michael Sarazen