Disentangled Interest Network for Out-of-Distribution CTR Prediction

TL;DR

The paper addresses the challenge of out-of-distribution generalization in CTR prediction caused by evolving user interests. It introduces DiseCTR, a causal framework that factorizes CTR into $P(Z)$, $P(X|Z)$, and $P(Y|X,Z)$ and learns disentangled, independent interest embeddings via a sparse attention-based encoder, a prototype-based disentangler with weak supervision, and an attentive aggregator. Empirical results on three real-world datasets show that DiseCTR achieves superior AUC/GAUC and logloss performance under OOD conditions, and its disentangled interests enable fast, data-efficient transfer with strong explainability. The work advances robust CTR systems for real-world deployments where distribution shifts are common and highlights avenues for future integration with sequential modeling and online testing.

Abstract

Click-through rate (CTR) prediction, which estimates the probability of a user clicking on a given item, is a critical task for online information services. Existing approaches often make strong assumptions that training and test data come from the same distribution. However, the data distribution varies since user interests are constantly evolving, resulting in the out-of-distribution (OOD) issue. In addition, users tend to have multiple interests, some of which evolve faster than others. Towards this end, we propose Disentangled Click-Through Rate prediction (DiseCTR), which introduces a causal perspective of recommendation and disentangles multiple aspects of user interests to alleviate the OOD issue in recommendation. We conduct a causal factorization of CTR prediction involving user interest, exposure model, and click model, based on which we develop a deep learning implementation for these three causal mechanisms. Specifically, we first design an interest encoder with sparse attention which maps raw features to user interests, and then introduce a weakly supervised interest disentangler to learn independent interest embeddings, which are further integrated by an attentive interest aggregator for prediction. Experimental results on three real-world datasets show that DiseCTR achieves the best accuracy and robustness in OOD recommendation against state-of-the-art approaches, significantly improving AUC and GAUC by over 0.02 and reducing logloss by over 13.7%. Further analyses demonstrate that DiseCTR successfully disentangles user interests, which is the key to OOD generalization for CTR prediction. We have released the code and data at https://github.com/DavyMorgan/DiseCTR/.

Figures (9)

• Figure 1: (a) OOD issue on a video recommendation dataset shown with histogram of $\Delta P$, where $\Delta P = P_{eval}(Y=1|X) - P_{train}(Y=1|X)$ is the difference of CTR on popular videos between training and evaluation data. About 60% of users have a distribution variation larger than 0.1, and only 40% of the data can be considered as IID. (b) Performance of three popular CTR prediction models on both IID and OOD data, showing their degraded performance when OOD issue occurs.
• Figure 2: Existing entangled methods directly learn $P(Y|X)$, while DiseCTR learns $P(Z|X)$ and $P(Y|Z)$ separately with disentangled interests. Under partial changes of interests (gray in the figure), existing methods tend to fail due to the large variation of $P(Y|X)$, while DiseCTR is more robust since variation of $P(Z|X)$ and $P(Y|Z)$ is relatively small.
• Figure 3: (a) Causal graph of features $X$, interactions $Y$ and interests $Z$. (b) Partial distribution shift .
• Figure 4: We first cluster $\tilde{\mathbf{Z}}$ towards a set of interest prototypes, then impose weak supervision by averaging the pair adaptively which makes $\mathcal{A}$ and $\bar{\mathcal{A}}$ capture the overlapped and distinguishing interests of the pair respectively. Cosine similarity loss is further minimized to reduce mutual information between different interests. (Best viewed in color).
• Figure 5: (a) Logloss of different methods on the Kuaishou dataset under OOD-hard protocol. (b) AUC of baseline methods with the affected feature (price) discarded on Amazon dataset.