New User Event Prediction Through the Lens of Causal Inference

TL;DR

The paper addresses cold-start next-event prediction for new users with limited history by reframing history as an intervention and user category as a confounder, enabling unbiased counterfactual estimation via inverse propensity weighting. It introduces a category-agnostic intensity framework and an alternating learning algorithm that updates propensity weights and model parameters, supported by a nonparametric history-transition estimator and a theoretical analysis of bias-variance through bin discretization. Empirical results on synthetic data and real datasets (Netflix ratings and Amazon seller contact) show consistent improvements over standard neural and Hawkes-based point processes, especially in heterogeneous and sparse regimes. The approach offers a principled, scalable path to robust, category-agnostic event prediction in large, dynamic user bases and cold-start situations.

Abstract

Modeling and analysis for event series generated by users of heterogeneous behavioral patterns are closely involved in our daily lives, including credit card fraud detection, online platform user recommendation, and social network analysis. The most commonly adopted approach to this task is to assign users to behavior-based categories and analyze each of them separately. However, this requires extensive data to fully understand the user behavior, presenting challenges in modeling newcomers without significant historical knowledge. In this work, we propose a novel discrete event prediction framework for new users with limited history, without needing to know the user's category. We treat the user event history as the "treatment" for future events and the user category as the key confounder. Thus, the prediction problem can be framed as counterfactual outcome estimation, where each event is re-weighted by its inverse propensity score. We demonstrate the improved performance of the proposed framework with a numerical simulation study and two real-world applications, including Netflix rating prediction and seller contact prediction for customer support at Amazon.

Figures (6)

• Figure 1: The goal is to predict the next event for a new user without knowing its category. The distribution of the next event is influenced by its category even with the same event history. Each event $x$ consists of its occurrence time $t$ and associated mark data $m$. The conditional intensity $\lambda(x|\mathcal{H}_t, C)$ represents the occurrence rate of the next event $x$ given its history $\mathcal{H}_t$ and the user's category $C$.
• Figure 2: Causal DAG between user category, history variables, and category-agnostic intensity. The lowercase notations represent the realizations of these variables. We use $n$ to denote the number of observed events in the history and $h(x)$ (or simply $h$) denotes the history when we observe the next ($n+1$)-th event at $x$.
• Figure 3: The histogram of $p(h_i | C)$ and the conditional transition probability $p(h_i | h_{i-1}, C)$ for the three categories $C=1, 2, 3$ when $1/\delta=20$ visually reveal distinct patterns. These differences in the conditional transition probabilities across user categories highlight their unique behavior pattern.
• Figure 4: Category-wise prediction MAE of synthetic experiments.
• Figure 5: Convergence analysis.

Theorems & Definitions (13)

• Definition 1: Random history variable
• Lemma 1: Markov property
• Definition 2: Dependence of users' future event
• Definition 3: Category-agnostic intensity
• Lemma 2: Category-agnostic probability density
• Proposition 1: Weighted maximum log-likelihood estimation
• Proposition 2: Improvement in binning
• Proposition 3: Optimal bin size
• Remark 1
• Remark 2