Adversarial Training Model Unifying Feature Driven and Point Process Perspectives for Event Popularity Prediction

Abstract

This paper targets a general popularity prediction problem for event sequence, which has recently gained great attention due to its extensive applications in various domains. Feature driven method and point process method are two basic thinking paradigms to tackle the prediction problem, but both of them suffer from limitations. In this paper, we propose PreNets unifying the two thinking paradigms in an adversarial manner. On one side, feature driven model acts like a 'critic' who aims to discriminate the predicted popularity from the real one based on a set of temporal features from the sequence. On the other side, point process model acts like an 'interpreter' who recognizes the dynamic patterns in sequence to generate a predicted popularity that can fool the 'critic'. Through a Wasserstein learning based two-player game, the training loss of the 'critic' guides the 'interpreter' to better exploit the sequence patterns and enhance prediction, while the 'interpreter' pushes the 'critic' to select effective early features that helps discrimination. This mechanism enables the framework to absorb the advantages of both feature driven and point process methods. Empirical results show that PreNets achieves significant MAPE improvement for both Twitter cascade and Amazon review prediction.

PREP: Pre-training with Temporal Elapse Inference for Popularity Prediction

“…• Deep learning based methods. With the rapid development of deep learning and the availability of abundant online data, deep learning based popularity prediction methods have emerged in recent years [3,4,6,11,12,24,25,31,37,[40][41][42]44]. Based on the powerful representation ability of deep learning, researchers propose various deep prediction models to automatically extract representations for user [37], content [25], temporal [12,31,41] and structures [3,4,6,24,43].…”

“…• Mean Absolute Percentage Error (MAPE) [32,40]. This metric measures the average deviation between predicted and real popularity, i.e., 𝑀𝐴𝑃𝐸 = 1…”

“…Instead of training a separate prediction model through downstream labels for each popularity prediction task, the proposed framework aims to pre-train a general deep representation model to capture the characteristics of popularity dynamics by learning the intrinsic knowledge about popularity dynamics from readily available diffusion data without any guidance of downstream labels. Note that, the recent development of deep learning based popularity prediction models [3,4,6,12,24,25,31,37,[40][41][42]44] provides an important basis for the achievement of such a general deep representation model. Considering that the temporal information has shown to be the dominant prediction factor [7] and the most general information that can be generalized across different platforms, in this paper, we focus on time-aware popularity prediction scenario and take a recent state-of-the-art deep popularity prediction model, i.e., temporal convolutional neural network (TCN) [1,31] as the basic representation model.…”

Capturing Dynamics of Information Diffusion in SNS: A Survey of Methodology and Techniques

“…• Sequential features describe numerical changes of some indicators over time during the diffusion process. These indicators can be the number of reposts/comments/views [25,202,210], time elapsed between the current and previous post [25,70], and so on. • Statistical features are secondary processing of sequential features, aiming to find some potential diffusion laws.…”

“…The effectiveness of these mathematical models has been verified in multiple scenarios. Some articles involve these mechanisms [17,23,24,202]. Although neural networks have strong approximation capabilities, we believe that introducing the knowledge of time-series techniques can help model optimization, such as reducing the search space.…”