INFLECT-DGNN: Influencer Prediction with Dynamic Graph Neural Networks

TL;DR

A novel profit-driven framework that supports decision-making based on model predictions that innovatively combines Graph Neural Networks and Recurrent Neural Networks with weighted loss functions, synthetic minority oversampling adapted to graph data, and a carefully crafted rolling-window strategy is introduced.

Abstract

Leveraging network information for predictive modeling has become widespread in many domains. Within the realm of referral and targeted marketing, influencer detection stands out as an area that could greatly benefit from the incorporation of dynamic network representation due to the continuous evolution of customer-brand relationships. In this paper, we present INFLECT-DGNN, a new method for profit-driven INFLuencer prEdiCTion with Dynamic Graph Neural Networks that innovatively combines Graph Neural Networks (GNNs) and Recurrent Neural Networks (RNNs) with weighted loss functions, synthetic minority oversampling adapted to graph data, and a carefully crafted rolling-window strategy. We introduce a novel profit-driven framework that supports decision-making based on model predictions. To test the framework, we use a unique corporate dataset with diverse networks, capturing the customer interactions across three cities with different socioeconomic and demographic characteristics. Our results show how using RNNs to encode temporal attributes alongside GNNs significantly improves predictive performance, while the profit-driven framework determines the optimal classification threshold for profit maximization. We compare the results of different models to demonstrate the importance of capturing network representation, temporal dependencies, and using a profit-driven evaluation. Our research has significant implications for the fields of referral and targeted marketing, expanding the technical use of deep graph learning within corporate environments.

Figures (7)

• Figure 1: Model architectures. (a) For each time point, the model takes the network as input, i.e., network connectivity, node & edge features. This input is first passed through GNN layer(s), which generate network embeddings in Euclidean space. These embeddings, together with the hidden state from the previous step (initially a tensor of ones), are then fed into RNN layer(s). This process is repeated for each time point. At the last time point, the output of the RNN model is fed into a fully connected layer to produce the final probabilities of a node being an influencer. The weights of these layers are learned by optimizing a loss function, using either class-balanced binary cross-entropy loss or class-balanced focal loss (see Section \ref{['loss_functions']}). (b) For each time point, the model takes node features as input (with the PageRank feature added as well in the corresponding configurations). These features, along with the hidden state from the previous step (initially a tensor of ones), are then fed into the RNN layer(s). This process is repeated for each time point. At the final time point, the output of the RNN model is fed into a fully connected layer to produce the final probabilities of a node being an influencer. The weights of these layers are learned by optimizing a loss function, using either class-balanced binary cross-entropy loss or class-balanced focal loss (see Section \ref{['loss_functions']}). (c) For each time point, the model takes the network as input, i.e., network connectivity, node & edge features. This input is passed through GNN layer(s) that generate network embeddings in Euclidean space. The embeddings are then passed to a Fully Connected Layer to produce the final probabilities of a node being an influencer. The weights of these layers are learned by optimizing a loss function, using either class-balanced binary cross-entropy loss or class-balanced focal loss (see Section \ref{['loss_functions']}). There is no temporal component in this configuration, and time points are used in a batch manner.
• Figure 2: Labelling intuition
• Figure 3: Train/validation/test setup
• Figure 4: Convergence plots, GIN-GRU models
• Figure 5: $p$ vs. $c$ analysis