Non-autoregressive Personalized Bundle Generation

TL;DR

This work proposes to perform the bundle generation via non-autoregressive mechanism and design a novel encoder-decoder framework named BundleNAT, which can effectively output the targeted bundle in one-shot without relying on any inherent order.

Abstract

The personalized bundle generation problem, which aims to create a preferred bundle for user from numerous candidate items, receives increasing attention in recommendation. However, existing works ignore the order-invariant nature of the bundle and adopt sequential modeling methods as the solution, which might introduce inductive bias and cause a large latency in prediction. To address this problem, we propose to perform the bundle generation via non-autoregressive mechanism and design a novel encoder-decoder framework named BundleNAT, which can effectively output the targeted bundle in one-shot without relying on any inherent order. In detail, instead of learning sequential dependency, we propose to adopt pre-training techniques and graph neural network to fully embed user-based preference and item-based compatibility information, and use a self-attention based encoder to further extract global dependency pattern. We then design a permutation-equivariant decoding architecture that is able to directly output the desired bundle in a one-shot manner. Experiments on three real-world datasets from Youshu and Netease show the proposed BundleNAT significantly outperforms the current state-of-the-art methods in average by up to 35.92%, 10.97% and 23.67% absolute improvements in Precision, Precision+, and Recall, respectively.

Figures (4)

• Figure 1: A toy example illustrating personalized bundle generation. Figure 1a: Given user $u$ and candidate items $i_{1},...,i_{6}$, the goal is to generate the preferred bundle consists of $i_{1}$, $i_{5}$, and $i_{6}$. Figure 1b shows the 4-step generating process based on sequence modeling, which is unaware of multiple optimal sequential orders and might result in inference failure when changing the sequential order. Figure 1c illustrates the proposed non-autoregressive generation which aims to output the size-3 bundle in 1 step by utilizing preference and compatibility information (An encoder-decoder architecture is adopted in our paper).
• Figure 2: The overall architecture of the proposed framework. $u_{t}$ stands for the target user, the items within the ground-truth bundle are marked by $i_{t}$.
• Figure 3: Overall training time comparison on all the three datasets. We compare the proposed BundleNAT with bundle-specific methods, i.e., Exact-k and BYOB.
• Figure 4: Influence of hyper-parameters on Netease (K=5, N=100) dataset.