Enhancing Recommendation with Denoising Auxiliary Task

TL;DR

Noise in historical user interaction sequences hinders accurate next-action prediction in recommender systems. The authors propose Auxiliary Task Joint Training (ATJT), a self-supervised approach that jointly trains a noise-recognition auxiliary task to reweight sequences, using artificially generated noise to learn robust weights $w_i$. By combining a noise-recognition objective with the main recommender objective, ATJT achieves adaptive sequence weighting during training. Experiments across MovieLens20M, Amazon Electro, and Yelp on six base models demonstrate consistent improvements in AUC, HR@5, and NDCG@5, highlighting ATJT’s robustness and broad applicability in denoising sequential data for recommendation.

Abstract

The historical interaction sequences of users plays a crucial role in training recommender systems that can accurately predict user preferences. However, due to the arbitrariness of user behavior, the presence of noise in these sequences poses a challenge to predicting their next actions in recommender systems. To address this issue, our motivation is based on the observation that training noisy sequences and clean sequences (sequences without noise) with equal weights can impact the performance of the model. We propose a novel self-supervised Auxiliary Task Joint Training (ATJT) method aimed at more accurately reweighting noisy sequences in recommender systems. Specifically, we strategically select subsets from users' original sequences and perform random replacements to generate artificially replaced noisy sequences. Subsequently, we perform joint training on these artificially replaced noisy sequences and the original sequences. Through effective reweighting, we incorporate the training results of the noise recognition model into the recommender model. We evaluate our method on three datasets using a consistent base model. Experimental results demonstrate the effectiveness of introducing self-supervised auxiliary task to enhance the base model's performance.

Figures (5)

• Figure 1: The ATJT method consists of two main components: 1. Training the noise recognition model (compose with Noise Recognition Training Only and Noise Recognition Training and Inference). 2. Training the recommender model using reweighted sequences (compose with Recommender Training, Noise Recognition Inference Only and Noise Recognition Training and Inference).
• Figure 2: The left figure shows the division of noisy and clean sequences within a batch in the noise recognition model, following a 1:1 ratio. The right figure illustrates the partitioning of training data for the recommender model and the noise recognition model (blue represents training data $\mathcal{B}^{R}$ for the recommender model, and green represents training data $\mathcal{B}^{D}$ for the noise recognition model). (a) Data Replacement. (b) Data Partition.
• Figure 3: When training the recommender model with the $i$-th set of data from $\mathcal{B}^{R}$, concurrently train the noise recognition model with data from $\mathcal{B}^{D}$ in the range $[i(N-1), (i+1)(N-1)-1]$ (green represents training data for the noise recognition model, and blue represents training data $\mathcal{B}^{R}$ for the recommender model).
• Figure 4: Analyzing the impact of using different number of sequences as noisy sequences in the training of noise recognition models (Base Model: DCN). (a) MovieLens20M. (b) Amazon (Electro). (c) Yelp.
• Figure : Overall Optimization Algorithm of Model Training