Personalized Negative Reservoir for Incremental Learning in Recommender Systems

TL;DR

This work addresses the challenge of incremental learning for graph-based recommender systems under implicit feedback, where naive fine-tuning leads to catastrophic forgetting and traditional negative samplers neglect time-evolving user interests. It introduces GraphSANE, a personalized negative reservoir that biases hard negative sampling toward item categories where a user shows declining interest, guided by a Dirichlet prior on per-user category distributions and end-to-end differentiable clustering of items into latent clusters. The method combines a hard-negative reservoir with standard knowledge-distillation-based incremental learning and demonstrates state-of-the-art improvements across six large datasets when integrated with three SOTA incremental models, with notable gains for users exhibiting high interest shift. The approach offers a scalable, plug-in enhancement to existing pipelines, providing faster convergence and more accurate recommendations in dynamic, production-scale settings.

Abstract

Recommender systems have become an integral part of online platforms. Every day the volume of training data is expanding and the number of user interactions is constantly increasing. The exploration of larger and more expressive models has become a necessary pursuit to improve user experience. However, this progression carries with it an increased computational burden. In commercial settings, once a recommendation system model has been trained and deployed it typically needs to be updated frequently as new client data arrive. Cumulatively, the mounting volume of data is guaranteed to eventually make full batch retraining of the model from scratch computationally infeasible. Naively fine-tuning solely on the new data runs into the well-documented problem of catastrophic forgetting. Despite the fact that negative sampling is a crucial part of training with implicit feedback, no specialized technique exists that is tailored to the incremental learning framework. In this work, we propose a personalized negative reservoir strategy, which is used to obtain negative samples for the standard triplet loss of graph-based recommendation systems. Our technique balances alleviation of forgetting with plasticity by encouraging the model to remember stable user preferences and selectively forget when user interests change. We derive the mathematical formulation of a negative sampler to populate and update the reservoir. We integrate our design in three SOTA and commonly used incremental recommendation models. We show that these concrete realizations of our negative reservoir framework achieve state-of-the-art results for standard benchmarks using multiple top-k evaluation metrics.

Figures (6)

• Figure 1: Toy example with daily model updates and 3 item categories: pets, car parts and coffee items. The figure shows a user's interactions with these categories over a week. (a) Cumulative user interest over time: A larger fraction of user $u$'s interactions comes from pet items at the start of the week, but his interests change over time and he interacts with more car part items by the end of the week. (b) The change of preference is reflected on the daily histograms of user-item-category interactions $\mathbf{H}_{u,t}$. (c) The histograms are normalized $\widehat{\mathbf{H}}_{u,t}$ and projected on the simplex $\Delta_2$. There we see the interest shift on the simplex representing the user's interest shift from $t=2,$ (Tuesday) to $t=3$, (Wednesday) and on the right the user's interest shift from $t=3,$ (Wednesday) to $t=4$, (Thursday). We see that the user exhibits a large change in interests on the second simplex (moves far away from current neighborhood in simplex denoted by a red dashed circle). (d) We propose to sample more negatives from pet category when fine tuning the model on Thursday to allow the model to quickly adjust to new user interests (in the figure the right bucket has many more pet items).
• Figure 2: The hierarchical model in equation \ref{['eq:prior']} & 8. From time $t-1$ to $t$, a user loses interest in the i-th category, this increases $\boldsymbol{\alpha}_i$. Thus $\boldsymbol{\theta}_i$ has a high prior value so we sample negatives with higher probability from that item category in equation \ref{['eq:sample_prob']}.
• Figure 3: Left: Boxplot of ranks of the algorithms across the 6 datasets. The medians and means of the ranks are shown by the vertical lines and the black triangles respectively; whiskers extend to the minimum and maximum ranks. Stars "*" represent methods that integrate our proposed negative reservoir. We abbreviate GraphSail to "GSAIL". Right: Comparison of recall performance and training time required per incremental block for various negative samplers of Yelp discussed in Table \ref{['tab:samplers_table']}. Our method (SANE) outperforms with time complexity in line with the baselines. Note that full batch retraining takes more than an hour.
• Figure 4: Case Study: Among the 15% of users with the highest interest shift on Taobao14 we observe that Graph-SANE increases the amount of old positives sampled as current negatives by 7 times. Our sampler strongly improves Recall@20.
• Figure 5: Data is separated to a base block with 60% of the data and 4 incremental blocks, each with 10% of the remaining data.