A User-Adaptive Layer Selection Framework for Very Deep Sequential Recommender Models

Abstract

Sequential recommender systems (SRS) have become a research hotspot in recent studies. Because of the requirement in capturing user's dynamic interests, sequential neural network based recommender models often need to be stacked with more hidden layers (e.g., up to 100 layers) compared with standard collaborative filtering methods. However, the high network latency has become the main obstacle when deploying very deep recommender models into a production environment. In this paper, we argue that the typical prediction framework that treats all users equally during the inference phase is inefficient in running time, as well as sub-optimal in accuracy. To resolve such an issue, we present SkipRec, an adaptive inference framework by learning to skip inactive hidden layers on a per-user basis. Specifically, we devise a policy network to automatically determine which layers should be retained and which layers are allowed to be skipped, so as to achieve user-specific decisions. To derive the optimal skipping policy, we propose using gumbel softmax and reinforcement learning to solve the non-differentiable problem during backpropagation. We perform extensive experiments on three real-world recommendation datasets, and demonstrate that SkipRec attains comparable or better accuracy with much less inference time.

One Person, One Model--Learning Compound Router for Sequential Recommendation

“…Then the whole routing space can be described by the product of these parameters. It is worth noting that our design can well generalize to existing works [14], [16] by only considering layer depth or embedding size in the routing space. Moreover, it can scale up to more complex parameters like layer-wise hidden size instead of being limited to the experimental setup.…”

“…In this section, we describe CANet in detail, which assigns suitable submodels to input user sequences adaptively to reduce computation cost with minimal degradation on recommendation performance. To the best of our knowledge, most previous works are focused on designing a general powerful architecture manually or automatically for all the inputs, while only a few attempts are devoted to modeling user preferences by personalized architectures [16]. We first present the whole framework then introduce the optimization strategies.…”

“…• SkipRec [16] is a user-specific depth selection framework where the number of network layers can be selected on a per-user basis. The max searching depth is 8.…”

“…To tackle such problem, embedding size searching methods [13], [14] are widely studied, aiming to search for suitable embedding sizes for input features in an automatic way to reduce parameters in embedding layer. But on the other hand, recent few studies point out that even the hidden layers of recommendation models can also be wider and deeper to gain a performance boost [8], [16]. How to achieve computation efficient in recommendation is becoming more and more important and some works have been devoted in this searching area.…”

“…Besides, embedding size search methods [13], [14] try to figure out a suitable embedding size for input features to achieve storage space efficiency in a automated schema [15]. Further in a view of dynamic architecture, SkipRec [16] utilizes the layer-skipping schema to reduce unimportant computation in the middle layers. Despite their effectiveness, we hold these methods are not scalable enough since they implement a fixed network or a single-dimensional adaptive architecture.…”