ASMG

ASMG stands for Adaptive Sequential Model Generation.

research paper

Peng et. al., “Learning an Adaptive Meta Model-Generator for Incrementally Updating Recommender Systems”. RecSys, 2021.

Recommender Systems (RSs) in real-world applications often deal with billions of user interactions daily. To capture the most recent trends effectively, it is common to update the model incrementally using only the newly arrived data. However, this may impede the model's ability to retain long-term information due to the potential overfitting and forgetting issues. To address this problem, we propose a novel Adaptive Sequential Model Generation (ASMG) framework, which generates a better serving model from a sequence of historical models via a meta generator. For the design of the meta generator, we propose to employ Gated Recurrent Units (GRUs) to leverage its ability to capture the long-term dependencies. We further introduce some novel strategies to apply together with the GRU meta generator, which not only improve its computational efficiency but also enable more accurate sequential modeling. By instantiating the model-agnostic framework on a general deep learning-based RS model, we demonstrate that our method achieves state-of-the-art performance on three public datasets and one industrial dataset.

The model $Θ_t$ updated from $Θ_{t-1}$ is obtained by minimizing loss on $D_t$:

$$Θ_t = \argmin_Θ \mathcal{L}(Θ|D_t,Θ_{t-1})$$

where $\mathcal{L}(Θ|D_t,Θ_{t-1})$ is the loss on $𝐷_𝑡$ with $Θ_{𝑡−1}$ as initialization (for recommendation tasks, the loss can be pointwise log loss or pairwise BPR loss). The initial model $Θ_0$ can be obtained by random initialization or pre-training with sufficient amount of historical data.

Now, we assume that mining the temporal trends exist in this sequence can have great potential in generating a better model for the next period’s serving. Let $\mathcal{M}_𝑡$ denote the meta generator that takes in a sequence of models until $Θ_𝑡$, the output model $Θ_t^*$ used for serving $𝑇_{𝑡+1}$ is obtained by:

$$Θ_t^* = \mathcal{M}_t(Θ_{1:t})$$

where, $Θ_{1:t}$ is a sequence of models of length t from $Θ_1$ to $Θ_t$. Note that $Θ_0$ is excluded from the input sequence, as it is obtained from random initialization or pre-training instead of regular incremental update.

To generate model that is particularly good at serving the next period, we need to ensure that past knowledge that is specially useful for the next period’s serving is extracted. To achieve this, we propose to update the meta generator by adaptively optimizing the output model towards the next period data. More specifically, let $\omega_𝑡$ denote the parameters of $\mathcal{M}_𝑡$, i.e., $\mathcal{M}_𝑡 = \mathcal{M}_{\omega_𝑡}$, the parameters $\omega_{𝑡+1}$ are updated from $\omega_𝑡$ by optimizing $Θ_𝑡^*$ towards $𝐷_{𝑡+1}$:

$$\omega_{t+1} = \argmin_{\omega} \mathcal{L}(Θ_𝑡^*|D_{t+1},\omega_t) = \argmin_{\omega} \mathcal{L}(\mathcal{M}_\omega(Θ_{1:𝑡})|D_{t+1},\omega_t)$$

To ensure that the meta generator is sufficiently trained before it can be deployed for online model generation, we perform 𝜏 periods of meta generator warm-up training prior to the online deployment. During the online deployment phase, the meta generator will still need to be adaptively updated in order to keep up with the most recent sequential trends.

Comparison between conventional incremental update and after applying ASMG framework.

Algorithm

Variants

ASMG-GRUfull

This variant trains GRU meta generator on the full sequence of historical models.

ASMG-GRUzero

This variant uses zero-initialized input hidden state for every period’s training instead of continuing on a previously learned hidden state.

ASMG-GRUmulti

The intuition behind is to enforce that the hidden states generated at different steps are on the right track and able to produce model that serves well for the respective next period. More specifically, when training $M_{𝑡+1}$, instead of optimizing the last model towards $𝐷_{𝑡+1}$ only, we optimize all concurrently. To take into account that the later data is less seen before and hence more informative, we assign greater weight 𝜆 to loss of the more recent data.

Training of GRU meta generator under the ASMG framework for three consecutive periods (i.e., training of M4, M5 and M6). Transition from (a) to (b) illustrates the first strategy, which trains the GRU meta generator on truncated sequence (here we use sequence length 𝑘 = 3) by continuing on a previously learned hidden state. Transition from (b) to (c) illustrates the second strategy, which performs concurrent training by optimizing data at multiple steps, with greater weights assigned to the more recent data. ASMG-GRUsingle and ASMG-GRUmulti are depicted by (b) and (c) respectively.

ASMG-GRUunif

This variant assigns uniform weights to the loss at different time steps. Illustration of this variant is the same as ASMG-GRUmulti, as the weights assigned are not explicitly shown in the diagram.

ASMG-GRUsingle

This variant performs single-step training at the last output model towards the newly collected data only.

Links

• https://arxiv.org/abs/2111.04282v1