Parameters Efficient Fine-Tuning for Long-Tailed Sequential Recommendation

Abstract

In an era of information explosion, recommendation systems play an important role in people’s daily life by facilitating content exploration. It is known that user activeness, i.e., number of behaviors, tends to follow a long-tail distribution, where the majority of users are with low activeness. In practice, we observe that tail users suffer from significantly lower-quality recommendation than the head users after joint training. We further identify that a model trained on tail users separately still achieve inferior results due to limited data. Though long-tail distributions are ubiquitous in recommendation systems, improving the recommendation performance on the tail users still remains challenge in both research and industry. Directly applying related methods on long-tail distribution might be at risk of hurting the experience of head users, which is less affordable since a small portion of head users with high activeness contribute a considerate portion of platform revenue. In this paper, we propose a novel approach that significantly improves the recommendation performance of the tail users while achieving at least comparable performance for the head users over the base model. The essence of this approach is a novel Gradient Aggregation technique that learns common knowledge shared by all users into a backbone model, followed by separate plugin prediction networks for the head users and the tail users personalization. As for common knowledge learning, we leverage the backward adjustment from the causality theory for deconfounding the gradient estimation and thus shielding off the backbone training from the confounder, i.e., user activeness. We conduct extensive experiments on two public recommendation benchmark datasets and a large-scale industrial datasets collected from the Alipay platform. Empirical studies validate the rationality and effectiveness of our approach.

Z. Lv and F. Wang—These authors equally contributed to this study.

This work was supported in part by National Natural Science Foundation of China (62006207, 62037001, U20A20387), Young Elite Scientists Sponsorship Program by CAST (2021QNRC001), Zhejiang Province Natural Science Foundation (LQ21F020020), Project by Shanghai AI Laboratory (P22KS00111), Program of Zhejiang Province Science and Technology (2022C01044), the StarryNight Science Fund of Zhejiang University Shanghai Institute for Advanced Study (SN-ZJU-SIAS-0010), and the Fundamental Research Funds for the Central Universities (226-2022-00142, 226-2022-00051).

Appendices

A Pseudo Code

The pseudo code of our proposed method is summarized in Algorithm 1.

B Experiments

1.1 B.1 Experiments Settings

Datasets

Movielens¹. MovieLens is a widely used public benchmark on movie ratings. In our experiments, we use movielens-1M which contains one million samples.

Amazon². Amazon Review dataset [15] is a widely-known recommendation benchmark. We use the Amazon-Books dataset for evaluation.

Alipay. We collect a larger-scale industrial dataset for online evaluation from the AliPay platform³. Applets such as mobile recharge service are treated as items. For each user, clicked applets are treated as positives and other applets exposed to the user are negatives.

The detailed statistics of these datasets are summarized in Table 3.

Table 3. Statistics of the evaluation datasets.

Evaluation Metrics

$$\begin{aligned} \begin{aligned} & \textrm{AUC}=\frac{\sum _{x_0\in \mathcal {D}_T} \sum _{x_1 \in \mathcal {D}_F}\mathbbm {1}[f(x_1)<f(x_0)]}{|\mathcal {D}_T||\mathcal {D}_F|}, \\ & \text {HitRate}@K = \frac{1}{|\mathcal {U}|}\sum _{u\in \mathcal {U}} \mathbbm {1}(R_{u,g_u}\le K), \end{aligned} \end{aligned}$$

(14)

where $\mathbbm {1}(\cdot )$ is the indicator function, f is the model to be evaluated, $R_{u,g_u}$ is the rank predicted by the model for the ground truth item $g_u$ and user u, and $\mathcal {D}_T$, $\mathcal {D}_F$ is the positive and negative testing sample set respectively.

Baselines

GRU4Rec [4] is one of the early works that introduce recurrent neural networks to model user behavior sequences in recommendation.

DIN [30] introduces a target-attention mechanism for historically interacted items aggregation for click-through-rate prediction.

SASRec [7] is a representative sequential modeling method based on self-attention mechanisms. It simultaneously predicts multiple next-items by masking the backward connections in the attention map.

To evaluate the effectiveness on tail user modeling, the following competing methods are introduced for comparison.

Agr-Rand [14] introduced a gradient surgery strategy to solve the domain generalization problem by coordinating inter-domain gradients to update neural weights in common consistent directions to create a more robust image classifier.

PCGrad [25] is a very classic gradient surgery model that mitigates the negative cosine similarity problem by projecting the gradients of one task onto the normal components of the gradients of the other task by removing the disturbing components to mitigate gradient conflicts.

Grad-Transfer [24] adjusts the weight of each user during training through resample and gradient alignment, and adopts an adversarial learning method to avoid the model from using the sensitive information of user activity group in prediction to solve the long-tail problem.

Implementation Details

Preprocessing. On the Alipay dataset, the dates of all samples in the dataset are from 2021-5-19 to 2021-7-10. In order to simulate the real a/b testing environment, we use the date to divide the dataset. We take the data before 0:00 AM in 2021-7-1 as the training set, and vice versa as the test set. On Movielens and Amazon datasets, we treat the labels of all user-item pairs in the dataset as 1, and the labels of user-item pairs that have not appeared as 0. We take the user’s last sample as the test set. On Movielens, we use positive samples in the training set: the ratio of negative samples = 1:4 to sample negative samples. In the test set, we refer to [8], so we use all negative samples of a user as the test set. In Amazon’s training set, we sample negative samples with the ratio of positive samples: negative samples = 1:4, and this ratio becomes 1:99 in the test set. We also filter out all users and items in Amazon with less than 15 clicks to reduce the dataset. On Alipay and Amazon datasets, we group by the number of samples of users. On the Movielens dataset, we group by the length of the user’s click sequence

Implementation. In terms of hardware, our models are trained on workstations equipped with NVidia Tesla V100 GPUs. For all datasets and all models, the batch size is set to 512. The loss function is optimized by the Adam optimizer with a learning rate of 0.001 for the gradient aggregation learning stage and 0.0001 for the plugin model learning stage. The training is stopped when the loss converges on the validation set.

Fig. 4.

Performance of Hit@1 on validation set with different epochs.

1.2 B.2 Results

According to Fig. 4, the larger the group number, the more active the user is, that is, the first group is the least active user group, and the fifth group is the most active user group. Training the plugin network for relatively inactive groups of users requires only a small number of epochs to be optimal (such as groups 1 and 2). The training curve of the user group with relatively high activity level has a more stable upward trend with the increase of epoch (such as group 3, group 4 and group 5). This is mainly due to the difference in the amount of personalized information in user groups with different levels of activity. For a user group with a large amount of data, more personalized information is required, and more epochs are needed to learn the personalized information. Otherwise, only a few epochs are required.

Toy Example. In Fig. 5, we give a toy example of long-tail effect of a real case in Alipay platform and show the improvement brought by our propose method. In this case, there are two groups of women, one is young women with high activity and the other is middle-aged women with low activity. They have common preferences such as clothes and shoes but also some different preferences. Due to the long-tail effect, the preferences of low-activity women are difficult to capture, and the model will recommend some popular products for them. To address this problem, we extract generalization information via the gradient aggregation module so that the model can recommend common preferences such as clothes and shoes to low-activity women, although sometimes not her favorite style. The model’s recommendation for the high-activity women and the low-activity women are more similar, so the performance of the model on high-activity women has decreased. Next, we train a plugin network for each group of women. The plugin network captures the group-specific personalization information such as different styles of clothes and shoes and other unpopular preference.

Fig. 5.

A toy example of the long-tail effect and the improvement brought by our method.