[RS sampling] a gain tuning dynamic negative sampler for recommendation (WWW 2022)

《Simplify and Robustify Negative Sampling》 NIPS 2020

This article experimentally observed that although False Negative and Hard Negative There will be larger Socre, however False Negative There is a lower prediction variance . So I propose a Simplify and Robustify Negative Sampling Method , In the training epoch $t$ when , According to the former 5 individual epoch My training record , High prediction score 、 The sample with large variance is taken as Hard Negative：

A Gain-Tuning Dynamic Negative Sampler for Recommendation (WWW 2022)

Existing excavation RS The method of hard negative samples only wants to mine samples with large gradient contribution in the training process （ There is a big gap between prediction and label ）, stay RS In this scenario, it is easy to choose False negative sample （False Negative、missing data）, This leads to over fitting the training data set .

This paper presents a sampler based on expected gain , In the training process, according to the expected change of the gap between positive and negative samples , Dynamically direct negative sampling , False negative samples can be identified .

Gain aware negative sampler

Measure an object $j$ Is it the user $u$ The method of true negative samples ：
$${\mathcal{H}}^t(u,j)={\mathbb{E}}_{i\sim {\Delta }_u}\sigma \left(r_{u,j}-r_{u,i}\right)$$
The formula calculates the expectation ,$t$ It's training epoch,${\Delta }_u$ A collection of items that users have interacted with ,$\sigma$ yes sigmoid function , In parentheses is the score of the negative sample minus the score of the positive sample .

The negative sample selected in this way is close to the positive sample , It can provide a relatively large gradient for the training process , To provide more information . The ideal is very good , But experiments have found that such hard negative samples are really few , Instead, it is likely to choose pseudo negative samples . The experiment also found that , True negative samples ${\mathcal{H}}^t(u,j)$ The degree of change is greater than that of pseudo negative samples , Therefore, a measurement method of gain perception is further proposed , Monitor samples with large changes ：
$${\mathcal{G}}_{u,j}^t=\alpha \cdot {\mathcal{G}}_{u,j}^{t-1}+(1-\alpha )\cdot \sigma \left(\frac{{\mathcal{H}}_{u,j}^{t-1}-{\mathcal{H}}_{u,j}^t}{{\mathcal{H}}_{u,j}^t+ϵ}\right)$$
This indicator measures ${\mathcal{H}}^t(u,j)$ The degree of decline , The author thinks that two epoch The expected gain in the middle is the signal that is more sensitive to detect the difference between negative samples and positive samples . among $\alpha$ Is the smoothing coefficient ,$ϵ$ Is to prevent the denominator from being 0.

This indicator can be understood as , In the last epoch in , Which sample ${\mathcal{H}}^t(u,j)$ The decline is the most , Choose who is the negative sample .

Grouping optimizer

Proposed a similar MCL、CPR Of loss
$$\mathcal{L}\left(u,{\Delta }_u,{\Delta }_u^{\prime }\right)=\sum _{i\in {\Delta }_u}\sum _{j\in {\Delta }_u^{\prime }}{\left|r_{u,j}-r_{u,i}+\gamma \right|}_{+}$$
${\Delta }_u,{\Delta }_u^{\prime }$ They are users $u$ Positive sample set and negative sample set , It means that each positive sample should be calculated separately for all negative samples loss, Equal to all positive samples share negative sample information , Instead of one-on-one optimization , More efficient , More information . and CPR and MCL It means very much .

experimental result

base The model is GMF：$r_{u,i}=W^{\top }\left(P_u\odot Q_i\right)={\sum }_{k=1}^d{w}_k\cdot p_{u,k}\cdot q_{i,k}$

Performance gains mainly come from grouping loss

The core of the article idea Mainly from this experimental diagram ：

Analyze real and false negative samples H and G The distribution of , It can be seen that in the process of training H Higher and higher are false negative samples , True negative samples G Higher and higher .