1.
Building a DL model that is lightweight, fast to train and estimate and optionally allow injection of knowledge such as query workload is non-trivial.
2.
Recently, there has been intense interest in applying DL for selectivity estimation.
3.
被广泛探索的任务
Time series forecasting has been a widely explored task that is of great importance in many applications.
4.
DL估计迫切需要的东西
desiderata for DL estimation
5.
推断它来获得一个整体的估计的结果
extrapolating it to get an overall estimate
6.
穷尽的枚举
exhaustive enumeration
7.
由于XX坏的影响
due to curse of dimensionality
8.
可能根据之前获得的样本而改变
In other words, each sample could have a different weight and the probability with which a new point is selected could vary based on previously obtained samples.
9.
However, naively implementing this idea results in biased and incorrect results.
10.
Let F be the probability density function based on query q such that if we sample points proportional to F, we might get accurate estimates.
11.
The solution is to proceed in stages and use the information collected from samples of previous stages to improve the next stage.
12.
高效实例化
an efficient instantiation of F
13.
a bad permutation
14.
We do away with this expensive step by choosing several random orderings of attributes.
As we shall show experimentally, this approach works exceedingly well in practice.
works exceedingly well
shall show experimentally
do away with this expensive step
15.
the powerful learning capacity of neural networks can readily learn even a challenging decomposition by increasing the depth of the MADE model
16.
Of course, different orderings result in different models with their corresponding estimate and associated accuracy.
17.
While a random ordering often provides good results, it is desirable to guard against the worst case scenario of a bad permutation.
18.
the autoregressive approach outlined above
19.
this falls under the umbrella of supervised learning methodologies using regression
20.
The impediment is the complex relationship between queries and their selectivities where simplifying assumption such as attribute value independence do not hold.