TapNet: Multivariate Time Series Classification with Attentional Prototypical Network

TapNet： Semi-supervised learning   20 AAAI The paper

Innovation points 1：

RDP: random dimension permutation 

Every time when training, not all variable All training , It's random groups , Train again .

Innovation points 2：

Semi supervised prototype learning ：

Time series through the network encoder Is a low dimensional representation vector . We get the centroid of each category . We use categories k For example , For categories k Centroid of . The center of mass passes

For each instance i Do attention calculations （ That is, the center of mass passes attention Mechanism , For different instance Have different attention ）.attention Find  

From the above , We can get the prototype of each class , After getting the prototype , We can know x The distribution of ,x The distribution of is shown in the formula

among ,ck Is the center of mass ,x after embedding Result , Calculation x Distance from the center of mass . But in the author's source code , The calculation is not as shown in the figure , Because loss It's calculation log(p_θ) Of , So after simplification , The author's source code has a -1/2 （-1/2 It is an absolute value, so it comes into being ）

Semi supervision ：

In the calculation When , Add unlabeled data , Because there is no label , It is also possible to calculate p_θ

That's formula 4.（x To k Distance of /x Distance to all other centroids ）