Trensking
commented
3 years ago I'm very grateful to the author Ruijia Li for his help. Without his help, I may spend more time thinking about these problems. Now let me explain some problems in the use of this algorithm.
- The core of LGD algorithm is based on the calculation results of the KNN algorithm, so it can't get desired results when the data is less discriminative calculated from the Euclidean distance in the high-dimensional space.
- Data input parameter Y is the real label of X and it is usually obtained by manual calibration. The metric ACC indicates the degree of matching between the clustering results and the actual results. The closer the value is to 1, the better the clustering result.
- The upper limit of the value of the parameter c is determined by the data distribution and the calculation result of KNN.
- The figure "Core points and border points" is the result of high-dimensional data after PCA is reduced to 2 dimensions. The same goes for other similar figures.
- The best value range of parameter k is an integer of [2,20] and tau is [0.4,0.6]. The larger the value of tau, the fewer core points.
- If you have another problems, please contact the author by email first.
I want to use this algorithm to perform cluster analysis on sift feature descriptors (128-dimensional vectors). But I don’t know how to choose the parameters k, tau and c, so the clustering result is very bad and it produces too much border points.
In addition, the value of parameter c seems to have an upper limit, which always reports an error during operation.
If you read my question, can you leave a contact information?QQ, WeChat or telegram are all ok and I can pay some remuneration. Thanks！