https://ieeexplore.ieee.org/document/8594834
Yishuai Du | Yimin Zheng | Kuang-chih Lee | Shandian Zhe
- MIT Licence
- Matlab as the software to run POST code
- Download POST repository
- Open Matlab, run the "main.m" file in each subfolder of POST/code.
There are 5 datasets:
- MovieLens, a binary three-mode (user, movie, week) tensor of size 400×400×31
- Twitter Topic, a binary three-mode (user, expert, topic) tensor, of size 500×500×20
- ACC, a continuous tensor which records the three-mode interactions (user, action, resource), of size is 3K×150×30K
- DBLP, a binary three-mode (author, conference, keyword) tensor, bibliography relationships, of size 3K×150×30K
- Kaggle-ctr, a contest for click-through-rate (CTR) prediction in online advertising, sponsored by Avazu Inc，build a four-mode binary tensor (banner_pos,site_id,app_id,device_model), of size 7×2854×4114×6061
1. Evaluation on Dynamic Tensor Increments
Twitter Topic and MovieLens followed a similar procedure to MAST to conduct the experiments. Randomly chose {50%, 80%, 90%} entries of the entire tensor as missing for each dataset. The dimension of embedding vectors(Rank), from {3, 5, 8, 10}. In incremental case, the batch size is training data in this increment.
Example:
-
Twitter Topic: RUN "twitter/main.m", you can set Rank, say "R" and tune "v", the initial variance of the embeddings. For random missing data, for example, for 90% missing, we have missing 90% with five mat file, say "omegaT_90_i.mat"
-
MovieLens: same as Twitter Topic.
2. Evaluation on Streaming Tensor Entries in Arbitrary Orders
Acc(continuous) and DBLP(binary) are examined POST when tensor entries stream in arbitrary orders.
Example:
-
Acc(continuous): RUN "acc/main.m" , you can set Rank, say "R" and tune "v" ,the initial variance of the embeddings. In this case, you can set batch size, say "batch_size"
-
DBLP(binary): same as Acc
3. Uncertainty Investigation
Example:
- Kaggle-ctr: samse as 2