Dongxu Zhang 2015-12-14

Last Week

---

• Run experiments with KNN on cosine distance.

 Found general improvements over different topic models. Proposed model is better on two dataset out of three.

• Run experiments with fix window(=5).

 Partial results shows that the accuracy is lower than document-wise window.
 Problem is that only 10 iteration is implemented, which may not converge to the best performance.

• Train multi-gram model in a joint way.

 No good results.

This Week

---

• Finish the fix window experiments on larger iteration(=50)
• Tried multi-gram training separately.