Monday, 22 December 2014
Sunday, 7 December 2014
Quora haqathon 2014
Quora haqathon today, from 11am to 7pm - Pacific standard time! Features 9 problems mixed with tradition algorithm tasks, machine learning and system programming tasks. Link to site.
Ontology
Linearize the tree - each query reduces to "in question q[x...y], how many of them start with prefix p?". Offline query + Partial Sum+ Trie. Linear time.
Wombats
Maximum closure.
Labeler
Use training set to calculate \(\text{Pr}[q_i \in t_k | w_j \in q_i ] \) for all question \(q_i\), topic \(t_k\) and word \(w_j\). Improve using bi-gram.
Duplicate
Use \( \text{Pr}[w \in \text{question_text}_i \text{ and } w \notin \text{question_text}_j ]\) as classifying criteria - 60% accuracy. Consider also \( \frac{\text{view_count}_i }{ \text{view_count}_j } \) improved it to 70%.
Ontology
Linearize the tree - each query reduces to "in question q[x...y], how many of them start with prefix p?". Offline query + Partial Sum
Wombats
Maximum closure.
Labeler
Use training set to calculate \(\text{Pr}[q_i \in t_k | w_j \in q_i ] \) for all question \(q_i\), topic \(t_k\) and word \(w_j\). Improve using bi-gram.
Duplicate
Use \( \text{Pr}[w \in \text{question_text}_i \text{ and } w \notin \text{question_text}_j ]\) as classifying criteria - 60% accuracy. Consider also \( \frac{\text{view_count}_i }{ \text{view_count}_j } \) improved it to 70%.
Wednesday, 19 November 2014
PhD 4: Peaky Blinders
Wednesday, 12 November 2014
PhD 3: Old news is so exciting
近來看到一個「早該知道」但昨天才知道的 fact: if \(X \subset L^d_2\) then \(X\) isometric embeds to \(L_1\)
如果只說 idea 很簡單 -- pick a random vector \(r\) in unit sphere \(S_{d-1}\), consider map \(f: x \mapsto \langle r, x\rangle\), then
\( \mathbb{E}[|f(x) - f(y)|] = \mathbb{E}[|\langle r,x \rangle - \langle r,y\rangle |]
= \mathbb{E}[ |\langle r, x-y \rangle|] \propto ||x-y||_2
\)
所以只要取"足夠"(無限)多的 sample 然後做 scaling, 就可以準確地 preserve 原本的 distance
加上因為 \(L_1^* \hookrightarrow L_1^{\binom{n}{2}}\), 所以是存在 finite dimension 的 \(L_1\) embedding
如果只說 idea 很簡單 -- pick a random vector \(r\) in unit sphere \(S_{d-1}\), consider map \(f: x \mapsto \langle r, x\rangle\), then
= \mathbb{E}[ |\langle r, x-y \rangle|] \propto ||x-y||_2
\)
所以只要取"足夠"(無限)多的 sample 然後做 scaling, 就可以準確地 preserve 原本的 distance
加上因為 \(L_1^* \hookrightarrow L_1^{\binom{n}{2}}\), 所以是存在 finite dimension 的 \(L_1\) embedding
Monday, 10 November 2014
PhD 2: Netflix
Netflix, 線上睇片網站, 月費 $9, unlimited streaming (嚴格來說 bounded by 2 screens * 30 days), 睇美劇一流, 而且似乎唔少歐洲電影
花了三個星期把 Breaking Bad 追完, 才剛發現原來在 imdb top TV 居首
(一開始不完全是因為出名才看, 而是看完簡介覺得劇情半重口味很合胃口)
1. 發現原來只用英文字幕也足夠, Netflix 練英文一流
2. 很成功地避免了所有劇透
睇完又將 "好劇" 的 threshold 提升, HBO 快點出 internet streaming 吧...
花了三個星期把
(一開始不完全是因為出名才看, 而是看完簡介覺得劇情半重口味很合胃口)
1. 發現原來只用英文字幕也足夠, Netflix 練英文一流
2. 很成功地避免了所有劇透
![]() |
| image from wikipedia |
Subscribe to:
Posts (Atom)


