吐槽

雖然說過「歐洲的城堡都一樣」，不過美國基本上沒有真正的城堡這點還是非常遺憾

還有很莫明奇妙的是，在 Espresso 旁的雪櫃有 85% 是 reduced fat milk，但同時又要落 syrup (罪大惡極)，耐人尋味

不過，至少還沒看到 canned coffee 這種不能稱為咖啡的物體.. 其實喝過「真正」的咖啡以後就完全鄙視像雀Ｘ的罐裝咖啡

但多得有 Amazon ，我早已在 Expresso 的路上奔馳

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%.