## moving pains

The school of computing, NUS, is relocating - and today was the first day for me at my new office ("ooooooo..."), except that it did not go down too well for me.

Let me see,

(1) supposed to pick up keys in the morning, but the lady in charge was in another part of NUS. Waited from 8-10:30am.

(2) made to move from one office to the adjacent one for admin reasons - had to go back to lady in charge to get new keys.

(3) had to set up computer, no internet connections!! even wifi was not working.

somewhat ironically, instead of doing all these internetz stuff at the "state-of-the-art" school, I am writing this in United Square, a shopping mall. Free Singapore-wide Wifi FTW.

## It's True!

Q: What is an American University?

A: This is a strange place where Russian professors teach Chinese students in English.

- post in Slashdot

## That just bites

One of the most fascinating things I have learnt lately is the Trivers-Willard hypothesis that essentially predicts that well-off parents have more sons, while poor parents of low status have more daughters, as children inherit wealth and social status of their parents. As such, sons tend to benefit more from rich parents of high social status than a daughter would.

This hypothesis has been documented around the globe. American presidents, vice presidents, and cabinet secretaries have more sons than daughters. Poor Mukogodo herders in East Africa have more daughters than sons. Church parish records from the 17th and 18th centuries show that wealthy landowners in Leezen, Germany, had more sons than daughters, while farm laborers and tradesmen without property had more daughters than sons. In a survey of respondents from 46 nations, wealthy individuals are more likely to indicate a preference for sons if they could only have one child, whereas less wealthy individuals are more likely to indicate a preference for daughters.

The generalized Trivers-Willard hypothesis goes beyond a family's wealth and status: If parents have any traits that they can pass on to their children and that are better for sons than for daughters, then they will have more boys. Conversely, if parents have any traits that they can pass on to their children and that are better for daughters, they will have more girls.

Physical attractiveness, while a universally positive quality, contributes even more to women's reproductive success than to men's. The generalized hypothesis would therefore predict that physically attractive parents should have more daughters than sons. Once again, this is the case. Americans who are rated "very attractive" have a 56 percent chance of having a daughter for their first child, compared with 48 percent for everyone else.

Couple this hypothesis with the trend of favouristism of males in Asian culture, and throw in government control of how many children every family can have in, say China, and you can see where this is going - the poor, who want sons, have a limit of how children they can have. Ironically, they are also predicted to have higher chances of a daughter which leads to more infanticide and abandonment of female babies. _sigh_

## I see

Islandwide Delivery Service

You Call

We Delivery

- seen on back of delivery van

## Partially-Observed "Singular Value Decomposition"

Singular Value Decomposition is a fanastic linear algebra operation which factorizes a matrix - that is, for a matrix M, it returns U, S, V such that M = U * S * V'. This can be viewed as a basis transformation from row space to column space. The main problem is that standard SVD requires a completely-observed matrix M, that is, all values of M are known.

SVD has been shown to work surprisingly well for collaborative filtering applications such as the netflix prize. However, in such problems, the matrix is usually *partially-observed*, or in other words, some entries of M are not known. The "simple" workaround has been to perform gradient descent in U and V to minimize the objective function:

f = \sum_{i=1}^N \sum_{j \in N(i)} (u_i * v_j' - M_{ij})^2 + \lambda \sum_{ij} ( ||u_i||^2 + ||v_j||^2 )

where N(i) are the observed entries in row i. This breaks down to the updates:

\delta_{ij} = u_i * v_j' - M_{ij}

u_{ik} = u_{ik} + \alpha( \delta_{ij} v_{jk} - \lambda u_{ik} )

v_{jk} = v_{jk} + \alpha( \delta_{ij} u_{ik} - \lambda v_{jk} )

Since we only work with known values of M, this simply avoids the requirement of M being completely-observed.

I am hoping to use this simple technique to perform data mining in other similar problems where "collaborative filtering" is assumed but tedious to compute - e.g., medical data where you have tons of clinical measurements and information about patients, together with a set of diagnosis. So given a set of patient info and clinical measurements, can you predict the diagnosis? Casting this information into a partially-observed matrix M is fairly simple, and I am having moderate success. Nevertheless, I would prefer to have Bayesian belief networks to incorporate some prior knowledge into a structured model.

## Moving Away

Change is inevitable, growth is intentional.

-Glenda Cloud

In more ways than I care to say right now, change is inevitable - but moving from paid hosting to free hosting while maintaining my domain is a nice change in my opinion.