rk2900 / dlf Goto Github PK

Hello, I want to know how the two curves ‘Survival Curve of Different Model’ and ‘Figure 2: Learning curves’ mentioned in your paper are drawn?
Thank you.

Hello, firstly, I'd like to say that I enjoyed the paper and the model's idea of predicting bidding landscape.

I'm trying to implement paper's model in PyTorch and I get some results with my implementation, but when calculating c-index, best values I can get on test dataset are around 68%. The dataset I'm using is 2259 iPinYou. On the train set I get results closer to the paper's 87%, that is 83% (score obtained using some checkpoint, before doing hparam search, so I think there is some room for closing the gap).

Maybe do you remember whether the c-index was calculated on the train dataset?

Hi, Kan. I read your paper, admire and find it promising of this model without distribution assumption. Here is two questions I have to expect your answer:

1. For the Equation(6), should the subscript of b be l instead of l-1? I found in your presentation on KKD the subscript had been corrected.

2. For the lose part of the 2nd loss L_lose in Equation(12), should the range of l in SUM be [0, index of bidding price]? In the previous equations the paper said "where li is the interval index of the true market price zi ∈Vli given the feature vector xi", however, the market price is censored in losing bids and li could not be acquired. In which way, I expect the range to be different from previous though they used the same notation here.

Thank you for your amazing work. However, I get clone error below when trying to git clone this repository. Would you please add some other ways to download "DLF-data.7z", e.g. baidu netdisk.

Downloading DLF-data.7z (1.3 GB)
Error downloading object: DLF-data.7z (b5c53e5): Smudge error: Error downloading DLF-data.7z (b5c53e513ba892ac455c1a7d461aa280eac8fa9db182ec5cb5e7fad1f043e9cd): batch response: This repository is over its data quota. Account responsible for LFS bandwidth should purchase more data packs to restore access.

Hi Kan,

Kudos for the excellent research and development work!

I just want to know that in the sample 'yzbx' dataset, we can see the values of z >= b when we have lost the cases in the train as well as the test dataset. Can you please let us know how we are getting those values as it is not possible to get the market price when we have lost the auction? Have they randomly generated the market price data (z >= b and z < MAX_SEQ_LEN) ?

Hi Ren,

I have read your paper and encountered some questions.

1.What's the winning condition in one auction?

it confused me that your paper gives a winning condition as z<b,

while you give uncensored condition $z \le b$ in README.

and I also found an interesting statement about this dataset in a previous published paper

so could an advertiser win this auction with z equals b ?

I assume not for now.

2.What does h denotes in your paper ?

the defination in the paper is attached here

so let's take h as 'the just winning probability given z>=b_{l-1}'.
but if z is in V_{l} = (bl ,bl+1] , it means z>bl , then how could we win by b_{l}?

since z need to be at least no greater than b could the advertiser win, I will give my understanding of h

and if you would confirm that one cannot win with z==b , I believe the defination of V_{l} should be [bl,bl+1).

Only in this way can you get the equation h_{l} = Pr(z \in V_{l-1} | z \ge b_{l-1}).

3.What's the exact meaning of each h produced by RNN Cell?

after the second question, I'm trapped by a even more bigger one reading your code.

Assume that we still have h as the just winning probability given z>=b_{l-1}, which should be consistent both in your paper and code.

by the code below you prod all 'h's tegother , which is unreasonble.

survival_rate_last_one = tf.reduce_prod(x[0:bid_len])
anlp_rate_last_one = tf.reduce_prod(x[0:market_len + 1])
anlp_rate_last_two = tf.reduce_prod(x[0:market_len])

this enlights me that output of RNN maynot be the 'h' in your paper.
instead it looks like the losing probability at b_{l} given z>=b_{l-1} ,not winning , i.e. the survival rate at b_{l}

so I wonder whether you have given a reverse defination in the paper about h?
if not, what's the output of RNN stand for?

thanks

what is MD5 code in
The full dataset can be downloaded at this link and the corresponding MD5 code is 841698b0dd8718b1b4a4ff2e54bb72b4.