Dear Zuo， I hope everything is fine with you, after saw your paper, I am really interested in you work, and learned a lot from it. However, I have a little question, how do you get the so small RMSE? While I get the RMSE on Financal data, is like Minimum RMSE: xx.xxxxx while in the paper, it is only 0.93. Are there some hyper-parameters about loss function need to adjustment? I set loss = 0 * event_loss + 0* pred_loss + se / scale_time_loss But it still doesn' t work, could you offer me some help? Thank you.

Sincerely yours, Luning Zhang

In the Paper the Loss function contains 3 parts:

The first part is the log-likelihood, the second is the event cross-entropy loss, and the third is the time RMSE loss.

I don't know if it is appropriate to mix these all together. As I expect, if the log-likelihood is applied, then to predict the next event's time, the only way is to calculate the expectation of PDF:

Can anyone help me make further clarification?

Hi, I have the following issue regarding calculating LL in Utils.py in line 49 inside the function "compute_integral_unbiased":

temp_hid = torch.sum(temp_hid * type_mask[:, 1:, :], dim=2, keepdim=True)

you have only considered the occurred events for calculating integral while according to the formula we should compute the integral of each event type. I would expect the output dimension of this function to be [Batch_size, Length, Num_types]. then we should sum over all num_types instead of reducing it to only occurred events.

I believe that this underestimation of this integral has led to your high overall LL compared to other studies.

looking forward to your clarifications

Hi, I tried to reproduce the transformer hawkes process on StackOverflow fold1. However, the results of accuracy and RMSE is as below.

![image](https://user-images.githubusercontent.com/56212725/173004512-ba357b4d-244a-4f73-9ca1-9d9535f3f1df.png

)

I think I have something missing. Compared to the relased code of Self-Attentive Hawkes process, I think it is not because of scaling factor. What does make the difference between the paper and this repository?

Suppose the transformer hidden state at event i is h_i, should the likelihood of this event be computed using h_i or h_{i-1}?

Using h_{i-1} makes more sense to me because this will encourage model to assign high intensity to the true next event, therefore learn to forecast.

But the implementation and the paper seem to be using h_i. The problem is that, since the transformer is given the true event i as part of the input, it can simply learn to output infinitely high intensity for the correct event type in order to maximize the likelihood. Still, the learned model will have no predictive power.

I feel I must have missed something. Any clarification is appreciated. Thanks.

Could you please provide additional details on how to obtain the 911-Calls and Earthquake datasets used in your paper? The CSV found at the provided webpage has 663,522 calls, all of which are in the EMS, fire, or traffic categories. For the 75 most frequent ZIP codes in this dataset, there are 582,045 total calls, which is considerably more than the 290,293 listed in Table 1 (see below code).

import pandas as pd

df = pd.read_csv("911.csv")
print(len(df))  # 663522
cats = ["EMS: ", "Fire: ", "Traffic: "]
in_cats = 0
for title in df["title"]:
    for cat in cats:
        if cat in title:
            in_cats += 1
            break

print(in_cats)  # 663522
zip_calls = (
    df.groupby("zip")
    .size()
    .reset_index(name="n_calls")
    .sort_values("n_calls", ascending=False)
)
print(zip_calls["n_calls"][:75].sum())  # 582045

The paper also states that:

An undirected edge exists between two vertices if their zipcodes are within 10 of each other.

Does this mean two vertices were considered neighbors if abs(ZIP_{1} - ZIP_{2}) <= 10?

For the Earthquake dataset, the provided website is in Chinese and seems to host a number of datasets. Could you provide precise instructions on where to find the specific earthquake dataset used in your paper?