Upload README.md
Browse files
README.md
ADDED
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Avazu_x4
|
2 |
+
|
3 |
+
+ **Dataset description:**
|
4 |
+
|
5 |
+
This dataset contains about 10 days of labeled click-through data on mobile advertisements. It has 22 feature fields including user features and advertisement attributes. Following the same setting with the [AutoInt](https://arxiv.org/abs/1810.11921) work, we split the data randomly into 8:1:1 as the training set, validation set, and test set, respectively.
|
6 |
+
|
7 |
+
The dataset statistics are summarized as follows:
|
8 |
+
|
9 |
+
| Dataset | Total | #Train | #Validation | #Test |
|
10 |
+
| :--------: | :-----: |:-----: | :----------: | :----: |
|
11 |
+
| Avazu_x4 | 40,428,967 | 32,343,172 | 4,042,897 | 4,042,898 |
|
12 |
+
|
13 |
+
|
14 |
+
- Avazu_x4_001
|
15 |
+
|
16 |
+
In this setting, we preprocess the data split by removing the ``id`` field that is useless for CTR prediction. In addition, we transform the timestamp field into three fields: hour, weekday, and is_weekend. For all categorical fields, we filter infrequent features by setting the threshold min_category_count=2 (performs well) and replace them with a default ``<OOV>`` token. Note that we do not follow the exact preprocessing steps in AutoInt, because the authors neither remove the useless ``id`` field nor specially preprocess the timestamp field. We fix **embedding_dim=16** following the existing [AutoInt work](https://arxiv.org/abs/1810.11921).
|
17 |
+
|
18 |
+
- Avazu_x4_002
|
19 |
+
|
20 |
+
In this setting, we preprocess the data split by removing the ``id`` field that is useless for CTR prediction. In addition, we transform the timestamp field into three fields: hour, weekday, and is_weekend. For all categorical fields, we filter infrequent features by setting the threshold min_category_count=1 and replace them with a default ``<OOV>`` token. Note that we found that min_category_count=1 performs the best, which is surprising. We fix **embedding_dim=40** following the existing [FGCNN work](https://arxiv.org/abs/1904.04447).
|
21 |
+
|
22 |
+
|
23 |
+
+ **Source:** https://www.kaggle.com/c/avazu-ctr-prediction/data
|
24 |
+
+ **Download:** https://huggingface.co/datasets/reczoo/Avazu_x4/tree/main
|
25 |
+
+ **RecZoo Datasets:** https://github.com/reczoo/Datasets
|
26 |
+
|
27 |
+
+ **Used by papers:**
|
28 |
+
- Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, Jian Tang. [AutoInt: Automatic Feature Interaction Learning via Self-Attentive Neural Networks](https://arxiv.org/abs/1810.11921). In CIKM 2019.
|
29 |
+
|
30 |
+
+ **Check the md5sum for data integrity:**
|
31 |
+
```bash
|
32 |
+
$ md5sum train.csv valid.csv test.csv
|
33 |
+
de3a27264cdabf66adf09df82328ccaa train.csv
|
34 |
+
33232931d84d6452d3f956e936cab2c9 valid.csv
|
35 |
+
3ebb774a9ca74d05919b84a3d402986d test.csv
|
36 |
+
```
|