File size: 4,894 Bytes
19154aa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
46558a7
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
---
pretty_name: "Live or Lie — Live Streaming Room Risk Assessment (May/June 2025)"
language:
  - zh
tags:
  - live streaming risk assessment
  - fraud detection
  - weak supervision
  - multiple-instance-learning
  - behavior sequence
license: other
---

# Dataset Card: Live Streaming Room Risk Assessment (May/June 2025)

## Dataset Summary
This dataset contains **live-streaming room interaction logs** for **room-level risk assessment** under **weak supervision**. Each example corresponds to a single live-streaming room and is labeled as **risky (> 0)** or **normal (= 0)**.

The task is designed for early detection: each room’s action sequence is **truncated to the first 30 minutes**, and can be structured into **user–timeslot capsules** for models such as AC-MIL.


## Languages
  - Predominantly **Chinese (zh)**: user behaviors are presented in Chinese, e.g., "主播口播:...", these action descriptions are then encoded as action vectors via a **Chinese-bert**.


## Data Structure
Each room has a label and a sequence of **actions**:

- `room_id` (`string`)
- `label` (`int32`, {0,1,2,3}))
- `patch_list` (`list` of tuples):
  - `u_idx` (`string`): user identifier within a room
  - `t` (`int32`): time index along the room timeline
  - `l` (`int32`): capsule index
  - `action_id` (`int32`): action type ID 
  - `action_vec` (`list<float16>` or `null`): action features encoded from masked action descriptions
  - `timestamp` (`string`): action timestamp
  - `action_desc` (`string`): textual action descriptions
  - `user_id` (`string`): user indentifier across rooms

## Action Space
The paper’s setup includes both viewer interactions (e.g., room entry, comments, likes, gifts, shares, etc.) and streamer activities (e.g., start stream, speech transcripts via voice-to-text, OCR-based visual content monitoring). Text-like fields are discretized as part of platform inspection/sampling.

## Data Splits
The paper uses two datasets (“May” and “June”), each with train/val/test splits.
| Split | #Rooms | Avg. actions | Avg. users | Avg. time (min) |
|------:|------:|-------------:|-----------:|----------------:|
| May train | 176,354 | 709 | 35 | 30.0 |
| May val   | 23,859  | 704 | 36 | 29.6 |
| May test  | 22,804  | 740 | 37 | 29.7 |
| June train| 80,472  | 700 | 36 | 30.0 |
| June val  | 10,934  | 767 | 40 | 29.1 |
| June test | 11,116  | 725 | 37 | 29.1 |

## Quickstart
Below we provide a simple example showing how to load the dataset.

We use LMDB to store and organize the data. Please install the Python package first:
```
pip3 install lmdb
```

Here is a minimal demo for reading an LMDB record:
```python
import lmdb
import pickle

room_id = 0  # the room you want to read

env = lmdb.open(
    lmdb_path,
    readonly=True,
    lock=False,
    map_size=240 * 1024 * 1024 * 1024,
    readahead=False,
)

with env.begin() as txn:
    value = txn.get(str(room_id).encode())
    if value is not None:
        data = pickle.loads(value)
        patch_list = data["patch_list"]  # list of tuples: (u_idx, t, l, action_id, action_vec, timestamp, action_desc, global_user_idx)
        room_label = data["label"]

# close lmdb after reading
env.close()
```


## Claim
To ensure the security and privacy of TikTok users, all data collected from live rooms has been anonymized and masked, preventing any content from being linked to a specific individual. In addition, action vectors are re-encoded from the masked action descriptions. As a result, some fine-grained behavioral signals are inevitably lost, which leads to a performance drop for AC-MIL. The corresponding results are shown below.

| Split | PR_AUC | ROC_AUC | R@P=0.9 | P@R=0.9 | R@FPR=0.1 | FPR@R=0.9 |
|------:|------:|-------------:|-----------:|----------------:|----------------:|----------------:|
| May   | 0.6518  | 0.9034 | 0.2281 | 0.2189 |   0.7527  | 0.3215 |
| June   | 0.6120  | 0.8856 | 0.1685 | 0.1863 |   0.7111  | 0.3935 |

---

## Considerations for Using the Data

Intended Use \
•	Research on weakly-supervised risk detection / MIL in live streaming \
•	Early-warning room-level moderation signals \
•	Interpretability over localized behavior segments (capsule-level evidence) 

Out-of-scope / Misuse \ 
•	Do not use this dataset to identify, profile, or target individuals. \
•	Do not treat predictions as definitive enforcement decisions without human review. 

Bias, Limitations, and Recommendations \
•	Sampling bias: negatives are downsampled (1:10); reported metrics and thresholds should account for this. \
•	Domain specificity: behavior patterns are platform- and policy-specific; transfer to other platforms may be limited. \
•	Weak supervision: only room-level labels are provided; interpretability at capsule level is model-derived.


## License
This dataset is licensed under CC BY 4.0:
https://creativecommons.org/licenses/by/4.0/