Corodic commited on
Commit
367abf4
·
verified ·
1 Parent(s): d1eb56d

add read_me and display histograms

Browse files
Files changed (2) hide show
  1. README.md +186 -3
  2. event_per_user_correction.csv +21 -0
README.md CHANGED
@@ -1,3 +1,186 @@
1
- ---
2
- license: cc-by-sa-4.0
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-sa-4.0
3
+ size_categories:
4
+ - 10M<n<100M
5
+ ---
6
+ # Dataset Documentation
7
+
8
+ ## Private Bidding Optimisation {#private-conversion-optimisation}
9
+
10
+ The advertising industry lacks a common benchmark to assess the privacy
11
+ / utility trade-off in private advertising systems. To fill this gap, we
12
+ are open-sourcing CriteoPrivateAd, the largest real-world anonymised
13
+ bidding dataset, in terms of number of features. This dataset enables
14
+ engineers and researchers to:
15
+
16
+ - assess the impact of removing cross-domain user signals,
17
+ highlighting the effects of third-party cookie deprecation;
18
+ - design and test private bidding optimisation approaches using
19
+ contextual signals and user features;
20
+ - evaluate the relevancy of answers provided by aggregation APIs for
21
+ bidding model learning.
22
+
23
+ ## Summary
24
+
25
+ This dataset is released by Criteo to foster research and industrial
26
+ innovation on privacy-preserving machine learning applied to a major
27
+ advertising use-case, namely bid optimisation under user signal loss /
28
+ obfuscation.
29
+
30
+ This use-case is inspired by challenges both browser vendors and AdTech
31
+ companies are facing due to third-party cookie deprecation, such as
32
+ ensuring a viable cookie-less advertising business via a pragmatic
33
+ performance / privacy trade-off. In particular, we are expecting to see
34
+ improvements of Google Chrome Privacy Sandbox and Microsoft Ad Selection
35
+ APIs via offline benchmarks based on this dataset.
36
+
37
+ The dataset contains an anonymised log aiming to mimic production
38
+ performance of AdTech bidding engines, so that offline results based on
39
+ this dataset could be taken as ground truth to improve online
40
+ advertising performance under privacy constraints. Features are grouped
41
+ into several groups depending on their nature, envisioned privacy
42
+ constraints and availability at inference time.
43
+
44
+ Based on this dataset, the intended objective is to implement privacy
45
+ constraints (e.g. by aggregating labels or by adding differential
46
+ privacy to features and/or labels) and then learn click and conversion
47
+ (e.g. sales) prediction models.
48
+
49
+ The associated paper is available [here](https://arxiv.org/abs/2502.12103)
50
+
51
+ As a leading AdTech company that drives commerce outcomes for media
52
+ owners and marketers, Criteo is committed to evaluating proposals that
53
+ might affect the way we will perform attribution, reporting and campaign
54
+ optimisation in the future. Criteo has already participated in testing
55
+ and providing feedback on browser proposals such as the Privacy Sandbox
56
+ one; see all our [Medium articles](https://techblog.criteo.com) Back in 2021, we also
57
+ organised a public challenge aiming to assess bidding performance when
58
+ learning on aggregated data: our learnings are available [here](https://arxiv.org/abs/2201.13123).
59
+
60
+
61
+
62
+ ## Dataset Description
63
+
64
+
65
+ This dataset represents a 100M anonymised sample of 30 days of Criteo
66
+ live data retrieved from third-party cookie traffic on Chrome. Each line corresponds to one impression (a banner)
67
+ that was displayed to a user. For each impression, we are providing:
68
+
69
+ - campaign x publisher x (user x day) granularity with respective ids, to match Chrome Privacy Sandbox scenarios and both
70
+ display and user-level privacy.
71
+
72
+ - 4 labels (click, click leading to a landing on an advertiser
73
+ website, click leading to a visit on an advertiser website -
74
+ i.e. landing followed by one advertiser event, number of sales
75
+ attributed to the clicked display).
76
+
77
+ - more than 100 features grouped in 5 buckets with respect to their
78
+ logging and inference constraints in Protected Audience API from
79
+ Chrome Privacy Sandbox (note that these buckets are generic enough
80
+ to cover other private advertising frameworks as we are mainly
81
+ providing a split between ad campaign features, single-domain &
82
+ cross-domain user features, and contextual features) :
83
+
84
+ - Features available in the key-value server with 12-bit logging
85
+ constraint (i.e. derived from current version of modelingSignals
86
+ and standing for single-domain user features).
87
+ - Features available in the key-value server with no logging
88
+ constraint (i.e. derived from Interest Group name / renderURL).
89
+ - Features available in browser with 12-bit constraint
90
+ (i.e. cross-domain features available in generateBid).
91
+ - Features from contextual call with no logging constraint
92
+ (i.e. contextual features).
93
+ - Features not available (i.e. cross-device and cross-domain
94
+ ones).
95
+
96
+ - `day_int` enabling (1) splitting the log into training, validation
97
+ and testing sets; (2) performing relevant model seeding.
98
+
99
+ - Information about conversion delay to simulate the way Privacy Sandbox APIs are working.
100
+
101
+ - `time_between_request_timestamp_and_post_display_event` (column name
102
+ in clear): time delta (in minutes) between the request timestamp and the
103
+ click or sale event. All displays are considered starting the day of
104
+ the event at 00:00 to avoid providing complete timelines.
105
+
106
+ - We include a display order from 1 to K for display on the same day
107
+ for the same user.
108
+
109
+ CriteoPrivateAd is split into 30 parquets (one per day from 1 to 30) in day_int={i} directory.
110
+ The displays-per-user histograms can be deduced from event_per_user_contribution.csv,< it is useful to build importance sampling ratios and user-level DP.
111
+ Please, see the companion paper for more details.
112
+
113
+
114
+ A precise description of the dataset and each column is available in [the
115
+ companion paper](https://arxiv.org/abs/2502.12103)
116
+
117
+
118
+
119
+ ## Metrics
120
+
121
+ The metrics best suited to the click and conversion estimation problems
122
+ are:
123
+
124
+ - the log-likelihood (LLH), and preferably a rescaled version named LLH-CompVN defined
125
+ as the relative log-likelihood uplift compared to the naive model
126
+ always predicting the average label in the training dataset;
127
+ - calibration, defined as the ratio between the sum of the predictions
128
+ and the sum of the validation labels. It must be close to 1 for a
129
+ bidding application;
130
+
131
+ We would like to point out that conventional classification measures
132
+ such as area under the curve (AUC) are less relevant for comparing
133
+ auction models.
134
+
135
+ The click-through rate is higher than the one encountered in real-world
136
+ advertising systems on the open internet. To design realistic bidding
137
+ applications, one must use a weighted loss for validation. We defer the
138
+ interested readers to the [associated companion paper](https://arxiv.org/abs/2502.12103) for more details
139
+
140
+ ## Baselines
141
+
142
+ The Training period has been fixed to 1->25 and Validation period to 26->30. The chosen loss is the LLH-CompVN with weighting as defined above. The Sales | Display is a product of the Landed Click | Display and the Sales | Landed Click.
143
+
144
+ | Task/CTR | 0.1% | 0.5% | 1% |
145
+ |-------------------------|-------|-------|-------|
146
+ | Landed Click \| Display | 0.170 | 0.186 | 0.234 |
147
+ | Sales \| Landed Click | 0.218 | 0.218 | 0.218 |
148
+ | Sales \| Display | 0.171 | 0.187 | 0.237 |
149
+
150
+ Note that our baseline results might be difficult to achieve because of the anonymisation of the dataset.
151
+
152
+ ## License
153
+
154
+ The data is released under the license. You are free to
155
+ Share and Adapt this data provided that you respect the Attribution and
156
+ ShareAlike conditions. Please read carefully the full license before
157
+ using.
158
+
159
+ ## Citation
160
+
161
+ If you use the dataset in your research please cite it using the
162
+ following Bibtex excerpt:
163
+
164
+ @misc{sebbar2025criteoprivateadrealworldbiddingdataset,
165
+ title={CriteoPrivateAd: A Real-World Bidding Dataset to Design Private Advertising Systems},
166
+ author={Mehdi Sebbar and Corentin Odic and Mathieu Léchine and Aloïs Bissuel and Nicolas Chrysanthos and Anthony D'Amato and Alexandre Gilotte and Fabian Höring and Sarah Nogueira and Maxime Vono},
167
+ year={2025},
168
+ eprint={2502.12103},
169
+ archivePrefix={arXiv},
170
+ primaryClass={cs.CR},
171
+ url={https://arxiv.org/abs/2502.12103},
172
+ }
173
+
174
+
175
+ ## Acknowledgment
176
+
177
+ We would like to thank:
178
+
179
+ - Google Chrome Privacy Sandbox team, especially Charlie Harrisson,
180
+ for feedbacks on the usefulness of this dataset.
181
+ - W3C PATCG group, notably for their public data requests to foster
182
+ work on the future of attribution and reporting.
183
+ - Criteo stakeholders who took part of this dataset release: Anthony
184
+ D'Amato, Mathieu Léchine, Mehdi Sebbar, Corentin Odic, Maxime Vono,
185
+ Camille Jandot, Fatma Moalla, Nicolas Chrysanthos, Romain Lerallut,
186
+ Alexandre Gilotte, Aloïs Bissuel, Lionel Basdevant, Henry Jantet.
event_per_user_correction.csv ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ cnt_displays;nb_users_CriteoPrivateAd;nb_users_original;sampling_weight
2
+ 1;81.59017519438582;31.461234309327686;0.00018312306581380486
3
+ 2;12.627252670118407;16.94631036619179;0.000637340393948341
4
+ 3;3.3731253041235503;11.037908922733743;0.001554030330413786
5
+ 4;1.2091565351024356;8.042709640813122;0.003158821037283753
6
+ 5;0.5263210799756821;6.192651422222794;0.0055876736015271124
7
+ 6;0.26267937088171844;4.817314665760953;0.008709319179265464
8
+ 7;0.1452702128816801;3.7343443553809625;0.012207955829026505
9
+ 8;0.0859987504408444;3.0199377298955095;0.016676734172371088
10
+ 9;0.05417405693062993;2.4569364892295944;0.021538116890158193
11
+ 10;0.0357337866530729;2.048885560260378;0.027229760448449734
12
+ 11;0.02422468916735485;1.7184218500846973;0.033688116639026405
13
+ 12;0.01686884080816877;1.512096206268133;0.04256957781762428
14
+ 13;0.011723383322452808;1.2980893644479816;0.05258438334231591
15
+ 14;0.00835714763265579;1.1345208344752733;0.0644702806246388
16
+ 15;0.00623896460408367;0.9944869496366027;0.07569927598273567
17
+ 16;0.004649353306123967;0.8766500878134798;0.08954449470814815
18
+ 17;0.0034337681959194883;0.7780990377271169;0.1076140716491054
19
+ 18;0.00268961192653149;0.712301325826477;0.1257706455148872
20
+ 19;0.00211818302857212;0.6379885420653203;0.14303902927982048
21
+ 20;0.0016415593795923728;0.5791123398383868;0.1675372494934399