# Avito Context Ad Clicks — Data

来源：https://www.kaggle.com/competitions/avito-context-ad-clicks/data

采集时间：2026-10-18T13:52:26Z

This competition has 8 relational datasets. All these files were encoded in UTF-8 and stored into tab separated format (.tsv). There is also an sqlite database (database.sqlite) alternative with all data available. Relationships between the datasets are captured in the following schema:
![](assets/b9731e8f8ef2467c75a17fba.png)
trainSearchStream.tsv, testSearchStream.tsv
These two files are the main datasets related to your predictive models.
trainSearchStream is a random sample of previously selected users' searches on avito.ru during at least 16 consecutive days from April'25 until the target impression. Different types of ads on the site are shown in the picture below:
![](assets/2d7a812b96447c0e22b34fc4.png)
Regular ads are shifted down constantly as new ads come in. (Normally, a visitor's search results are sorted by the time an ad is submitted to Avito). Each line in the file describes one "impression" (an ad that is shown to a particular user based on a search). testSearchStream shares the same format, except the target variable field "
IsClick"
is omitted.
SearchID - identifier for a visitors's search event.
AdID - identifier of the ad (see also ad description in AdsInfo.tsv).
Position - position of the ad in search result page (1 - is first ad on a page starting from the top). Only ads on position 1, 2, 6, 7, and 8 are logged.
ObjectType - type of the ad shown to user. The options are: 1 - regular free ads added by users; 2 - highlighted regular (owners have to pay fixed price to highlight them and stick to the top for some period of time); 3 - contextual ads (owners have to pay per visitor's click).
HistCTR - some naive history-based estimation of click-through rate for contextual ads, calculated when the ad is showed. For non-contextual ads this field equals NULL.
IsClick - 1 if there was a click on this ad. Otherwise 0. For non-contextual ads this field equals NULL. The goal of this competition is to make a click prediction model for contextual ads.
VisitsStream.tsv, PhoneRequestsStream.tsv
These are samples of users' visits to non-contextual ad landing pages and the corresponding phone request (if one occurred). Each ad's landing page shows the hidden seller's phone number. To be able to contact the seller, the user needs to click the request phone button:
![](assets/a1c5f13aa31f1765cc7ca28b.png)
Consequently, a user's phone request event could be considered a proxy for a user's response to the advertisement. We believe that clicking the phone request indicates a high level of interest in the ad.
UserID - anonymized identifier of visitor's cookie (see also UserInfo.tsv).
IPID - anonymized identifier of visitor's IP.
AdID - identity of the ad's landing page visited by user (see also AdsInfo.tsv).
ViewDate - date and time of viewing the ad.
PhoneRequestDate - date and time of the phone request event.
SearchInfo.tsv
SearchID - identifier of search event.
SearchDate - date and time of the search event.
UserID - anonymized identifier of visitor's cookie.
IsUserLoggedOn - whether user was logged on with his/hers login (1) or not (0).
IPID - anonymized identifier of visitor's IP.
SearchQuery - raw query text if it was specified while search. NULL otherwise.
SearchLocationID - identifier of the location where search was made (see also Location.tsv).
SearchCategoryID - category filter of the search (see also Category.tsv).
SearchParams - dictionary-like structure with optional filters of the search if they were specified while search. NULL otherwise.
AdsInfo.tsv:
AdID - identity of an ad.
LocationID - ad's geo-targeting. (references Location.tsv for regular ads). NULL for contextual ads.
CategoryID - ad's category according to avito classification model. (references Category.tsv)
Params - parameters for the context ad in dictionary-like format. NULL for some ads (where no parameters are available).
Price - price for an ad
Title - raw title text.
Note that both Params from AdsInfo.tsv and SearchParams from SearchInfo.tsv shares same dictionary (keys and values). The Params are semi-structured to reflect the nature of the search and the product. For example, the Params for clothing might be gender, size, color, brand; while the Params for houses might be size, # of bedrooms, # of bathrooms, etc.
UserInfo.tsv:
UserID - anonymized identifier of visitor's cookie.
UserAgentID - anonymized identifier of user's browser.
UserAgentFamilyID - anonymized identifier of user's browser family (Chrome, Safari, etc).
UserAgentOSID - anonymized identifier of user's OS derived from browser family.
UserDeviceID - anonymized identifier of user device type and model (Samsung GT-I9500, iPhone, etc.)
Location.tsv
LocationID - identifier of the location.
Level - level of search/impression (3 = city, 2 = region, 1 = country)
RegionID - identifier of the search/impression region.
CityID - identifier of the search/impression city.
Category.tsv
CategoryID - identifier of the category.
Level - level of category for search/impression/ad (3 = subcategory, 2 = category, 1 = total)
ParentCategoryID - identifier of parent category.
SubcategoryID - identifier of subcategory that has parent category.
Data Splitting
The dataset contains a sample of users and their behavior. For each user, one target impression between time point A (May, 12) and time point B (May, 20) was selected randomly. The task of this competition is to provide the probability that a user will click on the selected target ad, given all the information generated by the user from the Start time point (April, 25) until the target impression.
![](assets/935198febaff9361e4faa232.png)
Note that some users from the testSearchStream may not have any historical information in VisitStream, PhoneRequestsStream and trainSearchStream as some of them may be new registered or had no activity within Start - Test event time interval.