Loading repository data…
Loading repository data…
varepsilon / repository
ClickModels is a small set of Python scripts for the user click models initially developed at Yandex. A Click Model is a probabilistic graphical model used to predict search engine click data from past observations. This project is aimed to deal with click models used in Information Retrieval (see next README.md) and intended to be easy-to-read and easy-to-modify. If it's not, please let me know how to improve it :)
ClickModels is a small set of Python scripts for the user click models initially developed at Yandex. A Click Model is a probabilistic graphical model used to predict search engine click data from past observations. This project is aimed to deal with click models used in Information Retrieval (see next section) and intended to be easy-to-read and easy-to-modify. If it's not, please let me know how to improve it :)
If you are using this code for your research work, consider citing one of our papers when appropriate (see References section below).
If you are looking for a general-purpose framework to work with probabilistic graphical models you might want to examine Infer.NET. It should also work with IronPython.
cp clickmodels/config_sample.py config.pyvim config.pypython bin/run_inference.py < data/click_log_sample.tsv 2>inference.logMore details about the config and input data formats below.
If you wish, you can install the ClickModels core (parameter inference and click simulation) to a system-wide location:
sudo python setup.py install
Uninstall:
sudo pip uninstall clickmodels
Now, thanks to agrotov, the models can also be run in a click generation mode and predict relevance (DBN only). Check out ClickModel.get_model_relevances() and ClickModels.generate_clicks() methods.
N.B.: Use this code with care as it is not fully tested yet.
A small example can be found under data/click_log_sample.tsv. This is a tab-separated file, where each line has 7 elements. For example, the line 1dd100500 QUERY1 50 0.259109 ["http://1", "http://2", "http://3","http://4","http://5","http://6","http://7","http://8","http://9","Http://10"] [false, false, false, false, true, true, false, false, false, false] [0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0] has the following fields:
1dd100500 — some identifier (currently not used)QUERY1 — text of the query. It can contain any UTF-8 characters except tab sign \t50 — integer identifier of the region (country, city) of the user who submitted the query. If you don't want this, just put some constant (e.g., 0) in this column.
At Yandex user region is heavily used by ranking, so throughout the code the pair (query, region) is used to identify the query, i.e., if we have the same query string from two different region we consider them as two separate queries.0.259109 — float value, corresponding to the probability P(I = V) that user has a vertical intent V, i.e., he or she is interested more in the vertical documents than in organic web documents. If you do not want any of this intent stuff just put 0 in this column. We assume that the user has one of the two intents: vertical intent V with probability P(I = V) and regular web intent with probability 1 - P(I = V). For example, if we want to take into account user's interest in images and we somehow know the probability that the user is interested more in images than organic web results, we can make use of intent-aware click models. See Chuklin, A. et al. 2013. Using Intent Information to Model User Behavior in Diversified Search. ECIR (2013) for more details.false of the same length as the previous list.If you need more data to experiment with you can use any publicly available dataset and convert it to the format described above. For example, you can use a dataset provided by one of the Yandex challenges (you need to register to get access to the data):
This file.
Self explaining.
Python package installation file: python setup.py --help
Directory with the scripts.
Directory with the core code. This is the directory that gets installed by setup.py.
data/ directory contains an example of click log (see format description above) as well as two examples of result pages with fresh block included (see makeGluedSERP.py description above): data/serp_sample.json is used in an example above, while data/serp_sample2.json was used to create a picture in the paper Chuklin, A. et al. 2013. Using Intent Information to Model User Behavior in Diversified Search. ECIR (2013).
Automaticaly generated documentation. Also available online.
Directory with the CSS/JS files for bin/generate_serp.py; it outputs there a number of HTML files for you to examine to get an idea what kind of SERPs we address with the intent-aware click models.
{not used by the other, does not use other code}
Create html of the SERP containing fresh block item. This is used just to illustrate the notion of presentation types used by Intent-Aware models. Run as: bin/generate_serp.py < data/serp_sample.json. Output is placed in html/ directory.
WARNING: all previously generated html files in this directory will be removed
{not used by other scripts}
Script used to compare different models and output significance of the difference. The pair of models to compare is specified in the code by modifying TESTED_MODEL_PAIRS variable. Model pair is a text string which is mapped to the pair of functions returning model objects (see MODEL_CONSTRUCTORS dict for the mapping). E.g. UBMvsDBN is used to compare UBM model (UbmModel()) to the default DBN model (DbnModel((0.9, 0.9, 0.9, 0.9))). NB: we may have a list of models needed to be compared to each other. For this purpose the same notion of pair is abused. For instance, MODEL_CONSTRUCTORS['EB_UBM'] contains 3 algorithms to be compared to each other: UBM, EB_UBM, EB_UBM-IA.
Usage: bin/compare_click_models.py directory_with_click_logs 2>run.log
Input: directory_with_click_logs — directory containing files with click logs. Each file is in the format described above. These files are then sorted alphabetically and split into pairs where first file is used for training, the second one is used for testing. For example, if the directory contains files f01, f02, f03, f04 then f02 will be used to test models trained using f01, f04 will be used to test models trained using f03 and so on. Two models are evaluated on the test set and their performances (Average Perplexity or Log Likelihood) are compared using appropriate formula (see perpGain and llGain functions respectively). NB: Multiple train and test files are needed to calculate confidence interval for the gains (bootstrap.py is used for this purpose).
Output: some progress output is printend to sys.stderr which might be useful for a long run. Finally the gains of one model over another is output in the following format:
UBM (0, 1) [-0.0115, 0.0659, 0.075, 0.0778, 0.0623, 0.0403, 0.0593, -0.040] (0.0095, 0.0662)
It first outputs the name of the "pair" pairName specified in the TESTED_MODEL_PAIRS, then pair of indeces (i, j) which mean that the model compared are MODEL_CONSTRUCTORS[pairName][i] and MODEL_CONSTRUCTORS[pairName][j] which is UBM and UBM-IA in our example. Next is the list of gains of model j over model i for each pair of the train/test logs (in our example we had 8 pairs of files under directory_with_click_logs). The next element is the confidence interval according to bootstrap test (with 95% confidence level and 1000 bootstrap samples). This line will be printed for all the "pairs" listed under TESTED_MODEL_PAIRS and for both Average Perplexity (PERPLEXITY) and Log Likelihood (LL). For perplexity measure also the gains for individual position (rank) are printed. Like this: UBM POSITION PERPLEXITY GAINS: (0, 1) [[average_gain_for_pos1, confidence_interval_for_pos_1], …]
Run click model inference and evaluate the models:
bin/run_inference.py < data/click_log_sample.tsv 2>inference.logsys.stdin)TRAIN_FOR_METRIC = False): ModelName (LogLikelihood, Perplexity)This file contains implementation of all the click models, probabilistic inference and helper functions needed to work with them. This is the core of the codebase. More details about the classes/functions below.
Copyright © Ernesto P. Adorio: code we use to perform a bootstrap test.
Copy this file to config.py and modify it; it will be used by bin/run_inference.py and bin/compare_click_models.py.
This is the file with default settings for different parameters of the inference, input and output data.
MAX_ITERATIONS — maximum number of iterations in Expectation Maximization (EM) algorithm (applicable only for models using EM algorithm).DEBUG — perform some additional tests when running algorithm (makes it slower)PRETTY_LOG — make log output prettier. If False then more information is put into log.USED_MODELS — list of model names to be tested in __main__ section of the script. Possible names are ['Baseline', 'SDBN', 'UBM', 'UBM-IA', 'EB_UBM', 'EB_UBM-IA', 'DCM', 'DCM-IA', 'DBN', 'DBN-IA']. Please refer to the __main__ section of inference.py to see how these names are expressed in terms of our class hierarchy (all those nasty if 'XXX' in USED_MODELS).MIN_DOCS_PER_QUERY, MAX_DOCS_PER_QUERY – number of documents per query. Set to 10 by default as most of search engines return list of 10 doucments.SERP_SIZE - size of the search engi