clickmodels 1.0.2

README.md

This file.

LICENSE, AUTHORS, CHANES.txt

Self explaining.

setup.py

Python package installation file: python setup.py --help

bin/

Directory with the scripts.

clickmodels/

Directory with the core code. This is the directory that gets installed by setup.py.

data/

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).

doc/

Automaticaly generated documentation. Also available online.

html/

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.

bin/generate_serp.py

{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

bin/compare_click_models.py

{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], …]

bin/run_inference.py

Run click model inference and evaluate the models:

• Usage: bin/run_inference.py < data/click_log_sample.tsv 2>inference.log

• Input: click log in the format described above (sys.stdin)

• Output (assuming that TRAIN_FOR_METRIC = False): ModelName (LogLikelihood, Perplexity)

clickmodels/inference.py

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.

clickmodels/bootstrap.py

Copyright © Ernesto P. Adorio: code we use to perform a bootstrap test.

clickmodels/config_sample.py

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 engine result page (SERP). Used if we want to model clicks beyond the first result page. See the section named Beyond the First Result Page below for more details.

• EXTENDED_LOG_FORMAT - if set to True the urls, layout and clicks are dicts instead of lists (see Format of the Click Log section above). Example: data/click_log_sample_extended_format.tsv.

• TRANSFORM_LOG - transform the click log by inserting the fake documents for pagination button (currently works only with EXTENDED_LOG_FORMAT = True). See the section named Beyond the First Result Page below.

• QUERY_INDEPENDENT_PAGER - used to switch between SDBN(P) / SDBN(P-Q). Only used with TRANSFORM_LOG = True.

• TRAIN_FOR_METRIC – if True the model will be trained such that its parameters can be used in a metric (like Chuklin, A. et al. 2013. Click Model-Based Information Retrieval Metrics. SIGIR (2013).). See the section below for more details.

• PRINT_EBU_STATS — if True the parameters of the EBU metric will be printed first (Yilmaz, E. et al. 2010. Expected browsing utility for web search evaluation. CIKM. (2010)).

• DEFAULT_REL - default (prior) relevance (attractiveness, satisfaction) parameter values used in click models like DBN or EBU.

clickmodels/input_reader.py

A class used for reading input data in a click log format described above.

Click Models

The base class for all the click models is the class called ClickModel. In order to define a new click model you should create a clas inherited from it and re-define methods train and _getClickProbs.

• train function

Note, that test method is already implemented and uses _getClickProbs function. If you redefine __init__ method, then be sure to invoke the __init__ of the parent class to set the ignoreIntents and ignoreLayout parameters (they should be set to True unless you are using Intent Aware model)

ClickModel class by itself represents a baseline click model which sets probability 0.5 to any click event.

InputReader

This class intented to read input (click log) in the format described above. To save memory, it maps queries and urls to ids. It means, that you need to use the same instance of the InputReader class even if you read multiple click log files. Otherwise you will end up with two different ids assigned to the same query.

Input Format

The format in this case is similar to the one descirbed above with only difference that URLs should be replaced by the relevance grade of the corresponding document to the corresponding query. The query field will be ignored in that case. The relevance grade should take one of the following values:

• IRRELEVANT — lowest relevance scorn, document is not relevant to the query

• RELEVANT — combines marginally relevant and just relevant documents

• USEFUL — document is more than just relevant, it is really useful

• VITAL — highest relevance score, the document is essential

Please note, that if you also have PRINT_EBU_STATS set to True, then the parameters of the EBU / rrDBN metric will be printed out first (these ones can be computed directly without need to train a model).

Output

For each model corresponding parametres will be printed out:

• UBM — attractiveness probabilities alpha and position discount parameters gamma

• DCM — attractiveness probabilities alpha (named as urlRelevances in the code) and position discount parameters gamma

For more conceptual details about converting click models into evaluation metrics please refer to the paper Chuklin, A. et al. 2013. Click Model-Based Information Retrieval Metrics. SIGIR (2013).