def __init__(self, feature_count, arg_str):
"""
@param featur_count: the number of features
@param arg_str: "-h HISTORY_LENGTH -e NUM_CANDIDATES \
-s SELECT_CANDIDATE".
"""
ListwiseLearningSystem.__init__(self, feature_count, arg_str)
parser = argparse.ArgumentParser(prog=self.__class__.__name__)
parser.add_argument("-e", "--num_candidates", required=True, type=int,
help="Number of candidate rankers to explore in each round.")
parser.add_argument("-l", "--history_length", required=True, type=int,
help="Number of historic data points to take into account when "
"pre-selecting candidates.")
parser.add_argument("-s", "--select_candidate", required=True,
help="Method for selecting a candidate ranker from a ranker pool."
" Options: select_candidate_random, select_candidate_simple,"
" select_candidate_repeated, or own implementation.")
parser.add_argument("-b", "--biased", default="False",
help="Set to true if comparison should be biased (i.e., not use"
"importance sampling).")
parser.add_argument("-r", "--num_repetitions", type=int, default=1,
help="The number of repetitions for each ranker pair evaluation"
"(when the selection method is select_candidate_repeated).")
args = vars(parser.parse_known_args(split_arg_str(arg_str))[0])
self.num_candidates = args["num_candidates"]
self.select_candidate = getattr(self, args["select_candidate"])
self.history_length = args["history_length"]
self.biased = string_to_boolean(args["biased"])
logging.info("Initialized historical data usage to: %r" % self.biased)
self.num_repetitions = args["num_repetitions"]
self.history = []
def __init__(self, feature_count, arg_str):
"""
@param featur_count: the number of features
@param arg_str: "-h HISTORY_LENGTH -e NUM_CANDIDATES \
-s SELECT_CANDIDATE".
"""
ListwiseLearningSystem.__init__(self, feature_count, arg_str)
parser = argparse.ArgumentParser(prog=self.__class__.__name__)
parser.add_argument(
"-e",
"--num_candidates",
required=True,
type=int,
help="Number of candidate rankers to explore in each round.")
parser.add_argument(
"-l",
"--history_length",
required=True,
type=int,
help="Number of historic data points to take into account when "
"pre-selecting candidates.")
parser.add_argument(
"-s",
"--select_candidate",
required=True,
help="Method for selecting a candidate ranker from a ranker pool."
" Options: select_candidate_random, select_candidate_simple,"
" select_candidate_repeated, or own implementation.")
parser.add_argument(
"-b",
"--biased",
default="False",
help="Set to true if comparison should be biased (i.e., not use"
"importance sampling).")
parser.add_argument(
"-r",
"--num_repetitions",
type=int,
default=1,
help="The number of repetitions for each ranker pair evaluation"
"(when the selection method is select_candidate_repeated).")
args = vars(parser.parse_known_args(split_arg_str(arg_str))[0])
self.num_candidates = args["num_candidates"]
self.select_candidate = getattr(self, args["select_candidate"])
self.history_length = args["history_length"]
self.biased = string_to_boolean(args["biased"])
logging.info("Initialized historical data usage to: %r" % self.biased)
self.num_repetitions = args["num_repetitions"]
self.history = []
def setUp(self):
# initialize query
self.test_num_features = 6
test_query = """
4 qid:1 1:2.6 2:1 3:2.1 4:0 5:2 6:1.4 # highly relevant
1 qid:1 1:1.2 2:1 3:2.9 4:0 5:2 6:1.9 # bad
0 qid:1 1:0.5 2:1 3:2.3 4:0 5:2 6:5.6 # not relevant
0 qid:1 1:0.5 2:1 3:2.3 4:0 5:2 6:5.6 # not relevant
"""
self.query_fh = cStringIO.StringIO(test_query)
self.queries = query.Queries(self.query_fh, self.test_num_features)
self.query = self.queries['1']
# initialize listwise learner
self.learner = ListwiseLearningSystem(self.test_num_features,
"--init_weights 0,0,1,0,0,0 --delta 1.0 --alpha 0.01 --ranker "
"ranker.ProbabilisticRankingFunction --ranker_args 3 --ranker_tie "
"first --comparison comparison.ProbabilisticInterleaveWithHistory"
" --comparison_args \"--history_length 10 --biased true\"")
class TestListwiseLearning(unittest.TestCase):
def setUp(self):
# initialize query
self.test_num_features = 6
test_query = """
4 qid:1 1:2.6 2:1 3:2.1 4:0 5:2 6:1.4 # highly relevant
1 qid:1 1:1.2 2:1 3:2.9 4:0 5:2 6:1.9 # bad
0 qid:1 1:0.5 2:1 3:2.3 4:0 5:2 6:5.6 # not relevant
0 qid:1 1:0.5 2:1 3:2.3 4:0 5:2 6:5.6 # not relevant
"""
self.query_fh = cStringIO.StringIO(test_query)
self.queries = query.Queries(self.query_fh, self.test_num_features)
self.query = self.queries['1']
# initialize listwise learner
self.learner = ListwiseLearningSystem(self.test_num_features,
"--init_weights 0,0,1,0,0,0 --delta 1.0 --alpha 0.01 --ranker "
"ranker.ProbabilisticRankingFunction --ranker_args 3 --ranker_tie "
"first --comparison comparison.ProbabilisticInterleaveWithHistory"
" --comparison_args \"--history_length 10 --biased true\"")
def testRanker(self):
self.learner.get_ranked_list(self.query)
def _update_solution(self, outcome, clicks):
# Keep track of history
if self.history_length > 0:
if len(self.history) == self.history_length:
self.history.pop(0)
# store probability of the observed list under the source
# distribution so that it only has to be computed once
new_h_item = HistoryItem(self.current_l, self.current_context,
clicks, self.current_query)
new_h_item.p_list_source = self.comparison.get_probability_of_list(
self.current_l, self.current_context, self.current_query)
self.history.append(new_h_item)
# use inherited method for the actual update
return ListwiseLearningSystem._update_solution(self, outcome, clicks)
def _update_solution(self, outcome, clicks):
# Keep track of history
if self.history_length > 0:
if len(self.history) == self.history_length:
self.history.pop(0)
# store probability of the observed list under the source
# distribution so that it only has to be computed once
new_h_item = HistoryItem(self.current_l, self.current_context,
clicks, self.current_query)
new_h_item.p_list_source = self.comparison.get_probability_of_list(
self.current_l, self.current_context, self.current_query)
self.history.append(new_h_item)
# use inherited method for the actual update
return ListwiseLearningSystem._update_solution(self, outcome, clicks)
