Fix null convergence (#7)

* Add treatment for null convergence due precision limits. * add OS update to build. * update pip version on build.
WillianFuks · Feb 17, 2021 · 0aa8a73 · 0aa8a73
1 parent 09215c7
commit 0aa8a73
Show file tree

Hide file tree

Showing 9 changed files with 95 additions and 76 deletions.
diff --git a/Makefile b/Makefile
@@ -2,7 +2,7 @@
 
 flake8:
 	pip install -U flake8
-	flake8
+	flake8 pyClickModels
 
 isort:
 	pip install -U isort
@@ -30,4 +30,4 @@ publish:
 	sh ./scripts/build_wheels.sh
 	#twine upload --repository testpypi dist/*
 	twine upload dist/*
-	rm -fr build dist .egg *.egg-info
+	#rm -fr build dist .egg *.egg-info
diff --git a/pyClickModels/DBN.pyx b/pyClickModels/DBN.pyx
@@ -1,22 +1,24 @@
 # cython: linetrace=True
 
-import os
-from glob import glob
 import gzip
+import os
 import time
+from glob import glob
+
 import ujson
-from libcpp.vector cimport vector
-from libcpp.unordered_map cimport unordered_map
-from libcpp.string cimport string
-from libc.stdlib cimport rand, RAND_MAX, srand
-from libc.time cimport time as ctime
+
 from cython.operator cimport dereference, postincrement
-from pyClickModels.jsonc cimport(json_object, json_tokener_parse,
-                                 json_object_object_get_ex, json_object_get_string,
-                                 lh_table, lh_entry, json_object_array_length,
-                                 json_object_array_get_idx, json_object_get_int,
-                                 json_object_put)
+from libc.stdlib cimport RAND_MAX, rand, srand
+from libc.time cimport time as ctime
+from libcpp.string cimport string
+from libcpp.unordered_map cimport unordered_map
+from libcpp.vector cimport vector
 
+from pyClickModels.jsonc cimport (json_object, json_object_array_get_idx,
+                                  json_object_array_length,
+                                  json_object_get_int, json_object_get_string,
+                                  json_object_object_get_ex, json_object_put,
+                                  json_tokener_parse, lh_entry, lh_table)
 
 # Start by setting the seed for the random values required for initalizing the DBN
 # parameters.
@@ -125,12 +127,14 @@ cdef class Factor:
                             result *= (1 - self.gamma) * (1 - self.cr)
                         else:
                             result *= self.gamma * (1 - self.cr)
+        # Compute P(C_{>r},P_{>r} | E_{r+1})
         if not z:
             if self.last_r >= self.r + 1:
                 return 0
         else:
             if self.r < self.cp_vector_given_e[0].size():
                 result *= self.cp_vector_given_e[0][self.r]
+        # P(E_r=x | C<r, P<r)
         result *= (self.e_r_vector_given_CP[0][self.r] if x else
                    1 - self.e_r_vector_given_CP[0][self.r])
         return result
@@ -414,6 +418,7 @@ cdef class DBNModel():
         # Probability of clicks at positions greater than the last document in results
         # page is zero.
         X_r_vector[total_docs] = 0
+        gamma = self.get_param(b'gamma')
 
         for r in range(total_docs - 1, -1, -1):
             json_object_object_get_ex(
@@ -423,7 +428,6 @@ cdef class DBNModel():
             )
             doc = json_object_get_string(tmp)
             alpha = self.get_param(b'alpha', query, &doc)
-            gamma = self.get_param(b'gamma')
 
             X_r_1 = X_r_vector[r + 1]
             X_r = alpha[0] + (1 - alpha[0]) * gamma[0] * X_r_1
@@ -438,6 +442,10 @@ cdef class DBNModel():
 
         Mathematically: P(E_r = 1 | C_{<r}, P_{<r})
 
+        This is discussed in equation (24) in the blog post:
+
+        https://towardsdatascience.com/how-to-extract-relevance-from-clickstream-data-2a870df219fb
+
         Args
         ----
           clickstream: *json_object
@@ -465,8 +473,10 @@ cdef class DBNModel():
             # position r + 1 will be required later so add +1 in computation
             vector[float] e_r_vector_given_CP = vector[float](total_docs + 1 - idx, 0.0)
 
-        # First document has 100% of being Examined regardless of clicks or purchases.
+        # First document has 100% chance of being Examined regardless of clicks or
+        # purchases.
         e_r_vector_given_CP[0] = 1
+        gamma = self.get_param(b'gamma')
 
         for r in range(idx, total_docs):
             json_object_object_get_ex(
@@ -492,7 +502,6 @@ cdef class DBNModel():
 
             alpha = self.get_param(b'alpha', query, &doc)
             sigma = self.get_param(b'sigma', query, &doc)
-            gamma = self.get_param(b'gamma')
 
             if purchase:
                 return e_r_vector_given_CP
@@ -602,6 +611,10 @@ cdef class DBNModel():
 
         P(C_{>r}, P_{>r} | E_{r+1})
 
+        This is equation (25) from blog post:
+
+        https://towardsdatascience.com/how-to-extract-relevance-from-clickstream-data-2a870df219fb
+
         Args
         ----
           clickstream: *json_object
@@ -624,13 +637,10 @@ cdef class DBNModel():
 
         # Subtract 1 as E_{r+1} is defined up to r - 1 documents
         for r in range(total_docs - 1):
-
             e_r_vector_given_CP = self.build_e_r_vector_given_CP(clickstream, r + 1,
                                                                  query)
-
             cp_vector_given_e[r] = self.compute_cp_p(clickstream, r + 1, query,
                                                      &e_r_vector_given_CP, cr_dict)
-
         return cp_vector_given_e
 
     cdef int get_last_r(self, json_object *clickstream, const char *event=b'click'):
@@ -868,7 +878,7 @@ cdef class DBNModel():
         json_object_object_get_ex(doc_data, b'purchase', &tmp)
         purchase = json_object_get_int(tmp)
 
-        alpha = self.get_param(b'gamma', query, &doc)[0]
+        alpha = self.get_param(b'alpha', query, &doc)[0]
         sigma = self.get_param(b'sigma', query, &doc)[0]
         gamma = self.get_param(b'gamma')[0]
 
@@ -887,15 +897,19 @@ cdef class DBNModel():
             e_r_vector_given_CP,
             cp_vector_given_e
         )
+
         # Loop through all possible values of x, y and z, where each is an integer
         # boolean.
         for i in range(2):
             for j in range(2):
                 for k in range(2):
                     ESS_denominator += factor.compute_factor(i, j, k)
 
-        ESS_0 = factor.compute_factor(1, 0, 0) / ESS_denominator
-        ESS_1 = factor.compute_factor(1, 0, 1) / ESS_denominator
+        if not ESS_denominator:
+            ESS_0, ESS_1 = 0, 0
+        else:
+            ESS_0 = factor.compute_factor(1, 0, 0) / ESS_denominator
+            ESS_1 = factor.compute_factor(1, 0, 1) / ESS_denominator
 
         tmp_gamma_param[0][0] += ESS_1
         tmp_gamma_param[0][1] += ESS_0 + ESS_1

diff --git a/pyClickModels/__version__.py b/pyClickModels/__version__.py
@@ -1,3 +1 @@
-VERSION = (0, 0, 1)
-
-__version__ = '.'.join([str(e) for e in VERSION])
+__version__ = '0.0.2'
diff --git a/scripts/build_wheels.sh b/scripts/build_wheels.sh
@@ -1,10 +1,12 @@
 docker run -v $(pwd):/pyClickModels quay.io/pypa/manylinux1_x86_64 sh -c '''
+yum update
 yum install -y json-c-devel
 
 cd /pyClickModels
 
 for PYVER in /opt/python/*/bin/; do
     if [[ $PYVER != *"27"* ]]; then
+        "${PYVER}/pip" install -U pip
         "${PYVER}/pip" install -U setuptools
         "${PYVER}/pip" install -r requirements.txt
         "${PYVER}/python" setup.py sdist bdist_wheel

diff --git a/setup.py b/setup.py
@@ -2,13 +2,13 @@
 
 import os
 import sys
-import Cython.Compiler.Options
 from codecs import open
-from Cython.Distutils import build_ext
 from setuptools import setup
-from Cython.Build import cythonize
-from distutils.extension import Extension
 from setuptools.command.test import test as TestCommand
+from distutils.extension import Extension
+import Cython.Compiler.Options
+from Cython.Distutils import build_ext
+from Cython.Build import cythonize
 
 
 here = os.path.abspath(os.path.dirname(__file__))

diff --git a/tests/conftest.py b/tests/conftest.py
@@ -82,24 +82,3 @@ def build_DBN_test_data(users=10, docs=10, queries=2):
         for row in final_result[half_results:]:
             f.write(json.dumps(row).encode() + '\n'.encode())
     return persistence, params, tmp_folder
-
-
-@pytest.fixture
-def sessions():
-    sessions = [
-        {
-            'sessionID': [
-                {"doc": "doc0", "click": 0, "purchase": 0},
-                {"doc": "doc1", "click": 1, "purchase": 0},
-                {"doc": "doc2", "click": 1, "purchase": 1}
-            ]
-        },
-        {
-            'sessionID': [
-                {"doc": "doc0", "click": 0, "purchase": 0},
-                {"doc": "doc1", "click": 1, "purchase": 0}
-            ]
-        },
-
-    ]
-    return sessions
diff --git a/tests/fixtures/eighty_skus/judgments.gz b/tests/fixtures/eighty_skus/judgments.gz
diff --git a/tests/fixtures/null_test/judgments_test_null.gz b/tests/fixtures/null_test/judgments_test_null.gz
diff --git a/tests/test_cy_DBN.pyx b/tests/test_cy_DBN.pyx
@@ -1,16 +1,23 @@
+import gzip
 import tempfile
+
 import ujson
-import gzip
+
+from cython.operator cimport dereference, postincrement
 from libcpp.string cimport string
 from libcpp.unordered_map cimport unordered_map
 from libcpp.vector cimport vector
-from cython.operator cimport dereference, postincrement
+
 from pyClickModels.DBN cimport DBNModel, Factor
+
 from pyClickModels.DBN import DBN
-from pyClickModels.jsonc cimport(json_object, json_tokener_parse,
-                                 json_object_get_object, lh_table, json_object_put)
+
+from pyClickModels.jsonc cimport (json_object, json_object_get_object,
+                                  json_object_put, json_tokener_parse,
+                                  lh_table)
+
 from conftest import build_DBN_test_data
-from numpy.testing import assert_almost_equal, assert_allclose
+from numpy.testing import assert_allclose, assert_almost_equal
 
 ctypedef unordered_map[string, unordered_map[string, float]] dbn_param
 
@@ -49,7 +56,6 @@ cdef bint test_fit():
 
     # it = model.alpha_params.begin()
     while(it != model.alpha_params.end()):
-    #     prints keys
         # print(dereference(it).first)
         query = (dereference(it).first)
         dquery = extract_keys(query)
@@ -1886,6 +1892,22 @@ cdef bint test_export_judgments():
     return True
 
 
+cdef bint test_not_null_converence():
+    cdef:
+        DBNModel model = DBN()
+
+    model.fit('tests/fixtures/null_test', iters=1)
+    return True
+
+
+cdef bint test_long_list_null_converence():
+    cdef:
+        DBNModel model = DBN()
+
+    model.fit('tests/fixtures/eighty_skus', iters=2)
+    return True
+
+
 cpdef run_tests():
     assert test_get_search_context_string()
     assert test_compute_cr()
@@ -1906,25 +1928,29 @@ cpdef run_tests():
     assert test_update_gamma_param()
     assert test_fit()
     assert test_export_judgments()
+    assert test_not_null_converence()
+    assert test_long_list_null_converence()
 
 
 if __name__ == '__main__':
-    assert test_get_search_context_string()
-    assert test_compute_cr()
-    assert test_get_param()
-    assert test_build_e_r_vector(&alpha_params, &sigma_params, &gamma_param)
-    assert test_build_X_r_vector(&alpha_params, &sigma_params, &gamma_param)
-    assert test_build_e_r_vector_given_CP(&alpha_params, &sigma_params, &gamma_param)
-    assert test_build_cp_p(&alpha_params)
-    assert test_build_CP_vector_given_e(&alpha_params, &sigma_params, &gamma_param)
-    assert test_get_last_r()
-    assert test_update_tmp_alpha(&alpha_params, &sigma_params, &gamma_param)
-    assert test_update_tmp_sigma(&alpha_params, &sigma_params, &gamma_param)
-    assert test_compute_factor_last_click_lower_than_r()
-    assert test_compute_factor_last_click_higher_than_r()
-    assert test_update_tmp_gamma()
-    assert test_update_alpha_params()
-    assert test_update_sigma_params()
-    assert test_update_gamma_param()
-    assert test_fit()
-    assert test_export_judgments()
+    #assert test_get_search_context_string()
+    #assert test_compute_cr()
+    #assert test_get_param()
+    #assert test_build_e_r_vector(&alpha_params, &sigma_params, &gamma_param)
+    #assert test_build_X_r_vector(&alpha_params, &sigma_params, &gamma_param)
+    #assert test_build_e_r_vector_given_CP(&alpha_params, &sigma_params, &gamma_param)
+    #assert test_build_cp_p(&alpha_params)
+    #assert test_build_CP_vector_given_e(&alpha_params, &sigma_params, &gamma_param)
+    #assert test_get_last_r()
+    #assert test_update_tmp_alpha(&alpha_params, &sigma_params, &gamma_param)
+    #assert test_update_tmp_sigma(&alpha_params, &sigma_params, &gamma_param)
+    #assert test_compute_factor_last_click_lower_than_r()
+    #assert test_compute_factor_last_click_higher_than_r()
+    #assert test_update_tmp_gamma()
+    #assert test_update_alpha_params()
+    #assert test_update_sigma_params()
+    #assert test_update_gamma_param()
+    #assert test_fit()
+    #assert test_export_judgments()
+    #assert test_not_null_converence()
+    pass