def test_weighted_distance(self):
test_weights = [2, 2, 1, 1]
scaled_x1 = [ sqrt(w) * v for w, v in zip(test_weights, self.x1) ]
scaled_x2 = [ sqrt(w) * v for w, v in zip(test_weights, self.x2) ]
ref_res = euclidean(scaled_x1, scaled_x2)
test_res = weighted_euclidean(self.x1, self.x2, test_weights)
self.assertEqual(test_res, ref_res)
def test_unweighted_distnace(self):
"""Test unweighted(or euqally weighted) distance calculation
which effectively same to standard euclidean distance
"""
ref_res = euclidean(self.x1, self.x2)
test_res = weighted_euclidean(self.x1, self.x2)
self.assertEqual(test_res, ref_res)
def fitted_dist_func(self, x, y):
""" Returned the distance functions used in fitting model
Returns:
--------
func: {function} a function accept (x1, x2, *arg)
"""
if self._transform_matrix is not None:
w = self._transform_matrix
g = lambda x, y: weighted_euclidean(x, y, w)
return g(x, y)
def fitted_dist_func(self, x, y):
""" Returned the distance functions used in fitting model
Returns:
--------
func: {function} a function accept (x1, x2, *arg)
"""
if self._transform_matrix is not None:
w = self._transform_matrix
g = lambda x, y: weighted_euclidean(x, y, w)
return g(x, y)
def _fit(self, X, S, D=None):
""" Fit the model with given information: X, S, D
Fit the learning distance metrics: (1) if only S is given, all pairs of
items in X but not in S are considered as in D; (2) if both S and D
given, items in X but neither in S nor in D will be removed from
fitting process.
Parameters:
----------
X: {matrix-like, np.array}, shape (n_sample, n_features) matrix of
observations with 1st column keeping observation ID
S: {vector-like, list} a list of tuples which define a pair of data
points known as similiar
D: {vector-like, list} a list of tuples which define a pair of data
points known as different
Returns:
--------
_trans_vec: {matrix-like, np.array}, shape(n_features, n_features)
A transformation matrix (A)
_ratio: float
"""
# if isinstance(X, pd.DataFrame):
# X = X.as_matrix()
try:
# ids = X["ID"]
# X = X[[c for c in X.columns if c != "ID"]]
ids = X.index.tolist()
except ValueError:
print "Oops! No 'ID' column is found !"
# ids = [int(i) for i in X.ix[:, 0]]
# X = X.ix[:, 1:]
n_sample, n_features = X.shape
bnds = [(0, None)] * n_features # boundaries
init = [1] * n_features # initial weights
if D == None:
all_pairs = [p for p in combinations(ids, 2)]
D = get_exclusive_pairs(all_pairs, S)
else:
# if D is provided, keep only users not being
# covered either by S or D
covered_items = get_unique_items(S, D)
keep_items = [find_index(i, ids) for i in ids \
if i in covered_items]
X = X.ix[keep_items, :]
# Convert ids in D and S into row index, in order to provide them to
# a set of two distance functions, squared_sum_grouped_dist() and
# sum_grouped_dist()
S_idx = [(find_index(a, ids), find_index(b, ids)) for (a, b) in S]
D_idx = [(find_index(a, ids), find_index(b, ids)) for (a, b) in D]
def objective_func(w):
a = squared_sum_grouped_dist(S_idx, X, w) * 1.0
b = sum_grouped_dist(D_idx, X, w) * 1.0
return a - b
if self._is_debug:
try:
print "Examples of S: %s" % S[:5], len(S)
print "Examples of D: %s" % D[:5], len(D)
print "Examples of X: %s" % X[:5, :], X.shape
except:
print "Examples of S: %s" % S, len(S)
print "Examples of D: %s" % D, len(D)
print "Examples of X: %s" % X, X.shape
start_time = time.time()
fitted = minimize(objective_func, init, method="L-BFGS-B", bounds=bnds, options={'maxiter':10, 'disp' : True})
duration = time.time() - start_time
if self._report_excution_time:
print("--- %.2f seconds ---" % duration)
w = self._transform_matrix
self._transform_matrix = vec_normalized(fitted['x'])
# optimized value vs. value of initial setting
self._ratio = fitted['fun'] / objective_func(init)
self._dist_func = lambda x, y: weighted_euclidean(x, y, w)
return (self._transform_matrix, self._ratio)
def _fit(self, X, S, D=None):
""" Fit the model with given information: X, S, D
Fit the learning distance metrics: (1) if only S is given, all pairs of
items in X but not in S are considered as in D; (2) if both S and D
given, items in X but neither in S nor in D will be removed from
fitting process.
Parameters:
----------
X: {matrix-like, np.array}, shape (n_sample, n_features) matrix of
observations with 1st column keeping observation ID
S: {vector-like, list} a list of tuples which define a pair of data
points known as similiar
D: {vector-like, list} a list of tuples which define a pair of data
points known as different
Returns:
--------
_trans_vec: {matrix-like, np.array}, shape(n_features, n_features)
A transformation matrix (A)
_ratio: float
"""
# if isinstance(X, pd.DataFrame):
# X = X.as_matrix()
try:
# ids = X["ID"]
# X = X[[c for c in X.columns if c != "ID"]]
ids = X.index.tolist()
except ValueError:
print "Oops! No 'ID' column is found !"
# ids = [int(i) for i in X.ix[:, 0]]
# X = X.ix[:, 1:]
n_sample, n_features = X.shape
bnds = [(0, None)] * n_features # boundaries
init = [1] * n_features # initial weights
if D == None:
all_pairs = [p for p in combinations(ids, 2)]
D = get_exclusive_pairs(all_pairs, S)
else:
# if D is provided, keep only users not being
# covered either by S or D
covered_items = get_unique_items(S, D)
keep_items = [find_index(i, ids) for i in ids \
if i in covered_items]
X = X.ix[keep_items, :]
# Convert ids in D and S into row index, in order to provide them to
# a set of two distance functions, squared_sum_grouped_dist() and
# sum_grouped_dist()
S_idx = [(find_index(a, ids), find_index(b, ids)) for (a, b) in S]
D_idx = [(find_index(a, ids), find_index(b, ids)) for (a, b) in D]
def objective_func(w):
a = squared_sum_grouped_dist(S_idx, X, w) * 1.0
b = sum_grouped_dist(D_idx, X, w) * 1.0
return a - b
if self._is_debug:
try:
print "Examples of S: %s" % S[:5], len(S)
print "Examples of D: %s" % D[:5], len(D)
print "Examples of X: %s" % X[:5, :], X.shape
except:
print "Examples of S: %s" % S, len(S)
print "Examples of D: %s" % D, len(D)
print "Examples of X: %s" % X, X.shape
start_time = time.time()
fitted = minimize(objective_func,
init,
method="L-BFGS-B",
bounds=bnds,
options={
'maxiter': 10,
'disp': True
})
duration = time.time() - start_time
if self._report_excution_time:
print("--- %.2f seconds ---" % duration)
w = self._transform_matrix
self._transform_matrix = vec_normalized(fitted['x'])
# optimized value vs. value of initial setting
self._ratio = fitted['fun'] / objective_func(init)
self._dist_func = lambda x, y: weighted_euclidean(x, y, w)
return (self._transform_matrix, self._ratio)