def test_ials_overfit_cholesky(
test_interaction_data: Dict[str, sps.csr_matrix]
) -> None:
X = test_interaction_data["X_small"]
rec = IALSRecommender(X, n_components=4, alpha=0, reg=0, use_cg=False)
rec.learn()
assert rec.trainer is not None
uvec = rec.trainer.core_trainer.transform_user(X.tocsr().astype(np.float32))
ivec = rec.trainer.core_trainer.transform_item(X.tocsr().astype(np.float32))
X = X.toarray()
reprod = uvec.dot(ivec.T)
np.testing.assert_allclose(reprod, X, rtol=1e-2, atol=1e-2)
def test_ials_overfit_nonzero_alpha(
test_interaction_data: Dict[str, sps.csr_matrix]
) -> None:
ALPHA = 4.5
REG = 3
X = test_interaction_data["X_small"]
rec_chol = IALSRecommender(
X, n_components=4, alpha=ALPHA, reg=REG, use_cg=False, max_epoch=5
)
rec_chol.learn()
assert rec_chol.trainer is not None
uvec_chol = rec_chol.trainer.core_trainer.transform_user(
X.tocsr().astype(np.float32)
)
ivec_chol = rec_chol.trainer.core_trainer.transform_item(
X.tocsr().astype(np.float32)
)
rec_cg = IALSRecommender(
X,
n_components=4,
alpha=ALPHA,
reg=REG,
use_cg=True,
max_cg_steps=4,
max_epoch=5,
)
rec_cg.learn()
assert rec_cg.trainer is not None
uvec_cg = rec_cg.trainer.core_trainer.transform_user(X.tocsr())
ivec_cg = rec_cg.trainer.core_trainer.transform_item(X.tocsr())
np.testing.assert_allclose(uvec_chol, uvec_cg, atol=1e-3, rtol=1e-4)
np.testing.assert_allclose(ivec_chol, ivec_cg, atol=1e-3, rtol=1e-4)
def test_ials_overfit_cg(test_interaction_data: Dict[str, sps.csr_matrix]) -> None:
X = test_interaction_data["X_small"]
rec = IALSRecommender(
X, n_components=4, alpha=0, reg=1e-2, use_cg=True, max_cg_steps=4
)
rec.learn()
assert rec.trainer is not None
uvec = rec.trainer.core_trainer.transform_user(X.tocsr().astype(np.float32))
ivec = rec.trainer.core_trainer.transform_item(X.tocsr().astype(np.float32))
X_dense = X.toarray()
X_dense[X_dense.nonzero()] = 1.0
reproduced_user_vector = uvec.dot(ivec.T)
np.testing.assert_allclose(reproduced_user_vector, X_dense, rtol=1e-2, atol=1e-2)
X_reproduced = rec.get_score_cold_user(X)
np.testing.assert_allclose(X_reproduced, X_dense, rtol=1e-2, atol=1e-2)
with pytest.raises(ValueError):
_ = rec.compute_item_embedding(X.T)
reproduced_item_vector = rec.compute_item_embedding(X)
X_reproduced_item = rec.get_score_from_item_embedding(
np.arange(X.shape[0]), reproduced_item_vector
)
np.testing.assert_allclose(X_reproduced_item, X_dense, rtol=1e-2, atol=1e-2)
def test_ials_cg_underfit(test_interaction_data: Dict[str, sps.csr_matrix]) -> None:
X = test_interaction_data["X_small"]
rec = IALSRecommender(
X,
n_components=4,
alpha=0,
reg=1e-2,
use_cg=True,
max_cg_steps=1,
max_epoch=10,
)
with pytest.raises(RuntimeError):
_ = rec.core_trainer.user
rec.learn()
assert rec.trainer is not None
uvec = rec.core_trainer.user
ivec = rec.core_trainer.item
X_dense = X.toarray()
reprod = uvec.dot(ivec.T)
np.testing.assert_allclose(reprod, X_dense, rtol=1e-2, atol=1e-2)
def test_ials_overfit_ialspp(
subspace_dimension: int,
test_interaction_data: Dict[str, sps.csr_matrix]) -> None:
X = test_interaction_data["X_small"].copy()
ALPHA0 = 100
REG = 1.0
rec = IALSRecommender(
X,
n_components=4,
alpha0=ALPHA0,
reg=REG,
solver_type="IALSPP",
loss_type="ORIGINAL",
max_epoch=300,
n_threads=1,
ialspp_subspace_dimension=subspace_dimension,
nu=0,
)
rec.learn()
assert rec.trainer is not None
uvec = rec.get_user_embedding()
ivec = rec.get_item_embedding()
X = X.toarray()
X[X.nonzero()] = 1.0
reprod = uvec.dot(ivec.T)
np.testing.assert_allclose(reprod, X, rtol=1e-2, atol=1e-2)
def test_ials_overfit_cholesky_logscale(
test_interaction_data: Dict[str, sps.csr_matrix]
) -> None:
ALPHA = 1.0
REG = 0.1
EPSILON = 1.0
N_COMPONENTS = 1
X = test_interaction_data["X_small"]
rec_chol = IALSRecommender(
X,
n_components=N_COMPONENTS,
alpha=ALPHA,
reg=REG,
use_cg=False,
epsilon=EPSILON,
confidence_scaling="log",
max_epoch=200,
init_std=1e-1,
)
rec_chol.learn()
assert rec_chol.trainer is not None
uvec_chol = rec_chol.get_user_embedding()
ivec_chol_cold = rec_chol.compute_item_embedding(X)
print(X.shape)
grad_uvec_chol, grad_ivec_chol = ials_grad(
X, uvec_chol, ivec_chol_cold, REG, ALPHA, EPSILON
)
np.testing.assert_allclose(grad_ivec_chol, np.zeros_like(grad_ivec_chol), atol=1e-5)
np.testing.assert_allclose(grad_uvec_chol, np.zeros_like(grad_uvec_chol), atol=1e-5)
def test_ials_overfit_cholesky(
test_interaction_data: Dict[str, sps.csr_matrix]
) -> None:
X = test_interaction_data["X_small"]
rec = IALSRecommender(
X,
n_components=4,
alpha=0,
reg=0.001,
use_cg=False,
)
rec.learn()
assert rec.trainer is not None
uvec = rec.compute_user_embedding(X)
ivec = rec.compute_item_embedding(X)
X = X.toarray()
X[X.nonzero()] = 1.0
reprod = uvec.dot(ivec.T)
np.testing.assert_allclose(reprod, X, rtol=1e-2, atol=1e-2)
def test_ials_overfit_cholesky(
test_interaction_data: Dict[str, sps.csr_matrix]) -> None:
X = test_interaction_data["X_small"]
rec = IALSRecommender(
X,
n_components=4,
alpha0=100,
reg=1e-1,
solver_type="CHOLESKY",
loss_type="ORIGINAL",
max_epoch=100,
n_threads=1,
nu=0,
)
rec.learn()
assert rec.trainer is not None
uvec = rec.compute_user_embedding(X)
ivec = rec.compute_item_embedding(X)
X = X.toarray()
X[X.nonzero()] = 1.0
reprod = uvec.dot(ivec.T)
np.testing.assert_allclose(reprod, X, rtol=1e-2, atol=1e-2)
def test_ials_overfit_cholesky_logscale(
test_interaction_data: Dict[str, sps.csr_matrix]) -> None:
ALPHA0 = 2.4
REG = 1.1
EPSILON = 3.0
N_COMPONENTS = 5
X = test_interaction_data["X_small"]
rec_chol = IALSRecommender(
X,
n_components=N_COMPONENTS,
alpha0=ALPHA0,
reg=REG,
nu=0,
nu_star=0,
solver_type="CHOLESKY",
loss_type="ORIGINAL",
epsilon=EPSILON,
confidence_scaling="log",
max_epoch=200,
init_std=1e-1,
)
rec_chol.learn()
assert rec_chol.trainer is not None
uvec_chol = rec_chol.get_user_embedding()
ivec_chol_cold = rec_chol.compute_item_embedding(X)
grad_uvec_chol, grad_ivec_chol = ials_grad(X, uvec_chol, ivec_chol_cold,
REG, ALPHA0, EPSILON)
np.testing.assert_allclose(grad_ivec_chol,
np.zeros_like(grad_ivec_chol),
atol=1e-5)
np.testing.assert_allclose(grad_uvec_chol,
np.zeros_like(grad_uvec_chol),
atol=1e-5)
def test_ials_overfit_nonzero_alpha(
test_interaction_data: Dict[str, sps.csr_matrix]) -> None:
ALPHA = 4.5
REG = 3
X = test_interaction_data["X_small"]
rec_chol = IALSRecommender(
X,
n_components=4,
alpha0=1 / ALPHA,
reg=REG,
solver_type="CHOLESKY",
max_epoch=5,
)
rec_chol.learn()
assert rec_chol.trainer is not None
uvec_chol = rec_chol.compute_user_embedding(X.tocsr().astype(np.float32))
ivec_chol = rec_chol.compute_item_embedding(X.tocsr().astype(np.float32))
rec_cg = IALSRecommender(
X,
n_components=4,
alpha0=1 / ALPHA,
reg=REG,
solver_type="CG",
max_cg_steps=5,
max_epoch=5,
)
rec_cg.learn()
assert rec_cg.trainer is not None
uvec_cg = rec_cg.compute_user_embedding(X.tocsr())
ivec_cg = rec_cg.compute_item_embedding(X.tocsr())
np.testing.assert_allclose(uvec_chol, uvec_cg, atol=1e-3, rtol=1e-4)
np.testing.assert_allclose(ivec_chol, ivec_cg, atol=1e-3, rtol=1e-4)