def calculate_user_cf_recommendations(sparse_user_job,
reduced_sparse_user_job,
user_index,
n=10):
'''
Input:
user_index: the index of the user to calculate the recommendations for in the user-job matrix
Adds the first three applications made by the user as a new user to sparse matrix and generates recommendations with the user-cf algorithm.
Returns the recall and atLeastOne accuracy metrics calculated based on the remaining applications made by this user.
'''
alpha = 40
row = sparse_user_job[np.array([user_index]), :]
(_, nonzero_columns) = row.nonzero()
training_columns = nonzero_columns[:3]
test_columns = nonzero_columns[3:]
#ensures user_job_mat is the correct shape
reduced_sparse_user_job = reduced_sparse_user_job[:7812, :]
n_users, n_jobs = reduced_sparse_user_job.shape
train_ratings = [alpha for i in range(3)]
user_job_new_mat = reduced_sparse_user_job
user_job_new_mat.data = np.hstack(
(reduced_sparse_user_job.data, train_ratings))
user_job_new_mat.indices = np.hstack(
(reduced_sparse_user_job.indices, training_columns))
user_job_new_mat.indptr = np.hstack(
(reduced_sparse_user_job.indptr, len(reduced_sparse_user_job.data)))
user_job_new_mat._shape = (n_users + 1, n_jobs)
model = NearestNeighbors(metric='cosine', algorithm='brute')
model.fit(user_job_new_mat)
distances, indices = model.kneighbors(user_job_new_mat[n_users],
n_neighbors=n + 1)
raw_recommends = sorted(list(
zip(indices.squeeze().tolist(),
distances.squeeze().tolist())),
key=lambda x: x[1])[:0:-1]
recommended_jobs_indices = []
for _, (idx, _) in enumerate(raw_recommends):
user_row = reduced_sparse_user_job.getrow(idx)
(_, nonzero_columns) = user_row.nonzero()
recommended_jobs_indices = recommended_jobs_indices + nonzero_columns.tolist(
)
unique_recommendations = []
for x in recommended_jobs_indices:
if x not in unique_recommendations:
unique_recommendations.append(x)
final_recommendations = unique_recommendations[:n]
return recall(final_recommendations,
test_columns), at_least_one_metric(final_recommendations,
test_columns)
def calculate_job_cf_recommendations(sparse_job_user,
reduced_sparse_job_user,
user_index,
n=10):
'''
Input:
user_index: the index of the user to calculate the recommendations for in the user-job matrix
Adds the first three applications made by the user as a new user to sparse matrix and generates recommendations with the job-cf algorithm.
Returns the recall and atLeastOne accuracy metrics calculated based on the remaining applications made by this user.
'''
column = sparse_job_user[:, np.array([user_index])]
(nonzero_rows, _) = column.nonzero()
training_rows = nonzero_rows[:3]
test_rows = nonzero_rows[3:]
model = NearestNeighbors(metric='cosine', algorithm='brute')
model.fit(reduced_sparse_job_user)
#list of the lists of recommendations from each job in selected_jobs
possible_recommendations = []
for job_index in training_rows:
distances, indices = model.kneighbors(
reduced_sparse_job_user[job_index], n_neighbors=n + 1)
raw_recommends = sorted(list(
zip(indices.squeeze().tolist(),
distances.squeeze().tolist())),
key=lambda x: x[1])[:0:-1]
similar_jobs = []
for _, (idx, _) in enumerate(raw_recommends):
similar_jobs = similar_jobs + [idx]
possible_recommendations = possible_recommendations + [
(job_index, similar_jobs)
]
unique_recommendations = []
for i in range(10):
for (job_index, job_list) in possible_recommendations:
job = job_list[i]
if job not in unique_recommendations:
unique_recommendations.append(job)
final_recommendations = unique_recommendations[:n]
return recall(final_recommendations,
test_rows), at_least_one_metric(final_recommendations,
test_rows)
def calculate_mf_recommendations(mf_model,
sparse_user_job,
reduced_sparse_user_job,
user_index,
n=10):
'''
Input:
user_index: the index of the user to calculate the recommendations for in the user-job matrix
Adds the first three applications made by the user as a new user to sparse matrix and generates recommendations with the ALS algorithm.
Returns the recall and atLeastOne accuracy metrics calculated based on the remaining applications made by this user.
'''
alpha = 40
row = sparse_user_job[np.array([user_index]), :]
(_, nonzero_columns) = row.nonzero()
training_columns = nonzero_columns[:3]
test_columns = nonzero_columns[3:]
#ensures user_job_mat is the correct shape
reduced_sparse_user_job = reduced_sparse_user_job[:7812, :]
n_users, n_jobs = reduced_sparse_user_job.shape
train_ratings = [alpha for i in range(3)]
new_sparse_user_job = reduced_sparse_user_job
new_sparse_user_job.data = np.hstack(
(reduced_sparse_user_job.data, train_ratings))
new_sparse_user_job.indices = np.hstack(
(reduced_sparse_user_job.indices, training_columns))
new_sparse_user_job.indptr = np.hstack(
(reduced_sparse_user_job.indptr, len(reduced_sparse_user_job.data)))
new_sparse_user_job._shape = (n_users + 1, n_jobs)
recommended_index, _ = zip(*mf_model.recommend(
n_users, new_sparse_user_job, N=n, recalculate_user=True))
return recall(recommended_index,
test_columns), at_least_one_metric(recommended_index,
test_columns)
def main(_):
with tf.device('/gpu:0'):
# Input
x = tf.placeholder(tf.float32,
[None, model.time_step, model.num_input])
y_ = tf.placeholder(tf.float32, [None, model.num_class])
# Create lstm model
y_lstm, keep_prob = cross.lstm(x)
# Define loss
with tf.name_scope('loss'):
cross_entropy = tf.nn.softmax_cross_entropy_with_logits_v2(
labels=y_, logits=y_lstm)
cross_entropy = tf.reduce_mean(cross_entropy)
# Define optimizer
with tf.name_scope('adam_optimizer'):
train_step = tf.train.AdamOptimizer(learn_rate).minimize(
cross_entropy)
# Create the node to calculate ccc
with tf.name_scope('accuracy'):
correct_prediction = tf.equal(tf.argmax(y_lstm, 1),
tf.argmax(y_, 1))
correct_prediction = tf.cast(correct_prediction, tf.float32)
accuracy = tf.reduce_mean(correct_prediction)
# Create the node to calculate auc
with tf.name_scope('auc'):
labels = tf.reshape(
tf.slice(tf.cast(y_, dtype=tf.bool), [0, 0], [-1, 1]), [-1])
predictions = tf.reshape(
tf.subtract(tf.slice(y_lstm, [0, 0], [-1, 1]),
tf.slice(y_lstm, [0, 1], [-1, 1])), [-1])
# Min Max Normalization
Y_pred = (predictions - tf.reduce_min(predictions)) / (
tf.reduce_max(predictions) - tf.reduce_min(predictions))
roc_auc, roc_auc_update_op = tf.metrics.auc(labels,
Y_pred,
curve='ROC',
name='roc')
# Create the node to calculate acc
with tf.name_scope('metrics'):
acc, acc_op = tf.metrics.accuracy(tf.argmax(y_, 1),
tf.argmax(y_lstm, 1))
rec, rec_op = tf.metrics.recall(tf.argmax(y_, 1),
tf.argmax(y_lstm, 1))
all_pos = tf.reduce_sum(tf.argmin(y_lstm, 1))
all_neg = tf.reduce_sum(tf.argmax(y_lstm, 1))
fn, fn_op = tf.metrics.false_negatives(tf.argmax(y_, 1),
tf.argmax(y_lstm, 1))
fp, fp_op = tf.metrics.false_positives(tf.argmax(y_, 1),
tf.argmax(y_lstm, 1))
# Add ops to save and restore all the variables
saver = tf.train.Saver()
sess = tf.InteractiveSession()
with tf.Session(config=tf.ConfigProto(
allow_soft_placement=True)) as sess:
for seed in range(1, seeds_num + 1):
print('*' * 30, 'seed=', seed, '*' * 30)
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
summary_writer = tf.summary.FileWriter(
LOGS_DIRECTORY, graph=tf.get_default_graph())
sum_acc = 0
sum_auc = 0
sum_spec = 0
sum_recall = 0
record_fn = 0
record_fp = 0
training_accuracy_list = []
all_piRNA = input_data.read_all(TRAIN_IMAGES,
TRAIN_LABELS,
test_size=test_size,
seed=seed,
is_display=is_display)
test_accuracy_list = []
for fold in range(10):
print('fold %d:' % fold)
piRNA = input_data.read_CV_datasets(
fold, int(DATA_NUM * (1 - test_size)), all_piRNA)
for i in range(TOTAL_BATCH):
batch_x, batch_y = piRNA.train.next_batch(batch_size)
batch_x = batch_x.reshape(batch_size, model.time_step,
model.num_input)
step, training_accuracy = sess.run(
[train_step, accuracy],
feed_dict={
x: batch_x,
y_: batch_y,
keep_prob: 0.5
})
# print out results
if i % 50 == 0:
print('step %d, training accuracy %g' %
(i, training_accuracy))
training_accuracy_list.append(training_accuracy)
if i % 1000 == 0:
print('test accuracy %g' % accuracy.eval(
feed_dict={
x:
piRNA.test.images.reshape(
-1, model.time_step, model.num_input),
y_:
piRNA.test.labels,
keep_prob:
1.0
}))
auc, acc, recall, pred_neg, false_nega, false_posi, pred_pos = sess.run(
[
roc_auc_update_op, acc_op, rec_op, all_neg, fn_op,
fp_op, all_pos
],
feed_dict={
x:
piRNA.validation.images.reshape(
-1, model.time_step, model.num_input),
y_:
piRNA.validation.labels,
keep_prob:
1.0
})
# update specificity
current_fn = false_nega - record_fn
current_fp = false_posi - record_fp
true_nega = pred_neg - current_fn # fp_op accumulate every loop
spec = true_nega / (true_nega + current_fp)
record_fn = false_nega
record_fp = false_posi
test_accuracy = accuracy.eval(
feed_dict={
x:
piRNA.test.images.reshape(-1, model.time_step,
model.num_input),
y_:
piRNA.test.labels,
keep_prob:
1.0
})
test_accuracy_list.append(test_accuracy)
# Test Set
print('Test set accuracy %g' % test_accuracy)
# 10-CV metrices (acc, auc)
sum_acc = cv.acc(sum_acc, acc, fold, is_display=is_display)
sum_auc = cv.auc(sum_auc, auc, fold, is_display=is_display)
sum_spec = cv.spec(sum_spec,
spec,
fold,
is_display=is_display)
sum_recall = cv.recall(sum_recall,
recall,
fold,
is_display=is_display)
test_accuracy_average = cv.average(test_accuracy_list)
auc_average = cv.average(cv.auc_list)
acc_average = cv.average(cv.acc_list)
spec_average = cv.average(cv.spec_list)
recall_average = cv.average(cv.recall_list)
acc_list.append(acc_average)
auc_list.append(auc_average)
spec_list.append(spec_average)
recall_list.append(recall_average)
test_acc_list.append(test_accuracy_average)
if is_display:
print('*** Test accuracy is:', test_accuracy_list)
print('*** The average test accuracy is:%.3f' %
test_accuracy_average)
print('acc', acc_average)
print('auc', auc_average)
print('spec', spec_average)
print('recall', recall_average)
data_frame = pd.DataFrame({
'AUC': auc_list,
'ACC': acc_list,
'SP': spec_list,
'SN': recall_list,
'Test ACC': test_acc_list
})
data_frame.to_csv('human1vs3cross.csv',
index=True,
columns=['AUC', 'ACC', 'SP', 'SN', 'Test ACC'])