├── CML.py ├── LICENSE ├── README.md ├── citeulike-t ├── README.md ├── tag-item.dat ├── tags.dat └── users.dat ├── evaluator.py ├── imgs └── embedding.png ├── requirements.txt ├── sampler.py └── utils.py /CML.py: -------------------------------------------------------------------------------- 1 | import functools 2 | import numpy 3 | import tensorflow as tf 4 | import toolz 5 | from tqdm import tqdm 6 | from evaluator import RecallEvaluator 7 | from sampler import WarpSampler 8 | from utils import citeulike, split_data 9 | 10 | 11 | def doublewrap(function): 12 | """ 13 | A decorator decorator, allowing to use the decorator to be used without 14 | parentheses if not arguments are provided. All arguments must be optional. 15 | """ 16 | 17 | @functools.wraps(function) 18 | def decorator(*args, **kwargs): 19 | if len(args) == 1 and len(kwargs) == 0 and callable(args[0]): 20 | return function(args[0]) 21 | else: 22 | return lambda wrapee: function(wrapee, *args, **kwargs) 23 | 24 | return decorator 25 | 26 | 27 | @doublewrap 28 | def define_scope(function, scope=None, *args, **kwargs): 29 | """ 30 | A decorator for functions that define TensorFlow operations. The wrapped 31 | function will only be executed once. Subsequent calls to it will directly 32 | return the result so that operations are added to the graph only once. 33 | The operations added by the function live within a tf.variable_scope(). If 34 | this decorator is used with arguments, they will be forwarded to the 35 | variable scope. The scope name defaults to the name of the wrapped 36 | function. 37 | """ 38 | attribute = '_cache_' + function.__name__ 39 | name = scope or function.__name__ 40 | 41 | @property 42 | @functools.wraps(function) 43 | def decorator(self): 44 | if not hasattr(self, attribute): 45 | with tf.variable_scope(name, *args, **kwargs): 46 | setattr(self, attribute, function(self)) 47 | return getattr(self, attribute) 48 | 49 | return decorator 50 | 51 | 52 | class CML(object): 53 | def __init__(self, 54 | n_users, 55 | n_items, 56 | embed_dim=20, 57 | features=None, 58 | margin=1.5, 59 | master_learning_rate=0.1, 60 | clip_norm=1.0, 61 | hidden_layer_dim=128, 62 | dropout_rate=0.2, 63 | feature_l2_reg=0.1, 64 | feature_projection_scaling_factor=0.5, 65 | use_rank_weight=True, 66 | use_cov_loss=True, 67 | cov_loss_weight=0.1 68 | ): 69 | """ 70 | 71 | :param n_users: number of users i.e. |U| 72 | :param n_items: number of items i.e. |V| 73 | :param embed_dim: embedding size i.e. K (default 20) 74 | :param features: (optional) the feature vectors of items, shape: (|V|, N_Features). 75 | Set it to None will disable feature loss(default: None) 76 | :param margin: hinge loss threshold i.e. z 77 | :param master_learning_rate: master learning rate for AdaGrad 78 | :param clip_norm: clip norm threshold (default 1.0) 79 | :param hidden_layer_dim: the size of feature projector's hidden layer (default: 128) 80 | :param dropout_rate: the dropout rate between the hidden layer to final feature projection layer 81 | :param feature_l2_reg: feature loss weight 82 | :param feature_projection_scaling_factor: scale the feature projection before compute l2 loss. Ideally, 83 | the scaled feature projection should be mostly within the clip_norm 84 | :param use_rank_weight: whether to use rank weight 85 | :param use_cov_loss: use covariance loss to discourage redundancy in the user/item embedding 86 | """ 87 | 88 | self.n_users = n_users 89 | self.n_items = n_items 90 | self.embed_dim = embed_dim 91 | 92 | self.clip_norm = clip_norm 93 | self.margin = margin 94 | if features is not None: 95 | self.features = tf.constant(features, dtype=tf.float32) 96 | else: 97 | self.features = None 98 | 99 | self.master_learning_rate = master_learning_rate 100 | self.hidden_layer_dim = hidden_layer_dim 101 | self.dropout_rate = dropout_rate 102 | self.feature_l2_reg = feature_l2_reg 103 | self.feature_projection_scaling_factor = feature_projection_scaling_factor 104 | self.use_rank_weight = use_rank_weight 105 | self.use_cov_loss = use_cov_loss 106 | self.cov_loss_weight = cov_loss_weight 107 | 108 | 109 | self.user_positive_items_pairs = tf.placeholder(tf.int32, [None, 2]) 110 | self.negative_samples = tf.placeholder(tf.int32, [None, None]) 111 | self.score_user_ids = tf.placeholder(tf.int32, [None]) 112 | 113 | 114 | self.user_embeddings 115 | self.item_embeddings 116 | self.embedding_loss 117 | self.feature_loss 118 | self.loss 119 | self.optimize 120 | 121 | 122 | @define_scope 123 | def user_embeddings(self): 124 | return tf.Variable(tf.random_normal([self.n_users, self.embed_dim], 125 | stddev=1 / (self.embed_dim ** 0.5), dtype=tf.float32)) 126 | 127 | @define_scope 128 | def item_embeddings(self): 129 | return tf.Variable(tf.random_normal([self.n_items, self.embed_dim], 130 | stddev=1 / (self.embed_dim ** 0.5), dtype=tf.float32)) 131 | 132 | @define_scope 133 | def mlp_layer_1(self): 134 | return tf.layers.dense(inputs=self.features, 135 | units=self.hidden_layer_dim, 136 | activation=tf.nn.relu, name="mlp_layer_1") 137 | 138 | @define_scope 139 | def mlp_layer_2(self): 140 | dropout = tf.layers.dropout(inputs=self.mlp_layer_1, rate=self.dropout_rate) 141 | return tf.layers.dense(inputs=dropout, units=self.embed_dim, name="mlp_layer_2") 142 | 143 | @define_scope 144 | def feature_projection(self): 145 | """ 146 | :return: the projection of the feature vectors to the user-item embedding 147 | """ 148 | 149 | # feature loss 150 | if self.features is not None: 151 | # fully-connected layer 152 | output = self.mlp_layer_2 * self.feature_projection_scaling_factor 153 | 154 | # projection to the embedding 155 | return tf.clip_by_norm(output, self.clip_norm, axes=[1], name="feature_projection") 156 | 157 | @define_scope 158 | def feature_loss(self): 159 | """ 160 | :return: the l2 loss of the distance between items' their embedding and their feature projection 161 | """ 162 | loss = tf.constant(0, dtype=tf.float32) 163 | if self.feature_projection is not None: 164 | # the distance between feature projection and the item's actual location in the embedding 165 | feature_distance = tf.reduce_sum(tf.squared_difference( 166 | self.item_embeddings, 167 | self.feature_projection), 1) 168 | 169 | # apply regularization weight 170 | loss += tf.reduce_sum(feature_distance, name="feature_loss") * self.feature_l2_reg 171 | 172 | return loss 173 | @define_scope 174 | def covariance_loss(self): 175 | 176 | X = tf.concat((self.item_embeddings, self.user_embeddings), 0) 177 | n_rows = tf.cast(tf.shape(X)[0], tf.float32) 178 | X = X - (tf.reduce_mean(X, axis=0)) 179 | cov = tf.matmul(X, X, transpose_a=True) / n_rows 180 | 181 | return tf.reduce_sum(tf.matrix_set_diag(cov, tf.zeros(self.embed_dim, tf.float32))) * self.cov_loss_weight 182 | 183 | @define_scope 184 | def embedding_loss(self): 185 | """ 186 | :return: the distance metric loss 187 | """ 188 | # Let 189 | # N = batch size, 190 | # K = embedding size, 191 | # W = number of negative samples per a user-positive-item pair 192 | 193 | # user embedding (N, K) 194 | users = tf.nn.embedding_lookup(self.user_embeddings, 195 | self.user_positive_items_pairs[:, 0], 196 | name="users") 197 | 198 | # positive item embedding (N, K) 199 | pos_items = tf.nn.embedding_lookup(self.item_embeddings, self.user_positive_items_pairs[:, 1], 200 | name="pos_items") 201 | # positive item to user distance (N) 202 | pos_distances = tf.reduce_sum(tf.squared_difference(users, pos_items), 1, name="pos_distances") 203 | 204 | # negative item embedding (N, K, W) 205 | neg_items = tf.transpose(tf.nn.embedding_lookup(self.item_embeddings, self.negative_samples), 206 | (0, 2, 1), name="neg_items") 207 | # distance to negative items (N x W) 208 | distance_to_neg_items = tf.reduce_sum(tf.squared_difference(tf.expand_dims(users, -1), neg_items), 1, 209 | name="distance_to_neg_items") 210 | 211 | # best negative item (among W negative samples) their distance to the user embedding (N) 212 | closest_negative_item_distances = tf.reduce_min(distance_to_neg_items, 1, name="closest_negative_distances") 213 | 214 | # compute hinge loss (N) 215 | loss_per_pair = tf.maximum(pos_distances - closest_negative_item_distances + self.margin, 0, 216 | name="pair_loss") 217 | 218 | if self.use_rank_weight: 219 | # indicator matrix for impostors (N x W) 220 | impostors = (tf.expand_dims(pos_distances, -1) - distance_to_neg_items + self.margin) > 0 221 | # approximate the rank of positive item by (number of impostor / W per user-positive pair) 222 | rank = tf.reduce_mean(tf.cast(impostors, dtype=tf.float32), 1, name="rank_weight") * self.n_items 223 | # apply rank weight 224 | loss_per_pair *= tf.log(rank + 1) 225 | 226 | # the embedding loss 227 | loss = tf.reduce_sum(loss_per_pair, name="loss") 228 | 229 | return loss 230 | 231 | @define_scope 232 | def loss(self): 233 | """ 234 | :return: the total loss = embedding loss + feature loss 235 | """ 236 | loss = self.embedding_loss + self.feature_loss 237 | if self.use_cov_loss: 238 | loss += self.covariance_loss 239 | return loss 240 | 241 | @define_scope 242 | def clip_by_norm_op(self): 243 | return [tf.assign(self.user_embeddings, tf.clip_by_norm(self.user_embeddings, self.clip_norm, axes=[1])), 244 | tf.assign(self.item_embeddings, tf.clip_by_norm(self.item_embeddings, self.clip_norm, axes=[1]))] 245 | 246 | @define_scope 247 | def optimize(self): 248 | # have two separate learning rates. The first one for user/item embedding is un-normalized. 249 | # The second one for feature projector NN is normalized by the number of items. 250 | gds = [] 251 | gds.append(tf.train 252 | .AdagradOptimizer(self.master_learning_rate) 253 | .minimize(self.loss, var_list=[self.user_embeddings, self.item_embeddings])) 254 | if self.feature_projection is not None: 255 | gds.append(tf.train 256 | .AdagradOptimizer(self.master_learning_rate) 257 | .minimize(self.feature_loss / self.n_items)) 258 | 259 | with tf.control_dependencies(gds): 260 | return gds + [self.clip_by_norm_op] 261 | 262 | @define_scope 263 | def item_scores(self): 264 | # (N_USER_IDS, 1, K) 265 | user = tf.expand_dims(tf.nn.embedding_lookup(self.user_embeddings, self.score_user_ids), 1) 266 | # (1, N_ITEM, K) 267 | item = tf.expand_dims(self.item_embeddings, 0) 268 | # score = minus distance (N_USER, N_ITEM) 269 | return -tf.reduce_sum(tf.squared_difference(user, item), 2, name="scores") 270 | 271 | 272 | BATCH_SIZE = 50000 273 | N_NEGATIVE = 20 274 | EVALUATION_EVERY_N_BATCHES = 30 275 | EMBED_DIM = 100 276 | 277 | 278 | def optimize(model, sampler, train, valid): 279 | """ 280 | Optimize the model. TODO: implement early-stopping 281 | :param model: model to optimize 282 | :param sampler: mini-batch sampler 283 | :param train: train user-item matrix 284 | :param valid: validation user-item matrix 285 | :return: None 286 | """ 287 | sess = tf.Session() 288 | sess.run(tf.global_variables_initializer()) 289 | if model.feature_projection is not None: 290 | # initialize item embedding with feature projection 291 | sess.run(tf.assign(model.item_embeddings, model.feature_projection)) 292 | 293 | # sample some users to calculate recall validation 294 | valid_users = numpy.random.choice(list(set(valid.nonzero()[0])), size=1000, replace=False) 295 | 296 | while True: 297 | # create evaluator on validation set 298 | validation_recall = RecallEvaluator(model, train, valid) 299 | # compute recall on validate set 300 | valid_recalls = [] 301 | 302 | # compute recall in chunks to utilize speedup provided by Tensorflow 303 | for user_chunk in toolz.partition_all(100, valid_users): 304 | valid_recalls.extend([validation_recall.eval(sess, user_chunk)]) 305 | print("\nRecall on (sampled) validation set: {}".format(numpy.mean(valid_recalls))) 306 | # TODO: early stopping based on validation recall 307 | 308 | # train model 309 | losses = [] 310 | # run n mini-batches 311 | for _ in tqdm(range(EVALUATION_EVERY_N_BATCHES), desc="Optimizing..."): 312 | user_pos, neg = sampler.next_batch() 313 | _, loss = sess.run((model.optimize, model.loss), 314 | {model.user_positive_items_pairs: user_pos, 315 | model.negative_samples: neg}) 316 | 317 | losses.append(loss) 318 | 319 | print("\nTraining loss {}".format(numpy.mean(losses))) 320 | 321 | 322 | if __name__ == '__main__': 323 | # get user-item matrix 324 | user_item_matrix, features = citeulike(tag_occurence_thres=5) 325 | n_users, n_items = user_item_matrix.shape 326 | # make feature as dense matrix 327 | dense_features = features.toarray() + 1E-10 328 | # get train/valid/test user-item matrices 329 | train, valid, test = split_data(user_item_matrix) 330 | # create warp sampler 331 | sampler = WarpSampler(train, batch_size=BATCH_SIZE, n_negative=N_NEGATIVE, check_negative=True) 332 | 333 | # WITHOUT features 334 | # Train a user-item joint embedding, where the items a user likes will be pulled closer to this users. 335 | # Once the embedding is trained, the recommendations are made by finding the k-Nearest-Neighbor to each user. 336 | model = CML(n_users, 337 | n_items, 338 | # set features to None to disable feature projection 339 | features=None, 340 | # size of embedding 341 | embed_dim=EMBED_DIM, 342 | # the size of hinge loss margin. 343 | margin=1.9, 344 | # clip the embedding so that their norm <= clip_norm 345 | clip_norm=1, 346 | # learning rate for AdaGrad 347 | master_learning_rate=0.1, 348 | 349 | # whether to enable rank weight. If True, the loss will be scaled by the estimated 350 | # log-rank of the positive items. If False, no weight will be applied. 351 | 352 | # This is particularly useful to speed up the training for large item set. 353 | 354 | # Weston, Jason, Samy Bengio, and Nicolas Usunier. 355 | # "Wsabie: Scaling up to large vocabulary image annotation." IJCAI. Vol. 11. 2011. 356 | use_rank_weight=True, 357 | 358 | # whether to enable covariance regularization to encourage efficient use of the vector space. 359 | # More useful when the size of embedding is smaller (e.g. < 20 ). 360 | use_cov_loss=False, 361 | 362 | # weight of the cov_loss 363 | cov_loss_weight=1 364 | ) 365 | 366 | #optimize(model, sampler, train, valid) 367 | 368 | # WITH features 369 | # In this case, we additionally train a feature projector to project raw item features into the 370 | # embedding. The projection serves as "a prior" to inform the item's potential location in the embedding. 371 | # We use a two fully-connected layers NN as our feature projector. (This model is much more computation intensive. 372 | # A GPU machine is recommended) 373 | model = CML(n_users, 374 | n_items, 375 | # enable feature projection 376 | features=dense_features, 377 | embed_dim=EMBED_DIM, 378 | margin=2.0, 379 | clip_norm=1.1, 380 | master_learning_rate=0.1, 381 | # the size of the hidden layer in the feature projector NN 382 | hidden_layer_dim=512, 383 | # dropout rate between hidden layer and output layer in the feature projector NN 384 | dropout_rate=0.3, 385 | # scale the output of the NN so that the magnitude of the NN output is closer to the item embedding 386 | feature_projection_scaling_factor=1, 387 | # the penalty to the distance between projection and item's actual location in the embedding 388 | # tune this to adjust how much the embedding should be biased towards the item features. 389 | feature_l2_reg=0.1, 390 | 391 | # whether to enable rank weight. If True, the loss will be scaled by the estimated 392 | # log-rank of the positive items. If False, no weight will be applied. 393 | 394 | # This is particularly useful to speed up the training for large item set. 395 | 396 | # Weston, Jason, Samy Bengio, and Nicolas Usunier. 397 | # "Wsabie: Scaling up to large vocabulary image annotation." IJCAI. Vol. 11. 2011. 398 | use_rank_weight=True, 399 | 400 | # whether to enable covariance regularization to encourage efficient use of the vector space. 401 | # More useful when the size of embedding is smaller (e.g. < 20 ). 402 | use_cov_loss=False, 403 | 404 | # weight of the cov_loss 405 | cov_loss_weight=1 406 | ) 407 | optimize(model, sampler, train, valid) 408 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | {one line to give the program's name and a brief idea of what it does.} 635 | Copyright (C) {year} {name of author} 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | {project} Copyright (C) {year} {fullname} 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # CollMetric 2 | 3 | A Tensorflow implementation of Collaborative Metric Learning (CML): 4 | 5 | *Cheng-Kang Hsieh, Longqi Yang, Yin Cui, Tsung-Yi Lin, Serge Belongie, and Deborah Estrin. 2017. Collaborative Metric Learning. In Proceedings of the 26th International Conference on World Wide Web (WWW '17) ([perm_link](http://dl.acm.org/citation.cfm?id=3052639), [pdf](http://www.cs.cornell.edu/~ylongqi/paper/HsiehYCLBE17.pdf))* 6 | 7 | ** Note: the original Theano implementation is deprecated and is kept in the *old_experiment_code branch* 8 | # Features 9 | * Produces embedding that accurately captures the user-item, user-user, and item-item similarity. 10 | * Allows the exploitation of item features (e.g. tags, text, image features). 11 | * Outperforms state-of-the-art recommendation algorithms on a wide range of tasks 12 | * Enjoys an extremely efficient Top-K search using Fast KNN algorithms. 13 | 14 | # Utility Features 15 | * Parallel negative sampler that can sample the user-item pairs when the model is being trained on GPU 16 | * Fast recall evaluation based on Tensorflow 17 | 18 | # Requirements 19 | * python3 20 | * tensorflow 21 | * scipy 22 | * scikit-learn 23 | 24 | # Usage 25 | ```bash 26 | # install requirements 27 | pip3 install -r requirements.txt 28 | # run demo tensorflow model 29 | python3 CML.py 30 | ``` 31 | 32 | # Known Issue 33 | * AdaGrad does not seem to work on GPU. Try using AdamOptimizer instead 34 | * ~~the WithFeature version does not seems to perform as well as the Theano version. It is being investigated.~~ (The performance is actually slightly better (with AdamOptimizer) than the number reported in the paper now!) 35 | 36 | # Visuals 37 | ### An illustration of embbeding learning procedue of CML 38 | ![CML](http://portalparts.acm.org/3060000/3052639/core/fp0554.jpg) 39 | ### Flickr photo recommendation embedding produced by CML (compared to original ImageNet features) 40 | ![Embedding](https://github.com/changun/CollMetric/blob/master/imgs/embedding.png?raw=true) 41 | # TODO 42 | * Model Comparison. 43 | * TensorBoard visualization 44 | -------------------------------------------------------------------------------- /citeulike-t/README.md: -------------------------------------------------------------------------------- 1 | ### The following text is from http://www.wanghao.in/CDL.htm ### 2 | This dataset, *citeulike-t*, was used in the paper **[Collaborative Topic Regression with Social Regularization](https://www.cs.ucsb.edu/~binyichen/IJCAI13-400.pdf)** 3 | [Wang, Chen and Li]. 4 | It was collected from CiteULike and Google Scholar. 5 | CiteULike allows users to create their own collections of articles. 6 | There are abstracts, titles, and tags for each article. 7 | It is collected by us independently from the dataset citeulike-a. 8 | We manually select 273 seed tags and collect all the articles with at least one of these tags. 9 | We also crawl the citations between the articles from Google Scholar. 10 | Note that the final number of tags associated with all the collected articles is far more than the number (273) of seed tags. 11 | 12 | 13 | The text information (item content) of citeulike-a is preprocessed by following the same procedure as that in citeulike-a. After removing the stop words, we choose the top 20000 discriminative words according to the tf-idf values as our vocabulary. 14 | 15 | Some statistics are listed as follows: 16 | 17 | * users 7947 18 | * items 25975 19 | * tags 52946 20 | * citations 32565 21 | * user-item pairs 134860 22 | 23 | ## DATA FILES ## 24 | * citations.dat citations between articles 25 | * tag-item.dat articles corresponding to tags, one line corresponds to articles relating to the same tags (note that it is different from the other dataset citeulike-a and that this is the version prior to preprocess thus would have more tags than used in the paper) 26 | * mult.dat bag of words for each article 27 | * rawtext.dat raw data 28 | * tags.dat tags, sorted by tag-id's 29 | * users.dat rating matrix (user-item matrix) 30 | * vocabulary.dat corresponding words for file mult.dat 31 | 32 | ```BibTex 33 | @inproceedings{DBLP:conf/ijcai/WangCL13, 34 | author = {Hao Wang and 35 | Binyi Chen and 36 | Wu-Jun Li}, 37 | title = {Collaborative Topic Regression with Social Regularization 38 | for Tag Recommendation}, 39 | booktitle = {IJCAI}, 40 | year = {2013}, 41 | ee = {http://www.aaai.org/ocs/index.php/IJCAI/IJCAI13/paper/view/7006}, 42 | crossref = {DBLP:conf/ijcai/2013}, 43 | bibsource = {DBLP, http://dblp.uni-trier.de} 44 | } 45 | ``` 46 | ``` 47 | @proceedings{DBLP:conf/ijcai/2013, 48 | editor = {Francesca Rossi}, 49 | title = {IJCAI 2013, Proceedings of the 23rd International Joint 50 | Conference on Artificial Intelligence, Beijing, China, August 51 | 3-9, 2013}, 52 | booktitle = {IJCAI}, 53 | publisher = {IJCAI/AAAI}, 54 | year = {2013}, 55 | isbn = {978-1-57735-633-2}, 56 | bibsource = {DBLP, http://dblp.uni-trier.de} 57 | } 58 | ``` 59 | -------------------------------------------------------------------------------- /evaluator.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | from scipy.sparse import lil_matrix 3 | 4 | 5 | class RecallEvaluator(object): 6 | def __init__(self, model, train_user_item_matrix, test_user_item_matrix): 7 | """ 8 | Create a evaluator for recall@K evaluation 9 | :param model: the model we are going to evaluate 10 | :param train_user_item_matrix: the user-item pairs used in the training set. These pairs will be ignored 11 | in the recall calculation 12 | :param test_user_item_matrix: the held-out user-item pairs we make prediction against 13 | """ 14 | self.model = model 15 | self.train_user_item_matrix = lil_matrix(train_user_item_matrix) 16 | self.test_user_item_matrix = lil_matrix(test_user_item_matrix) 17 | n_users = train_user_item_matrix.shape[0] 18 | self.user_to_test_set = {u: set(self.test_user_item_matrix.rows[u]) 19 | for u in range(n_users) if self.test_user_item_matrix.rows[u]} 20 | 21 | if self.train_user_item_matrix is not None: 22 | self.user_to_train_set = {u: set(self.train_user_item_matrix.rows[u]) 23 | for u in range(n_users) if self.train_user_item_matrix.rows[u]} 24 | self.max_train_count = max(len(row) for row in self.train_user_item_matrix.rows) 25 | else: 26 | self.max_train_count = 0 27 | 28 | def eval(self, sess, users, k=50): 29 | """ 30 | Compute the Top-K recall for a particular user given the predicted scores to items 31 | :param users: the users to eval the recall 32 | :param k: compute the recall for the top K items 33 | :return: recall@K 34 | """ 35 | # compute the top (K + Max Number Of Training Items for any user) items for each user 36 | 37 | _, user_tops = sess.run(tf.nn.top_k(self.model.item_scores, k + self.max_train_count), 38 | {self.model.score_user_ids: users}) 39 | recalls = [] 40 | for user_id, tops in zip(users, user_tops): 41 | train_set = self.user_to_train_set.get(user_id, set()) 42 | test_set = self.user_to_test_set.get(user_id, set()) 43 | top_n_items = 0 44 | hits = 0 45 | for i in tops: 46 | # ignore item in the training set 47 | if i in train_set: 48 | continue 49 | elif i in test_set: 50 | hits += 1 51 | top_n_items += 1 52 | if top_n_items == k: 53 | break 54 | recalls.append(hits / float(len(test_set))) 55 | return recalls -------------------------------------------------------------------------------- /imgs/embedding.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/changun/CollMetric/d9026cfce7c6e8dd2640b842ad524b61031b29ac/imgs/embedding.png -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | scipy 2 | numpy 3 | tensorflow 4 | tqdm 5 | toolz 6 | scikit-learn 7 | -------------------------------------------------------------------------------- /sampler.py: -------------------------------------------------------------------------------- 1 | import numpy 2 | from multiprocessing import Process, Queue 3 | from scipy.sparse import lil_matrix 4 | 5 | 6 | def sample_function(user_item_matrix, batch_size, n_negative, result_queue, check_negative=True): 7 | """ 8 | 9 | :param user_item_matrix: the user-item matrix for positive user-item pairs 10 | :param batch_size: number of samples to return 11 | :param n_negative: number of negative samples per user-positive-item pair 12 | :param result_queue: the output queue 13 | :return: None 14 | """ 15 | user_item_matrix = lil_matrix(user_item_matrix) 16 | user_item_pairs = numpy.asarray(user_item_matrix.nonzero()).T 17 | user_to_positive_set = {u: set(row) for u, row in enumerate(user_item_matrix.rows)} 18 | while True: 19 | numpy.random.shuffle(user_item_pairs) 20 | for i in range(int(len(user_item_pairs) / batch_size)): 21 | 22 | user_positive_items_pairs = user_item_pairs[i * batch_size: (i + 1) * batch_size, :] 23 | 24 | # sample negative samples 25 | negative_samples = numpy.random.randint( 26 | 0, 27 | user_item_matrix.shape[1], 28 | size=(batch_size, n_negative)) 29 | 30 | # Check if we sample any positive items as negative samples. 31 | # Note: this step can be optional as the chance that we sample a positive item is fairly low given a 32 | # large item set. 33 | if check_negative: 34 | for user_positive, negatives, i in zip(user_positive_items_pairs, 35 | negative_samples, 36 | range(len(negative_samples))): 37 | user = user_positive[0] 38 | for j, neg in enumerate(negatives): 39 | while neg in user_to_positive_set[user]: 40 | negative_samples[i, j] = neg = numpy.random.randint(0, user_item_matrix.shape[1]) 41 | result_queue.put((user_positive_items_pairs, negative_samples)) 42 | 43 | 44 | class WarpSampler(object): 45 | """ 46 | A generator that, in parallel, generates tuples: user-positive-item pairs, negative-items 47 | 48 | of the shapes (Batch Size, 2) and (Batch Size, N_Negative) 49 | """ 50 | 51 | def __init__(self, user_item_matrix, batch_size=10000, n_negative=10, n_workers=5, check_negative=True): 52 | self.result_queue = Queue(maxsize=n_workers*2) 53 | self.processors = [] 54 | for i in range(n_workers): 55 | self.processors.append( 56 | Process(target=sample_function, args=(user_item_matrix, 57 | batch_size, 58 | n_negative, 59 | self.result_queue, 60 | check_negative))) 61 | self.processors[-1].start() 62 | 63 | def next_batch(self): 64 | return self.result_queue.get() 65 | 66 | def close(self): 67 | for p in self.processors: # type: Process 68 | p.terminate() 69 | p.join() 70 | -------------------------------------------------------------------------------- /utils.py: -------------------------------------------------------------------------------- 1 | from collections import defaultdict 2 | 3 | import numpy as np 4 | from scipy.sparse import dok_matrix, lil_matrix 5 | from tqdm import tqdm 6 | 7 | 8 | def citeulike(tag_occurence_thres=10): 9 | user_dict = defaultdict(set) 10 | for u, item_list in enumerate(open("citeulike-t/users.dat").readlines()): 11 | items = item_list.strip().split(" ") 12 | # ignore the first element in each line, which is the number of items the user liked. 13 | for item in items[1:]: 14 | user_dict[u].add(int(item)) 15 | 16 | n_users = len(user_dict) 17 | n_items = max([item for items in user_dict.values() for item in items]) + 1 18 | 19 | user_item_matrix = dok_matrix((n_users, n_items), dtype=np.int32) 20 | for u, item_list in enumerate(open("citeulike-t/users.dat").readlines()): 21 | items = item_list.strip().split(" ") 22 | # ignore the first element in each line, which is the number of items the user liked. 23 | for item in items[1:]: 24 | user_item_matrix[u, int(item)] = 1 25 | 26 | n_features = 0 27 | for l in open("citeulike-t/tag-item.dat").readlines(): 28 | items = l.strip().split(" ") 29 | if len(items) >= tag_occurence_thres: 30 | n_features += 1 31 | print("{} features over tag_occurence_thres ({})".format(n_features, tag_occurence_thres)) 32 | features = dok_matrix((n_items, n_features), dtype=np.int32) 33 | feature_index = 0 34 | for l in open("citeulike-t/tag-item.dat").readlines(): 35 | items = l.strip().split(" ") 36 | if len(items) >= tag_occurence_thres: 37 | features[[int(i) for i in items], feature_index] = 1 38 | feature_index += 1 39 | 40 | return user_item_matrix, features 41 | 42 | 43 | def split_data(user_item_matrix, split_ratio=(3, 1, 1), seed=1): 44 | # set the seed to have deterministic results 45 | np.random.seed(seed) 46 | train = dok_matrix(user_item_matrix.shape) 47 | validation = dok_matrix(user_item_matrix.shape) 48 | test = dok_matrix(user_item_matrix.shape) 49 | # convert it to lil format for fast row access 50 | user_item_matrix = lil_matrix(user_item_matrix) 51 | for user in tqdm(range(user_item_matrix.shape[0]), desc="Split data into train/valid/test"): 52 | items = list(user_item_matrix.rows[user]) 53 | if len(items) >= 5: 54 | 55 | np.random.shuffle(items) 56 | 57 | train_count = int(len(items) * split_ratio[0] / sum(split_ratio)) 58 | valid_count = int(len(items) * split_ratio[1] / sum(split_ratio)) 59 | 60 | for i in items[0: train_count]: 61 | train[user, i] = 1 62 | for i in items[train_count: train_count + valid_count]: 63 | validation[user, i] = 1 64 | for i in items[train_count + valid_count:]: 65 | test[user, i] = 1 66 | print("{}/{}/{} train/valid/test samples".format( 67 | len(train.nonzero()[0]), 68 | len(validation.nonzero()[0]), 69 | len(test.nonzero()[0]))) 70 | return train, validation, test 71 | --------------------------------------------------------------------------------