Add TriRank model (#525)

* Add TriRank model

* Add max_iter in online training

README.md

@@ -138,6 +138,7 @@ The recommender models supported by Cornac are listed below. Why don't you join
 |      | [Visual Bayesian Personalized Ranking (VBPR)](cornac/models/vbpr), [paper](https://arxiv.org/pdf/1510.01784.pdf) | [requirements.txt](cornac/models/vbpr/requirements.txt) | [vbpr_tradesy.py](examples/vbpr_tradesy.py)
 | 2015 | [Collaborative Deep Learning (CDL)](cornac/models/cdl), [paper](https://arxiv.org/pdf/1409.2944.pdf) | [requirements.txt](cornac/models/cdl/requirements.txt) | [cdl_exp.py](examples/cdl_example.py)
 |      | [Hierarchical Poisson Factorization (HPF)](cornac/models/hpf), [paper](http://jakehofman.com/inprint/poisson_recs.pdf) | N/A | [hpf_movielens.py](examples/hpf_movielens.py)
+|      | [TriRank: Review-aware Explainable Recommendation by Modeling Aspects](cornac/models/trirank), [paper](https://wing.comp.nus.edu.sg/wp-content/uploads/Publications/PDF/TriRank-%20Review-aware%20Explainable%20Recommendation%20by%20Modeling%20Aspects.pdf) | N/A | [trirank_example.py](examples/trirank_example.py)
 | 2014 | [Explicit Factor Model (EFM)](cornac/models/efm), [paper](http://yongfeng.me/attach/efm-zhang.pdf) | N/A | [efm_exp.py](examples/efm_example.py)
 |      | [Social Bayesian Personalized Ranking (SBPR)](cornac/models/sbpr), [paper](https://cseweb.ucsd.edu/~jmcauley/pdfs/cikm14.pdf) | N/A | [sbpr_epinions.py](examples/sbpr_epinions.py)
 | 2013 | [Hidden Factors and Hidden Topics (HFT)](cornac/models/hft), [paper](https://cs.stanford.edu/people/jure/pubs/reviews-recsys13.pdf) | N/A | [hft_exp.py](examples/hft_example.py)

cornac/models/__init__.py

@@ -58,6 +58,7 @@
 from .skm import SKMeans
 from .sorec import SoRec
 from .svd import SVD
+from .trirank import TriRank
 from .vaecf import VAECF
 from .vbpr import VBPR
 from .vmf import VMF

cornac/models/trirank/__init__.py

@@ -0,0 +1 @@
+from .recom_trirank import TriRank

cornac/models/trirank/recom_trirank.py

@@ -0,0 +1,332 @@
+# Copyright 2018 The Cornac Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+import numpy as np
+from scipy.sparse import csr_matrix
+from tqdm.auto import tqdm
+
+from ..recommender import Recommender
+from ...utils import get_rng
+from ...utils.init_utils import uniform
+from ...exception import ScoreException
+
+
+EPS = 1e-10
+
+
+class TriRank(Recommender):
+    """TriRank: Review-aware Explainable Recommendation by Modeling Aspects.
+
+    Parameters
+    ----------
+    name: string, optional, default: 'TriRank'
+        The name of the recommender model.
+
+    alpha: float, optional, default: 1
+        The weight of smoothness on user-item relation
+
+    beta: float, optional, default: 1
+        The weight of smoothness on item-aspect relation
+
+    gamma: float, optional, default: 1
+        The weight of smoothness on user-aspect relation
+
+    eta_U: float, optional, default: 1
+        The weight of fitting constraint on users
+
+    eta_P: float, optional, default: 1
+        The weight of fitting constraint on items
+
+    eta_A: float, optional, default: 1
+        The weight of fitting constraint on aspects
+
+    max_iter: int, optional, default: 100
+        Maximum number of iterations to stop online training. If set to `max_iter=-1`, \
+        the online training will stop when model parameters are converged.
+
+    trainable: boolean, optional, default: True
+        When False, the model is not trained and Cornac assumes that the model already \
+        pre-trained (R, X, Y, p, a, u are not None).
+
+    verbose: boolean, optional, default: False
+        When True, running logs are displayed.
+
+    init_params: dictionary, optional, default: None
+        List of initial parameters, e.g., init_params = {'R':R, 'X':X, 'Y':Y, 'p':p, 'a':a, 'u':u}
+
+        R: csr_matrix, shape (n_users, n_items)
+            The symmetric normalized of edge weight matrix of user-item relation, optional initialization via init_params
+
+        X: csr_matrix, shape (n_users, n_aspects)
+            The symmetric normalized of edge weight matrix of user-aspect relation, optional initialization via init_params
+
+        Y: csr_matrix, shape (n_items, n_aspects)
+            The symmetric normalized of edge weight matrix of item-aspect relation, optional initialization via init_params
+
+        p: ndarray, shape (n_items,)
+            Initialized item weights, optional initialization via init_params
+
+        a: ndarray, shape (n_aspects,)
+            Initialized aspect weights, optional initialization via init_params
+
+        u: ndarray, shape (n_aspects,)
+            Initialized user weights, optional initialization via init_params
+
+    seed: int, optional, default: None
+        Random seed for parameters initialization.
+
+    References
+    ----------
+    He, Xiangnan, Tao Chen, Min-Yen Kan, and Xiao Chen. 2014. \
+    TriRank: Review-aware Explainable Recommendation by Modeling Aspects. \
+    In the 24th ACM international on conference on information and knowledge management (CIKM'15). \
+    ACM, New York, NY, USA, 1661-1670. DOI: https://doi.org/10.1145/2806416.2806504
+    """
+
+    def __init__(
+        self,
+        name="TriRank",
+        alpha=1,
+        beta=1,
+        gamma=1,
+        eta_U=1,
+        eta_P=1,
+        eta_A=1,
+        max_iter=100,
+        verbose=True,
+        init_params=None,
+        seed=None,
+    ):
+        super().__init__(name)
+        self.alpha = alpha
+        self.beta = beta
+        self.gamma = gamma
+        self.eta_U = eta_U
+        self.eta_P = eta_P
+        self.eta_A = eta_A
+        self.max_iter = max_iter
+        self.verbose = verbose
+        self.seed = seed
+        self.rng = get_rng(seed)
+
+        # Init params if provided
+        self.init_params = {} if init_params is None else init_params
+        self.R = self.init_params.get("R", None)
+        self.X = self.init_params.get("X", None)
+        self.Y = self.init_params.get("Y", None)
+        self.p = self.init_params.get("p", None)
+        self.a = self.init_params.get("a", None)
+        self.u = self.init_params.get("u", None)
+
+    def _init(self):
+        # Initialize user, item and aspect rank.
+        if self.p is None:
+            self.p = uniform(self.train_set.num_items, random_state=self.rng)
+        if self.a is None:
+            self.a = uniform(
+                self.train_set.sentiment.num_aspects, random_state=self.rng
+            )
+        if self.u is None:
+            self.u = uniform(self.train_set.num_users, random_state=self.rng)
+
+    def _symmetrical_normalization(self, matrix: csr_matrix):
+        row = []
+        col = []
+        data = []
+        row_norm = np.sqrt(matrix.sum(axis=1).A1)
+        col_norm = np.sqrt(matrix.sum(axis=0).A1)
+        for i, j in zip(*matrix.nonzero()):
+            row.append(i)
+            col.append(j)
+            data.append(matrix[i, j] / (row_norm[i] * col_norm[j]))
+
+        return csr_matrix((data, (row, col)), shape=matrix.shape)
+
+    def _create_matrices(self, train_set):
+        from time import time
+
+        start_time = time()
+        if self.verbose:
+            print("Building matrices started!")
+        sentiment_modality = train_set.sentiment
+        n_users = train_set.num_users
+        n_items = train_set.num_items
+        n_aspects = sentiment_modality.num_aspects
+
+        X_row = []
+        X_col = []
+        X_data = []
+        Y_row = []
+        Y_col = []
+        Y_data = []
+        for uid, isid in tqdm(
+            sentiment_modality.user_sentiment.items(),
+            disable=not self.verbose,
+            desc="Building matrices",
+        ):
+            for iid, sid in isid.items():
+                aos = sentiment_modality.sentiment[sid]
+                aids = set(aid for aid, _, _ in aos)  # Only one per review/sid
+                for aid in aids:
+                    X_row.append(iid)
+                    X_col.append(aid)
+                    X_data.append(1)
+                    Y_row.append(uid)
+                    Y_col.append(aid)
+                    Y_data.append(1)
+
+        # Algorithm 1: Offline training line 2
+        X = csr_matrix((X_data, (X_row, X_col)), shape=(n_items, n_aspects))
+        Y = csr_matrix((Y_data, (Y_row, Y_col)), shape=(n_users, n_aspects))
+
+        # Algorithm 1: Offline training line 3
+        X.data = np.log2(X.data) + 1
+        Y.data = np.log2(Y.data) + 1
+
+        # Algorithm 1: Offline training line 4
+        if self.verbose:
+            print("Building symmetric normalized matrices R, X, Y")
+        self.R = self._symmetrical_normalization(train_set.csr_matrix)
+        self.X = self._symmetrical_normalization(X)
+        self.Y = self._symmetrical_normalization(Y)
+
+        if self.verbose:
+            total_time = time() - start_time
+            print("Building matrices completed in %d s" % total_time)
+
+    def fit(self, train_set, val_set=None):
+        """Fit the model to observations.
+
+        Parameters
+        ----------
+        train_set: :obj:`cornac.data.Dataset`, required
+            User-Item preference data as well as additional modalities.
+
+        val_set: :obj:`cornac.data.Dataset`, optional, default: None
+            User-Item preference data for model selection purposes (e.g., early stopping).
+
+        Returns
+        -------
+        self : object
+        """
+        Recommender.fit(self, train_set, val_set)
+        self._init()
+
+        if not self.trainable:
+            return self
+
+        # Offline training: Build item-aspect matrix X and user-aspect matrix Y
+        self._create_matrices(train_set)
+        return self
+
+    def _online_recommendation(self, user):
+        # Algorithm 1: Online recommendation line 5
+        p_0 = self.train_set.csr_matrix[[user]]
+        p_0.data.fill(1)
+        p_0 = p_0.toarray().squeeze()
+        a_0 = self.Y[user].toarray().squeeze()
+        u_0 = np.zeros(self.train_set.csr_matrix.shape[0])
+        u_0[user] = 1
+
+        # Algorithm 1: Online training line 6
+        if p_0.any():
+            p_0 /= np.linalg.norm(p_0, 1)
+        if a_0.any():
+            a_0 /= np.linalg.norm(a_0, 1)
+        if u_0.any():
+            u_0 /= np.linalg.norm(u_0, 1)
+
+        # Algorithm 1: Online recommendation line 7
+        p = self.p.copy()
+        a = self.a.copy()
+        u = self.u.copy()
+
+        # Algorithm 1: Online recommendation line 8
+        prev_p = p
+        prev_a = a
+        prev_u = u
+        inc = 1
+        while True:
+            # eq. 4
+            u_denominator = self.alpha + self.gamma + self.eta_U + EPS
+            u = (
+                self.alpha / u_denominator * self.R * p
+                + self.gamma / u_denominator * self.Y * a
+                + self.eta_U / u_denominator * u_0
+            ).squeeze()
+            p_denominator = self.alpha + self.beta + self.eta_P + EPS
+            p = (
+                self.alpha / p_denominator * self.R.T * u
+                + self.beta / p_denominator * self.X * a
+                + self.eta_P / p_denominator * p_0
+            ).squeeze()
+            a_denominator = self.gamma + self.beta + self.eta_A + EPS
+            a = (
+                self.gamma / a_denominator * self.Y.T * u
+                + self.beta / a_denominator * self.X.T * p
+                + self.eta_P / a_denominator * a_0
+            ).squeeze()
+
+            if (self.max_iter > 0 and inc > self.max_iter) or (
+                np.all(np.isclose(u, prev_u))
+                and np.all(np.isclose(p, prev_p))
+                and np.all(np.isclose(a, prev_a))
+            ):  # stop when converged
+                break
+            prev_p, prev_a, prev_u = p, a, u
+            inc += 1
+
+        # Algorithm 1: Online recommendation line 9
+        return p, a, u
+
+    def score(self, u_idx, i_idx=None):
+        """Predict the scores/ratings of a user for an item.
+
+        Parameters
+        ----------
+        u_idx: int, required
+            The index of the user for whom to perform score prediction.
+
+        i_idx: int, optional, default: None
+            The index of the item for which to perform score prediction.
+            If None, scores for all known items will be returned.
+
+        Returns
+        -------
+        res : A scalar or a Numpy array
+            Relative scores that the user gives to the item or to all known items
+
+        """
+        if self.train_set.is_unk_user(u_idx):
+            raise ScoreException("Can't make score prediction for (user_id=%d" & u_idx)
+        if i_idx is not None and self.train_set.is_unk_item(i_idx):
+            raise ScoreException("Can't make score prediction for (item_id=%d" & i_idx)
+
+        item_scores, *_ = self._online_recommendation(u_idx)
+        # Set already rated items to zero.
+        item_scores[self.train_set.csr_matrix[u_idx].indices] = 0
+
+        # Scale to match rating scale.
+        item_scores = (
+            item_scores
+            * (self.train_set.max_rating - self.train_set.min_rating)
+            / max(item_scores)
+            + self.train_set.min_rating
+        )
+
+        if i_idx is None:
+            return item_scores
+        else:
+            return item_scores[i_idx]

docs/source/models.rst

@@ -153,6 +153,11 @@ Hierarchical Poisson Factorization (HPF)
 .. automodule:: cornac.models.hpf.recom_hpf
    :members:
 
+TriRank: Review-aware Explainable Recommendation by Modeling Aspects (TriRank)
+--------------------------------------------
+.. automodule:: cornac.models.trirank.recom_trirank
+   :members:
+
 Explicit Factor Model (EFM)
 --------------------------------------------
 .. automodule:: cornac.models.efm.recom_efm

examples/README.md

@@ -50,6 +50,8 @@
 
 [cvae_example.py](cvae_example.py) - Collaborative Variational Autoencoder (CVAE) with CiteULike dataset.
 
+[trirank_example.py](trirank_example.py) - TriRank with Amazon Toy and Games dataset.
+
 [efm_example.py](efm_example.py) - Explicit Factor Model (EFM) with Amazon Toy and Games dataset.
 
 [hft_example.py](hft_example.py) - Hidden Factor Topic (HFT) with MovieLen 1m dataset.