Tag-aware recommender systems (TRS) utilize rich tagging information to better depict user portraits and item features. Existing methods fail to capture multi-aspect user preferences and lack of exploration of tags intersection. In this work, we propose attentive intersection model (AIRec) to address these issues. User representations are constructed via a hierarchical attention network, where the item-level attention di erentiates the contributions of interacted items and the preference-level attention discriminates the saliencies between explicit and implicit preferences. Besides, the tags intersection is exploited to enhance the learning of conjunct features. Finally, we combine factorization machines (FM) with BPR for score prediction. Experiments on two real-world datasets demonstrate signi cant improvements of AIRec over state-of-the-art methods for tag-aware top-n recommendation.
Social tagging systems, also known as folksonomies, are widely used in various websites, where users can freely annotate online resources (e.g., movies, artists) with arbitrary tags. These tags are composed by laconic words or phrases, which can not only indicate user preferences, but also summarize features of items. Consequently, user-de ned tags can be introduced into recommender systems for alleviating the cold-start problem and improving recommendation quality.
To solve the problem of sparsity, ambiguity and redundancy in tag space,
some neural networks-based methods are proposed by converting the tag space
into dense latent space, such as CFA [
Prediction Layer
ො Training BPR Factorization Machines
Concat
In this paper, we focus on developing solution to address the drawbacks mentioned above and propose an Attentive Intersection Recommendation model (AIRec) for TRS. Compared to the previous models, our method not only takes both explicit and implicit preferences into consideration for capturing more accurate user portrait via hierarchical attention network, but also makes full use of the tags intersection to improve performance. 2
In this section, we will present the architecture of our proposed AIRec model and explain the training procedure. Figure 1 illustrates the structure of our model. Preference-level 1
2 Item-level
Element-wise ⨉
Element-wise + Hybrid user model ··· ··· Hidden Layers
MLP
Shared parameters
Shared parameters MLP
MLP
Input Layer and Hidden Layers The user feature vector is constructed as xu = (p1u; p2u; :::; pVu ), where V is the size of tag set and pju = jf(u; i; tj ) 2 Aji 2 Igj is the number of times that user u annotates items with tag tj . Similarly, the item feature vector can be represented as yi = (q1i; q2i; :::; qVi ).
To solve the problem of sparsity and high-dimension, xu and yi are fed into the multi-layer perceptrons (MLPs) with shared parameters. Sharing parameters can not only obtain better generalization capability and less computational overhead, but force networks to use the same feature space to describe user and item. The latent representations of user and item are x~1u = h(xu) and y~i = h(yi). Hybrid User Model To capture multi-aspect user preferences, we should consider not only the explicit preferences x~1u re ected by user's own tagging behaviors, but also the implicit preferences x~2u conveyed by the historical interacted items. In this part, we elaborate a hybrid user model with hierarchical attention AIRec: Attentive Intersection Model for Tag-Aware Recommendation network. The item-level attention aims to depict user implicit preferences x~2u by di erentiating contributions of historical items, while the preference-level attention dynamically discriminates the saliencies between explicit tagging behaviors and implicit preferences for obtaining hybrid user representation x~uH .
In the item-level attention, we leverage an additive attention network to di erentiate contributions of items by investigating the similarities between item representations and explicit preferences x~1 . Suppose the historical items set of u user u is Iu, the representation of kth item ik 2 Iu is y~k. The attention weight (u; k) can be interpreted as the contribution of the kth item to the implicit preferences, which is shown as: (u; k) = sof tmax(v1T ReLU (W0x~1u + W1y~k + b1)); (1) where matrices W0, W1 and vectors b1, v1 are the trainable parameters. Finally, the implicit preferences x~2u can be represented as x~2u = Pik2Iu (u; k)y~k.
The hybrid user representation can be obtained by fusing x~1u with x~2 . Difu ferent from manually setting a hyper-parameter for all users to determine the trade-o , we design a self-attentive fusion mechanism for complying with individual diversity. Similarly, the attention weight (u; k) of the kth part is: (u; k) = sof tmax(v2T ReLU (W2x~ku + b2)): (2) And the hybrid user representation is formulated as x~uH = (u; 1)x~1u + (u; 2)x~2u. Intersection Module Item features are multi-dimensional and have diverse attractions for di erent users. The intersection of user and item tags reveals the deep reason why the user focuses on the item and which are the vital dimensions when modeling this transaction. Motivated by this observation, we elaborate an intersection module to extract the intersection for further enhancing the recommendation performance. Firstly, we calculate the tags intersection by iiu = yi \xu = (r1iu; r2iu; :::; rViu), where rjiu = min(qji ; pj ) means the minimum u occurrences of tag tj . Then iiu is fed into a MLP that shares parameters with the previous MLPs for further training the networks. At last, the latent representation ~iiu is added to the user/item representations, that is, y~i = y~i iiu and ~ x~uH = x~uH ~iiu, where operation means element-wise addition.
Due to the shared parameters, intersection module can constrain MLPs to focus on the conjunct features, obtaining more concrete user/item representations under a certain user-item transaction scenario.
Training Details At the prediction stage, feature vectors y~i and x~uH are
concatenated into a single vector z = y~i; x~uH , and passed through a prediction
layer consisting of a factorization machine [
We optimize the model with the BPR framework and the loss function is
L = Phu;i+;i i ln (y^ui+ y^ui ), where i+ and i are the positive and
negative items of user u respectively. The negative items are randomly sampled from
a uniform distribution. Besides, dropout is also used to prevent over tting.
We conduct experiments on two public datasets: Last.Fm and Delicious and
adopt the same preprocessing as [
We compare the performance of AIRec with FM[
In this work, we propose a novel tag-aware top-n recommendation model AIRec. We design a hybrid user model with a hierarchical attention network for better user modeling and leverage the tags intersection for constraining neural networks to focus on the conjunct features. Extensive experiments shows that AIRec signi cantly outperforms the state-of-the-art baselines.