Intent Disentanglement and Feature Self-supervision for Novel Recommendation

Initial collaborative features. We set up a lookup table to transform the one-hot representations of each user and item into low-dimensional vector. After transformation, $\bm{p}_{u}\in\mathbb{R}^{D}$ and $\bm{q}_{v}\in\mathbb{R}^{D}$ are the latent factor representation of the user $u$ and the item $v$, respectively. We denote them as collaborative features because they are learned from user-item interaction behaviors.

Initial content features. Besides the interactions, each user/item is equipped with side information that describes its own characteristics, such as attribute information, text description, and image. To be general, we term them as content features. Specifically, in our scenario, each user or item is associated with a set of attributes from different fields. Below is an example of user attributes:

Similar to the transformation of collaborative features, we set up an attribute encoding matrix of users and items. The attribute encoding of a user $u$ and an item $v$ is denoted as $[\bm{a}_{u}^{1},\bm{a}_{u}^{2},...,\bm{a}_{u}^{k}]$ and $[\bm{a}_{v}^{1},\bm{a}_{v}^{2},...,\bm{a}_{v}^{k}]$, respectively, where $\bm{a}^{i}\in\mathbb{R}^{D}$.

New released items can be a good way to surprise users. However, these cold-start items do not contain any historical interaction data, so the valuable collaborative features cannot be learned from the traditional method. In this paper, since we extend the standard long-tail recommendation to cold-start items, we have to face such a collaborative feature missing problem. Below we elaborate our new paradigm on solving this challenge.

The content information of the new item has been proven to be important in solving the cold-start problem. Different from simply fusing the auxiliary content information into the item representation, we believe that it is critical to establish the correlation between the content features and the collaborative features for cold-start items. We term this the feature self-supervision between two views of the target item. For example, animation movies are the mainstream entertainment among teenage children. The animation movies (content features) and the behaviors of their audience (collaborative features) have a strong correlation, while the love movies (other content features) and the behaviors of teenage children (collaborative features) have a weak correlation.

In light of this, we first set up the mapping between collaborative and content features, and then generate the missing collaborative features of cold-start items when making recommendations. In the field of representation learning, self-supervised learning leverages input data to construct pre-training tasks and thus can benefit most types of downstream tasks. In our recommendation scenario, we consider the self-supervision as an auxiliary task, and design a collaborative-content feature self-supervision module.

As shown in Figure 2(b), our self-supervision module has two key components: variational autoencoder (VAE) and mutual information maximization (MIM). Both of them are used to establish the strong correlation between the behaviors of items and their related contents. Specifically, we adopt VAE to encode content features into the latent space, so that the latent representation has the same distribution as the original one. We further maximize the mutual information between the generated collaborative features and the inherent content features in order to enhance the representation’s discriminative ability.

The VAE structure assumes that data are generated from underlying latent representation. Given the data, the posterior distribution over a set of unobserved variables is approximated by a variational distribution. VAE consists of a generation part and an inference part. Take the item $v$ as an example, we denote its content features as the concatenation of all of its attribute encoding, i.e., $\bm{x}_{v}=\bm{a}_{v}^{1}\oplus...\oplus\bm{a}_{v}^{k}$. In the generation part, the reconstructed content features $\bm{x}^{\prime}_{v}$ is generated from its latent variables $\bm{z}_{v}$ through a generation network like MLP parameterized by $\theta$:

In the inference part, variational inference approximates the true intractable posterior of the latent variable $\bm{z}_{v}$ by introducing an inference network parameterized by $\phi$  [36].

The objective of variational inference is to optimize the free variational parameters so that the KL-divergence $D_{KL}(q(\bm{z}_{v})\|p(\bm{z}_{v}\|\bm{x}_{v}))$ is minimized. With the reparameterization trick, we sample $\bm{\epsilon}\sim N(0,\bm{I})$ and reparameterize $\bm{z}_{v}=\textmu_{\phi}(\bm{x}_{v})+\bm{\epsilon}\odot\sigma_{\phi}(\bm{x}_{v})$. In this case, the gradient towards $\phi$ can be back-propagated through the sampled $\bm{z}_{v}$. In summary, the loss function in VAE is defined as follows:

Mutual information (MI) is a basic concept in statistics, which measures dependencies between random variables. When encoding the the content features into latent space, to increase the discriminative ability, we further identify the correlation between the latent content features and its corresponding collaborative features using the MIM strategy.

Intuitively, the item $v$’s latent content features $\bm{z}_{v}$ is more relevant to its own collaborative features $\bm{q}_{v}$, compared with the collaborative features of other items. Therefore, we devise the MIM loss function as:

where $(a,v,v^{\prime})$ are the anchor item attributes, the positive item, and the negative item. We uniformly sample the item $v^{\prime}$ as the negative one, which does not contain any attributes of $a$ from the entire item space. $f_{D}({\cdot})$ is the discriminator function that takes two vectors as the input and then scores the agreement between them. We simply implement it as the dot product between two representations to exclude additional parameters  [42].

Finally, the whole loss function of the auxiliary feature self-supervision task can be defined as:

Recall that we analyze the users’ different levels of conformity in the introduction part, which are ignored by existing novel recommendation methods. In this subsection, we present our intent disentanglement module to tackle this problem.

To distinguish the reason why users interact popular or niche items, we set the intents as popularity and preference factors and model the contribution of specific intents to the user-item interactions. In this way, we alleviate the dominance of popularity and increase the novelty naturally. Our proposed intent disentanglement module is shown in Figure 2(c), which has three unique properties.

• •

  We present an intent prototype strategy which explicitly disentangles the intents into two factors (popularity and preference), rather than the parameterized $K$ channels in other disentanglements methods [30, 29, 27, 28, 31].

• •

  Both collaborative and content features take effects when disentangling intents, which tallies better with the conformity theory than previous casual model and negative sampling strategy [32].

• •

  Our disentangled factors are naturally integrated into the end-to-end novel recommendation framework, where the popularity and preference factors account for the accuracy and novelty, respectively.

Intent prototype strategy. We take the user $u$ as an example for illustration. Formally, the user $u$’s overall feature list is denoted as $[\bm{a}_{u}^{0},\bm{a}_{u}^{1},\bm{a}_{u}^{2},...,\bm{a}_{u}^{k}]$, where $\bm{a}_{u}^{0}$ is equal to her collaborative feature $\bm{p}_{u}$ and the rest are $k$ content features.

We start by defining the prototypical intent representations, i.e., the popularity prototype $\bm{c}^{(pop)}\in\mathbb{R}^{D}$ and the preference prototype $\bm{c}^{(pref)}\in\mathbb{R}^{D}$, which are part of model parameters and reflect the position of specific intent in the latent space. We then cluster the initial features according to their distance to two intent prototypes:

where $i=0,1,2,...,k$. We adopt dot product to measure the similarity between a given input feature vector and an intent prototype.

Intent aggregation based on collaborative and content features. The clustering weight $p$ describes the correlation between a specific feature vector and the intent prototype. We now aggregate all features and obtain the integrated representation of the user $u$ under each disentangled intent:

where $\text{FFN}(\bm{x})=\bm{W}_{f}\bm{x}+\bm{b}_{f}$ is a feed-forward network. The item $v$’s popularity and characteristic representation $\bm{q}_{v}^{(pop)},\bm{q}_{v}^{(pref)}$ can be obtained in the same way. So far, we have disentangled the user/item representation under specific intents. We will later explain how the disentangled intents contribute to the whole framework in the optimization subsection.

After obtaining the user and item representations, we take traditional recommendation methods as a plug-in backbone layer to model the relationship between user and item, which ensures the generality of our framework in enhancing recommendation novelty.

Traditional recommendation methods can be roughly grouped into two types: point-wise and pair-wise [47]. Note that some methods use the list-wise objective, but they are not widely adopted due to the expensive computational cost.

Point-wise objective. The basic idea in point-wise methods is that they only consider the individual user-item pair for ranking prediction. They do not relate other observed user-item pairs to the learning process. This kind of methods takes a triplet $(u,v,y_{uv})$ as the input, where $y_{uv}=1$ indicates that the user $u$ has interacted with the item $v$ (implicit preference), otherwise $y_{uv}=0$. The general form of the point-wise loss function can be defined as:

where $intent\in\{pop,pref\}$ and $\varOmega(\theta)$ is a regularization term. A series of traditional methods adopt this kind of optimization strategy and we adopt LFM [48] and NCF [49] as representative base models. During this process, we feed the user/item representation under specific intents into the backbone layer, then obtain the predicted score $\hat{y}_{ui}$ and finally calculate the ranking loss according to these base models.

Pair-wise objective. The pair-wise approaches try to construct the set of item pairs by concerning their relative ordering. It is usually considered to be more suitable for optimizing the top-$N$ recommendation since it compares the relevance of samples instead of getting one ranking score. This kind of methods inputs a triplet $(u,v^{+},v^{-})$, where $v^{+},v^{-}$ are the implicitly liked and not observed item by the user $u$. The general form of the pair-wise loss function can be defined as:

We chose the classic CML [50] as a representative base model of this kind of approaches.

After feeding the disentangled user and item representations into the backbone layer, we can obtain the specific loss $\mathcal{L}^{(intent)}$ ($intent\in\{pop,pref\}$). However, if simply summing the losses without distinction, the learned representation of different intents cannot well reflect the popularity and preference factors. Thus we design a novelty-weighted loss to control the impacts of different intents.

Firstly, inspired by  [6], the item novelty can be measured by its popularity degree, namely an item is more likely to be novel when more users have never interacted with it. The novelty score of the item $v$ can be defined as:

After that, the recommendation loss takes a weighted sum of two intents through the novelty score, where $\sigma(.)$ is sigmoid function for normalization:

Even if it is a common practice to regard the novelty score as weight, we argue that it is an effective way to align the intents to the popularity and preference factors.

Below we discuss the training strategy for point-wise and pair-wise objectives based on the disentangled user and item representations. (1) For the point-wise learning, the model takes a triplet $(u,v,y_{uv})$ as input. Assuming that the novelty of the item $v$ is high, e.g. $\alpha_{v}>0.5$, the users’ interaction to $v$ is reasonably due to her preference intent. Therefore, the loss $\mathcal{L}^{(pref)}$ has a larger weight and it would have more impacts on $\bm{p}_{u}^{(pref)},\bm{q}_{v}^{(pref)}$. Conversely, the popularity intent becomes the leading factor when the user interacts hot items.

(2) For the pair-wise learning, the model takes a triplet $(u,v^{+},v^{-})$ as input. Assuming that the novelty of the item $v^{+}$ is higher than $v^{-}$, i.e., $\alpha_{v^{+}}>\alpha_{v^{-}}$, the user $u$ is more likely to interact with the item $v^{+}$ for personal preference than $v^{-}$. Hence $\mathcal{L}^{(pref)}$ has a larger weight, and vice versa.

Finally, the overall loss of our IDS4NR framework contains the loss for the main task of novel recommendation and that for the auxiliary task of feature self-supervision:

where $\gamma$ is a constant weighting factor.

We use three publicly available datasets from different domains. MovieLens²²2https://grouplens.org/datasets/movielens/ is a widely adopted dataset in the application domain of recommending movies to users, and we employ the MovieLens-100K version. Music and Beauty are chosen from Amazon³³3http://jmcauley.ucsd.edu/data/amazon/links.html. Following the evaluation settings in  [51, 52], we take the 5-core version for experiments, where each user or item has at least five interactions.

In order to be consistent with the implicit feedback setting [49, 53], we transform the detailed rating into a value of 0 or 1, indicating whether a user has rated an item. The statistics of the datasets are shown in Table I.

We evaluate the performance of novel recommendation in terms of accuracy, coverage, novelty and the trade-off.

We adopt Recall@N (Rec@N)  [50, 5, 7] as the accuracy metric, which considers whether the ground-truth is ranked amongst the top-$N$ items (normally @5 or @10):

where $\mathcal{I}_{u}^{Te}$ is ground-truth list of user $u$ and $\mathcal{P}_{u}$ is user $u$’s top-$N$ recommendation list.

Coverage@N (Cov@N)  [2, 5, 7] is the ratio of the total number of distinct recommended items to the total items:

Following the Pareto principle  [54], the head items are the 20% most popular items, and the rest are long-tail novel items. We extend the NovAccuracy@N (Nov@N) [5, 6, 7] to measure how many novel items are in each top-$N$ recommendation list:

where $\mathcal{I}_{Nov}$ is novel item set including both long-tail and cold-start items.

Lastly, we employ F-score as a harmonic mean of conflicting accuracy (recall), novelty, and coverage:

To demonstrate the effectiveness of our proposed IDS4NR model, we compare it with the following state-of-the-art baseline methods. In addition, we adopt different classical base models as backbone layer to show the generality of our framework.

Long-tail recommendation baselines:

• •

  GANC [5] presents a re-ranking framework to integrate the learned user long-tail preference for accuracy, coverage, and novelty oriented recommendation.

• •

  PPNW [6] proposes a loss weighting approach by leveraging users’ and items’ novelty information for end-to-end novel recommendation.

• •

  TailNet [7] designs a preference mechanism which determines users’ preference on popular or niche items in session-based long-tail recommendation. We remove its GRU encoder that is irrelevant to our experiments.

• •

  MIRec [8] is a dual transfer learning framework which collaboratively transfers knowledge from both model-level and item-level and from head items to tail items.

Cold-start recommendation baselines:

• •

  DropoutNet [24] incorporates content and collaborative information with a neural model. It presents the dropout technique to tackle the cold-start issues.

• •

  HERS [17] employs users’ social relations as user-user graph and builds item-item graph based on the common tags between two items. It aggregates user and item relations for addressing cold-start problem.

• •

  MetaEmb [23] is a meta-learning approach to cold-start problem. It trains an embedding generator for new IDs through gradient-based meta-learning technique.

• •

  AGNN [18] develops a graph neural network variant for attribute graph and designs an extended VAE structure for strict cold-start recommendation.

Classical base models:

• •

  LFM [48] is a classical matrix factorization method for collaborative filtering, which learns the latent factors by alternating least squares.

• •

  NCF [49] presents a neural collaborative filtering method that combines multi-layer perceptron with generalized matrix factorization to encode non-linearities.

• •

  CML [50] is a metric learning method, which encodes the user and item into a joint metric space and measures the user-item pair by euclidean distance.

To simulate the strict item cold-start scenario, we choose 20% items which have the latest average interactions chronologically as cold-start items. Note all their interactions are put into the testing set such that they are unseen during training. In addition, to be consistent with normal training scenario, we randomly select 10% of historical interactions for each user and add them into the testing set, and the remaining data are treated as training set. During the test phrase, our evaluation protocol ranks all unobserved items in the training set for each user [5, 55].

For the baselines, we follow the same hyper-parameter settings if they are reported by the authors and we fine-tune the hyper-parameters if they are not reported. For our IDS4NR, we set the batch size = 128, the initial learning rate = 0.001, the number of max epoch = 40, 40, 30 for the base model LFM, NCF, and CML, respectively. We use Adam [56] as optimizer to self-adapt the learning rate. We sample 4 unobserved items as negative samples for each positive user-item pair [49]. The embedding dimension is set to 50 for both our IDS4NR and all the baselines, to ensure the comparable trainable parameters.

Given the appealing performance of IDS4NR, we further examine how the individual components in our model contribute to the improved performance on novel recommendation. Concretely, we conduct a series of ablation studies to investigate the effects of explicit intent disentanglement and collaborative-content self-supervision:

• •

  IDS4NR${}_{\text{w/o-SS}}$: This variant removes the feature self-supervision module from the IDS4NR framework.

• •

  IDS4NR${}_{\text{w/o-SS, Exp}}$: Our IDS4NR uses an item novelty score $\alpha_{v}$ to control the impacts of disentangled representation of popularity and preference intents. This variant is based on IDS4NR${}_{\text{w/o-SS}}$ and further replaces $\alpha_{v}$ in Eq. 13 with 0.5 for treating all samples equally.

• •

  IDS4NR${}_{\text{w/o-SS, ID}}$: This variant removes the intent disentanglement module completely. To ensure utilizing the content information as the standard IDS4NR, we take the average of collaborative feature and all kinds of attributes of users/items as the input of backbone layer on the basis of IDS4NR${}_{\text{w/o-SS}}$.

For clarity, we choose the CML backbone as the representative to perform the ablation and other parameter studies. We first present The ablation results on three datasets are shown in Table III.

Firstly, we investigate the impacts of feature self-supervision module. We find that the overall trade-off performance for IDS4NR${}_{\text{w/o-SS}}$ decline on all three datasets compared with the standard IDS4NR. This indicates that the feature self-supervision component captures the relationship between the collaborative feature and content feature and well models the user preference for cold-start items, thus it is an essential part of IDS4NR.

Next, we examine the impacts of explicit intent disentanglement module. It is clear that both the coverage and novelty of IDS4NR${}_{\text{w/o-SS, Exp}}$ decrease a lot if compared with IDS4NR${}_{\text{w/o-SS}}$. The reason might be that the intent disentanglement without any constraint cannot learn the specific factors explicitly and will limit the ability of learning representations for users’ intents on popular/niche items.

Finally, the overall performance of the variant IDS4NR${}_{\text{w/o-SS, ID}}$ decreases dramatically along with a few increased accuracy. After removing the intent disentanglement, the model degrades into a naive content-based method, and focuses on the items having similar contents with users’ historical interactions. With the limited recommendation scope, the accuracy of the model may increase a bit but the coverage and novelty performance become extremely poor.

We investigate the impacts of hyper-parameters in IDS4NR, including the dimension of the latent vector $D$ and the weighting factor $\gamma$ for feature self-supervision loss. We first fix $\gamma$ to 0.01 and vary $D$, and then fix $D$ to 50 and vary $\gamma$. The results of IDS4NR with CML as backbone layer for the varying parameter $D$ and $\gamma$ are shown in Figure 3 and Figure 4, respectively.

We can see from Figure 3 that accuracy and novelty show opposite trends in all datasets. This is consistent with the problem setting and the results in previous studies. Movielens is a small dataset with relatively dense user-item interactions. The model can already learn a good feature representation even with a small dimensionality ($D$=10). In contrast, Music and Beauty are even sparser. The larger dimension improves the representation ability of features with more latent factors of users and items, and thus the model reaches a better accuracy, but its novelty declines at the same time. Hence we set $D$ as 50 for a better balance in most cases.

Our IDS4NR takes feature self-supervision learning as an auxiliary task to enhance the collaborative features of cold-start items, and the weighting factor $\gamma$ controls the proportion of self-supervision task. As we can seen in Figure 4, with the increase of $\gamma$, the accuracy on different datasets gradually decreases and the novelty increases, and this trend is more obvious when $\gamma$ exceeds 0.01. When $\gamma$ is small, it can improve the performance for cold start items as an auxiliary task and thus enhances the recommendation novelty. However, when $\gamma$ is too large, the self-supervision task dominates the loss and interfere the learning of the main task, leading to a sharp decline of recommendation accuracy.

To investigate the training process of model, we plot the model’s training curves in Figure 5. As can be seen, the loss declines quickly first and gradually slows down. It finally reaches the convergence when the training epoch increases, which proves that our model is stable and easy to train.

We further compare the model’s space complexity by presenting the trainable parameter number of our model and that of base models in Table IV. The parameter number of our IDS4NR and its different base models mainly depends on the user/item embedding matrix. Compared to the base models, our framework introduces additional user/item attributes, so the increased cost lies primarily in the attribute embedding. For the intent disentanglement and self-supervision module in the framework, they need a few weight matrix at a lower cost. Note that NCF has the largest parameter scale on Music and Beauty because it uses the independent embedding layer in its GMF and MLP module, which is equivalent to expanding the dimension of embedding vectors. When we take it as the backbone layer, these two modules are shared by the disentangled user/item representation and thus decrease the parameter number.