2.1 Side information

To solve the sparse data and cold start problems of a recommender system, many recommendation algorithms have been proposed, which utilize the profile information of users or items (such as social network and item category) and have high effectiveness in improving the recommendation performance. In recent years, a recommender system not only uses the interactive historical data of users and items but also adds considerable side information to estimate users’ preferences.

Side information can be divided into two categories: structured data and nonstructured data. Structured data can be divided into flat features, feature hierarchies, and knowledge graphs. For nonstructured data, text features, image features, and video features are included. The above data can be integrated into a recommender system through a unique deep learning training model, which has also attracted the researchers’ attention. He and Xiangnan et al. [17] modeled the implicit feedback information (interaction matrix) of user projects and improved the traditional collaborative filtering model. The multilayer perceptron (MLP) is used to model the user-project interaction matrix. Man, Tong et al. [18] proposed a cross-domain recommended embedded mapping framework called EMCDR. The proposed EMCDR framework is different from the existing cross-domain recommendation model in two aspects. First, embedding learning is realized by employing the implicit factor model in each domain to learn the entities’ specific features in each domain. Second, a graphing technique is used to supplement sparse data in different domains. Zheng et al. [19] proposed using two parallel CNNs to process text information corresponding to comments at the user level and comments at the project level to mine users’ behavioral characteristics and the attribute characteristics of projects. The ConvMF model proposed by Kim et al. [20] also uses text information to alleviate the scoring matrix’s sparsity by using text information as auxiliary edge information. The Wide&Deep learning model proposed by Cheng et al. [21] uses the characteristics of users and projects to improve recommendation accuracy. For the Ripp-MKR model presented in this paper, KG is mainly used to fuse the recommender system’s side information.

2.2 KGE for RS

A recommender system refers to analyzing historical data to recommend products that users are likely to like or buy, the core of which is the users and items. Furthermore, the key to recommender systems includes three parts: user preference, item features, and interactions.

The common approaches for embedding-based methods are divided into two steps. The first step is to embed all entities and relationships with knowledge graph embedding (KGE) technology, such as TransE [22], TransR [23], TransH [24], and other Trans series algorithms. The second step is to fuse the entity and relationship embedded vectors with item-user interactions and then model the knowledge graph.

For instance, Zhang et al. proposed CKE [25], which unifies various side information types in the CF framework. They fed an item’s structural knowledge (items attributes represented with a knowledge graph) and content (textual and visual) knowledge into a knowledge base embedding module. First, they adopted TransR to calculate the embedding knowledge graph, and each entity embedding was extracted as the structural vector of the item. Then, the SDAE model was adopted to obtain the descriptive text of the items. Finally, the SCAE model was adopted to obtain the visual embedding of the item-related images. Therefore, an item representation vector is the sum of the original vector and the new information vector from the three aspects. However, CKE does not utilize the relationship information among entities in a knowledge map, and user embedding has not been updated. Even though the additional information is multifaceted, adding a layer of weight could be beneficial.

Wang et al. proposed DKN [26] for news recommendations. It models the news by combining the textual embedding of sentences learned with Kim’s CNN [27] and the knowledge-level embedding of news content entities via TransD. DKN introduces a knowledge graph to measure user preferences based on the user’s past interactions. However, only one hop is selected as the analysis of context information, and the relational information in the map is not involved at all.

The other type of embedding-based method directly builds a user-item graph, where users, items, and their related attributes function as nodes. Zhang et al. proposed CFKG [28], which constructs a user-item KG. User behaviors (purchases and mentions) are regarded as one relation type among entities in this user-item graph. Multiple types of item side information (review, brand, category, bought together, etc.) are included. Wang et al. proposed SHINE [29], which takes the celebrity recommendation task as the sentiment link prediction task among entities in a graph. Besides, this model uses social networks, personal data information, and sentiment networks as side information. Dadoun et al. proposed DKFM for POI recommendations. DKFM [30] applies TransE over a city KG to enrich the representation of a destination.

Yang et al. introduced a GAN-based model, KTGAN [31], for movie recommendations. In the first phase, KTGAN learns the knowledge embedding for movies by incorporating the Metapath2Vec model [32] with a movie’s KG and tag embedding with the Word2Vec model [33] on a movie’s attributes. Later, Ye et al. proposed BEM [34], which uses two types of graphs for items, the knowledge-related graph (containing item attribute information, such as brand and category) and behavior graph (containing item interaction-related information, including cobuy, corate, and coadd to cart) for recommended. In Hongwei Wang’s paper on MKR [35] (multitask feature learning for knowledge graph enhanced recommendations), the main idea is that there is overlap among items in a recommender system and entities in a knowledge graph, multitask learning frameworks, two separate but related tasks, and alternating learning. MKR’s model framework includes recommender systems tasks and knowledge graph feature learning tasks. The recommended inputs are characteristic representations of users and items, and the estimated click-through rate is the output. In the feature learning part of the knowledge graph, the head node and relation of triples are used as input, and the predicted tail node is used as output. The MKR model applies a pattern of alternating training between the knowledge graph and recommender systems.

Path-based methods are primarily based on the inherent correlation structure of knowledge graphs, and the recommended pattern, namely, "metapath ", is defined artificially in advance. For example, in a movie recommendation task, a metapath can be defined as a "user-item-actor-item" or "user-item-director-item" and then modeled on the metapath.

Xiao Yu, Xiang Ren, Ycursed Sun, et al. proposed the PER [36] model published in 2014. In this paper, the concept of a metapath was proposed. Based on each type’s predefined paths, the authors extracted all paths that fit the definition of the metapath for each pair of user items and designed an evaluation function to measure the possibility of user-item interaction. Then, the classic MF algorithm is used to model and solve for the user and item vectors. Similar to PER, Huan Zhao et al. [37] designed metapaths for Yelp and Amazon datasets and obtained L evaluation functions. However, the FM model is adopted in the modeling, and the FM-group lasso is used for further optimization. Sun et al. proposed a recurrent knowledge graph embedding (RKGE) approach that mines the path relation between a user and an item automatically, without manually defining metapaths. Wang et al. proposed a knowledge-aware path recurrent network (KPRN) [38] solution. KPRN constructs the extracted path sequence with both the entity embedding and the relation embedding.

For unified methods, the RippleNet model is the leading representative and the basis for improving this article. RippleNet proposed that users’ preferences spread like waves on the knowledge graph, which is also known as one-hop, two-hop, etc. For an embedding knowledge graph, user embedding is realized through a hop communication of an item that has interfaced with the user. In other words, starting from an item entity with historical interactions, embedding is diffused into three layers, and an entity embedding weighted sum is realized at each layer. Finally, an embedding of the three layers is added.

4.1 Framework

The framework of the Ripp-MKR model is shown in Fig 2. The framework can be divided into a recommendation module, KGE module, and cross and compress units framework.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. The framework of Ripp-MKR.

https://doi.org/10.1371/journal.pone.0251162.g002

As shown in Fig 2, the tail vectors are predicted by the knowledge graph’s head vectors and the relation vectors (represented by the blue module). The tail vectors represent the attribute values of the item. According to users’ historical interaction information, perform ripple propagation gets user preferences for item’s attributes, aggregate these tail vectors, and form a preliminary user vector representation (green module representation). Through ripple propagation in different hops, multiple user vector representations can be obtained and merged to form the final user vector (the rightmost part of the figure). During this period, the cross-training of item vectors (the leftmost part of the figure) and head vectors in the knowledge graph is performed through the cross-compression unit, and the knowledge graph and the recommender system are merged again to iterate the above process. On the one hand, the user vector is represented by the knowledge graph and the user’s historical interaction information. Then cross-training is performed by combining the head vector of the knowledge graph with the item vector in the recommender system. Converged by the loss function, and the final vector representation is obtained and recommended.

The KGE module uses multiple layers to extract features from the head and relation of a knowledge triple. The predicted tail characteristics can be output under the control of the loss function. In this paper’s model, the training of the KGE module is similar to that of the MKR model, while for the recommended module, the training is improved based on the MKR model. In the MKR model, the feature learning of users and items is extracted through multiple layers. In the Ripp-MKR model, the idea of RippleNet is utilized to obtain users’ interest propagation lists through the historical interaction information between users and items. In this process, the knowledge graph’s t vector is combined to express users’ feature vectors instead of the one-hot method. The cross and compress units are another bridge between the KGE module and the recommendation module; they can automatically learn the high-order feature interactions of items in recommender systems and entities in the KG.

4.2 Cross and compress unit

The method of the MKR model combining the KG model and RS model has been described previously. For the Ripp-MKR model, the cross and compress units are represented by a red rectangle in Fig 2, which will be described in this section. The cross and compress unit generates a cross feature matrix from item and entity vectors by cross operation, and outputs their vectors for the next layer by compress operation.

The cross and compress units are shown in Fig 3. This unit is the link module between item v and one of its associated entities e. For the latent feature of latent feature v_l∈R^d and the latent feature of latent feature e_l∈R^d, we construct C_l, representing the cross feature matrix of layer L. is the possible feature interaction between item v and its associated entity e is modeled explicitly in the cross feature matrix of layer L.

(1)

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. The framework of cross and compress units.

https://doi.org/10.1371/journal.pone.0251162.g003

Eq 1 describes the cross operation in cross and compress units. In cross and compress units, there are compress operations in addition to cross operations. The feature vectors of items and entities for the next layer are output by projecting the cross feature matrix into the latent representation spaces, as shown in Eq 2. (2) where w_ï∈R^d and b_i∈R^d are trainable weight and bias vectors. The cross and compress unit can be denoted as the formula: (3)

We use a suffix [v] or [e] to distinguish its two outputs in this paper’s following. Through cross and compress units, Ripp-MKR can adaptively adjust the weights of knowledge transfer and learn the relevance between the two tasks.

4.3 KGE modul

The purpose of knowledge graph embedding is to represent the entities and relationships as vectors and map the KG triples to low-dimensional space under the supervision of functions while preserving their structure. In the recommendation model design process, a knowledge graph is often used as side information because a triple of the knowledge graph has a high correlation with the items and users of the recommender system. When using a KG as side information for the recommender system, item attribute characteristics can be represented as triples of the knowledge graph.

In the KGE module of Ripp-MKR, a head vector and a relation vector are taken as inputs. The head and relation features are extracted using the multilayer perceptron (MLP) and cross and compress units, respectively. The head embedding of the KG triple corresponds to the item ID in the recommender system. The relation embedding corresponds to the item attribute, and the tail embedding corresponds to the specific item attribute value. The process of obtaining a k-layer MLP for predicting tail t is as Eq 4: (4)

S(h) stands for the association set of h in the knowledge graph, v stands for the item ID corresponding to h in the recommender system’s data, and C^L is the cross and compress unit. M(x) = σ(Wx+b) is a fully connected neural network layer with weight W, bias b, and nonlinear activation function σ(∙). For simplicity, we use the exponent L in Eq (4) and the following equation throughout the rest of this paper but note that the L layer parameters are different. M^K is a fully connected neural network with K layer. And is the predicted vector of tail t. h_L and r_L are respectively latent features of h and r. Finally, the score of triple (h, r, t) is calculated using a score (similarity) function f_KG.

(5)

In this model, f_KG is defined in the form of the inner product using the same treatment as in the MKR model: (6)

4.4 Recommendation modul

The Ripp-MKR model of the KGE module part is consistent with the conventional MKR model, and no great innovation has been made In contrast, the recommendation module is the main innovation and improvement of this paper. Among the models that take knowledge graphs as side information, knowledge graphs are mainly carried out in two ways: joint training or alternating training. The Ripp-MKR model integrates these two methods, which is bound to improve the mining degree of latent information of the knowledge graph.

The input of the recommendation module in MKR consists of two primary feature vectors, u and v, that describe user u and item v. In the input process, a one-hot method is used to encode the user and the item by transferring the V vector and the E vector of the KGE model to the cross and compress unit and the recommender system’s deep fusion, and the knowledge graph is completed.

The Ripp-MKR model proposed in this paper combines the knowledge graph processing method of the RippleNet model with the knowledge graph processing method of the MKR model. The RippleNet model uses t vectors combined with the user’s interaction history and the items and represents the user’s vectors through the t vector set of the knowledge graph. As an improvement point, the ripple propagation idea of RippleNet is applied to the user representation method of the recommendation module in the MKR model to eliminate the original one-hot coding.

The Ripp-MKR model uses user U’s interaction history V_u and item V as input in the recommendation module, t. Start with the item of user U’s history click as the seed collection (using these items as the user’s existing preference information). The following logic exists: (7)

Therefore, any member of V_u must be a member of V. Therefore, V_u is a subset of V. In a KG, each item has its corresponding triplet (item, attribute, and attribute value), so the user’s item interaction history set can be changed into the set of corresponding attribute values, which correspond to the t vector in the knowledge graph.

For the input user u, the historical set of interests V_u is treated as seeds in the KG and then extended along links to form a set of k-hop relevant entities for user u . The relational entity of user U can be defined as: (8)

When the value of k is 0, the value of is the collection of historical interaction items of user U, which can be seen as the seed set of user u in the KG.

The k-hop ripple set for user U is defined as a related triplet with k−1 relevant entities as heads. is the set of knowledge triples that are k-hop(s) away from seed set V_u. Consistent with the RippleNet model, we can define the ripple set as: (9)

Through this recursive form, ripple simulation is carried out; water wave transmission is carried out in the knowledge graph. The transferred entity sequence is composed of the set of knowledge graphs, and the corresponding user vector representation is completed. To avoid the ripple set being too large, a maximum length is usually set as a cutoff. On the other hand, the constructed knowledge graph is a directed graph that considers only the degree of the point’s output.

As shown in Fig 2, user embedding is realized by adding vectors represented by the green rectangle in the figure. The vector represented by the first green rectangle needs to use the ripple set of 1-hop (the first item set that extends outward in the KG). Given the item embedding v and the 1-hop ripple set of user u, each triple(h_i, r_i, t_i) in is assigned a relevance probability p_i by comparing item v to the h_i and r_i in this triple. The relevance probabilities are as follows: (10)

Finally, the embedding vector represented by the first green rectangle is weighted for all corresponding t, is the set of knowledge triples that are 1-hop(s) away from seed set V_u, means the result of the user interest after the 1-hop: (11)

The procedure can be performed iteratively on user u’s ripple sets for i = 1,…,H and the result of user embedding u is: (12)

Given user u’s raw feature vector u, we use an L-layer MLP to extract the feature of user u, M^K is a fully connected neural network with K layer: (13)

For item v, we use L cross and compress units to extract its latent feature: (14) where S(v) is the set of associated entities of item v. After acquiring the latent feature of user u and item v, the final predicted probability of user u engaging with item v can be obtained through the prediction function: (15)

We can obtain the final representation by formulas (7)~(12): (16) where is the sigmoid function.

6.1 Dataset

The data set used in this article is shown below.

1. MovieLens-1M [39] is a commonly used recommender system dataset that mainly includes user data, movie data, and rating data, which consists of approximately 1 million explicit ratings (ranging from 1 to 5) on the MovieLens website. The knowledge graph section dataset contains the movie’s attributes and tags.
2. Book-Crossing [40] dataset contains 1,149,780 explicit ratings (ranging from 0 to 10) of books in the Book-Crossing community. The dataset contains binary feedback between users and books, and the KG for each dataset is built by mapping books to the corresponding entities in Satori, DBpedia, or Freebase.
3. Last.FM [41] is the most popular dataset for music recommendation. The dataset comprises information about users and their music listening records from the Last.FM online music system.

Because MovieLens-1M and Book-Crossing contain explicit feedback data, we convert their data to implicit feedback data; each item that is marked as 1 indicates that the user has rated the item (the threshold of rating is 4 for MovieLens-1M, while no threshold is set for Book-Crossing due to its sparsity), and each sample looked at by each user has a UN tag set to 0, this is a rating of the same size. An unwatched set is sampled for each user; each sample is marked as 0 and is of the same size as the watched set.

A similar approach is used for the experimental process of the RippleNet model. We use Microsoft Satori to construct the knowledge graph for each dataset. For MovieLens-1M and Book-Crossing, we first select a subset of triples from the whole KG whose relation name contains "movie" or "book" and whose confidence level is more significant than 0.9. Given the sub-KG, we collect the IDs of all valid movies/books by matching their names with the tail of the triples (head, film.film.name, tail) or (head, book.book.title, tail). For simplicity, items with no matched or multiple matched entities are excluded. We then match the IDs with the head and tail of all KG triples, select all well-matched triples from the sub-KG, and extend the set of entities iteratively up to four hops. The basic statistics settings for the three datasets is shown as Table 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Basic statistics settings for the three datasets.

https://doi.org/10.1371/journal.pone.0251162.t003

6.2 Baseline

To demonstrate the reliability of our algorithms, we use other models that incorporated knowledge graph techniques as a baseline. These models and the Ripp-MKR model proposed in this paper jointly use the same dataset to conduct experimental verification in the sense of the AUC value and ACC value of the model.

MKR [35]: This model is the basis of the Ripp-MKR model proposed in this paper. We set the number of high-level layers to K = 1, f_RS is the inner product, and λ₂ = 10⁻⁶, L =1, d = 8 and λ₁ = 0.5. In this model, all attributes are added to the KGE unit, training is conducted in the knowledge graph unit, and only project-user rating is used as the training input in the RS unit. λ₁ is the weight of L₁ regularization and λ₂ is the weight of L₂ regularization. L is the number of low layers.

PER [36]: PER treats the KG as a heterogeneous information network and extracts metapath-based features to represent the connectivity between users and items. In this paper, we use manually designed user-item-attribute-item paths as features.

DKN [26]: DKN makes use of entity embedding and word embedding as multiple channels and combines them in a CNN for CTR prediction. In this paper, we use movie names as the textual input for DKN. The dimension of word embedding and entity embedding is 64, and the number of filters is 128 for window sizes 1, 2, and 3. Although DKN is an in-depth recommendation model designed for news, it represents the combination of knowledge graph and recommendation system.

Wide&Deep [21]: This is a deep recommendation model that combines a (wide) linear channel with a (deep) nonlinear channel. We concatenate the raw features of users and items as well as the corresponding averaged entity embeddings learned from TransR as input. The dimensions of the user, item, and entity are 64, and we use a two-layer deep channel with dimensions of 100 and 50 as well as a wide channel.

RippleNet [7]: This model is the basis of the Ripp-MKR model proposed in this paper. The parameters we set for the dataset MovieLens-1M are as follows: d = 16, H = 2, λ₁ = 10⁻⁷, λ₂ = 0.01, and η = 0.02. The parameters we set for the dataset are as follows: d = 4, H = 3, λ₁ = 10⁻⁵, λ₂ = 0.01, and η = 0.001. The hyperparameter settings for Last.FM are d = 8, H = 2, λ₁ = 10⁻⁶, λ₂ = 0.01, and η = 0.02. λ₁ is the weight of L₁ regularization and λ₂ is the weight of L₂ regularization. H is the maximum hops.

6.3 Experiments setup

In Ripp-MKR, we set the number of high-level layers to K = 1, f_RS is the inner product, and λ₂ = 10⁻⁶ for all three datasets, and other hyperparameters are given in Table 4. The settings of the hyperparameters are determined by optimizing the AUC on a validation set. For each dataset, the ratio of the training, validation, and test sets is 6:2:2. Each experiment is repeated three times, and the average performance is reported. We evaluate our method in two experimental scenarios: (1) In click-through rate (CTR) prediction, we apply the trained model to each piece of interaction in the test set and output the predicted click probability. We use the AUC and accuracy to evaluate the performance of CTR prediction. (2) In the top-K recommendation, we use the trained model to select K items with the highest predicted click probability for each user in the test set and choose Precision@K and Recall@K to evaluate the recommended sets.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Basic statistics settings for the three datasets.

https://doi.org/10.1371/journal.pone.0251162.t004

6.4 Result

Figs 4 and 5 and Table 5 show the experimental comparison results of the Ripp-MKR model and other baseline models on three different datasets. Evaluation indexes such as the AUC, ACC, Precision@K, and Recall@K were demonstrated. Table 5 shows the results of the AUC and accuracy in CTR prediction.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. The results of Precision@K in the top-K recommendation.

https://doi.org/10.1371/journal.pone.0251162.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. The results of Recall@K in the top-K recommendation.

https://doi.org/10.1371/journal.pone.0251162.g005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. The results of the AUC and accuracy in CTR prediction.

https://doi.org/10.1371/journal.pone.0251162.t005

As shown in Table 5, PER performs poorly on movie, book, and music recommendations because user-defined metapaths can hardly be optimal in reality. Moreover, because the text length in the data set is relatively short, the DKN model results are also unsatisfactory on these three data sets. This is because the DKN model is a recommender model involved in news recommendations. It mainly processes news titles with knowledge graph and conducts sequential training, so it has strong pertinence and limitations, leading to unsatisfactory recommendation results The performance of the Wide&Deep model is not as good as those of the MKR model and the RippleNet model because this model only splices attributes and does not integrate semantic analysis into training as side information like the other two models. For the RippleNet and MKR models, the results are excellent. However, compared with the Ripp-MKR model, there are still some deficiencies. The main reason is that the Ripp-MKR model is the fusion of the RippleNet and MKR models, which makes up for the deficiencies in the architecture of these two models. We can reduce the RippleNet model’s sensitivity to the degree of data sparseness by incorporating the idea of compress units. With the concept of ripple propagation, the MKR model can dig deeper into the connections between items and improve the logic of the knowledge map structure. In general, our Ripp-MKR performs best among all methods on the three datasets.

Ripple-MKR also achieves outstanding performance in the top-K recommendation, as shown in Figs 4 and 5.

The size of the ripple set in each hop is discussed next. We vary the size of a user’s ripple set in each hop to further investigate RippleNet and Ripp-MKR’s robustness. The AUC results for the two datasets are presented in Table 6.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. The results of the AUC w.r.t. different sizes of a user’s ripple set.

https://doi.org/10.1371/journal.pone.0251162.t006

The Ripp-MKR model has the same sensitivity to the ripple set’s size as the RippleNet model, and the accuracy increases with increasing set size. With the increase in the ripple set’s size, the performance of Ripp-MKR is improved at first because a more extensive ripple set can encode more knowledge from the KG.

We vary the ratio of the training set of MovieLens-1M from 100% to 20% (while the validation and test set remain fixed) and report the results of the AUC in CTR prediction for all methods. The results are shown in Table 7. We observe that the performance of all methods deteriorates with a reduction in the training set. The results show that in less training data, the algorithm proposed in this paper is still superior to other models.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Results of the AUC on MovieLens-1M in CTR prediction with different ratios of training set r.

https://doi.org/10.1371/journal.pone.0251162.t007