Patient-Oriented Herb Recommendation System Based on Multi-Graph Convolutional Network

Abstract

The presented herb recommendation system aims to analyze the patients’ symptoms and recommends a set of herbs as the prescription to treat diseases. In addition to symptoms, the patients’ personal properties and induced diagnoses are also essential for treatment making. Specifically, for different age groups, the treatments are different. However, the existing studies only use symptoms to represent patients and ignore the patients’ multidimensional features modeling. Thus, these models are insufficiently personalized. Meanwhile, most of these existing herb recommendation models based on graphs have not distinguished the effects of different node types. To address the above limitations, we propose a model named Patient-Oriented Multi-Graph Convolutional Network-based Herb Recommendation system (PMGCN). The prediction model contains two effective modules, patient portraits modeling and herb interactions modeling, to learn representations for patients and enhance herb interactions. First, we depict the patient portrait to enrich the individualized features. To distinguish personal properties, symptoms, and diagnoses, we adopt the type-aware attention mechanism, thereby improving the accuracy of personalized herb recommendation. Next, we build two herb-interaction graphs and design type-aware multigraph convolution networks to capture the interactions of herbs and patient features. In this way, our model emphasizes the impact of the patient portrait on diagnosis induction and herb selection. Experimental studies demonstrate that our method outperforms the compared methods and confirms the significance of patient portraits. In conclusion, this research proposes type-aware multigraph convolution networks and adds patient portraits modeling to simulate TCM prescriptions making.

1. Introduction

Traditional Chinese medicine (TCM) is a very comprehensive system that contributes to the Yin-Yang and Five Phases theories. TCM prescriptions have been used and recorded by doctors for thousands of years. Each complete clinical prescription contains three parts: patient information, diagnosis, and a set of herbs with their dosages. TCM prescription describes the process whereby doctors diagnose a specific patient and use a set of herbs as the treatment.

In clinical practice, doctors base their diagnoses and treatment on TCM theories and individual experience. Figure 1 shows the therapeutic process in TCM. The first step is symptom and patient-feature collection. The doctor examines the patient’s symptoms and considers personal properties, such as “female” and “fatigue”, for diagnosis. The second step is diagnosis induction. The doctor makes a diagnosis after an overall analysis of the patient. In this case, the primary syndrome is “damp-heat”, and the primary disease is “dermatitis”. The last step is treatment determination. The doctor chooses a set of herbs and gives their dosage in the TCM prescription to cure the disease.

There are many challenges for prescription making. First, diagnoses of patients are related to both symptoms and personal properties. For instance, for middle-aged people and children with the same symptom “Diarrhea”, the treatments are different. A herb such as “herba ecliptae” is available for middle-aged people but cannot be used for children. There are also many studies confirm that the gender and age of patients affect the diagnosis and classification of diseases [1,2,3]. Second, the compatibility of TCM is comprehensive [4]. The chemicals of herbs vary in different prescriptions because of the chemical reactions among various herbs. Therefore, it is difficult to select a proper herb set for a specific patient.

The first motivation of this research is to incorporate individualized features in patient portraits, which is essential for personalized herb recommendation. The second motivation is to extract interactions between herbs and patient features. In this way, we could simulate the process of diagnosis and treatment. In addition to the above, we have found that most of the current herb recommendation models ignore the impact of distinguishing node types. This shortcoming makes information in graphs lack of focused information. It is also impractical, for instance, the importance of personal properties, symptoms and diagnoses in prescription making is different.

To resolve the challenges of prescription making, this paper incorporates diagnosis information and personal properties into patient portrait modeling. This research extracted patients’ multidimensional features for diagnosis induction. Then, our model captured the interactions between the patient portrait and herbs for treatment. Using type-aware multigraph convolutional networks as graph encoders, this research work learns the representations of personal properties, symptoms, diagnosis, and herbs. Finally, we use the fusion and prediction module for making herb recommendations.

In summary, we propose a Patient-oriented Multi-Graph Convolutional Network-based herb recommendation system (PMGCN). This model contains three components: patient portraits modeling, herb interactions modeling and the fusion and prediction layers. Compared with the most recent state-of-the-art studies, the innovation of this research work lies in the patient portrait modeling and the adoption of type-aware attention mechanism in multigraph convolutional networks, which improves the message passing for patients’ and herbs’ graph learning in TCM recommendation. These contributions make our model more personalized than the other baselines.

We conduct experiments on medical and movie datasets. The experiments demonstrate the effectiveness of our model both in TCM and the common domain. Our model also outperforms the most recent state-of-the-art studies.

The main contributions of this paper are as follows:

• This research innovatively designs patient portrait modeling, which uses individualized characteristics to improve herb recommendation performance. This study divides different types of patient features and constructs patient portraits, which improve the quality of representation learning. Through the fusion of multidimensional features from patient portraits, we obtain better representations for patients.
• We propose the type-aware multigraph convolutional network model based on patient portraits and herb interactions, named PMGCN, for the patient-oriented herb recommendation system. The model unifies multiple graph convolutional networks (GCNs) and adopts a type-aware attention mechanism in comprehensive interaction modeling. Except for patient portrait modeling, we built herb–herb and patient–herb graphs to extract the interactions among patient features and traditional Chinese medicines. In addition, this is the first study to introduce an attention-based GCN method for making herb recommendations. Thus, the model could distinguish the effect of different node types, which is conformable to the practice needs in prescription making.
• Experimental results demonstrate that our model achieves the state of the arts on the clinical dataset. Additionally, this research confirms the contribution of patient portraits to herb recommendations.

2. Related Work

Originally, TCM recommendation studies focused on traditional data-mining methods and topic models, which are limited in semantic representation.

As deep-learning models have a strong expression ability in complex relationships, researchers have adopted them for producing herb recommendations. With the development of graph modeling, researchers have organized TCM prescriptions into graphs. However, existing studies use only symptoms for patients’ representations, thereby making these studies’ models insufficiently personalized.

The methods used for making herb recommendations are as follows.

2.1. Traditional Data Mining

Referring to the surveys [5,6,7], early studies employed mainly association rule analysis, decision tree and other technologies in mining symptoms and traditional Chinese medicines.

Early studies are devoted to dealing with famous TCM prescriptions in order to extract medication rules. Siqi [8] used statistical analysis to study the common rule of famous prescriptions for treating diabetes.

Considering the limitation of statistical analysis methods in extracting implicit knowledge, some studies use association rule analysis, Bayesian networks, decision trees and other data mining methods to summarize the patterns of TCM. Chen [9] uses the association rule to analyze the symptom–symptom and drug–drug relationships in the record of “Wind-Warmth”. Xie [10] establishes a decision tree model to diagnose the yin and yang syndrome for primary osteoporosis. Pan [11] constructs the attribute partial-order diagram to capture the relationship between syndromes and dosage for knee osteoarthritis.

The above studies focus only on special disease prescriptions, and the size of the dataset is limited, thus greatly restricting performance.

2.2. Topic Model

Topic models consider the co-occurrence of “symptoms” and “herbs” to explore the direct symptom–herb relationship, which is limited in semantic representation.

Early studies regarded prescriptions as documents and regarded herbs and symptoms as words. However, symptoms and herbs are two types of entities. Thus, traditional methods are inappropriate for symptom and herb representations. Ji [12] proposes multicontent LDA to solve this representation problem. The LDA proposed by the author regards pathogenesis as the hidden theme connecting symptoms and herbs. Lin [13] depicts the relationship between explicit symptoms and herbs and between implicit diagnosis and treatment in the framework of the topic model.

To enrich the representation, Zhang [14] employs the word vectors of the name of symptoms and herbs in the topic model. Chen [15] integrates a TCM knowledge graph into a topic model to enhance the association of entities. Wang [16] adopts TransE [17] on a TCM knowledge graph to enrich the representation of herbs and symptoms.

The deficiency of the above methods is that the employment of the multivariate heterogeneous relationships in TCM is too simple. The expression ability of topic models is weaker than that of deep-learning models.

2.3. Sequence Model

Early researchers used sequence generation models to derive the corresponding drug from the symptom descriptions and records. Li [18] adopts the Seq2seq [19] model with an attention mechanism. Similarly, Ua [20] employs an attention network for treatment making. Muhammad [21] proposes a deep neural network-based model with a hierarchal mechanism to extract drug–drug interactions from text sentences.

As the sequence relationship in records is very weak, Wei [22] proposes a prescription-level language model based on bi-GRNN. The model improves the representation of traditional Chinese medicines with prescription context.

The deficiency of the above methods is that the high-order relationship modeling between entities in the TCM domain is still insufficient.

2.4. Graph Representation Learning

The graph-based recommendation model has developed rapidly in recent years and employs graphs to model the high-order relationships in TCM prescriptions.

Most studies propose meta path-based models to learn graph representation [23]. However, these models cannot extract neighbor information completely, thereby leading to information loss.

To address the above limitation, HERec [24] employs a heterogeneous network recommendation method, which improves the fusion strategy of node information. Through deepwalk, HetGNN [25] fuses the neighbors’ representation via different contexts through bi-LSTM to enrich the representation of entities. Vidya [26] investigates multi-modal gene disease and disease drug networks via link prediction algorithms to select drugs to treat skeletal muscle atrophy.

Recently, a group of studies have confirmed the effectiveness of GCNs to model complex relationships in graphs. SMGCN [27] uses bipar-GCN on herb–herb homographs and symptom–symptom homographs to capture the correlation between symptoms and herbs. However, it has not considered patients’ multidimensional features and their comprehensive correlation. Additionally, it does not distinguish between the different types of nodes in message passing, which is not realistic in practice.

In contrast, we introduce multi-GCNs into patient portrait modeling and more complex herb interaction modeling. Additionally, we use type-aware attention GCN to distinguish the effect of different node types. This could improve the effect of message passing and aggregation for graph learning. In this research, we emphasize that patient portraits are essential for a more personalized recommendation system. We add patient portrait modeling for a more reliable representations of patients, instead of only considering symptoms.

In summary, the development of herb recommendations contains four stages: traditional statistical analysis and data-mining methods, topic models, sequence models and graph representation learning models. These studies are still in the initial stage, thus making them valuable for further exploration.

3. Problem Definition

Herb recommendations aim to predict the specific herb set by relying on the patient’s information, such as symptoms and diagnosis. First, we define the patient set $U=\{u_1,u_2,...,u_z\}$, where each u indicates a patient. Then, we define the patient feature set as $PF=\{p{f}_1,p{f}_2,...,p{f}_m\}$ where $pf$ denotes patient features, including diagnosis d, symptom s, and personal properties p. Thus, $PF$ contains more than one type of entity. Next, we define the herb set as $Herb=\{h_1,h_2,...,h_n\}$, where h indicates herbs. For each patient u, there is a patient portrait; the patient portrait, which is denoted as $p{f}_{set}$, contains multiple patient features $pf$.

Thus, each prescription is denoted by $pr=(p{f}_{set};h_{set})$, where $h_{set}$ denotes the herb set in the prescription, which is the input of model. In a diagnosis process, the doctor needs to consider overall symptoms and personal properties to induce the diagnosis for each patient to generate an appropriate herb set.

Given a specific patient portrait set $p{f}_{set}$, the task of our model is to compute an N-dimensional probability vector. The value of dimension i in the output vector represents the probability that herb i is selected. The prediction function is defined as ${\widehat{y}}_{p{f}_{set}}=f(p{f}_{set},H;\theta )$, where $\widehat{y}$ represents the probability vector and $\theta$ is all the trainable parameters. The function is used to generate the probability vectors for all herbs from H.

4. Model

4.1. Overview of Proposed Approach

In this study, we solve two problems in practice. The first is how to divide different types of patient features in representation learning. The second is how to use multidimensional features in patient portraits to model the comprehensive interactions between patient features and herbs.

To resolve these problems, we propose the PMGCN. Figure 2 shows the overall framework of our model. The proposed TCM recommendation model takes a patient portrait set $p{f}_{set}=\{p{f}_1,p{f}_2,...,p{f}_m\}$ and herb set $h_{set}=\{h_1,h_2,...,h_n\}$ as input, where each $pf$ can be diagnosis d, symptom s or personal property p. Thus, $p{f}_{set}$ can also be declared as $\{s_1,...,s_i;p_1,...,p_j;d_1,...,d_m\}$. The predicted probability vector $\widehat{y}$ in dimension $\left|H\right|$ is the output, where the value at position i indicates the probability that $h_i$ is used for patient features $p{f}_{set}$.

-

Multigraph Convolutional Network-Based Recommendation. We apply improved GCN models in the framework, which contains the parts: patient portrait and herb interaction modeling, fusion and a prediction module. We construct multi-GCNs and use type-aware attention to improve the effectiveness of message-passing. We add the fusion layer to better aggregate multidimensional patient features. Finally, the framework outputs the recommendation result via the prediction layer.

-

Patient Portrait and Herb Interaction Modeling. First, we construct the patient portrait graph, patient–herb graph, and herb–herb graph. Then, we introduce multi-GCNs as the graph encoder to encode synergy information of the patient portrait and herb interaction. Moreover, we use the type-aware attention mechanism to distinguish the message passing of different neighbors. Finally, we obtain a representation of patient features and herbs.

-

Fusion and Prediction Module. We use an MLP layer to fuse all the features in the patient portrait to improve the patient representation. Then, the fused embedding and items’ feature embedding are fed to the prediction layer to obtain a TCM recommendation.

4.2. Graph Construction

We construct the patient portrait heterogeneous graph, herb-herb homograph, and patient-herb related graph based on clinical data. The constructed graphs are undirected and unweighted initially. Figure 3 shows an example.

The weights of edges are computed by attention score in GCN encoder when the model is training.

4.2.1. Patient Portrait Construction

Unlike the previous study, we divide user features into three types, personal properties (including age and gender), symptoms (including symptoms and syndromes), and diagnosis (including diseases in both Western and Chinese doctors’ systems); these features are the critical basis of diagnosis in TCM. Therefore, we constructed the patient portrait graph, which contains three types of nodes. In addition, these graphs do not contain specific patient nodes, while we still learn the patients’ representation by aggregating features from the patient portrait.

Similar to prior studies [16,27], patients’ personal properties, symptoms, and disease are linked to each other if they appear in the same prescription. This is the appropriate way to describe the pattern of diagnosis. Patients’ symptoms are related to personal properties and diseases. The diagnosis of disease is based on personal properties and symptoms.

We can obtain the patterns and rules of diagnosis based on clinical data. Therefore, the patterns in our graph show not only knowledge but also experience in the diagnosis process.

4.2.2. Herb Interaction Graphs Construction

This research uses statistics on the co-occurrence information in prescriptions to construct the herb–herb relations, which indicates the compatibility of traditional Chinese medicine. In this way, we construct a herb–herb graph as a part of the overall graphs.

Then, we construct an herb–patient interaction graph to simulate TCM prescription making. Similarly, patient portrait features are linked to the corresponding herbs if they both appear in the same prescription. This link is realistic because in TCM, prescription making is based on the patient portrait.

4.3. Multigraph Convolutional Network

Recent studies on recommendation systems demonstrate the performance of the GCN encoder on user–item graph representation. However, most previous works on TCM recommendation do not considered the influence of patient portraits. On the contrary, these studies consider only the association between herbs and symptoms. Additionally, these studies do not divided the different types of patient features in the herb interaction graph but treat all neighbors as the same. Therefore, we model the patient portrait and introduce multi-GCNs on graphs to capture the multidimensional features of patients and extract the comprehensive interactions between patients’ features and herbs.

4.3.1. Multigraph Modeling

To model the comprehensive interactions among patients’ multidimensional features and herbs, we construct multiple graphs and apply multi-GCNs into our framework in Figure 2. Specifically, the multi-GCNs employ type-aware attention-guided graph convolutional operations on heterogeneous graphs to capture the influence of neighbors with different node types for patient portrait representation. Then, we use an MLP to fuse these features’ embedding to represent patients and make recommendations.

Model Initialization

The TCM prescription is denoted as $pr=\{p{f}_{set};h_{set}\}$, where $p{f}_{set}$ indicates the patient portrait set and $h_{set}$ indicates the corresponding herb set. Formally, $p{f}_{set}=\{p{f}_1,p{f}_2,...,p{f}_n\}$; $h_{set}=\{h_1,h_2,..,h_m\}$. The patient features $pf$ are defined as three types: symptom s, diagnosis d, and personal property p.

The links in the constructed graphs are initialized as $M_{pf,h}$.

$$M_{p{f}_i,h_j}=\left\{\begin{array}{cc}1,\hfill & if(p{f}_i,h_j)co-occur\hfill \\ 0,\hfill & otherwise\hfill \end{array}\right.$$

(1)

The same strategy is used on ($p{f}_i$,$p{f}_j$) and ($h_i$,$h_j$).

Concerning the graph nodes, we use $e_u$, $e_{pf}$ and $e_h$ to represent the embeddings of patients, patient features, and herbs, respectively. Following the prior works [24,25,27], we use this equation for graph construction and randomly initialized embedding vectors $e_{pf}$ and $e_h$ for patient features and herbs, respectively.

Suppose that in the k-th layer of the GCN encoder, the embedding of herb h is $e_h^k$, and the embedding of patient feature $pf$ is $e_{pf}^k$ (including $e_s^k$, $e_d^k$, and $e_p^k$). The weighted matrix $W^k$ is initialized as $W_h^k$, $W_s^k$, $W_d^k$, $W_p^k$, in which $W_h$ indicates the herb’s weight, $W_s$ indicates the symptom’s weight, $W_d$ indicates the diagnosis’s weight, and $W_p$ indicates the personal property’s weight. To propagate neighbor nodes, we set $m_{N_{h_i}}^k$ as the message passing from neighbors of herb $h_i$ in the k-th layer.

Patient Portrait Modeling

As shown in Figure 4, patient portrait modeling employs type-aware attention-guided convolution operations on the patient portrait graph to learn representations for patient features (i.e., symptoms, diagnosis and personal properties). Then, we adopt a fusion layer for the overall representations for patients; this layer will be introduced in the next section, titled “Fusion and Prediction Module”. Specifically, this research work aggregate the message from one-hop neighbors in each layer and extend the propagation rule to the multiple layers. We adopt type-aware attention to capture the influence of multitype neighbors in message passing, thereby improving the performance of patient portrait modeling.

The relations in the patient portrait are more complex than those in the herb–herb homograph. However, the general GCN operators share the same parameters on all nodes of different types (i.e., symptoms and herbs). In this work, we consider the heterogeneity of the patient portrait graph.

To discriminate the influence strengths of different neighbors, we incorporate the type-aware attention into the GCN encoder. We need to distinguish the different effects of different node types. Refer to the attention calculation in GAT [28], we use the sum of their exponential functions as the denominator, while we use the target node’s exponential function as the numerator. Besides, we change the equation by grouping and dividing the type of nodes. It could represent the effect of neighbor nodes under the same type of nodes. Thus, for target node i, the attention score for one-hop symptom neighbor j is computed as follows,

$${\beta }_{i,j}=\frac{exp(e_i^T\ast e_j)}{{\sum }_{z\in N_i^s}exp\left(e_i^T\ast e_z\right)}$$

(2)

where $N_i^s$ indicates the symptom neighbors of node i in the patient portrait. If neighbor j is a diagnosis or personal property node, we also use the same type of neighbors in $N_i$ to calculate the denominator. In this way, the contribution of neighbors is adjusted automatically.

Therefore, for each node $s_i$ in the patient portrait graph, we aggregate the neighbor symptoms, diagnosis, and personal properties by using different attention scores. To stress the type-aware attention scores more obviously, we use $\lambda$, $\theta$, and $\gamma$ instead of $\beta$ for representation.

Suppose the target node is symptom r. Thus, the message passing is obtained as

$$\begin{array}{c}\hfill m_{N_r}^k=\sum _{s_i\in N_r^s}{\lambda }_{r,s_i}{e}_{s_i}^k{W}_s^k+\sum _{d_j\in N_r^d}{\theta }_{r,d_j}{e}_{d_j}^k{W}_d^k+\sum _{p_k\in N_r^p}{\gamma }_{r,p_k}{e}_{p_k}^k{W}_p^k\end{array}$$

(3)

where $N_r$ indicates the neighbor set of target node r, and $N_r^s$ denotes the symptom neighbors. Similarly, the superscript d indicates diagnosis nodes and superscript p indicates personal property nodes. The common message is calculated as $eW$ which means the production of node embedding and the weight of neural network layer. In our model, this message passing equation adds the attention scores to distinguish different nodes, which is normal and similar to prior works [28].

Similar to message passing, the types of neighbors also influence high-order propagation. To extend the message passing from one layer to multiple layers, we use the aggregation of neighbors in the former layer to update the embedding of target node in the next layer.

Suppose the aggregation of neighbor nodes for target node r in the k-th layer is $e_{N_r^k}$. By averaging the sum of different neighbor nodes’ messages, and through the activation function $\Theta (·)$, the k-th layer defines the message aggregation as

$$e_{N_r}^k=\Theta (\frac{1}{\parallel N_r\parallel }{m}_{N_r}^k)$$

(4)

where $\parallel N_r\parallel$ indicates the number of neighbors.

After aggregation operation, we update the embedding of symptom node r in the next layer $k+1$,

$$e_r^{k+1}=\Theta (W_s^{k}cat(e_r^k,e_{N_r}^k))$$

(5)

where $cat(·)$ indicates the concatenation operation of embeddings.

Herb Interaction Modeling

As is shown in Figure 2, herb interaction modeling contains a patient–herb GCN and a herb–herb GCN. To enrich the representation of herbs, we built two herb interaction graphs and constructed multiple GCN models. This research used the herb–herb GCN to capture the compatibility of traditional Chinese medicine. We adopted the patient–herb GCN to model the relationship of patient features and herbs.

The details of herb interaction modeling are shown in Figure 5.

As $h_i$ denotes the target herb node, $N_h^1$, $N_p^1$, $N_d^1$ and $N_s^1$ denote the one-hop herb and patient portrait neighbors, respectively. The message passing and aggregation from neighbors is influenced by type-aware attention weights. Finally, we merge the embeddings from herb–herb GCN and patient–herb GCN in the fusion layer.

For the herb–herb GCN, we apply the weighted graph convolution operation on the herb–herb graph to aggregate the herb neighbors $N_h$ for each herb h. Suppose r indicates the target node, and $h_i$ indicates one of the herb neighbors in $N_r$. The message-passing mechanism in herb–herb GCN is defined as follows. Equation (6) defines the message passing from a neighbor node r to the target $h_i$. Equation (7) indicates the message passing from all the neighbors.

$$m_{r,h_i}^k=e_{h_i}^k{W}_h^k$$

(6)

$$m_{N_r}^k=\sum _{h_i\in N_r}{e}_{h_i}^k{W}_h^k$$

(7)

Concerning the patient–herb GCN, to model the interaction between patient features and herbs, we use type-aware attention to distinguish different types of neighbors in the graph. The mechanism of type-aware attention in the patient–herb GCN is shown in the upper right of Figure 5.

The message passing for target r is defined as follows.

$$\begin{array}{c}\hfill m_{N_r}^k=\sum _{s_i\in N_r^s}{\lambda }_{r,s_i}{e}_{s_i}^k{W}_s^k+\sum _{d_j\in N_r^d}{\theta }_{r,d_j}{e}_{d_j}^k{W}_d^k+\sum _{p_k\in N_r^p}{\gamma }_{r,p_k}{e}_{p_k}^k{W}_p^k+\sum _{h_z\in N_r^h}{\beta }_{r,h_z}{e}_{h_z}^k{W}_h^k\end{array}$$

(8)

Similarly, we use the aggregation of neighbors in the former layer to update the embedding of node t in the next layer. The update operation of nodes is also defined as Equation (5).

4.3.2. Fusion and Prediction Module

Until now, we have obtained the patient features’ embeddings from patient portrait modeling. Therefore, we employ a fusion layer to aggregate multidimensional features in the patient portrait. This research work utilizes an average pooling operation and an MLP layer to obtain an overall representation for patients. To fuse all the features in the patient portrait, we define each patient’s overall representation as

$$e_{u_i}=RelU(W·Mean\left(e_{pf}\right)+b)$$

(9)

where $e_{pf}$ indicates all the features of the patient portrait, b indicates the bias setting and W is the weight matrix in the MLP layer.

Additionally, we obtain two types of embeddings for each herb from herb interaction modeling. Thus, we use element-wise addition to merge herbs’ embeddings, which is a normal option for embedding merging. This research work defines each herb’s representation as

$$e_{h_j}=e_{h_j}^{ph}\oplus e_{h_j}^{hh}$$

(10)

where $hh$ indicates that the embedding $e_{h_j}^{hh}$ is from the Herb-Herb GCN and $ph$ indicates that $e_{h_j}^{ph}$ is from the Patient-Herb GCN. ⊕ indicates the addition operation.

Then, we use this overall representation for patient and herb representations as input for the prediction layer. Formally, the probability of herb $h_j$ being used for patient $u_i$ is defined as $f(u_i,h_j)$.

$$f(u_i,h_j)=e_{u_i}{e}_{h_j}^T$$

(11)

In practice, we use a frequency-based parameter to summarize all the $f(u_i,h_j)$ for $h_j\in h_{set}$.

4.3.3. Model Training

To estimate model parameters, this research requires multilabel loss to train the set2set recommendation model. We adopt the weighted mean-square loss [29] to quantify the difference between the output probability vector and ground truth vector, which is similar to prior work [27]. In this research work, we use $L_2$ regularization to prevent overfitting. Formally, the loss function is defined below,

$$Loss=\sum _{u_i\in U}WMSE(h_{set},f(u_i,H))+\lambda {∥\theta ∥}_2^2$$

(12)

where the $\theta$ denotes all the trainable parameters, and the $\lambda$ indicates the hyperparameter to control the $L_2$ regularization.

5. Experiment

Our experiments are designed to answer the following research questions:

RQ1.

Can our model outperform the state of the art on TCM recommendation?

RQ2.

How do patient portraits influence TCM recommendation?

RQ3.

Can our model perform well in a common domain?

5.1. Experiment Settings

We adopt Precision@N, Recall@N, and F1-score@N to measure the results of herb recommendations.

5.1.1. Dataset

We use three datasets: Real-sg, Ali, and MovieLens-1M.

Table 1. Statistics of Dataset.

Name	User	User Features	Feature Type	Item	Triples
Real-sg [30]	4176	701	3	340	9707
Ali [31]	1892	2354	2	1148	51,964
MovieLens-1M [32]	6050	48	4	3953	23,166

shows the statistics.

-

Real-sg [30]. Real-sg is a TCM clinical dataset from the collaborating hospital. We extract over 9000 triples as the knowledge graph for other baselines with the doctors’ instructions. In addition, we transform the age value into the age range number by dividing the age value by range (range equals 5).

-

Ali [31]. The Ali dataset is provided by the Alibaba cloud on the Tianchi platform. We adopt the dataset, which is a TCM manual dataset, into our recommendation task, whose user features, except symptoms, are less than 20.

-

MovieLens-1M [32]. MovieLens-1M is a public dataset in the common domain. We adopt the MovieLens-1M dataset for our task. We use preferred genres as users’ demand and movie records as recommendation lists (the size of records varies from dozens to thousands).

5.1.2. Baseline

We compare our model with the following methods.

-

BPRMF [33] BPRMF is a representable matrix factorization model whose loss calculation is made by dividing positive and negative item pairs. This method has not considered the graph information.

-

HERec [24] HERec is a deepwalk-based heterogeneous graph representation method for recommendation. It uses a type-aware neighbors aggregation mechanism. This method uses a type-aware strategy instead of an attention mechanism. Meanwhile, the deepwalk method is limited in graph information aggregation.

-

HetGNN+ATT [25] HetGNN is a GNN model for heterogeneous graph representation. However, this model uses only one overall heterogeneous graph. Meanwhile, it cannot distinguish the different effects of nodes. We integrate the patient portrait graph, herb–herb graph, and patient–herb graph into one heterogeneous graph for HetGNN. To improve the performance, we add attention mechanism and multi-label loss. In this way, the difference between our model and the improved HetGNN lies in patient portrait fusion modules and multigraph convolutional networks design.

-

SMGCN [27] SMGCN is a bipar-GCN model using a symptom–symptom graph and herb–herb graph. In previous models, SMGCN performed best on TCM recommendation.

In contrast, our model applies the multi-GCNs on the patient portrait heterogeneous graph and herb interaction graphs, which include the herb–herb graph and herb–patient portrait graph. In addition, we introduce the type-aware attention mechanism into our model, which is not used in SMGCN.

5.1.3. Parameter Setting

We implement our model and baselines on TensorFlow. For model training, we use the Xavier initializer [34] and Adam optimizer [35] with a batch size of 1024. The embedding size is 64. In addition, our model is constructed with 2 GCN layers. The optimal parameter settings are summarized in

Table 2. Optimal Parameter Settings.

Model	Parameters
Our Model	lr = 0.007, dropout = 0.3, regs = 0.04
HetGNN+ATT [25]	lr = 0.05, dropout = 0.2, regs = 0.02
HERec [24]	lr = 0.02, regs = 1.0, $\beta${e,h,p,w,b} = {0.1, 0.1, 2, 0.1, 0.01}
SMGCN [27]	lr = 0.01, dropout = 0.3, regs = 0.06
BPRMF [33]	$\alpha$ = 0.005, $\lambda$ = 0.01

.

To simplify the model by reducing its complexity, we set a regularization term in the model and use the dropout ratio hyperparameter to control the ratio of neurons dropped during training. Specifically, we adopt the message dropout operation introduced in the research [36] to randomly drop information during message passing. The time complexity is $O\left(Ed\right)$, where E indicates the number of all edges and d indicates the dimension of the node vector.

The averaged model training time is 45 s when we set 500 epochs.

5.2. Experiment Result

We evaluate our model on three different datasets to solve research questions 1 to 3. On the Real-sg and Ali datasets, we check the performance of our model in making herb recommendations. Furthermore, we test the generalization ability by the experiment on the common dataset MovieLens-1M. Finally, we conduct a case study to directly check the performance.

5.2.1. Quantitative Result

Experiment on TCM Recommendation

The overall experimental result in

Table 3. Experiment on Real-sg dataset [30] (TCM dataset).

	Pre@10	Pre@15	Pre@20	Rec@10	Rec@15	Rec@20	F1@10	F1@15	F1@20
BPRMF [33]	0.4427	0.4108	0.3623	0.2785	0.3876	0.4558	0.3419	0.3989	0.4037
HERec [24]	0.4553	0.4147	0.3605	0.2944	0.4015	0.4619	0.3549	0.4048	0.4019
HetGNN+ATT [25]	0.4650	0.4113	0.3678	0.3029	0.3978	0.4716	0.3668	0.4044	0.4133
SMGCN [27]	0.5058	0.4442	0.3906	0.3235	0.4271	0.4989	0.3946	0.4355	0.4382
Our model	0.5173	0.4633	0.4098	0.3305	0.4439	0.5242	0.4033	0.4534	0.4600
imp. BPRMF [33]%	16.9	12.8	13.1	18.6	14.5	15.0	17.9	13.7	13.9
imp. HERec [24]%	13.6	11.7	13.7	12.3	10.6	13.4	13.6	12.0	14.5
imp. HetGNN+ATT [25]%	11.2	12.6	11.4	9.10	11.6	11.2	9.94	12.1	11.3
imp. SMGCN [27]%	2.27	4.31	4.93	2.16	3.92	5.07	2.20	4.11	4.99

demonstrates that our model outperforms baselines on the Real-sg dataset (in this table, imp. in the table indicates improved rate). In Figure 6, the superiority of our model is observed more clearly.

First, the models based on the graphs outperform matrix factorization, thereby confirming that graph information enriches the representation of entities. The proposal of models based on graphs, such as HERec, aims to enrich entity representation by collecting graph information. However, HERec is very weak in the aggregation of high-order information because it is based on the deepwalk method. Our model, SMGCN, and HetGNN+ATT outperform HERec, thereby confirming that the GCN captures more comprehensive information on herb interactions, which is because GCN plays a vital role in graph representation with special message-passing mechanism. Meanwhile, the result of HetGNN+ATT is worse than SMGCN and our model, confirming that using multiple graphs and adopting multigraph convolutional networks are beneficial in herb recommendation. This is probably because the multi-GCN framework can learn a more flexible and expressive model to some extent compared with the unified graph-based GCN.

Additionally, our model achieves the state of the art on the clinical dataset, which demonstrates that our design is suitable for herb recommendation task. This model outperforms SMGCN, which confirms that the patient portrait and herb-interaction fusion module and the type-aware message-passing mechanism are effective in graph representation learning.

In addition, the overall result initially confirms that the patient portrait is essential in TCM recommendation. We conduct further experimentation on the influence of patient portrait in the third section of the quantitative experiment.

Table 4. Experiment on Ali dataset [31] (TCM dataset).

	Pre@10	Pre@15	Pre@20	Rec@10	Rec@15	Rec@20	F1@10	F1@15	F1@20
BPRMF [33]	0.3825	0.3007	0.2548	0.3576	0.4217	0.4763	0.3697	0.3511	0.3319
HERec [24]	0.3758	0.2903	0.2421	0.3836	0.4356	0.4818	0.3538	0.3484	0.3223
HetGNN+ATT [25]	0.4073	0.3652	0.3025	0.3708	0.5159	0.5521	0.3882	0.4276	0.3908
SMGCN [27]	0.4448	0.3719	0.3104	0.4379	0.5327	0.5706	0.4413	0.4380	0.4021
Our model	0.4449	0.3749	0.3112	0.4379	0.5369	0.5740	0.4414	0.4415	0.4036

confirms the effectiveness of our model on the Ali dataset.

Our model still outperforms baselines. As the comparison of the experimental results in Table 3 and Table 4 shows, all the models perform worse on the Ali dataset. The main reason should be the sparsity of data. Although the number of triples is larger than that in the other datasets, many entities have only one triple. Thus, the accuracy of the neural network and matrix factorization prediction models declines under such long-tailed and sparse training data.

The second finding is that our model was not much more improved over SMGCN. This is because the Ali dataset has too few personal features (fewer than 0.8% in all user features), except for symptoms. Therefore, the patient portrait modeling and the type-aware message-passing mechanism do not work best under this condition.

Experiment on Common Dataset

In this experiment, we tested our model on the MovieLen-1M dataset.

Table 5. Experiment on MovieLens-1M dataset [32] (Common dataset).

	Pre@30	Pre@40	Pre@50	Rec@30	Rec@40	Rec@50	F1@30	F1@40	F1@50
BPRMF [33]	0.1827	0.1744	0.1692	0.0320	0.0408	0.0495	0.0545	0.0661	0.0765
HERec [24]	0.1890	0.1793	0.1692	0.0411	0.0510	0.0597	0.0580	0.0672	0.0735
HetGNN+ATT [25]	0.4679	0.4521	0.4378	0.0440	0.0619	0.0782	0.0803	0.1088	0.1327
SMGCN [27]	0.4699	0.4561	0.4430	0.0434	0.0619	0.0801	0.0794	0.1090	0.13567
Our model	0.4722	0.4567	0.4435	0.0442	0.0623	0.0801	0.0808	0.1096	0.13568

shows that the recall and f1-score decrease. The reason, mentioned in the dataset description, is that the movie list size for specific users varies from dozens to thousands. The ground-truth set for each user is larger than that of the other datasets, so we adjust the parameter N (size of recommended set).

Our model still outperforms most baselines, thereby verifying its applicability in a common domain. Our model, SMGCN, and HetGNN+ATT outperform the others, thus confirming the superiority of GCN in graph representation learning.

In addition, the MovieLens-1M dataset is asymmetric in the user feature set and item set. The movie set is much more extensive than the user features set. This difference may be the main reason why BPRMF and HERec cannot work well. The result also indicates that the matrix factorization and deepwalk do not fit in such a dataset.

Experiment on Patient Features’ Influence

In this section, we discuss whether the patient portrait improves the performance of herb recommendations. In

Table 6. Experiment of Patient Features’ Influence on Real-sg dataset [30].

Patient Features	F1@10	F1@15	F1@20
Base(Symptoms)	0.37645	0.40738	0.41694
+Age and Gender	0.38090	0.41252	0.42418
+Age and Gender+ Diseases	0.40330	0.45340	0.46004

, “+ Age and Gender” means that these features, in addition to symptoms, are involved in the patient portrait in the second experiment. Furthermore, “+ Age and Gender+ Diseases” means that diseases, in addition to personal properties and symptoms, are involved in the patient portrait.

The findings show that regardless of personal properties or diagnosis, the patient portrait is beneficial in our task. Thus, the results confirm that patient portraits have a strong connection with treatment decisions; this finding is realistic. For instance, prescriptions are usually different for people of different ages. Elderly people often take warm tonic Chinese medicines for health, while children have a body of pure “Yang" and generally cannot utilize these medicines.

Ablation Experiment

In this section, we perform an ablation experiment to determine whether the type-aware attention mechanism improves the performance in our model.

The experimental results in

Table 7. Ablation Experiment of Attention on Real-sg dataset [30].

	F1@10	F1@15	F1@20
without ATT	0.39841	0.44150	0.44720
with ATT	0.40330	0.45340	0.46004

show that the employment of the type-aware attention mechanism improves the effect of the model. This is mainly because the mechanism distinguishes the effects of different node types, thereby improving the message passing and aggregation. The mechanism plays a positive role in the heterogeneous graph modeling. This effect has also been confirmed in other studies in the field of graph modeling. For example, the research of HERec takes type-aware strategy on deepwalk method, in which the benefit of distinguishing different nodes is also demonstrated.

5.2.2. Qualitative Result

We will now discuss the performance and output of our model intuitively. This test evaluates our model by comparing the output with the SMGCN to show the advantage of our model. We select the instances in

Table 8. Case Study of our model and SMGCN [27]. (The herb names are hard to read and understand, so we use the color and bold form to stress the right results of prediction.)

Ground Truth	PMGCN	SMGCN [27]
hedyotis diffusa	hedyotis diffusa	hedyotis diffusa
cortex dictamni	cortex dictamni	cortex dictamni
rhizoma anemarrhenae	rhizoma anemarrhenae	rhizoma anemarrhenae
baikal skullcap	baikal skullcap	baikal skullcap
cortex moutan	cortex moutan	cortex moutan
sophora flavescens	sophora flavescens	sophora flavescens
dandelion	dandelion	dandelion
coptis chinensis	coptis chinensis	coptis chinensis
smilax glabra	smilax glabra	smilax glabra
kochia scoparia	kochia scoparia	angelica sinensis
puncture vine	puncture vine	asparagus cochinchinensis
chrysanthemum	honeysuckle	honeysuckle
folium mori	polygonatum odoratum	polygonatum odoratum
pyrrosia lingua	rehmannia glutinosa	rehmannia glutinosa
	the root of red-rooted salvia	the root of red-rooted salvia
	ligustrum lucidum	ligustrum lucidum
	phellodendron	phellodendron
	folium isatidis	folium isatidis
	radix scrophulariae	radix scrophulariae
	selfheal	white atractylodes rhizome

of diseases that occur frequently. The input feature set of the models is {female, 81, wind-heat, accumulated dampness-heat, eczema, uraemia pruritus}. The content in bold is the correct result in this prescription.

The table shows that our model outperforms SMGCN. Furthermore, doctors confirm that the recommended results are reasonable.

6. Conclusions

In this paper, we investigate herb recommendations from the perspective of considering patient portrait modeling. We use type-aware multi-GCNs on patient portrait and herb interaction graphs to enrich the representation of patients and herbs. Specifically, we develop the patient portrait modeling and herb interaction modeling to extract comprehensive interactions among patients’ features and herbs. Furthermore, we aggregate different neighbors’ information by adopting type-aware attention for patient portrait and herb interaction modeling. The extensive experiments validate the effectiveness of our design and demonstrate its superiority. The main contribution is that it proves the significance of patient portraits in herb recommendations and proposes a graph based patient modeling method innovatively. The second innovation of this research work lies in the adoption of type-aware attention design in multigraph convolutional networks, which distinguishes the effect of different nodes in graph learning and improves the message passing for patients’ and herbs’ graph learning in TCM recommendation.

In fact, this work aims to comprehensively consider global information to model local prescription information. In another perspective, it can be understood as the discovery of associations between entities such as herbs and diseases. So, this research can be extended to solve the drug discovery problems, and the results can be helpful in field of herb detection in TCM. Meanwhile, this research has practical significance. As mentioned above, the inheritance of traditional Chinese medicine is difficult, because it is hard to summarize the experience of famous and old doctors. Studies on herb recommendation systems could help mitigate this issue.

Future Work

For herb recommendation models, most existing work, including our method, ignores the profound principles of traditional Chinese medicine in treating diseases. In future work, we will introduce the chemical information of herbs and adopt an attribution analysis method to improve our model.

In TCM theories, the roles of herbs in prescription are different and significant for TCM compatibility. Meanwhile, our study only captures the interactions between herbs and patient features, and have not considered the roles and effect of different herbs. With the help of doctors, we will introduce more reliable knowledge and classify the different roles of herbs in each prescription.

TCM recommendation studies are incomplete because they only make recommendations on herbs and ignore the dosage, which is important in prescription making. Recently, we studied on the dosage prediction task and made a preliminary exploration by adopting ensemble learning methods to incorporate multiple regression models.