GACE: Learning Graph-Based Cross-Page Ads Embedding For Click-Through Rate Prediction

The CTR prediction task in online advertising recommendation systems is to predict the click probability of a given user for a given advertisement, which plays an increasingly important role in various personalized online advertising recommendation services. In recent years, deep learning methods based on deep neural networks can interact and represent features in complex ways, and a series of deep neural network models have been proposed. The Wide&Deep[4] model attempts to capture high-level feature interactions through a simple multilayer perceptron (MLP)[9] network. DeepFM[11] and DCN [27] are proposed to handle complex feature interactions based on product operations. AutoInt[22] uses the Multi-head Self-Attention mechanism to produce high-level composite features. In addition, in the scenario with sequence characteristics, a series of neural networks based on sequence characteristics are proposed: Deep Interest Network (DIN)[30], Behavior Sequence Transformer (BST)[3], Deep Session Interest Network (DSIN)[8] and Bert4Rec[24] etc.

However, the performance of these models depends mainly on the quality of the input item and user embedding and cannot effectively solve the cold start problem of new advertisements. They often fail to perform satisfactorily in the sparse page scenario or when distributing ads not in the training set.

For item embedding, the first concept is to establish a one-hot vector for item id representation. The embedding learning methods based on collaborative filtering and matrix decomposition were subsequently proposed, but these methods have weak generalization ability and poor performance for items with sparse historical behavior. Researchers try to extend the concept of word embedding in the field of natural language process (NLP)[5] to the recommendation field, and propose Item2Vec[1] based on the concept of Word2Vec [6]. With the development of graph neural networks, embedding learning methods based on graph networks have also been proposed, such as DeepWalk[19], GraphSage[12], etc. However, new advertising embedding without historical interaction cannot be learned via these methods. In particular, for the embedded learning of cold-start new items, our standard solution is to put forward a series of neighbor embedding methods based on manifold learning. Some researchers have put forward some generative methods from the perspective of meta-learning, such as the MetaEmbedding model[21], autodis[10], etc. However, these methods only play a role in learning the embedded representation of new advertisements, and it is difficult to improve the recommendation efficiency of existing old items.

The task of an online advertising system is to establish a prediction model to estimate the click probability of a specific user for a specific ad. Each instance includes multiple-field information: user information (’User ID’, ’City’, ’Age’, etc.), item knowledge base, combined with the label from historical user interaction feedback.

We proposed a graph-based cross-page ads embedding learning method (GACE) by fully exploiting useful information in the item knowledge base. Our model contains two steps: graph creation and pre-training embedding learning. In the graph creation step, three entities’ embedding vectors of each item are extracted or encoded from the item knowledge base, representing latent knowledge in each domain. Then we established a weighted undirected graph structure based on the semantic knowledge and page knowledge and set the splicing of entities embeddings in the item knowledge base as the initial embedding for the graph node. In the pre-training embedding learning step, we proposed a pre-training task based on an improved graph auto-encoder for the weighted undirected graph established in the first step to learn the potential embedding representation of each advertising node in the graph. Accordingly, we obtain the optimization parameters of the ad embedding encoder and the potential vector representation of ad items. The graph creation and pre-training embedding learning steps will be further discussed in sections 4.2 and 4.3.

Typically, for a graph $\mathcal{G=(V,E)}$, where $\mathcal{V}=\big{\{}v_{1},\ldots,v_{n}\big{\}}$ is a set of ad nodes and $\mathcal{E}$ is a set of edges with weightings, we combined and set the splicing of three entities’ embeddings in the item knowledge base as the ad node’s initial vector $v$. We proposed an adjacent weighting matrix $\mathbf{A}\in\mathbb{R}^{n\times n}$, where $\mathbf{A}_{i,j}\geq 0$ to represent the graph structure. If $\mathbf{A}_{i,j}>0$, then there is an edge between ads item $i$ and ads item $j$, where $\mathbf{A}_{i,j}$ represents the edge weighting. If $\mathbf{A}_{i,j}=0$, then there is no connection between ads item $i$ and ads item $j$. The Adjacent weighting matrix is calculated based on semantic knowledge and page knowledge. Since the semantic knowledge of advertising is mainly composed of sentences, here we use the output of the sentence transformer[7] as the semantic knowledge vectors $\mathbf{s}\in\mathbb{R}^{k}$. As mentioned in section 2.1.3, we count the total number of ads placed on specific page channels and aggregate the average value of user interaction knowledge of historical ads placed on a specific page channel as Page knowledge vectors $\mathbf{s}\in\mathbb{R}^{d}$. The graph creation process is shown in Fig 2.

We use dot product as the semantic knowledge similarity $\alpha$ and page similarity $\beta$, and calculate the advertising weight matrix $\mathbf{A}$ based on $\alpha$ and $\beta$ as follows:

where $\alpha_{(}i,j)$ refers semantic knowledge similarity between ads item $i$ and ads item $j$, $s_{i}$ and $s_{j}$ are semantic knowledge vectors for ads item $i$ and ads item $j$. Similarly, $\beta_{(}i,j)$ is page similarity between ads item $i$ and ads item $j$, $p_{i}$ and $p_{j}$ are Page knowledge vectors for ads item $i$ and ads item $j$. $ReLU$ is a non-linear activation function.

In this section, to integrate the knowledge of different entities of ads nodes and their neighbors, we designed a variational graph auto-encoder for an elaborately designed self-encode task as a pre-training module to recover the undirected weighting graph structure, as shown in Fig 3. After the pre-training task, the encoder can generate embedding representations $Z$ for new and old ads.

Encoder: We designed an encoder based on variational encoder and graph attention neural network, which can adaptively adjust the weighting between ads nodes. Latent nodes vector $z_{i}$ for ads item $i$ are further introduced as:

where $x_{i}\in\mathbb{R}^{F}$. is the node vector of ads item $i$, and $F$ is the dimension of each node feature. $\mathbf{W}\in\mathbb{R}^{F^{\prime}\times F}$ is the trainable weighting matrix to distill useful information, and $F^{\prime}$ is the transformation size. $\mathcal{N}(i)$ is the set of neighbors of node $i$, and the cross-page attention weighting $\gamma_{i,j}$ is calculated as:

where $LeakyReLU$ and $softmax$ are non-linear activation functions. A feed-forward neural network is proposed here and is parameterized by a weight vector $a\in\mathbb{R}^{2F^{\prime}}$. $\mathbf{\Theta}_{i,j}$ is the enhanced and normalized edge weighting between ads item $i$ and ads item $j$.

Decoder: For non-probabilistic variant of the GAE model, we reconstruct the adjacent weighting matrix $\mathbf{\hat{A}}$ as:

where $\sigma(\cdot)$ is the ReLU non-linear activation function.

Learning: For GACE pre-training, we designed $\mathcal{L}_{r}$ to optimize the reconstruction of the adjacent weighting structure, and $\mathcal{L}_{n}$ ensures that the generated vector follows the normal distribution.

where $KL[q(\cdot)||p(\cdot)]$ is the Kullback-Leibler divergence[13] between distribution $q(\cdot)$ and $p(\cdot)$. We take $p(Z)=\prod\nolimits_{i}p(z_{i})$ as a Gaussian prior, and take Kullback-Leibler divergence for $\mathcal{L}_{r}$ as the reconstruction loss to retain the weighting distribution information. We perform full-batch gradient descent[23] and use the reparameterization trick[28] for training.

Considering cross-page information is significant, we elaborately selected the public datasets that have page-level information. Experiments are conducted on the public dataset AliEC[25] from RecBole[29] which has two pages (regarded as different resources PageID (PID) location) and a real-world CTR prediction dataset collected from three pages of Alipay.

We designed offline comparison experiments, evaluated performance on different pages respectively, and then conducted an online A/B test. The statistics of experimental datasets are shown in Table 1

In this section, we conduct model comparison experiments on the public AliEC Dataset and evaluate model performance on two different page scenarios.

Table 2 shows the results from the public AliEC Dataset. Models with deep feature interactions perform better than the original wide&deep models. GACEEmb stands out significantly among all the other item-embedding competitors, which shows in the two pages evaluated results respectively. We owe this to the neighbor information aggregation generation mechanism considering the similar content and similar distribution pages. Through improved variational graph auto-encoding, GACE obtained the enhanced representation of ad items.

In this section, we conducted offline comparison experiments and evaluate model performance on different Alipay pages respectively. Typically, we evaluated the impact of the GACE embedding generation framework on the recommendation effect of new and old ads in different CTR prediction models.

In order to address the limitations of offline evaluation and demonstrate the practical value of GACE, we further deployed the GACE model in an actual online environment, that owns tens of millions of users access it daily.

An online A/B test is designed to further evaluate the performance of GACE. We conducted a week-long rigorous A/B testing with three objectives: to cover more users, cover more cold start items, and collect more reliable results. Bert4Reg with N2VEmb is the main model already serving in the recommendation system. We allocated 10% of Bert4Rec traffic with N2VEmb as the baseline and compared 10% of Belt4Rec traffic with GACEEmb. We evaluated the new and old items that appeared this week separately.

Table 5 illustrates the cumulative relative improvement of the experimental model compared to the baseline model. The average CTR of using GACE has increased by 3.61%, 2.13%, and 3.02% respectively. For cold-start ads distribution, the CTR has increased by 9.96%, 7.51%, and 8.97% respectively. It is worth noting that both improvements have statistical significance (p-value less than 0.05). This practice-oriented experiment demonstrates the effectiveness of our model in real-world recommendation scenarios.