DaGzang: a synthetic data generator for cross-domain recommendation services

DaGzang architecture

DaGzang was implemented in the Java programming language that contains the web service in the front end and background service in the back end. In the background service, we deployed the MySQL and Java server pages (JSP) with the Spring Framework. Then, we applied the model-view-controller (MVC) model to handle the large datasets generated by multiple thread processing. In addition, our system has a simple process and exemplary performance in extensive data analysis geared toward adaptivity with many datasets and users. The Spring Web MVC framework is designed around a DispatcherServlet that handles all the HTTP requests and responses. The request processing workflow of the Spring Web MVC DispatcherServlet is illustrated in the following diagram.

As shown in Fig. 1, the sequence of events corresponding to an incoming HTTP request to the DispatcherServlet is described as follows: (1) The DispatcherServlet consults the HandlerMapping for calling the Controller after receiving an HTTP request. (2) The Controller receives the request and calls the service methods (GET or POST) that set model data based on defined business logic and returns the view name to the DispatcherServlet. (3) The DispatcherServlet interacts with the ViewResolver to select the defined view for the request. (4) The DispatcherServlet proceeds to transfer the model data to the view that is finally shown on the browser.

In designing the front end of DaGzang, the interaction with users is carefully considered to construct the simplest process for users to generate suitable datasets. Specifically, our system is web-based, and reducing latency in every process is the most important goal. Furthermore, to prevent the loading of all datasets in the system, DaGzang applies the pagination method in the displayed results and follows three rules in designing the user interface (Mandel, 1997), which are (i) placing users in control, (ii) decreasing user memory loading, and (iii) making a consistent interface. Following these rules, we use one template for all pages to guarantee the consistency of the user interface. Therefore, regardless of their actions, people only remember the same signs (e.g., fonts, keys, and items).

Consequently, user comfort with the system is paramount. The interface is also the same for all actions (e.g., using the same approach to view results). The system also clearly describes the processes for generating the data, allowing users to easily recognize where step they are staying, what they are processing, and what they need to do next. In addition, DaGzang provides users with roll-back procedures to return to past steps if they have incorrect actions.

Figure 2 shows the interface of the data generation function in our DaGzang system. Based on each purpose, the users can decide which parameters are injected from the list in the configure functions on the left. This list of parameters includes the number of users, items, and similar users as well as the sparsity, distribution, scale, and selection of the data. All these parameters were depicted as follows.

• Number of users or items: number of users or items in the output dataset.

• Data sparsity: how many rating values are empty in the output dataset.

• Scale: the range value of ratings.

• Data distribution: this parameter decides which distribution the output dataset has.

• Dataset selection: this parameter is used to select a real-world dataset to extract the rating patterns that can transfer to the output dataset. This function is used when the user wants to generate datasets based on a specific domain feature.

• Number of common users: This parameter indicates how many common users are between a real-world dataset and the output dataset. Particularly, the system uses this parameter to calculate the number of common users in the synthetic dataset corresponding with the rating patterns extracted from the input dataset.

Then, when the user clicks on the generate data button, our data generation algorithms will perform the calculations and display the synthetic data results in the output screen on the bottom right.

DaGzang functionalities

The key function of DaGzang is to generate synthetic datasets according to the users’ configurations and adapt to their demands. Creating synthetic datasets involves three main steps: (i) extracting data patterns (i.e., statistical distributions, sparsity, and rating patterns) from the selected real-world datasets, (ii) injecting user-customized parameters, and (iii) executing the generator based on these customized parameters combined with the extracted data patterns.

To generate a realistic synthetic dataset, we integrate all the steps described above and construct the algorithm. Firstly, the real-world dataset was extracted pattern (rating patterns, statistical distributions, sparsity) and set as initial input. This step aims to create the initial dataset and increases them depending on user purposes according to the pattern extracted from the real-world dataset. In the next step, the ratings for all users are randomly generated for this dataset in order to construct the user preferences. Then, in the last step, the domain adaptation was used to map the preferences between common/similar users. In this step, the rating was updated again based on the injecting user-customized parameters combined with the extracted patterns (ratings) in the first step. In particular, given a real-world dataset, by $\mathbf{A}=\left\{\mathit{D},\mathit{S},{\mathit{R}}_p\right\}$, we indicate the set of all attributes, where D and S represent the distribution and sparsity, respectively; R_p denotes the rating pattern of the real-world dataset. Let I_l with $lϵ\left(1,L\right)$ be the set of all possible items in the l-th domain. For each user, there is a maximum of n ratings for n items in domain I_l . Then, it is possible to denote a set of ratings in I_l of the form ${\mathit{R}}_l=\left\{r_{l_1},r_{l_2},\dots ,r_{l_n}\right\}$, where r_ln is the rating of the user for the n-th item in the domain I_l. Therefore, for each user, we have two sets of the item and rating represented by I_l and R_l, respectively, such that item_rating = $\left\{{\mathit{I}}_{\mathit{l}},{\mathit{R}}_{\mathit{l}}\right\}$. Algorithm 1 presents the procedures to generate synthetic data that is described as follows

1. Initial extraction of the attributes from a real-world datasets to get A.

2. Generate randomly for a set of item_rating = $\left\{{\mathit{I}}_{\mathit{l}},{\mathit{R}}_{\mathit{l}}\right\}$ for all users.

3. Define number of ratings will be created in the synthetic dataset based on distribution D and sparsity S.

4. Generate dataset based on parameters from step 2 and 3.

5. Update ratings for sh_u users.

In order to allow users to customize the output dataset, the system provides several options consisting of the data distribution based on the user-defined, domain-specific feature data and real-world input data models. The user-defined data distribution allows the end-user to select the desired data distribution to generate, while domain-specific feature data allows the user to choose the domain-specific data that the system is expected to generate based on the specific domain.

The real-world input data model option directs the system to extract the data model from real-world input data to create the synthetic data generation rule. Adapting to these requirements, DaGzang is designed with two separate modules. The first is the module to generate synthetic datasets with user input parameters. The second module is a pool that stores many existing well-known datasets, allowing users to choose between using an existing dataset and creating a new one for themselves.

Experimental setup

To conduct the experiments, we first collect 13 real-world datasets from Amazon (https://jmcauley.ucsd.edu/data/amazon/) and insert them into the DaGzang platform. Then, we classify these datasets into different types of the domain according to the information relevant to these datasets. The detailed information of these datasets is shown in Table 3

DaGzang extracts the overlap association between these real-world datasets in the next step. This overlap association is defined as the rating pattern of users and is labeled in our database. We then generate the synthetic dataset by combining this overlap with some parameter configurations in the input of DaGzang, and the platform loops this step according to the parameter changes to obtain a list of synthetic datasets. For this experiment, we generate 90 synthetic datasets with several parameter configurations, as illustrated in Table 4. Finally, each synthetic dataset generated from DaGzang is used for the recommendation functions in DakGalBi (the CF CDRS we built). The recommendation methods used for this experiment are the CF, including the item-based CF, user-based CF, and user-item (matrix factorization) CF. The mean absolute error (MAE) and root mean squared error (RMSE) metrics were used to calculate the accuracy of these output recommendations. The formulations of MAE and RMSE are described as follows: (1)${\displaystyle \mathbf{MAE}=\frac{1}{n}\sum _{i=1}^n\left|y_i-y_i^p\right|\mathrm{and}\mathbf{RMSE}=\sqrt{\frac{1}{n}\sum _{i=1}^n{\left(y_i-y_i^p\right)}^2}}$ where n is the number of items, y_i is the real ratings and $y_i^p$ is the predicted ratings. The results of MAE and RMSE range from 0 to infinity. Infinity is the maximum error according to the scale of the measured values.

We first use 10 of 90 datasets generated from DaGzang, including multiple domains for each experiment, to calculate the MAE and RMSE. The number of datasets increases each time, running ten until the entire dataset consists of 90. The experimental results are used to evaluate the accuracy of recommendation methods in DakGalBi CDRS. We then discuss the efficiency and suitability of these generated synthetic datasets with the CF CDRS to clarify whether DaGzang adapts to our purpose. To accomplish this, we use the bipartite graph and binary search to find the best matching. We observed that the number of common users in the rating patterns represented the number of common users in the synthetic and real datasets. In other words, these values show how much overlap exists between these datasets. Besides, we compare the proposed method with another, CART (Drechsler & Reiter, 2011). For this experiment, we use the 10 synthetic datasets generated from each method. Then we use these two datasets to input the DakGalBi cross-domain recommendation system with CF-based recommendation methods. Finally, we calculated the MAE and RMSE each time running the recommendations from DakGalBi. All experimental results and discussions are presented in the next section.

Experimental results and discussion

After generating 90 datasets following the randomization of the changing parameters shown in Table 4, we separate these datasets into nine parts consisting of [1-10], [1-20], [1-30], [1-40], [1-50], [1-60], [1-70], [1-80], and [1-90]. Nine groups of datasets are provided for the second experiment embedded in the DakGalBi CDRS to build the recommendations based on CF algorithms. The results of the second experiment are shown in Table 5. As shown in the experimental results, as the number of synthetic datasets increases, the overlap association between domains decreases and the accuracy of the recommendation methods is also reduced. Following each metric, MAE and RMSE, we can see in Figs. 3 and 4, the errors increase depending on the number of synthetic data used. In most cases from 10 to 90 synthetic datasets, these results are linear.

However, considering the case of ten synthetic datasets, the performance of the recommendation methods is very impressive, even with the traditional user-based CF method. This demonstrates that the proposed method to generate a synthetic dataset in DaGzang is suitable for evaluating the recommendation algorithms. Furthermore, we deploy the third experiment using the bipartite graph and binary search to clarify the overlap between these synthetic datasets. The numerical results of this experiment are shown in Table 6. In addition, for the last experiment that compared the proposed method with CART, the proposed method showed outperformance. This experimental result is described in Table 7

The experimental results show that the generated synthetic datasets adapt to the evaluation purpose. The datasets generated from DaGzang can be used to evaluate CDRS algorithms, although the accuracy is not sufficiently high. However, the main purpose of this study is to overcome the scarcity of collected real datasets for the CDRS. DaGzang provides functions to handle requirements from users based on several scenario purposes, such as (i) generating datasets for single-domain, (ii) multi-domain, and (iii) cross-domain implementations. To accomplish this, in DaGzang, we allow users to configure the parameters according to their purposes, highlighting our platform’s critical performance.