2.1 Live Interactive Behavior Model

During the live streaming process on short video platforms, user interactions with a network anchor include behaviors such as sending comments, giving rewards, etc. By recording a user's interaction history ^[7], a behavior model can be constructed for the user's behavior in different live streams. Based on the behavior model, the user's interest in the network anchor can be determined, which provides a basis for intelligent recommendations ^[8].

In live streaming on short video platforms, users have diverse needs, and the main means of monetizing user traffic is giving rewards. The comments sent by users while interacting with the network anchor can reflect the user's level of interest in the network anchor to some extent. But the number of words in the bullet comments is usually small, and the interest and emotional tendencies contained are also relatively fragmented, making it difficult to quantify and fully reflect the user's level of interest in the network anchor ^[9].

Rewards are a type of positive feedback from users to the content released by a network anchor, and the amount and frequency of rewards can be quantitatively and statistically measured. In addition, users' viewing time for a network anchor can also reflect their level of interest in that network anchor and can also be easily quantified and statistically measured ^[10]. Therefore, we used indicators such as reward behavior and viewing time to construct the model of a user's live interactive behavior. The corresponding formula is:

where $R_{ij}$ is the interval between the time point of the user's latest reward interaction with network anchor $j$ and the statistical time point $d_{i,end}$ of user $i$, and $d_{i,start}$ is the start time point of user $i$'s behavior record selected for the interaction model construction ^[11]. $d_{ij,r}$ is the time point of user $i$'s latest reward interaction with network anchor $j$, $F_{ij}$ is the frequency of user $i$'s rewarding interaction with network anchor $j$ in the selected time period, $M_{ij}$ is the total amount of user $i$'s reward to network anchor $i$, and $m_{ij,d}$ is the amount of user $i$'s reward to the network anchor at time point $d$. $T_{ij}$ is the total duration of user $i$ watching the live streaming of network anchor $j$ ^[12], and $t_{ij,d}$ is the duration of user $i$ watching the living streaming of network anchor $j$ at time point $d$. The value of $a_{ij,d}$ is 1 if user $i$ rewards network anchor $j$ at time point $d$; otherwise, it is 0.

2.2 Recommendation Algorithm based on Live Interactive Behavior Model

We used collaborative filtering ^[13] to recommend network anchors to users. However, when constructing the rating matrix in the collaborative filtering algorithm, using only the viewing time of live broadcasts as a measure is not comprehensive enough. Therefore, we used the live interactive behavior model to calculate the ratings in a rating matrix ^[14]. The basic flow of the recommendation algorithm is shown in Fig. 1.

Fig. 1. The basic process of the recommendation algorithm based on the live interactive behavior model.

A web crawler is used to retrieve the viewing history of users on a short video platform for network anchors from the interface ^[15]. The corresponding indicators that can measure a user's preference for network anchors are calculated based on Eq. (1) to construct the user's live interactive behavior model. The rating value of each user for different network anchors is calculated based on the corresponding indicators in the user's live interactive behavior model to construct the rating matrix between users and network anchors ^[16]. The rating calculation formula for the matrix is:

where $S_{ij}$ is the comprehensive score for network anchor $j$ of user $i$. $R'_{ij},F'_{ij},M'_{ij}$, and $T'_{ij}$ represent the standardized indicators of the live interaction behavior model, and $\omega _{r},\omega _{f},\omega _{m}$, and $\omega _{t}$ represent the weights of the corresponding standardized indicators. In the rating matrix, if user $i$ has not watched the live streaming of network anchor $j$, $S_{ij}$ is directly denoted as $NA$ without using Eq. (2) for calculation.

The similarity between different users is calculated by the rating matrix ^[17]:

where $sim(x,y)$ represents the similarity between users $x$ and $y$, and $\boldsymbol{x}$ and $\boldsymbol{y}$ represent the rating vectors of users $x$ and $y$. The dimensions of the rating vector depend on the number of network anchors on the platform. Each dimension of the rating vector represents a network anchor on the platform, and its value represents the rating given by the user to that network anchor ^[18].

A set of network anchors is selected from the nearest neighbor set that does not overlap with the historical records of the target user. Then, the predicted rating of the target user for the selected network anchors is calculated based on the rating given by the nearest neighbors to the selected network anchor set and the similarity between the target user and the nearest neighbors. The calculation formula is:

where $P_{x,j}$ is the predicted rating of user $x$ for network anchor $j$, $\overline{r}_{x}$ and $\overline{r}_{y}$ are the average ratings of users $x$ and $y$ for all network anchors, and $r_{y,j}$ is the rating of user $y$ for network anchor $j$, where user $y$ belongs to the nearest neighbor user set $KNN(x)$ of target user $x$. For target user $x$, network anchors selected from $KNN(x)$ are ranked from high to low according to the predicted rating. The top $N$ network anchors are recommended to target user $x$.

3.1 Experimental Data

This study focused on the TikTok short video platform, where web crawlers were used on users' historical records. The crawler software finds the data from the platform through the external interface of the platform. TikTok is a short video community platform for all ages built by ByteDance. On this platform, users can select songs, combine them with their own videos, and publish their work. In addition to uploading short videos, TikTok also provides live streaming functions.

A user can choose to watch multiple network anchors, and the number of users and network anchors on TikTok is enormous, so the amount of data a crawler can obtain is also enormous ^[19]. If all the historical data of all users on the platform were calculated and processed, the data volume would be too large. Therefore, this study only used the crawled historical records of 800 users as experimental data. Also, to avoid privacy issues, sensitive information was deleted from the data and replaced. The data were collected from January 1, 2022 to December 31, 2022.

Due to limited space, only a portion of the user's historical records are displayed, as shown in Table 1. The data include the user identify (ID), the ID of the corresponding network anchor that the user watched, the type of live content, and the four indicators of the live interaction behavior model. Fig. 2 shows a screenshot of a live show. The part marked by the red box is the reward function. A user can reward the network anchor, and the network anchor can obtain a proportional share of the amount after receiving the props.

Fig. 2. Screenshot of a live show with an interactive function.

3.2 Experimental Design

In the process of using the collaborative filtering algorithm, the top $k$ nearest neighbors with the highest similarity to the target user were selected for screening. The recommendation performance of the proposed algorithm was tested under different $k$ values of 5, 10, 15, 20, and 25 When mining association rules with an association rule-based recommendation algorithm, the support and confidence values are required ^[20]. In this study, they were set through orthogonal experiment as 0.1 and 0.5, respectively. The orthogonal experiment compared the recommendation performance of association rule recommendation algorithms under different combinations of support and confidence, and the best combination was selected. For the collaborative filtering-based recommendation algorithm, only the user's viewing time was used to construct the rating matrix, and the number of nearest neighbors k was set as 15. The proposed recommendation algorithm also used collaborative filtering, but the live interactive behavior model was used to calculate the rating. $k$ was also set as 15 in this case.

The performance of the algorithm was measured using the precision, recall rate, and F value. The confusion matrix is shown in Table 2. The corresponding calculation formulas are:

where $P$ stands for precision, $R$ stands for the recall rate, and $F$ stands for the harmonic mean of precision and recall rate.

3.3 Experimental Results

Fig. 3 shows the performance of the combined recommendation algorithm with different numbers of nearest neighbors. When the number was set as 15, the precision, recall rate, and F-value of the algorithm were the highest (89.6\%, 88.5\%, and 87.4\%, respectively). The reason is that the similarity between the nearest neighbors and the target user was high. As the number of nearest neighbors increases, target users can receive more network anchor recommendations from the nearest neighbors, making it easier to find network anchors that match their interests. However, the number of users with high similarity to a target user is limited. If the number of nearest neighbors continues to increase, the number of users with different interests will also increase.

Table 3 shows some specific recommended content for some users. For the same user, the recommendations of the algorithms were different. The recommended results of the association rule recommendation algorithm had a higher level of overlap, while the recommended results of the other two recommendation algorithms were very different. In addition, when comparing the category to which each recommended network anchor belongs (i.e., the main live content), the recommendation algorithm based on association rules had a large difference in the recommended network anchor type for each user. However, the recommendation algorithm based on collaborative filtering had a small difference in the recommended network anchor type with only one or two differences. The combined recommendation algorithm provided relatively consistent recommendations for each user.

Fig. 4 shows the performance comparison results of the association rule-based, collaborative filtering-based, and live interactive behavior model-based recommendation algorithms. Table 4 shows the significance of the differences in the performance indicators among the three recommendation algorithms. The p value indicates the significance level of the difference between two statistics. The smaller the p value is, the larger the statistical difference will be. Generally, p {\textless} 0.05 indicates that two statistics have a significant difference. For each performance indicator, the p value between two of the three recommendation algorithms was less than 0.05, indicating that the differences in the performance indicators were significant. Fig. 4 and Table 4 show that the recommendation performance of the live interactive behavior model-based algorithm was the best, followed by the collaborative filtering-based algorithm, while the association rule-based algorithm performed the worst.

Fig. 3. Performance of the combined recommendation algorithm with different numbers of nearest neighbors.

Fig. 4. Performance comparison of three recommen-dation algorithms.