腾讯 wxg 应用研究员: 微信游戏推荐技术

Author

Reads 458

African American woman working on a laptop in a modern office setting, emphasizing professional and technology themes.
Credit: pexels.com, African American woman working on a laptop in a modern office setting, emphasizing professional and technology themes.

Meet the Tencent wxg Application Researcher, who's working on a revolutionary technology to recommend games within the popular WeChat platform. This expert is responsible for developing algorithms that can learn users' gaming preferences and suggest new games they might enjoy.

As a gamer myself, I can attest to the frustration of discovering new games through trial and error. Tencent's wxg researcher is tackling this problem head-on, using machine learning techniques to create a personalized gaming experience.

One of the key challenges in game recommendation is handling the vast number of games available. According to the researcher, there are over 10,000 games on the WeChat platform, making it difficult for users to find hidden gems.

Object Detection

Object detection is a crucial step in the image recognition process, and it's essential to understand its role and challenges.

The object detection algorithm is the first step in the image recognition process, and it's responsible for locating the object of interest in the user's field of view and separating it from the background.

Geometric Shapes Objects on Blue Surface
Credit: pexels.com, Geometric Shapes Objects on Blue Surface

The detection algorithm is a critical component of the WeChat image recognition framework, and it plays a crucial role in both object detection and category prediction.

In the context of WeChat image recognition, the detection algorithm is designed to handle open-set detection problems, where the model needs to adapt to user input requests that may include novel or unseen classes.

The detection algorithm is not a simple task, and it faces several challenges, including attention problems, open-set detection problems, annotation costs, and multi-source heterogeneity problems.

Attention problems arise when the detection algorithm needs to focus on the object of interest in the user's field of view, reducing unnecessary user interactions.

Open-set detection problems occur when the model needs to detect objects that are not present in the training data, which is a common scenario in WeChat image recognition.

Annotation costs are a significant concern in object detection, as labeling objects in images can be time-consuming and expensive.

Multi-source heterogeneity problems are one of the most challenging aspects of object detection, as the model needs to handle images from different sources, such as indoor and outdoor environments, with varying lighting conditions, resolutions, and angles.

Credit: youtube.com, 腾讯RoboticsX实验室研发多模态感知技术,实现机器狗自主完成动态环境任务

The detection algorithm also needs to handle different categories with varying degrees of complexity, such as products, animals, and plants, each with its own unique characteristics and requirements.

The challenges of object detection are further exacerbated by the need to balance precision and recall, as well as the need to handle long-tail detection problems, where the model needs to detect objects that are rare or novel.

Selection and Prediction

In the Tencent wxg app, the selection and prediction process is crucial for accurate object detection. The app uses a box strategy module to focus on the user's visual attention center, which is simulated by combining the position, confidence, and area of detection boxes.

The module is designed to output the user's most likely target object, but it requires a lot of time to adjust parameters and fine-tune details.

To improve this process, the team plans to adopt the GFocal method, which redefines confidence as a combined indicator of box location, label classification, and area, making it more adaptive for guiding model training.

For offline cataloging mode, the visual center is defined based on the brand of the sold goods, and the app needs to match text titles with detection boxes to effectively complete the cataloging operation.

WeChat Game Recommends

Overhead view of a woman working at a desk with a laptop and books, late at night.
Credit: pexels.com, Overhead view of a woman working at a desk with a laptop and books, late at night.

WeChat's game recommendation algorithm is based on user behavior and preferences, which means it learns what you like and don't like over time.

The algorithm takes into account factors such as the number of times you've played a game, your average playtime per session, and your overall progress in a game.

If you've spent a lot of time playing a particular game, WeChat's algorithm assumes you enjoy it and recommends more games like it.

WeChat also considers your friends' gaming habits and preferences when making recommendations, so if your friends are into a certain type of game, you're more likely to see it in your recommendations.

The algorithm is constantly updated to reflect changes in user behavior and preferences, ensuring that recommendations stay relevant and accurate.

By analyzing your gaming history and preferences, WeChat's algorithm can predict which games you're likely to enjoy, making it easier to discover new games.

3.6 Selection Strategy and Prediction

Crop estate agent working with computer in office
Credit: pexels.com, Crop estate agent working with computer in office

We need to focus on two key aspects when it comes to object detection in scanning mode. Our system needs to prioritize the most visually significant objects in the user's field of vision, and we also have to handle the issue of detecting objects that are not part of a closed set.

To address the first issue, we've added a selection strategy module to our system, as shown in Figure 3.4. This module takes into account the location, confidence, and area of each detection box and combines them to produce a final ranking score for each box.

The module is designed to simulate the user's visual attention and prioritize the most likely object of interest. However, this module is not part of the network's training and requires a lot of time and effort to fine-tune its parameters.

In offline cataloging mode, the visually significant object is defined based on the brand of the item being sold, such as a skirt in the case of a clothing store. We plan to use the CLIP model to optimize the selection strategy for this mode.

Focused man in office working on computers, engaged in technical research and communication.
Credit: pexels.com, Focused man in office working on computers, engaged in technical research and communication.

We've also optimized the selection strategy by combining the detection confidence and retrieval distance into a single network, as shown in Figure 3.5. This allows us to get a more accurate ranking of the detected objects.

The selection strategy is a critical component of our system, and we're continually working to improve it.

Evaluation and Experience

The new version of WeChat's image translation feature, 2.0, has significantly improved in terms of text paragraph recognition, translation quality, and graphical layout.

Compared to the 1.0 version, users experienced a notable boost in text segment recognition, translation quality, and graphical layout.

GSB evaluation results show a significant improvement in the 2.0 version, with a notable increase in translation accuracy.

The 2.0 version has been rolled out to the WeChat iOS client, allowing users to access image translation through the chat box, photo selection, and scan-to-translate features.

Results from an AB test on real users showed that the 2.0 version increased the average number of translated text segments per user by 21% compared to the 1.0 version.

This indicates that the improved quality of the 2.0 version has led to increased user activity in translation, with more users engaging with the feature.

You might enjoy: 腾讯 Layout

Wm Kling

Lead Writer

Wm Kling is a seasoned writer with a passion for technology and innovation. With a strong background in software development, Wm brings a unique perspective to his writing, making complex topics accessible to a wide range of readers. Wm's expertise spans the realm of Visual Studio web development, where he has written in-depth articles and guides to help developers navigate the latest tools and technologies.

Love What You Read? Stay Updated!

Join our community for insights, tips, and more.