For cross-border e-commerce merchants, the sales data of a product in the store is a result, while whether it has been worn, photographed, and posted by influencers on overseas social media is often an earlier signal. To use this signal for product selection decisions, merchants need a clear answer: exactly which influencers have featured this product in which content.
As a leading domestic fashion data AI platform, Zhiyi Technology is committed to promoting the deep integration of AI and big data technologies within the apparel industry chain. The company focuses on core scenarios such as trend discovery, best-seller mining, and content marketing, providing comprehensive digital intelligence solutions for clothing enterprises. Its business spans both domestic and international markets, serving nearly ten thousand clients. The company owns a variety of data intelligence SaaS and AI innovative applications, including Zhiyi, Zhikuan, Douyi, Overseas Trend Exploration, FD, and Zhixiaoyi.
In services aimed at cross-border e-commerce merchants, Zhiyi Technology defines the above demand as an image retrieval task: converting product images from e-commerce platforms and influencer images from overseas popular social media into feature vectors, and establishing associations between the two through batch similarity matching. The technical path of the solution is clear, but large-scale implementation is constrained by the computing power and cost of the retrieval process. To address this, Zhiyi Technology collaborated with Alibaba Cloud's Big Data AI platform to replace the original Proxima retrieval solution with the vector retrieval capability provided by MaxCompute, completing the reconstruction of the retrieval process while retaining the in-house model pipeline.
01 Business Requirement: From Keyword Search to Full Image Matching
1.1 Coverage Bottleneck of Manual Retrieval
Before the introduction of image matching capabilities, the main way for merchants to obtain off-site information was keyword searches combined with manual browsing: operators would go through content on social media platforms one by one using brand keywords and category keywords. The output of this method heavily depended on human effort, and the coverage was determined by search results—if influencers did not mention the brand name when posting content, the related content could not be discovered; monitoring competitors was even more difficult, as merchants often could not determine the search terms in advance. The pool of overseas influencer content is on the scale of millions, and the proportion that can be covered by manual sampling is very limited.
1.2 Three Engineering Challenges Faced by Image Matching
Image main subject positioning. Scenario images posted by influencers contain people, backgrounds, and accessories, with the clothing subject only occupying part of the image; product-side images are mostly white-background images or model images. Directly calculating similarity between these two types of images can easily be affected by non-main-subject elements. Therefore, it is necessary to perform clothing area detection on both sides of the images first, crop out the main product, and then proceed to the feature extraction stage.
The computational pressure of full-scale cross-matching. The issue of clothing comparison accuracy for a single image has been solved by Zhiyi Technology through its self-developed CV model. The limiting factor for large-scale implementation lies on the retrieval side: on the social media side, posts cover multiple images, while on the product side, main and detail images from multiple e-commerce platforms are included. Incremental products need to be cross-matched with all images of the selected influencers. Taking one retrieval round as an example, there are about 10 million 512-dimensional vectors on the product side, which need to be fully matched with about 2 million query vectors on the social media side.
Search costs determine the boundaries of business coverage. Constrained by computing power costs, content that includes videos currently only extracts the first frame image for matching. Such trade-offs will exist for a long time in business evolution, and the unit cost of the retrieval process directly determines the scale of the content pool that can be covered and the frequency of data updates.
The three challenges point to three clear technical requirements: first, an engine capable of handling full-scale batch vector retrieval; second, a sufficiently low retrieval unit cost to support daily full-scale reruns, avoiding trade-offs between data freshness and coverage; third, the ability to interface with existing self-developed model pipelines without needing to rewrite algorithms for migration.
02 Solution: MaxCompute Offline Vector Retrieval for Handling Batch Matching
Zhiyi Technology and the Alibaba Cloud Big Data AI platform jointly organized the data processing workflow, forming a division of labor where 'self-developed models are responsible for understanding, and MaxCompute is responsible for retrieval.'
2.1 Data Cleaning and Image Preprocessing
After the publicly available blog posts on overseas well-known social media platforms are centrally cleaned, social media images and product images are separately subjected to clothing area detection, cropping out the main product area, and excluding people, backgrounds, and accessories to prevent interference with subsequent feature extraction.
2.2 Feature Vector Generation: Retain Proprietary Model Assets
The cropped images are labeled and feature vectors are extracted by Zhiyi Technology's self-developed CV model. The model is independently trained and deployed, running on a GPU cluster, allowing Zhiyi Technology to fully retain its accumulated model assets and optimization experience in the field of clothing recognition.
2.3 Batch Vector Retrieval: Migrating to MaxCompute
Batch similarity matching between product vectors and social media vectors is completed on MaxCompute, covering three stages: vector storage, index construction, and full-batch retrieval computation. The original scheme was built based on the Proxima engine provided by MaxCompute, requiring independent management of the index and retrieval clusters, resulting in high operational costs. After migrating to the new version of vector retrieval, the above stages are handled uniformly by the platform: the vector data itself remains in MaxCompute without the need for cross-platform migration, only requiring writing into a newly created VECTOR type column; the upstream link for generating embeddings from external models remains completely unchanged. The retrieval query switches to the standardVECTOR_SEARCHSQL functions and indexes are automatically built and maintained by the system. Zhiyi Technology completed the switch from validation to production in about two weeks, no longer maintaining a separate vector retrieval infrastructure, with resource consumption significantly reduced under the same data scale.
In addition, brand mentions in the body of the blog post (such as @BrandName) serve as auxiliary signals to confirm some content with obvious promotional features, while the main basis for matching judgment remains image similarity.
03 Application Scenarios: Supplementing off-site evidence for product selection decisions
After the new retrieval link goes online, merchants can directly view the association results between products and influencer content in Zhiyi Technology's products: which influencers' content features their own products, how competitor products are exposed outside the platform, and whether there is KOL content support behind best-selling products.
After this information enters the product selection phase, it changes the basis for judgment. Previously, merchants mainly relied on on-site sales data to evaluate a style, and off-site popularity could only be supported by scattered screenshots; now, whether a style has received spontaneous promotion by influencers on overseas social media has become an indicator that can be continuously monitored. For inventory decisions, 'good on-site data' and 'good on-site data along with existing influencer content off-site' are two signals of different strength.
04 Cooperation Results
Under the same retrieval scale of approximately 10 million 512-dimensional vectors on the product side and about 2 million on the social media side, Zhiyi Technology compared the resource consumption of MaxCompute vector retrieval with the original Proxima solution (historical version). Taking the average of Proxima's last 14 runs as the baseline, the CPU consumption of MaxCompute offline vector retrieval decreased by about 73%, and memory consumption decreased by about 93%.
|
Comparison item |
Historical versions (average of the last 14 times) |
MaxCompute Vector Search |
Change |
|
CPU Consumption (CU·min) |
17813.52 |
4790.06 |
Approximately -73% |
|
Memory Consumption (GB·min) |
24016.50 |
1712.65 |
Approximately -93% |
The significant reduction in resource overhead directly impacts the business: with the same budget, the same round of full matching can cover a larger content pool, or be rerun more frequently, allowing merchants to see product-influencer association results that remain up-to-date. The retrieval process is no longer a trade-off between content pool size and data freshness.
Regarding the core issues resolved in this collaboration, the technical team at Zhiyi Technology has a more direct understanding:
Having a complete match between product images and overseas social media content is something we have wanted to achieve for a long time. The real bottleneck lies in retrieval costs—in the original solution, we could only choose between data scale and computing power costs. With this cooperation with Alibaba Cloud's big data platform, large-scale batch vector retrieval is handled by the platform, allowing us to focus on extracting feature vectors of clothing styles. The teams from both sides conducted multiple rounds of joint optimization around index construction and batch retrieval, reducing CPU and memory usage by approximately 73% and 93%, respectively, significantly easing the cost pressure in the retrieval process. Next, we hope to extend this capability to more e-commerce platforms and a larger scale of social media images, making external real-world popularity a regular metric for cross-border product selection.
/ END /
