Serokell introduced AI-powered features for a leading online marketplace.
One of the major retailers approached Serokell to enhance the existing e-commerce platform. The ultimate goal was to achieve the clients’ satisfaction by introducing improved search for similar items and fine-tuning the product recommendation system.
AN ONLINE MARKETPLACE
We developed an ML model that matches products based on their descriptions. Our team integrated this system into the existing platform that suggests users alternative products if their desired item is unavailable.
It was necessary to merge the databases of several supermarket chains, devise an AI solution that could assess the similarity between products with differing properties.
Challenges we faced
At first, everything seemed to work fine. But, a slight delay in processing transaction information gradually turned into a serious problem.
In order to compare product similarity, we needed to split the features in two parts: numerical (e.g. package weight) and categorial (e.g. package type).
Due to several categories having less than 50 samples, using advanced supervised machine learning methods was not feasible.
To address these issues, Serokell ML experts implemented a name-based matching approach involving the following:
Manually labeling data and performing subsequent active learning.
Using the FastText open-source library, which converts text into continuous vectors, to feed the model.
These data mining techniques helped us design machine learning models for effective learning based on vector analysis
Fine-tuned the model’s performance with the help of the Siamese neural network with their minor dataset size requirements.
Established metrics to evaluate the similarity of items: for numerical feature vectors, we introduced a weighted sum of cosines, while for categorical feature vectors, we used a Jaccard-like score that measures the similarity between two sets.
As a next step of cooperation, Serokell participated in the improvement of the marketplace personal recommendation system.
Incorporate new user-specific features into the existing model.
Conduct a code review.
Implement codebase refactoring.
Ensuring proper documentation to support the updates.
The marketplace had been collecting statistical data, including the top five categories visited by buyers in the past month. We retrieved the data from the company's SQL-like storage and prepared it for the ML model training.
Our experts modified the model by adding a preprocessing block, changing input features, and adjusting parameters.
We achieved a statistically significant improvement in the model’s accuracy. The A/B tests confirmed the success of the implemented changes, resulting in increased user satisfaction.
As an additional functionality, we created a report classification system.
Reports included both relevant and irrelevant messages, without any clear distinction between the two.
We streamlined the report and dismissed overly emotional or spam-like pieces from analysis.
Serokell ML experts manually divided some samples into two groups: useful data and noise.
We trained an ML model on binary classification tasks with pre-trained FastText embeddings and a linear model, employing the data we had collected earlier. Next, we fine-tuned the performance via the Siamese neural network
The customer has acquired the marketplace solution with integrated advanced features, including a coherent recommendation system and well-structured customer feedback reports free from spam.