Girish Kurup
6 min readMay 27


Recommendation Models for Retail

In retail when a customer visits a website he search for his items among millions of options and clicks buy for few of them

There is pattern in which customers search, buys or returns

Sampling of items bought by set of customers is called positive sampling or Positive article embedding

Sampling of items that customers did not buy are called negative sampling or negative article embedding

Recommendation for a similar customer based on his profile is done based on the cosine similarity between positive and negative embedding

Negative sampling is a technique that is used in recommendation engine algorithms to efficiently train them on large datasets. Traditional recommender systems typically rely on Collaborative filtering, which requires the availability of user-item interaction data. However, in many real-life scenarios, it can be difficult to obtain sufficient user-item interactions. This results in a data sparsity problem, which can negatively impact the performance of traditional recommendation systems.

Negative sampling helps to overcome this problem by generating implicit or explicit negative samples, which are then used to balance out the dataset. In this way, negative samples can help the recommendation engine to learn from both positive and negative feedback, even when the majority of the feedback is positive. Negative sampling can be applied in various ways, such as by generating random negatives or by using pre-defined negative samples.

Overall, negative sampling is a useful technique for improving the performance and accuracy of recommendation engines. By balancing out datasets with negative samples, these algorithms can more effectively learn from user-item interactions and generate more relevant recommendations for users.

However, as with any machine learning algorithm, negative sampling can also be impacted by imbalanced datasets. When defining negative samples for a recommendation engine, the quality of selected negatives can significantly impact its performance. Therefore, it is important to utilize appropriate methods for sampling negative examples that cater to the imbalanced data problem. FitNET’s recommendation engine is a great example of the application of negative sampling in machine learning models. For instance, FitNET defines positive samples as the items clicked by users. Additionally, FitNET generates negative samples by sampling from the set of final…



Girish Kurup

Passionate about Writing . I am Technology & DataScience enthusiast. Reach me