Improving AI-powered matchmaking: Using Sentence-BERT to better understand participants’ profiles

Improving AI-powered matchmaking: Using Sentence-BERT to better understand participants’ profiles
No items found.

As technology continues to advance, the realm of event management is witnessing a significant transformation. AI-powered event matchmaking is one example. It’s now helping event organizers connect like-minded individuals leading to more meaningful relationships being established at their events.

In this technical article, we’ll delve into the concept of using Sentence-BERT (Bidirectional Encoder Representations from Transformers) as a way to understand the context of participants’ profiles so Grip’s platform can better connect the right people.

The idea

The fundamental premise behind this experiment to use Sentence-BERT for event matchmaking is to create clusters of interested parties based on their profiles. By embedding custom profile fields into a vector space, AI algorithms can group individuals who share similar interests, preferences or goals, facilitating better networking opportunities during events.

The problem

While Sentence-BERT has proven to be effective in capturing semantic information within English sentences, initial attempts to utilize this embedding system for event matchmaking did not yield satisfactory results. The challenge lies in optimizing the embeddings further to focus on compatibility rather than just similarity.

To tackle this issue, retraining Sentence-BERT using triplet loss was applied. Unlike the initial approach, which aimed to bring similar profiles close together in the embedding space, the triplet loss training aimed to identify and group compatible individuals. However, the results obtained from this approach showed poor results on the chosen examples, raising several questions.

Possible challenges and solutions

  1. Heterogeneity of the group: The lack of convergence in the embedding space could be attributed to the heterogeneity of the group under consideration. Identifying reasons for compatibility amongst a heterogenous group is hard; and It is essential to assess whether the diversity of profiles within the dataset impacts the effectiveness of the clustering approach. Focusing on specific types of events, such as networking-style gatherings, may provide a more homogeneous dataset for testing and optimization.
  2. The completeness of the used profiles. It is known that profiles vary both in information presented and in accuracy. Perhaps better cleaning of the dataset would have produced far better results, however it was designed to work in the real world where messy data is a fact of life.
  3. Evaluating the approach: The choice to optimize the embeddings using triplet loss for mutually interested parties raises questions about the suitability of this approach for achieving the desired outcome. It might be necessary to scale back the aim towards more accurately identifying similarities between attendees. This would be less useful for the cold start problem, but may provide better content embeddings for recommendations in general.
  4. Code implementation: Another possibility is that the issue lies within the code implementation itself. Double-checking the code and ensuring its correctness is essential. Additionally, it might be beneficial to seek feedback from the AI community and domain experts to identify potential pitfalls or improvements in the codebase.

Future direction

  1. Target networking events: To better understand the effectiveness of the approach, narrowing down the focus to networking events can provide valuable insights. By streamlining the dataset and tailoring the embedding process specifically for networking-style events, it may be possible to achieve more accurate and relevant results.
  2. Test on easier examples: Experimenting with simpler examples can serve as a stepping stone to validate the approach's effectiveness and identify any potential shortcomings. By starting with less complex datasets and gradually scaling up, it becomes easier to pinpoint specific challenges and address them accordingly.
  3. Consider session data: While the initial emphasis was on profile-based clustering, profiles contain lots of structured data which is quite unlike the sentences that the BERT model was originally trained on. Applying the Sentence-BERT approach to session-based data could be an interesting avenue to pursue. Sessions often have a more content rich sentence-like description, making them a potentially suitable dataset for this approach. By considering session-level embeddings, it may be possible to enhance the matchmaking capabilities within specific sessions or tracks.


AI-powered event matchmaking holds immense promise in revolutionizing how people connect and interact at events. While employing Sentence-BERT as an embedding system for Grip custom profile fields presents initial challenges, further exploration and optimization can lead to breakthroughs in clustering compatible individuals. By refining the methodology, experimenting with different event types, and considering session data, event organizers can leverage the power of AI to create more meaningful B2B relationships at their event.

For more on Grip’s AI-powered matchmaking capabilities, check out this product page.