Multimodal Channel Recommendation System
The process of collaboration in large organizations involves multiple means of communication between employees, teams, and departments, such as group meetings, personal chats, and public channels.
Channels are often related to specific topics, such as a project or an event, and may serve as a source of information for anyone interested in it.
However, messenger workspaces in large organizations often have so many channels that employees who can benefit from certain channels may be unaware of their existence. Thus, recommendation systems suggesting channels without any query specified by the user may help them discover relevant information.
The more information is available about users, the more accurate recommendations we can achieve. Moreover, combining multimodal data from all collaboration platforms could give a better representation of social and technical communication inside organizations. For example, knowledge of organizational structure could trigger position-based recommendations, or, in the case of IT companies, projects and source code could provide information on users' professional interests. However, such data is usually distributed across multiple platforms and is difficult to collect, so most of the existing approaches only use unimodal data.
The objective of this study is to build a channel suggestion system working with multimodal data and analyze the effects of multimodality on quality of the recommendations.
Multimodal Recommendation of Messenger Channels
Ekaterina Koshchenko, Egor Klimov, Vladimir Kovalenko
Collaboration platforms, such as GitHub and Slack, are a vital instrument in the day-to-day routine of software engineering teams. The data stored in these platforms has a significant value for datadriven methods that assist with decision-making and help improve software quality. However, the distribution of this data across different platforms leads to the fact that combining it is a very timeconsuming process. Most existing algorithms for socio-technical assistance, such as recommendation systems, are based only on data directly related to the purpose of the algorithms, often originating from a single system. In this work, we explore the capabilities of a multimodal recommendation system in the context of software engineering. Using records of interaction between employees in a software company in messenger channels and repositories, as well as the organizational structure, we build several channel recommendation models for a software engineering collaboration platform, and compare them on historical data. In addition, we implement a channel recommendation bot and assess the quality of recommendations from the best models with a user study. We find that the multimodal recommender yields better recommendations than unimodal baselines, allows to mitigate the overfitting problem, and helps to deal with cold start. Our findings suggest that the multimodal approach is promising for other recommendation problems in software engineering.