Including Multimodal Data into Channel Suggestions

Nowadays collaborative platforms, such as GitHub and Slack, are a vital instrument in a day-to-day routine of every IT employee. As a result, data that is aggregated by these platforms has a significant value for algorithms of a socio-technical assistance that improve employees’ working conditions. However, the distribution of this data across different platforms leads to the fact that combining it is a very time-consuming process, and therefore the existing algorithms for socio-technical assistance (such as recommendation systems for channels in messengers) are based only on data directly related to the purpose of the algorithms. In this work, I collected a dataset containing data on interactions between JetBrains employees in channels and repositories on the Space platform, repositories on GitHub, as well as the positions of users in the company, obtained from Space. Several types of Space channel recommendation systems, based on different combinations of data modalities, were built using the collected dataset. The best of the constructed models were combined into a resulting multimodal channel recommendation system. Even though the evaluation of the performance of this multimodal system on historical data showed the worse metrics compared to the best of the unimodal algorithms (matrix factorization on channels), Space users rated the channels recommended by the multimodal system significantly better than the once suggested by matrix factorization. Based on the results of this work, it was concluded that using the data on the organization structure and employees’ technical repositories allows to mitigate the overfitting problem, and solves the problem of users’ cold start. Key words: socio-technical assistance algorithms, channel recommendation system, multimodal data.