Sequential Recommendation with Temporal Context via Convolutional Sequence Embedding
Aruna B. Bhat, Rishabh Chandra, Kanika Kanika · 2021
Human behavior consists to a large extent of repeated temporal patterns. What a person is going to do next often depends on what time it is as well as what things she/he has interacted with in recent past. Often times, in many fields it is required to predict what actions or items a user is most likely to engage with in near future. To solve this problem, there are many deep learning approaches such as RNN, LSTM [2], [3] but they are usually computationally more demanding as well as difficult to train. Other sequential recommendation methodologies like CASER [1] does not incorporate temporal context. To overcome these limitations, this paper has proposed a convolutional neural network based deep learning approach to solve this problem. The experiments on public dataset demonstrated that our model which takes in account the temporal context performs better than CASER [1] without temporal context. Also, this research work has demonstrated the application of the proposed model as an on-device solution to enable voice assistants or other host applications to proactively provide recommendations and suggestions based on users' past activity, routine and current timestamp.