| 研究課題/領域番号 |
17K12784
|
|---|
| 研究機関 | 筑波大学 |
|---|
研究代表者 |
于 海涛 筑波大学, 図書館情報メディア系, 助教 (30751052)
|
|---|
| 研究期間 (年度) |
2017-04-01 – 2020-03-31
|
| キーワード | User modelling, / Embedding, / Result diversification |
|---|
| 研究実績の概要 |
This year I analyzed the Yandex query log and found that the rank bias for decoding users’ click behavior has not been investigated when using neural networks. The method for representing queries and documents is also not effective. Moreover, I found that an in-depth study of the cluster-based paradigm for search result diversification is needed, which helps to design effective ranking algorithms for the target interactive information retrieval system. To cope with the above challenges, I designed new models using neural networks to decoding users' click behaviour based on the above query log data. Meanwhile, I published one paper in the journal of Information Processing and Management, which provides an in-depth exploration of the cluster-based paradigm for search result diversification.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
1: 当初の計画以上に進展している
理由
The proposed research is conducted quite smoothly. In particular, in order to identify the underlying patterns among items, such as queries, documents and users, I designed and implemented several neural click models. The experimental results are promising and outperform the state-of-the-art baseline method in terms of standard metrics. Furthermore, industrial dataset has been deployed to test the proposed models. Moreover, one paper titled as ‘Revisiting the cluster-based paradigm for implicit search result diversification’ has been finished and accepted by the journal of Information Processing & Management. It is important for designing effective ranking algorithms for the target interactive information retrieval system. This part of work is progressing smoothly than the initial plan.
|
| 今後の研究の推進方策 |
For the future work, on one hand, I will explore how to perform interactive ranking by incorporating the previously proposed diversification algorithms and deep learning techniques. In particular, deep learning-to-rank algorithms will be implemented and tested based on standard collections. On the other hand, a series of user studies will be conducted to validate the performance from the perspective of real users.
|
