Fore Co., Ltd. Fore Co., Ltd. starts operating “Twitter Kotodama Search”, a semantic user recommendation system

Four Co., Ltd.
For Co., Ltd. begins operation of semantic user recommendation system “Twitter Kotodama Search”

Twitter Kotodama Search uses a method that vectorizes each user’s text by embedding words and evaluates the similarity between users based on the distance between vectors. In other words, it is possible to recommend close users based only on the information in the tweet text sent and published by the user. You can find potential customers of [Image

In recent years, recommendation algorithms based on natural language processing have come to be used in a variety of services, including marketing tools and social networks.
Currently, mainstream recommendation algorithms include content-based filtering, which refers to personal attribute information and text information and recommends users and products that are similar to those, and user preferences and unspecified number of users. There is also a “collaborative filtering” that performs an assessment based on a comparison of the preference records of people, and a certain effect has been recognized.
However, methods such as “content-based filtering” and
“collaborative filtering” are not effective because of the “cold start problem” in which users with little activity history cannot be accurately extracted, and sufficient user information cannot be obtained for small-scale services. The problem was that it was difficult to assess.
To address these issues, a team led by Associate Professor Tomohiro Shirakawa of the Intelligent Informatics Laboratory at Nagaoka University of Science and Technology used a word embedding method [TS1] that converts sentences into vectors, We have proposed a method (*1) [TS2] that makes recommendations based on the similarity between texts determined by distance, and is a recommendation system that does not depend on the number of connections between users, activity history, service scale, etc. has been engaged in research on In a demonstration experiment (*1) to measure the effect of this recommendation system, it was confirmed that a higher engagement rate was obtained than the existing recommendation method, so this time, it will be made available to the public as a web system called “Twitter Kotodama Search”. have become.
[Overview of Twitter Kotodama Search]
1. database
It consists of a list of about 26 million Japanese users collected by Fore- Co., Ltd. and tweet texts by these users.
2. Recommendation based on distance between vectors
In Twitter Kotodama Search, when you enter your Twitter screen name (@XXX), the database is referenced, the distance between the target user’s vector and the user vector in the database is calculated [TS1] and similarity is evaluated. Users whose cosine similarity exceeds a certain number are displayed as display candidates, and the three users with the highest cosine similarity are displayed.

[Measurement of effectiveness of recommendation algorithm]
As a result of a demonstration experiment targeting Twitter users, the engagement rate for recommendations by Kotodama Search (the rate of actions such as likes, replies, retweets, and follows for all users displayed by the recommendations) exceeds 15%. was shown. Although the results cannot be said to be definitive yet due to the small number of subjects, it can be said that an extremely high engagement rate has been achieved, considering that the engagement rate on normal SNS is generally several percent.

【the next deployment】
“Twitter Kotodama Search” developed by Associate Professor Shirakawa and others has achieved a very high engagement rate compared to existing recommendation methods. Although this verification was relatively small-scale and further verification is necessary in the future, it can be said that a certain level of effectiveness was demonstrated. This method of selecting recommendation targets based on publicly available user text has the following advantages, including potential for future development.
・Recommendations can be made based only on publicly available text information, so they are not affected by the amount of the user’s action history.
・Because it uses publicly available text information, it is possible to make recommendations that cross services and are not limited to specific SNSs.
・Since multilingual word embedding (vectorization) algorithms already exist, the method of this research is not limited to a single language. plan to do so).
・It is also possible to search for users who are “not at all similar or the exact opposite” of yourself. By doing so, it is thought that it will be possible to daringly find people who have different opinions from one’s own and prevent the timeline from becoming an echo chamber. ・(When operating an official account of a company, etc.) Users who are similar to the followers of the operating account can be searched and used as a “potential customer list”.
・Detect similar users with a probability of “one in a million” and enable extremely rare encounters (similarly, it is also possible to find “one in a million antis”) .
Fore Co., Ltd. [TS1] provided technical support for the creation of a database of Twitter users and the development of the WEB system in the development of “Twitter Kotodama Search”. In the future, through joint research with Associate Professor Shirakawa and his team, we will expand the scale of the experiment, support further verification, and work on updating the recommendation system and expanding the functions of the service.
Twitter Kotodama Search

Overview of Four Co., Ltd.
Company name: Four Co., Ltd.
Location: 2-8-16 Higashi-Kanda, Chiyoda-ku, Tokyo 101-0031
Established: August 2017
Representative: Yuto Tanaka
Business description:
・Development and sales of market analysis systems using AI
・Consulting on AI and IT, consulting on data processing
・System development business

Details about this release:


Leave a Reply

Your email address will not be published.

%d bloggers like this: