Yunseok Noh

A combination of temporal and general preferences for app recommendation

User preferences in various kinds of recommendations are in general made from the contents of recommending targets or the patterns that the targets are consumed in. As a result, a great number of previous works have focused on designing a good user preference. However, one important thing that is missed in the previous studies on user preference is that user preferences are affected by time. That is, it is of importance to capture the change of user preferences over time for better recommendations. This phenomenon is salient especially in using mobile apps. Therefore, this paper presents a time-based personalized application recommendation system which captures temporal changes in user preference. The proposed recommendation system can recommend dynamically the apps from an application market by considering the user preference and time. In order to recommend apps, the app descriptions are used to recommend new apps to users, and user preference is modeled using a probabilistic topic model from the descriptions. In order to incorporate time to the topic model, the proposed temporal topic model considers the usage of mobile apps over time for a specific user. The main problem of this temporal topic model is that it is not well trained when the number of apps that the user has used is small, and it can be remedied by incorporating a normal LDA-based topic model. As a result, the final recommendation model is a combination of temporal and LDA-based topic models. The proposed method is validated through a series of experiments. For app usages of three users for 35 days on average, it is compared with LDA-based topic model and the model that uses only temporal topic model. According to the experimental results, the proposed method outperforms the two baseline models up to 18% point in nDCG. This result proves that the proposed method is effective in content-based app recommendation.

Location Comparison through Geographical Topics

With the increasing interest in location-based services, location comparison gains more and more attentions. One of the best ways to represent a location is to use topics that are generated near the location. In order to compare locations through such geographical topics, two conditions need to be met. One is that the topic set should be fixed but cover various aspects of all possible locations, and the other is that geographical topics often depend on each other. This paper proposes Probabilistic Explicit Semantic Analysis (PESA) that meets these conditions. PESA represents a location as a weighted topic vector where each topic is a Wikipedia concept. The number of Wikipedia concepts is fixed, but their enormous quantity allows PESA to be used to compare various locations. In addition, link information within Wikipedia articles is used to compute prior probabilities of topics considering their dependencies. That is, it enables PESA to model the topic dependency. PESA was evaluated using eighteen locations in three distinct geographical categories and compare it with LDA and ESA. The experimental results that PESA outperformed LDA and ESA highlighting its superiority in location comparison.

Expanding knowledge source with ontology alignment for augmented cognition

Augmented cognition on sensory data requires knowledge sources to expand the abilities of human senses. Ontologies are one of the most suitable knowledge sources, since they are designed to represent human knowledge and a number of ontologies on diverse domains can cover various objects in human life. To adopt ontologies as knowledge sources for augmented cognition, various ontologies for a single domain should be merged to prevent noisy and redundant information. This paper proposes a novel composite kernel to merge heterogeneous ontologies. The proposed kernel consists of lexical and graph kernels specialized to reflect structural and lexical information of ontology entities. In experiments, the composite kernel handles both structural and lexical information on ontologies more efficiently than other kernels designed to deal with general graph structures. The experimental results also show that the proposed kernel achieves the comparable performance with top-five systems in OAEI 2010.

Regularizing Korean Conversational Model by Applying Denoising Mechanism

A conversation system is a system that responds appropriately to input utterances. Recently, the sequence-to-sequence framework has been widely used as a conversation-learning model. However, the conversation model learned in such a way often generates a safe and dull response that does not provide appropriate information or sophisticated meaning. In addition, this model is also useless for input utterances appearing in various forms, such as with changed ending words or changed word order. To solve these problems, we propose a denoising response generation model applying a denoising mechanism. By injecting noise into original input, the proposed method creates a model that will stochastically experience new input made up of items that were not included in the original data during the training process. This data augmentation effect regularizes the model and allows the realization of a robust model. We evaluate our model using 90k input utterances-responses from Korean conversation pair data. The proposed model achieves better results compared to a baseline model on both ROUGE F1 score and qualitative evaluations by human annotators.

Detecting multiple userids on Korean social media for mining TV audience response

Possession of multiple userids by a single user happens when more than two userids actually belong to the same user. In analysis of audience response of TV program, it is so important to detect these multi-id users because they often use the multiple ids to manipulate audience response or to take illegal profits. Detecting multiple userids of a single user has similiar nature with authorship attribution in terms of identifying authorships for given arbitrary texts. The conventional supervised techniques for authorship attribution, however, are difficult to be employed directly to the problem of multiple userids detection. This is because we do not know real authors and multiple userids may belong to the same author in the task of multiple userids detection. In addition, since we can not have all authors in advance, userids can not be treated as classes. This paper proposes a method of learning the element-wise differences between multiple userids. Each userid is represented as a feature vector from their postings on web social media. Then the similarity vector between two userid vectors can be obtained by performing their element-wise difference. With the similarity vectors, we train the similarity patterns for detecting if multiple userids belong to the same user or not. In order to solve the problem successfully, we present six features which are effective for Korean social media. We conducted comprehensive experiments on the Korean social media dataset. The experimental results show that the proposed similarity learning method with all presented features is successful for detecting multiple userids on Korean social media.

Keyword Extraction from Dialogue Sentences Using Semantic and Topical Relatedness

Dialogue reflects interests of the participants at that moment. Thus, it is desirable to extract keywords from each dialogue sentence as soon as they are spoken, because the keywords from dialogue can be used for various fields importantly such as personal assistant services, advertisement, and so on. This paper proposes a novel method of keyword extraction from dialogue sentences. The proposed method determines a word as a keyword by using semantic information of words in a dialogue sentence. That is, the proposed method extracts the keywords that are more semantically important within their sentences and more topically related to the dialogue. In the experiments on the ICSI meeting corpus, the proposed method achieves the state-of-the-art performance.