Alfred Krzywicki

Collaborative filtering for people to people recommendation in social networks

Predicting people other people may like has recently become an important task in many online social networks. Traditional collaborative filtering approaches are popular in recommender systems to effectively predict user preferences for items. However, in online social networks people have a dual role as both “users” and “items”, e.g., both initiating and receiving contacts. Here the assumption of active users and passive items in traditional collaborative filtering is inapplicable. In this paper we propose a model that fully captures the bilateral role of user interactions within a social network and formulate collaborative filtering methods to enable people to people recommendation. In this model users can be similar to other users in two ways – either having similar “taste” for the users they contact, or having similar “attractiveness” for the users who contact them. We develop SocialCollab, a novel neighbour-based collaborative filtering algorithm to predict, for a given user, other users they may like to contact, based on user similarity in terms of both attractiveness and taste. In social networks this goes beyond traditional, merely taste-based, collaborative filtering for item selection. Evaluation of the proposed recommender system on datasets from a commercial online social network show improvements over traditional collaborative filtering.

Interaction-based collaborative filtering methods for recommendation in online dating

We consider the problem of developing a recommender system for suggesting suitable matches in an online dating web site. The main problem to be solved is that matches must be highly personalized. Moreover, in contrast to typical product recommender systems, it is unhelpful to recommend popular items: matches must be extremely specific to the tastes and interests of the user, but it is difficult to generate such matches because of the two way nature of the interactions (user initiated contacts may be rejected by the recipient). In this paper, we show that collaborative filtering based on interactions between users is a viable approach in this domain. We propose a number of new methods and metrics to measure and predict potential improvement in user interaction success, which may lead to increased user satisfaction with the dating site. We use these metrics to rigorously evaluate the proposed methods on historical data collected from a commercial online dating web site. The evaluation showed that, had users been able to follow the top 20 recommendations of our best method, their success rate would have improved by a factor of around 2.3.

Reciprocal and heterogeneous link prediction in social networks

Link prediction is a key technique in many applications in social networks, where potential links between entities need to be predicted. Conventional link prediction techniques deal with either homogeneous entities, e.g., people to people, item to item links, or non-reciprocal relationships, e.g., people to item links. However, a challenging problem in link prediction is that of heterogeneous and reciprocal link prediction, such as accurate prediction of matches on an online dating site, jobs or workers on employment websites, where the links are reciprocally determined by both entities that heterogeneously belong to disjoint groups. The nature and causes of interactions in these domains makes heterogeneous and reciprocal link prediction significantly different from the conventional version of the problem. In this work, we address these issues by proposing a novel learnable framework called ReHeLP , which learns heterogeneous and reciprocal knowledge from collaborative information and demonstrate its impact on link prediction. Evaluation on a large commercial online dating dataset shows the success of the proposed method and its promise for link prediction.

A BDI agent architecture for dialogue modelling and coordination in a smart personal assistant

We discuss the architectural aspects of a smart personal assistant (SPA) system that enables users to access a range of applications from a range of devices using multi-modal natural language dialogue. Each back-end application is a personal assistant specializing in one specific task such as e-mail or calendar management, and typically each has its own user model, enabling it to adapt to the users changing preferences. The PDA interface to the SPA must present the system as a single unified set of back-end applications, enabling the user to conduct a dialogue in which it is easy to switch between these applications. Furthermore, the systems interaction with the user must be tailored to their current device. The SPA is implemented using an agent platform and includes a special BDI coordinator agent with plans both for coordinating the actions of the individual assistants and for encoding the systems dialogue model. The plan-based dialogue model is at a high level of abstraction, enabling the domain-independent plans in the dialogue model to be reused in different SPA systems.

People-to-People recommendation using multiple compatible subgroups

People-to-people recommendation aims at suggesting suitable matches to people in a way that increases the likelihood of a positive interaction. This problem is more difficult than conventional item-to-people recommendation since the preferences of both parties need to be taken into account. Previously we proposed a profile-based recommendation method that first uses compatible subgroup rules to select a single best attribute value for each corresponding value of the user, then combines these attribute value pairs into a rule that determines the recommendations. Though this method produces a significant improvement in the probability of an interaction being successful, it has two significant limitations: (i) by considering only single matching attribute values the method ignores cases where different attribute values are closely related, missing potential candidates, and (ii) when ranking candidates for recommendation the method does not consider individual behaviour. This paper addresses these two issues, showing how multiple attributes can be used with compatible subgroup rules and individual reply rates used for ranking candidates. Our experimental results show that the new approach significantly improves the probability of an interaction being successful compared to our previous approach.

An adaptive calendar assistant using pattern mining for user preference modelling

In this paper, we present SmartCal, a calendar assistant that suggests appointment attributes, such as time, day, duration, etc., given any combination of initial user input attributes. SmartCal uses closed pattern mining to discover patterns in past appointment data in order to represent user preferences and adapt to changing user preferences over time. The SmartCal interface is designed to be minimally intrusive: users are free to choose or ignore suggestions, which are dynamically updated as users enter new information. The user model as a collection of patterns is intuitive and transparent: users can view and edit existing patterns or create new patterns based on existing appointments. SmartCal was evaluated in a user study with four users over a four week period. The user study shows that pattern mining makes appointment creation more efficient and users regarded the appointment suggestion feature favourably.

Incremental E-Mail Classification and Rule Suggestion Using Simple Term Statistics

In this paper, we present and use a method for e-mail categorization based on simple term statistics updated incrementally. We apply simple term statistics to two different tasks. The first task is to predict folders for classification of e-mails when large numbers of messages are required to remain unclassified. The second task is to support users who define rule bases for the same classification task, by suggesting suitable keywords for constructing Ripple Down Rule bases in this scenario. For both tasks, the results are compared with a number of standard machine learning algorithms. The comparison shows that the simple term statistics method achieves a higher level of accuracy than other machine learning methods when taking computation time into account.

Learning to make social recommendations: a model-based approach

Social recommendation, predicting people who match other people for friendship or as potential partners in life or work, has recently become an important task in many social networking sites. Traditional content-based and collaborative filtering methods are not sufficient for people-to-people recommendation because a good match depends on the preferences of both sides. We proposed a framework for social recommendation and develop a representation for classification of interactions in online dating applications that combines content from user profiles plus interaction behaviours. We show that a standard algorithm can be used to learn a model to predict successful interactions. We also use a method to search for the best model by minimising a cost based on predicted precision and recall. To use the model in real world applications to make recommendations, we generate candidate pairs using the selected models and ranked them using a novel probabilistic ranking function to score the chance of success. Our model-based social recommender system is evaluated on historical data from a large commercial social networking site and shows improvements in success rates over both interactions with no recommendations and those with recommendations generated by standard collaborative filtering.

A Large-Scale Evaluation of an E-mail Management Assistant

EMMA is an e-mail management assistant based on ripple down rules, providing a high degree of classification accuracy while simplifying the task of maintaining the consistency of the rule base. A naive Bayes algorithm is used to improve the usability of EMMA by suggesting keywords to help the user define rules. In this paper, we report on an experimental evaluation of EMMA on 16 998 pre-classified messages. The aim of the evaluation was to show that the ripple down rule technique used in EMMA applies to large-scale data sets in realistic organizational contexts. The results showed conclusively that EMMA attained the agreed success criteria for the evaluation and that the knowledge acquisition method used in EMMA outperforms standard machine learning methods.

Closed Pattern Mining for the Discovery of User Preferences in a Calendar Assistant

We use closed pattern mining to discover user preferences in appointments in order to build structured solutions for a calendar assistant. Our choice of closed patterns as a user preference representation is based on both theoretical and practical considerations supported by Formal Concept Analysis. We simulated interaction with a calendar application using 16 months of real data from a user’s calendar to evaluate the accuracy and consistency of suggestions, in order to determine the best data mining and solution generation techniques from a range of available methods. The best performing data mining method was then compared with decision tree learning, the best machine learning algorithm in this domain. The results show that our data mining method based on closed patterns converges faster than decision tree learning, whilst generating only consistent solutions. Thus closed pattern mining is a better technique for generating appointment attributes in the calendar domain.