000059647 Time-constrained Item Sequences Recommender System for Mobile Users (Hệ thống đề xuất chuỗi mục có giới hạn thời gian dành cho người dùng di động)

000059647 Time-constrained Item Sequences Recommender System for Mobile Users (Hệ thống đề xuất chuỗi mục có giới hạn thời gian dành cho người dùng di động)

2 IE a

OSES Omens ys an hin ie 18 User

HANOI UNIVERSITY

Time-constrained Item Sequences

Recommender System for Mobile Users

Hanoi University for the honor degree of

Bachelor of Computer Science

By

Ma Kim Binh (Computer Science)

Supervisor: Dr Nguyen Nhat Quang

g THU VIEN DAI HOC HANOI | HANOI UNIVERSITY LIBRARY

Almost current RSs (both those for web users and those for mobile ones) have been implemented to recommend individual items (i.e., individual products or individual

services) In reality, however, there are some application problems where the users would

like to receive suggestions on sequences of items (ordered sets of items), instead of individual ones In those application problems, not only the individual items but also their order in each recommendation needs to match with the user’s preferences and his request context My research work, reported in this thesis, focuses on addressing the interesting

problem of recommending time - constrained sequences of items personalized for mobile

users

To solve the research problem, we need to address the two main sub-problems: 1) computing for each item category the individual items appropriate for the user’s preferences and his request context; 2) computing item sequences such that the order of

iii

items in each sequence is suitable for the user’s preferences and request context For the

first sub-problem, we propose integrating the user’s long-term preferences and the request context to solve the first sub-problem, and for the second one using the Case— Based Reasoning (CBR) problem-solving approach The proposed methodology should also take into account the mobile usage environment’s characteristics and the mobile user’s usage behavior Moreover, the critique-based conversational recommendation approach is applied in order to help the system better understand the user’s session- specific preferences and compute more suitable recommendations for him The recommender system iDo, which implements the proposed methodology to provide

personalized recommendation of time-constrained leisure activities for mobile users, is

presented

iv

Acknowledgements

I would like to express my deepest gratitude to Dr Nguyen Nhat Quang, the instructor in Hanoi University of Science and Technology, without whose patient and enthusiastic guidance, my graduation project would not have been possible I am also heartily thankful to Dr Nguyen Xuan Hoai, the vice Dean of Faculty of Information Technology - Hanoi University for his education and motivation

| offer my regards to my parents, my friends and my partner who have made available their support in a number of ways

Ma Kim Binh

Table of Contents

ADSLTACH, HH HH An n1 11 0110.01117101010.0111, iii

ACKNOWIEABEMENS! <scessssscrsrnvssronsncsnsososnnenesogasorosnsonssvenonoaianesnertbestensnoensanosasananstenteeeneeneasonen v THblẽ Of COETẨSe:sscsrtcsssoiccicsoSG00101506113110156308650d0813353354ê355344866468ê786i6u83ê.3458438.88.ï6i8040013êd888188 vi Lii§t:Gf FIBDES) csszci2nttudtdttl6t6oclipltcoHBi§§b,fi,RE@uyqussse viii Chapter 1- Introduction .cccessecsessessessesseeseeseessessesssesseessessecssssessesssecsseessesnseseesseessesseeane 1 Vs: (O]€€LVES socsnnnoniioniitiiggttiiaridtdililiSeGlgd4GSl61851g8040â68418364566êØ18664ê3844840615664638TSE4u8840nê9 1

3 The Proposed Methodology c.ssecssssssssessssssessssesscsssssesssesssessseseeseesneesnesseerscenseaneesnens 3

4 Overview of the Rest of the Thesis

Chapter 2 - The Background Knowledge

Chapter 3 — The Proposed Recommendation Methodology

1 The Formal Representation c sssssssssssssssesssssssseessssceneesssecessecssccesnecsssecsasecsnecssne 18 L.1 Item Representation cssscsessesssesssssseseesesneesesnseneeneeaesneesnseneeeneeneenneennense 18

vi

1⁄2 The User Profile Representations iscccssssescssssssssassosnsssvoncsssancassesonsasvcassvvavenssees 19 1.3 The User Query Representation .scccscssecseessesneesncsnecnsesneeneeneeneenneenneenes 20

2 The Recommendation Process ssessesssessesseeseesecsesssesseesesesesssessennesseeseeneesneesneese 21 2.1 The First Phase: Compute the Set of Candidate Individual Items 2 2.2 The Second Phase: Compute the Recommendation List of Item Sequences 25 2.3 The Third Phase: Adaptation of the User Query Representation through the

Chapter 4 - The Recommender System ¿Ï20 c5 cctceertrrrrrrerierrrreerree 34

vii

L ist of Figures

Figure 1 The Case-Based Reasoning problem solving cycle [ Í 8] - Figure 2 The Recommendation Process s.c.cisssscssssessssisestsisseosscssssnsavasssvoaseasonsnevsvarsonensoenes Figure 3 The main functions of iDo

viii

Chapter 1- Introduction

1 Objectives The main objective of this research is to develop an appropriate approach to the problem

of recommending time-constrained item sequences for mobile users To illustrate the proposed recommendation methodology, we build the mobile RS called iDo, aimed at assisting users to find sequences of leisure activities suitable for their free-time period and their preferences

2 Motivation Recommender Systems (RSs) are intelligent decision support tools that help users make the best decisions, and RSs are especially useful in cases that the user is overwhelmed by

a large number of options to consider, or the user does not have enough of the domain- specific knowledge and experience to make autonomous selection decisions Almost all existing RSs (i.e., both those for web users and those for mobile ones) are designed to

recommend individual items such as recommending books in Amazon.com, or

recommending movies in Netflix, etc In practice, however, there are some application problems where the users would like to receive suggestions on sequences (i.e., ordered sets) of items that meet users’ needs and preferences For example, the user might want to choose a sequence of leisure activities to do in their free time or the users would like to have a travel plan To our knowledge, there are so far no published results on the solutions for such kind of problem Therefore, in this research, we focus on proposing a recommendation approach to address the problem of recommending time-constrained

1

item sequences for mobile users For each computed recommendation of the system, not

only items included in that recommendation satisfy the user’s time constraint and

preferences, but also the order of those items needs to meet users’ preferences and his current request context Particularly, in the ICT (Information and Communication Technology) era, the mobile phone is becoming a more and more popular personal item; hence, building a RS that can be implemented for mobile users will be highly applicable and bring users significant benefits The proposed approach should take into account the challenges of mobile RSs such as the limitations of mobile devices, the limitations of wireless telecommunication network, the special characteristics of mobile users

Especially, for visualizing the proposed methodology, we build an application system named iDo which recommends leisure activities for mobile users to do in their free time People nowadays are often under high pressure work, which can easily cause stress Therefore, people are now having a trend to find and select entertainment activities to do

in their free time to reduce stress and enjoy the lifetime If the users use the traditional search engines, they have to spend much time and effort to browse much information,

compare and make decisions, or even there are not any available options that make users

pleased Besides, at the initialization of the searching process, the users are often required

to input their needs and preferences With the limitation of the mobile usage environment and the mobile user’s behaviors, such traditional systems do not prove their efficiency Dealing with those problems, the idea of a mobile RS which proposes sequences of leisure activities satisfying users’ preferences so that they can do them in their free time

is invented.

3 The Proposed Methodology The researchers must carry out the literature review to acquire the knowledge of RSs, what main techniques are proposed by other researchers, especially in the specific problem of context constrained RSs, mobile RSs and activity RSs After that, the research problem is carefully studied to find out the suitable and reasonable approach, and then to develop the proposed recommendation methodology The proposed methodology is then applied to implement a real prototypical system iDo — the recommender system aimed at recommending personalized leisure activity sequences for mobile users

In order to solve the problem of recommending item sequences for mobile users, there are 2 main sub-problems that need to be addressed The first one is to determine, for each item category, a candidate set of individual items which satisfy the time-constraint and fit the user’s preferences and the request context The second sub-problem is to compute the recommendation list of item sequences most appropriate to the user’s preferences and the request context For the first sub-problem, we propose to exploit the user’s long-term

preferences, his session-specific preferences and his request’s contextual information to

determine appropriate individual items For the second sub-problem, we apply the Case— Based Reasoning (CBR) approach [2] to exploit the knowledge contained in the past

recommendation cases, and adapt it to build the solution (i.e., the recommendation list)

for the current case Moreover, the critique-based conversational approach, which has been presented in [32], is exploited to allow the system to interact with the user to get more his preferences, and hence to compute better recommendations for him The system’s target users are mobile users (i.e., those interact with the system using some

kinds of mobile devices such as mobile phones, PDAs, etc) Therefore, the proposed methodology must take into account the limitations of mobile devices and the

characteristics of mobile usage environment like small screen, computation limitation,

short-time interaction and the mobile user’s behaviors such as preferences of fast and simple interactions, dislike of initial input requirement, etc

There is in this research a partnership between this thesis’s author and Nguyen Quang Bach — an undergraduate student in Hanoi University of Science and Technology The main concern of this thesis is representing the methodology to address the second sub- problem: computing item sequences

4 Overview of the Rest of the Thesis The remainder of this document is organized as follows In Chapter 2, the background knowledge and related works (related researches to this topic) which help the readers to understand the fundamental concepts about this research problem is presented The comparison and evaluation are also made to help readers have the insights into their success and limitations also In Chapter 3, we represent the recommendation methodology which includes the formal representation and the recommendation process

In Chapter 4, the recommender system iDo which uses the proposed recommendation approach is described Chapter 5 summaries all the main points in the research, mentions the shortcomings and the future work Finally, the reference refers all the involved

researched, documents, projects, websites, books and e-books that the researcher used in

the research process

Chapter 2 - The Background Knowledge

1 Overview of Recommender Systems With the development of the Internet and the bloom of websites, the use of Internet is engaged in people’s everyday activities like news browsing, music listening, book reading, learning or even online purchasing Particularly, with the strong support of search engines such as Google, Yahoo, Bing , information searching becomes an essential behavior of the most Internet users nowadays to make a decision on everything

In many situations, however, search engines cannot give full support to satisfy user’s requirements, Take an example of buying a camera, customers can search products using search engines and a large variety of cameras will be available However, it makes them sunk because they have many criteria to care about when buying a camera: price, brands, lens type, etc How can the customers select the most suitable one for their needs? The

RS are different from search engine beacause in a search engine, the user must explicitly and precisely formulate at the beginning of the session his search query about what he wants to find In RS, however, the user doesnot need to specify precisely/ completely his

needs and preferences Moreover, for the search engine, the system finds those items that

satisfy the user’s indicated query.But for RS, the system returns those items that are most suitable (but not nessesarily satisfy completely) the user’s needs and preferences RS is a kind of exploring/ discovery tool

Recommender system (RSs) are decision support tools which help users to choose optimal solutions when there are a huge number of available (product or service) options but the user has a limited time or limited (or no) knowledge about the current problem

domain to make selection decisions by himself [33] It means that RSs are useful in case that user has a lack of knowledge about the topic, i.e., the user is not aware of the range

of available options; RSs suggest options to the users by exploiting its knowledge about the user’s needs and preferences; the information about each option is also provided to help the user select the most suitable ones For example, user does not know what to search and the user just wants to experience the system, not for any initial purpose Via user — system interaction, RSs get user’s information and base on users’ preferences, past buying behavior to provide the suitable recommendations for users In other words, the RSs learn users’ preferences over time and automatically suggest products that fit the learned user model Moreover, RSs are particularly used to solve the problem that users are overwhelmed by a large number of alternative options RSs provide limited number

of options that are ranked which save time and effort of the users

2 Main Recommendation Techniques

There are a variety of techniques proposed [34, 5] The recommendation techniques can

be distinguished on the basis of their knowledge sources For the aim of providing the readers with the basic knowledge about RSs, this part briefly introduces three main recommendation techniques Collaborative Filtering, Content — based recommendation approach, Knowledge - based and finally, the Hybrid Method which combines different recommendation techniques

recommendation technique [18] In this approach, the system tries to predict the opinion

the user will have on the different items and be able to recommend the best items to each user based on the user’s previous likes and the opinion of other like-minded users [34] In CF-based recommendation approach, the system needs a list of m Users and n Items One way to visualize it is the matrix m x n which contains the ratings of users on items Secondly, each user has a list of items he/she expressed the opinion about (via ranking or via purchase record) Then, we need a metric to measure the similarity between users and

a method for selecting a subset of neighbors for prediction Finally, a method for predicting a rating for items not currently rated by active user is needed CF — based recommendation approach has been implemented in [40, 3, 28] For example, Movielens

(http://www.movielens.org/login) is a CF system for movies A user will rate the movies

with the scale from | to 5, Movielens then uses the ratings of the community to “predict” the rating user might give to a movie and recommend movies that user might like The disadvantage of CF technique is that it requires a large number of users’ ratings on

products, which becomes a hard problem when the users or the items are new to the

system,

In the content-based recommendation approach, the system tries to recommend items that are similar to those that a given user has liked in the past [26] The system generates recommendations from two sources: the features associated with the products and the ratings the user has given them So it can be seen that the system analyzes the features/content of items which user has rated in the past Therefore, the very first important thing to do is to represent the item to easily model the users’ interest [26] Content-based recommendation is particularly used for text-based products (like web

pages) What we need to do to eliminate the unknown rating is finding a model of current user, based on that model predicting ratings of unrated items For example, Yahoo! News

observes what online news the user has read and learns to present the user with news he/she may like to read Some systems using Content- Based recommendation approach are [10, 25]

Both the collaborative filtering-based and content-based recommendation approaches suffer from the cold- start problem (34, 35] This is the well- known problem of handling new items or new users For instance, CF-based recommender systems can not recommend to a user who has not given any ratings on products or if an item is new to the catalog, it is weaker to consider compared with more rated products

The idea of the knowledge-based approach is that the knowledge structure about the users and the application domain are used to infer what products fit the user’s preferences [7] The most advantage is that this approach does not depend on the customer’s rates Knowledge can be expressed as a user model, a model of the selection process or a description of the items that will be suggested [18] However, the complex knowledge retrieval techniques and complicated knowledge representation are the main limitations

of this approach Case-based recommender systems [2, 18] are knowledge-based ones that apply the Case-Based Reasoning (CBR) problem-solving strategy, i.e., a machine learning approach, to exploit the knowledge contained in past recommendation cases Case-based reasoning (CBR) [2] is a machine-learning technique which is widely applied in many RSs [4, 7, 15, 23] The idea of this technique is to retrieve the knowledge

encoded in the previous cases to apply to the current case In CBR, it is assumed that the similar problems have similar solutions; therefore, the new problem can be solved by retrieving similar problems and adapting retrieved solutions When the system is presented with a new problem, it searches for the most similar case(s) in the case base and reuses an adapted version of retrieved solution to solve the new problem In CBR approach, each case is composed of 2 parts: the problem description and the solution By comparing the problem description parts of 2 cases, the system finds out the past case most similar to the current case The solution of that past case, which is accepted by the user, is reused or extracted to apply to the current case CBR is a cyclic problem solving process which is composed of 4 main steps: retrieve, reuse, adaption and retain [18] In the retrieval phase, the system firstly retrieves from memory experience about the

similar situations to the current situation After the retrieval phase, in the reuse phase, the

solution is considered and the system evaluates if that solution can be reused for the

current problem or what part of the solution can be reused In the next phase revise, the

reused case is adapted to fit the current problem In the review step, the solution is evaluated by applying it to the current case and understanding why it fails and making

corrections [18] After the recommendation ends, the new case is retained which means

being stored in the case base for the future use in the retain phase [2] This technique shows the advantage of modeling the relationship between the user’s selection and the user’s preferences and the information of the request context

Input New Problem

Figure 1 The Case-Based Reasoning problem solving cycle [18]

Finally, the Hybrid recommendation approach [5] is understood as the method of combining multiple recommendation techniques together in a variety of ways, producing recommendation and improving the recommendation performance Hybrid system gain better performance by eliminating the limitations of each traditional recommendation approach and exploiting their advantages It is noted that there is no standard hybrid recommendation system

3 Mobile RSs

Recently, mobile RSs are widely applied for some main tasks and functions as introduced here The mobile RSs are used for the purpose of tourist guides such as recommending the tourist attractions like museums, art galleries, sky routes or finding relevant services

10

like restaurants, hotels, transportation services [13, 24] Mobile RSs are also useful in route recommendation [37] Based on the position of the users detected by the system, the recommendation about the directions to reach a destination is given to user There are mobile RSs serving for the purpose of information recommendation such as news

recommendation, multi-media content recommendation [17]

Mobile RSs is still a challenging field and worth consideration and research In [30], the main obstacles of mobile environment need to be overcome are the limitations of mobile

devices (small screen, limited keypad ), the limitations of wireless network (low bandwidth network connection), influences of external environment (the contextual

factors: weather, position ), and the behavioral characteristics of mobile users (being less patient, preferring simple interactions )

The first shortcoming we need to take into account in mobile RSs is the limitation of user interface Different from RSs for computer, mobile RSs must be designed in such a way that is suitable for mobile users who often do not have much time and patience The small size of the mobile screen may significantly affect the user’s behavior due to the reason that it is not convenient for users to scroll down to view all the results compared with browsing the large screen Besides, users when using small screen tend to not fulfill the assigned tasks (ex., input initial information) therefore the system may not get the needed information

l1

Moreover, mobile devices provide very limited input and interaction capabilities The popular 12 numeric keypad causes difficulties in text — input Requirements of entering queries based on keyword or product feature values using keypad becomes hard to fulfill Mobile devices also have small number of control keys assisting users with navigation and scrolling tasks The simple devices include just two soft keys with variable functions Besides, the time for browsing the mobile devices of user is very short, maybe a few minutes Recently, there are some kinds of RSs interface which have been developed to support better browsing and direct manipulation approach of information access rather than the search model [30] There are some approaches ever used to address those

problems raised in mobile environment, some of which are listed below: Starfield

Display [12, 36], Displaying Similar Searches [9], Critique - and Preference-based

Interfaces [32] For further details about each approach, reader can refer to [30] for the

approach overview

4 Important issues in RSs

A notable characteristic of RSs is “personalization” “Personalization” can be understood

as “the ability to provide content and services tailored to individuals based on knowledge about their preferences and behavior” [16] The inputs about the user (user's preferences

or tastes, past buying behavior or the context within which the application is applied) and feedback of user to any recommendation can be “learnt” by the system to generate additional inputs for future recommendation It means that the system results are not merely the things that should be filtered out but those things are computed to be most

12

suitable to the user’s needs and preferences This characteristic helps to make the users get stick with the system and be able to recommend new items which are likely preferred

by the user Therefore, personalized systems make the users feel convenient and satisfied

with the recommendation results

The user profile, which stores the user’s long-term preferences, is used in RSs for the purpose of “personalization” The two types of information of user’s profile are described in [26]

a The user model is a description of the types of items that interest the user User model can be represented in many possible alternative ways as the content: user’s preferred keywords for text-based products, for example The user model typically describes stable preferences of users because it is built by mining all

system-user interaction (ratings and queries) The stable and short-term

preferences can be combined in such a way that the selection of items satisfying user’s short term preferences can be sorted due to stable preferences

b History of user’s interaction with RS It may contain the items that user has

viewed; user’s interaction (select that item, rate that item; user’s queries)

In the research of [32], the user’s preferences are distinguished with 2 kinds: long -term preference and session-specific preference The first one indicates the stable preferences that the system can acquire by mining the past sessions of user-system interaction For example, the user always likes Italian food or action movies While the second one is temporary and characterizes the user request context, for example, the user unexpectedly prefers pizza rather than noodle as usual or the recommended restaurant must be not too

13

far from the user’s current position to satisfy the space-time constraints Integrating 2

kinds of preferences makes the recommendations become flexible, which means being suitable for user’s preferences but also suitable for specific context

Conversational approach in RSs is an innovative approach and has been recently received much attention of the researchers [37, 19, 8] Instead of providing the users with recommendations and ending the session, this approach reduces the user’s effort to input their information at the beginning of the session but continuing interact with the users The system shows the recommendations to the user, the user early notices the benefit the system brings and is willing to interact with the system However, the computed recommendations, sometimes, can not satisfy the user totally because there might be differences between the system’s representation about the user’s preferences and the user’s true preferences and needs at the current request context Therefore, after showing the first recommendation list to the user, the system, during the interaction, must acquire more information; to better represent the user’s preferences more precisely and compute new more suitable recommendations Conversational approach is proved to perform better than the traditional single-shot approach [32]

Conversational recommender systems implement the asking/ answering conversation mode, the proposing/ criticizing conversation mode, or both of them The conversation recommender systems can interact with the users through typing/ clicking or with natural language interface [22] A popular kind of conversational approach is critique-based approach There are some RSs which exploit the critique based approach allows users to

14

make critiques on the item level For example, after receiving recommendations, the user can give rates on the item Other systems allow the users to make critiques on the item feature level such as (37, 6]

The term Context plays an important role in mobile RSs “Context is any information that

can be used to characterize the situation of an entity” [11] Entity can be a person, a place

or an object Due to the mobility characteristic; context involves all the factors such as the time of request, the user’s position at the time of request, the weather condition, the companion, etc The contextual information is one of the knowledge sources which is involved in the system recommendation procedure For example, [1] represents an approach of integrating the CF recommendation approach and the user’s location, the time of request and the types of user’s needs The system collects the user’s position, time and needs and filters out the items that are not closed to the user’s current position Then the system uses that information to find the similar users with the target user

User query encodes the system’s understanding of the user’s session-specific preferences and is used by the system for computing the recommendation list For example, a user query can be a set of preferred features for a product to serve for finding suitable products

in the catalogue

5 Item Sequences Recommender Systems

For activity mobile RSs, the contextual information such as the user’s position, the time

of the request plays a very important role Zheng et al [39] proposed a method of using GPS data and the user community’s comments about a location to recommend locations

15

to the target user It can also suggest the activities that can be taken in the surrounding area The RS proposed in [38] exploits the user’s position to recommend the locations for

people who want to go shopping The system Magitti [3] applies the hybrid approach,

combining two traditional recommendation techniques, i.e., collaborative filtering and content-based filtering, and the system also integrates contextual factors such as time of day, location, weather condition to recommend leisure activities for mobile users

Besides, Magitti infers the user’s favorite activity categories (Eat, See, Do, Read) based

on the content of the emails sent/received, the web pages visited, user’s working calendar In the approach introduced by Pinyapong and Kato [27], the system exploits the time of request, location and the purpose of the user (e.g., to find activity for relaxation, or to find place for shopping or observation) and the preferences (e.g., user like cars, like shopping, etc.) to compute appropriate recommendations for the user In our approach, we exploit four contextual factors, including the user’s position, the time of request, the weather condition, and the user’s companion The user’s companion may be his girlfriend/ her boyfriend, or his friends, or his family, or alone Obviously, this contextual factor has an impact on the user’s items selection decision

In the recent researches, the system almost represents and stores in the User Profile the user’s preferences about the features of items [27, 33, 3] In our proposed approach, the

system stores the user’s preferences (about the features of items) along with the contextual information (i.e., the user’s position, the time of request, the weather condition,

and his companion) Representing the user’s preferences associated with their collected

16

contexts enables the system to better “understand” the user’s preferences in specific contexts, and therefore, to compute more suitable recommendations for the users

For the problem of recommending item sequences, to our knowledge, there have not been any researches aiming at addressing this problem There have been some systems

introduced in the literature, which recommend sets (i.e., not sequences) of items For

instance, in [39] the system tries to find the correlation between the activities such as: food and drink, shopping, sports, movies and show in an area by building an activity- activity correlation matrix to recommend set of activities to users Another system is DieToRecs [15], which helps a user to plan a travel (i.e., to select a set of travel products,

services) in a selected destination It is noted that those above mentioned systems

recommend sets of items, but the order of the items are not considered In this paper, we introduce our proposed approach to the problem of recommending sequences (i.e., ordered sets) of items The system uses single item recommendation technique (SIR) to generate the content queries to search in the catalogue for products that match the user’s preferences The second recommendation technique applied in DieToRecs is Travel

Completion (TC) which recommends additional travel products or services to complete

the travel plan of the user It is noted that those systems recommend a set of items but the

order of the items has not been cared about Therefore, in this paper, we focus on

presenting an approach to recommend the sets of ordered items

17

Chapter 3 — The Proposed Recommendation Methodology

In this chapter, I describe in details the methodology to address the problem declared in the first chapter, which is recommending sequences of time — constrained items for mobile users There are 2 main points which are included in this part are 1) the formal representation, 2) the recommendation process

1 The Formal Representation

1.1 Item Representation

In our proposed recommendation methodology, the system represents an item as a feature

vector : x= (x), X2, ., Xn), where the value x; of the feature f; can be numeric, nominal, or

a set of nominal values We note that different item categories are represented in different vector spaces (i.e., are represented by different features) For example, in the system iDo,

an activity of watching movie “Fast Five” is represented as follows: x = (‘Fast Five”, {"Action”, “Crime”}, (" Vin Diesel”, “Paul Walker”, “Dwayne Johnson”}, “Justin Lin”, “Universal Pictures, USA”, “English”, “ Dominic and his crew find themselves on the wrong side of the law once again as they try to switch lanes between a ruthless drug lord and a relentless federal agent.”, “Megastar”, 5, “24/5/2011, 13h-15h") \t means

that the movie has title x,=“Fast Five”, genre x= {“Action”, “Crime”}, stars x3={‘ Vin Diesel”, “Paul Walker”, Dwayne Johnson’}, director xj= “Justin Lin”, producer x;=

“Universal Pictures, USA”, language xo=“English”, description x7=“* Dominic and his

crew find themselves on the wrong side of the law once again as they try to switch lanes between a ruthless drug lord and a relentless federal agent.”, cinema x3="*Megastar”,

18

price xo=5 ($), show time x1o="24/5/2011, 13h-15h” Among the item’s features, there

are features just for displaying information (ex Movie’s title, description, food’s name ) but there are critique-enabled features which not only contain the information about item but also can be used to make critiques (ex Ticket price, cinema, food style,

service level ) In the system iDo, leisure activities are classified into the three categories: “Movie”, “Food” and “Drink”; and each category has a different vector space

representation

1.2 The User Profile Representation

The user’s long-term preferences are stored in the User Profile, and represented as a vector u = (uy, U2, ., Un); Where u, indicates the user’s long-term preference for feature fj,

and u, is represented as: 4= /(9, cụ, Sj), where v; is a possible value of feature fi, cx is a

context attached with fj, and sj is the user’s selection frequency for value v, of feature f; at context c, For example, a user’s long-term preferences for the feature Genre (3) of the

activity category “Movie” is described as:

uz = {(“Horror”, “Morning”, 3), (“Romance”, “Morning”, 10), (“Action”, “Afternoon”,

6), (‘Horror”, “Sunny”, 3), (“‘Action”, “Rainy”, 12), (“Romance”, “Girl-friend”, 7),

(‘Comedy”, “Friends”, 7)}

It means that for the feature Genre (f2), the user used to choose horror movies 3 times in

the morning, romance movies 10 times in the morning, action movies 6 times in the

afternoon, horror movies 3 times when it was sunny, action movies 12 times when it was

rainy, romance movies 7 times when he was going with his girl-friend, comedy movies 7

preferences are associated with the contextual information The reason for this

19

representation decision is that for the same user at different contexts he may have different preferences on a feature For instance, for the action movie genre, in different contexts (e.g., when the user goes to see movie with friends, or when he goes with his girlfriend), the user may have different preference levels Therefore, if the system only stores the item features’ values, it cannot understand precisely the user’s preferences at different contexts, and therefore it may compute recommendations that are not suitable for the user at his specific context

1.3 The User Query Representation

The user query encodes the system’s current understanding of the user’s preferences (in the current recommendation session) We represent Q to capture/ encode completely and precisely the user’s session-specific preferences The user query Q is represented as Q =

(t, c, 0, p, w) where:

© f =(tsart, tsp) tepresents the time constraint which all the system’s recommended

sequences of items must satisfy For example, in the system iDo, t = (15:20, 18:20) means that all the activity sequences recommended by the system must start after (or

at) 15:20 and end before (or at) 18:20

© c=(cj, ,¢,) represents the contextual factors that models the user’s request context

In the system iDo, it exploits four contextual factors including the user’s position, the time of request, the weather condition and the user’s companion

® o =(0), , 0: represents the user’s preference on the order of the item categories (i-e.,

z is the number of the item categories), and ø, = 0 if the item category i does not

20

appear in the sequence For example, in the system iDo, there are three activity

categories, and the order o = (3, /, 2) means that (in the current recommendation

session) the user likes the order of activities as follow: go to eat first, then to drink, and finally to watch a movie

preferences, p is a vector in the same vector space of x, where p, is the system’s current guess of the user’s preference for feature f; ø, -” ?” means that the user`s preferences on the feature f; is not known

w = (W),W2, W,, represents the importance weights of the features for the user w;

is the importance weight of feature /; for the user (w; € [0,1]) For instance, in iDo,

the vector w = (0,0,

0,0.3,0.6,0,0.1,0,0) for the activity of watching movies means that for the user the language of the movie is the most important, followed by the

movie publisher and the cinema; the user is indifferent to other features

2 The Recommendation Process

A recommendation session starts when a mobile user requests the system for recommendation and ends when the user selects a recommended item sequence or when

he quits the session A recommendation session evolves in cycles In our approach, each recommendation cycle consists of the steps where the user browses the recommendation list to see the details of a recommended item sequence and then criticizes that sequence (ie., the user may criticize to the features of an item in the sequence, or the order of the items) The user’s critiques help the system adapt its representation of the user’s needs

21

and preferences and compute a new recommendation list containing more suitable item

sequences for him The recommendation process is illustrated in Figure 2

Figure 2 The Recommendation Process

At the beginning of the session (ie.,, when the user requests the system for

recommendations), the system builds the initial user query representation (Q°) by exploiting the knowledge sources of the user request’s contextual information, the user’s long-term preferences and the past recommendation cases The system, in the first phase, determines for each item category the set of candidate individual items where each item satisfies the user’s time constraint and is suitable for his preferences and current request context In the second phase, the system computes the sequences of items to build the

recommendation list, where each sequence is a set, in an order suitable to the user’s

preferences, of individual items generated in the first phase Then, the recommendation list is showed to the user, and the user can browse them to see the details of each recommendation If the user likes, but does not completely satisfy a recommended item

22