Using Persuasion Profile Information
Jelle de Boer
10540075
Supervisors:
Leonard Wolters & Maarten van Someren
Bachelor Artificial Intelligence
The Netherlands
Abstract
In this research, information from an adaptive persuasive system is incorporated in a collab-orative filtering-based recommendation system. In order to make recommendations, similarities between users are based on both the user-item correlation and the user-principle correlation. To obtain these similarities, the widespread cosine-similarity measure is employed, as well as a diffusion-based approach. From the experiments can be seen that the cosine-based variant without the extra persuasive profile information performs better than both similarity measures with the information included.
Contents
Acknowledgements 4 1 Introduction 5 1.1 Research question . . . 5 1.2 Hypothesis . . . 5 2 Background 6 2.1 Content-based filtering . . . 6 2.2 Collaborative filtering . . . 72.2.1 User- and item-based filtering . . . 7
2.2.2 Memory- and model-based . . . 7
2.2.3 Cold start problem and sparsity . . . 8
2.2.4 Context-aware recommendation . . . 8
2.3 Persuasive profiles . . . 8
2.4 Related work . . . 9
2.4.1 Tags and tripartite graphs . . . 9
3 Contributions 10 4 Methods 10 4.1 Similarity measures . . . 10
4.1.1 Diffusion-based similarity . . . 10
4.1.2 Cosine-based similarity . . . 11
4.2 Preference and recommendation . . . 11
4.3 Data set . . . 11
4.3.1 Principle type . . . 12
4.3.2 Implicit feedback . . . 12
4.3.3 Pre-processing . . . 12
4.4 Preparing matrices . . . 12
5 Experiments and results 14 5.1 ARC-score . . . 14 5.2 Experiment A . . . 14 5.2.1 Results . . . 14 5.3 Experiment B . . . 15 5.3.1 Results . . . 16 6 Conclusion 16
7 Discussion 17 7.1 Future work . . . 18
Acknowledgements
Hereby, I want to thank Leonard Wolters for letting me write my thesis at his company Crobox. Besides the fact that he took care of providing me the indispensable data set, it was a pleasure to work with him. Furthermore, I want to acknowledge Sander van Splunter for the practical guidance during the thesis project, and Maarten van Someren for checking on my progress so now and then.
1
Introduction
People are said to have had to deal with an in-formation overload since the rise of the internet. This phenomenon makes it difficult for internet users to differentiate the most relevant informa-tion to fulfill their search needs (Isinkaye et al., 2015). Methods such as search engines have been developed in order to facilitate this pro-cess. However, filtering information by means of search queries results in the same informa-tion for all users, regardless of their personal preferences and manners (Shang et al., 2010). With the objective to take this user data into account and therefore make filtered information more personalized, recommendation systems are becoming an important part of the internet, es-pecially in the field of online media-providing systems and e-commerce websites (Adomavicius & Tuzhilin, 2015; Burke, 2002).
Recommendation systems produce individ-ualized recommendations as output or guides the user in a personalized way to interesting or useful objects in a large space of possible op-tions (Burke, 2002). The most commonly used technique to obtain such recommendations is collaborative filtering, which can be described as people-to-people correlation (Schafer et al., 1999). Logged historical activities - input from the user - are used to find similarities between users’ preferences and tastes for certain items to obtain personalized recommendations. This data usually consists of the interaction between users and items, such as purchased products by the user in an online web shop, or ratings for movies on Netflix (Hu et al., 2008). In many cases however, it can be useful to consider ex-tra information in order to recommend items to users under certain circumstances (Adomavicius & Tuzhilin, 2015).
Zhang et al. (2010) introduce Collaborative tagging systems (CTS) as an source of extra in-formation to improve the accuracy of traditional recommendation systems. This system allows people to assign tags to items, which can both express users’ preferences and describe the ob-jects’ contents. Shang et al. (2010) combines this tag information together with the user-item interaction in collaborative filtering.
Analogously, this research uses the same ap-proach, although the third component differs. Instead of the tag information, persuasion pro-files are introduced as a third component in ad-dition to the item-user connection. These pro-files are derived from a persuasive system inten-tionally designed to change a person’s attitude or behavior (Fogg, 1999).
1.1 Research question
The goal of this research is to determine if this information - the persuasive profiles - can be used in order to improve the performance of a recommendation system. Therefore the research question is as follows:
Will a collaborative filtering algorithm with diffusion-based similarity which integrates per-suasive profiles, perform better than traditional collaborative filtering?
1.2 Hypothesis
Since the persuasive profile information says something about the user, a positive effect on the accuracy of the recommendation system is expected. Moreover, it is said (Kaptein, 2011) that individual users respond differently to in-fluence strategies which might make it possible to differentiate among users in order to increase similarities between actual relevant users. This
expectation implies that people with a certain persuasive profile are looking for the same sorts of items.
Furthermore, as Zhang et al. (2010) evalu-ated, collaborative filtering based on diffusion-similarity with the use of tag information per-forms better in two ways. First, it showed better results than cosine-based collaborative filtering which includes the same tag information. Sec-ondly, the diffusion-based variant on its own per-formed better with tag information than with-out. In addition, in contrast to the explicit data set of (Shang et al., 2010), the data set in this research is implicit - which is data without any explicit feedback; it contains click behaviour -. Since the diffusion-based approach is indepen-dent from the substantive form, it anticipated that diffusion-based will perform better, even with implicit data.
In the following section, the phenomena mentioned earlier will be further explained. Hereafter the contributions of this research are mentioned in section 3. In the method section the pre-processing steps of the data set and the different similarity measures will be explained. Furthermore, the executed experiments are de-scribed and the results will be evaluated to-gether with the explanation of the evaluation metric in section 5. This thesis is summarized in section 6, and some possible improvements are addressed in the final section.
2
Background
Recommendation techniques can be differenti-ated based on the knowledge sources; where does the knowledge needed to make recommen-dations on come from? Basically, this informa-tion comes from two sources, i.e. knowledge
that comes from interaction with the system, e.g. rating products or movies, and informa-tion that is located in the item itself (Burke, 2005). Consequently, among other, two of the most commonly used algorithms can be distin-guished based on these types of sources, i.e., content-based filtering and collaborative filter-ing (Candillier et al., 2009).
Figure 1: Content-based filtering
2.1 Content-based filtering
A content-based recommendation system con-structs a profile of a user based on its prefer-ence for certain features in an item (see figure 1). Based on this profile and the features of a new item (e.g.: in case of a movie, the genre could be one of the features), the system de-cides if this item will be recommended to the user (Candillier et al., 2009; Burke, 2002).
Despite the ability to recommend an item to a user by means of its features, this system deals with some disadvantages. The content-based approach is restricted by the fact that a lot of features in an item haves to be present in order to be recommended. If some charac-teristics in an item are not explicitly defined in contrast to other items, it becomes difficult to make good recommendations (Candillier et al., 2009; Burke, 2002). Another drawback is that
a trained content-based recommendation system is not able to suggest cross-genre, e.g. a system is trained on the preference of jazz for someone who likes jazz, but also folk music. This person will not receive any recommendations for folk music (Burke, 2002).
2.2 Collaborative filtering
On the other hand, collaborative filtering is not burdened by the shortcomings of the content-based approach. In contrast to this type of fil-tering, collaborative techniques are independent of any machine-readable representation of the objects being recommended, and work well for complex objects such as music where variations in taste are responsible for the variation in pref-erences (Burke, 2002). Additionally, for collabo-rative filtering only past behaviour of the user is sufficient in order to deduce unknown relation-ships between users and item (Hu et al., 2008).
(a) User-based (b) Item-based
Figure 2: Collaborative filtering
To suggest new items to users, collaborative filtering makes use of users’ opinions about items (Candillier et al., 2009). These opinions can be explicitly given by the user in the form of a rat-ing, or can be derived from implicit feedback, i.e. purchase records or click behaviour (Lee et al., 2007; Sarwar et al., 2001). The users’
ap-preciation for items can be used in two ways to deduce recommendations.
2.2.1 User- and item-based filtering The first of two types of collaborative filtering is based on the similarity between items (see figure 2 b). This approach looks into the set of items the target users prefer and computes how similar they are to a target item. Based on this simi-larity often the K most items are recommended (Sarwar et al., 2001).
However, the most successful and widespread technique is based on the similar-ity between users instead of items (see figure 2 a). This technique locates peer users with a similar appreciation history to the target user; recommendations are subsequently made based on this neighbourhood (Burke, 2007). Because this type of recommendation system will be used in this research, a detailed explanation is given in section 4.
2.2.2 Memory- and model-based
In the field of collaborative filtering research, two types of methods can be distinguished, namely the memory-based and the model-based approach. The term collaborative filtering is sometimes used to refer only to the memory-based method (Yu et al., 2004).
Memory-based collaborative filtering com-pares users with each other using correlation or other kinds of measures (Burke, 2002). As ear-lier mentioned, the most similar neighbours are used to predict the users’ preference for a certain item.
In contrast, model-based algorithms use ma-chine learning or data mining techniques to build a model in order to recommend in a faster way (Isinkaye et al., 2015). This kind of
algo-rithm analyses a user-item matrix to get a pic-ture of the relations between items. By means of matrix completion techniques, unknown val-ues within the user-item matrix are predicted (as cited in Isinkaye et al., 2015).
2.2.3 Cold start problem and sparsity One of the main challenges in recommendation systems is the cold start problem, which is re-lated to the sparsity of information available. As Lika et al. (2014) evaluated, there are three sorts of cold start problems, namely:
1. recommendations for new users;
2. recommendations for new items;
3. recommendations on new items for new users
The aggregated burden is that both new users and new items presented to a recommendation system, lack information, explicit or implicit feedback, which leads to poor recommendations. Furthermore, sparsity as such is a problem to overcome to make effective predictions of which items a user potentially prefers (Adomavicius & Tuzhilin, 2015). Usually, the obtained informa-tion in recommendainforma-tion systems often is very little compared to the large amount of available items that can therefore be recommended; e.g. when a user has an unusual taste compared to the rest of the population within the recommen-dation system, there are barely any similarities which also leads to bad recommendations.
2.2.4 Context-aware recommendation To overcome the problem of sparsity and there-fore implicitly the cold-start problem, extra in-formation can be taken into account. Instead of
only using users and items as dimensions, con-textual information such as time could be used as a third dimension in addition to the former two (see figure 3). The three tables in figure 3 define the sets of users, items and times re-spectively, which together represent the prefer-ence of the user for a certain item at a certain time; e.g., a user might prefer a movie in the weekend but a TV-show on a weekday (Ado-mavicius & Tuzhilin, 2015). This insight is in line with observations in behavioural research which assumes that decision making is contin-gent on contextual aspects. Therefore, it can be said that the performance of a recommendation algorithm builds on the degree to which contex-tual information is incorporated (Adomavicius & Tuzhilin, 2015).
Figure 3: Multidimensional model for the User x Item x Time recommendation space. Adapted from Adomavicius & Tuzhilin (2015)
.
2.3 Persuasive profiles
Persuasive technologies are designed to change a person’s behaviour, and use a broad range of methods to attain this (Kaptein, 2011). As Kaptein (2011) mention, one of these implemen-tations is found in e-commerce, where webshops make use of a combination of technology and
principles from social sciences in order to in-crease consumer satisfaction and revenue. The extra information employed in the recommenda-tion algorithm in this research is derived from such a system; hence it is worthwhile to address this - the information - persuasion profiles fur-ther.
To attain behavioural change, a lot of influ-ence strategies are described by different the-orists, among other, Cialdini (1987). He ad-dresses the following six principles of persuasion: 1. Reciprocation: According to this rule, people should repay what someone has provided them with; therefore a future obligation is made; e.g. free samples are a marketing trick that way (Cialdini, 1987);
2. Commitment and consistency: It is peo-ples’ desire to be consistent with what we have done in the past. People convince themselves to act in line with what they believe in (Cialdini, 1987);
3. Social proof: This means that it is more appropriate to do things when a lot of other people are acting in the same man-ner. People believe that fewer mistakes are made by acting in line with social evidence (Cialdini, 1987);
4. Liking: People are susceptible to requests made by people they know or like. Often this technique is used by other than ac-quaintances and as long as someone does not have the feeling being manipulated, it is a very useful persuasive strategy (Cial-dini, 1987);
5. Authority: If a request comes from a recognized authority, people will comply sooner (Cialdini, 1987);
6. Scarcity: When people make decisions, they tend to be more motivated by the thought of losing something (Cialdini, 1987)
As Kaptein & Eckles (2010) evaluate, these influence strategies can be used to establish be-haviour change. Among other, in e-commerce these principles are embedded in a so called mean-adaptive persuasive system. An im-portant property of such a system is end-independence, which means that the knowledge gained about people can be used independently from the end goal. Given the differences be-tween people in sensitivity to these principles, these persuasive systems can respond to these distinctions and could create individual persua-sive profiles, to indicate which principles are ex-pected to be most effective to persuade.
2.4 Related work
2.4.1 Tags and tripartite graphs
As mentioned earlier, extra information can be added to the user-item connection in favour of improving the performance of the algorithm. As a new source of such information, Zhang et al. (2010) propose to employ Collaborative tagging system, which allows people to assign tags to their collected items. This tag information can be seen as abstracted content of items which implicitly says what the user thinks about the tagged item (Zhang et al., 2010).
Zhang et al. (2010) and Shang et al. (2010) both incorporate this information into a tripar-tite graph together with the user and item infor-mation, and subsequently calculate item-based similarity and user-based similarity respectively (figure 4) by means of a diffusion process. The diffusion-based user similarity measure is
em-Figure 4: Diffusion-based similarity on a tripartite graph. Adapted from Shang et al. (2010)
ployed in this research, and will be explained in more detail in section 4.
3
Contributions
This research will cover the following matters: 1. The diffusion-based similarity measure is
used with persuasion profile information in order to validate the method of (Shang et al., 2010) with another kind of data set; 2. In contrast to the explicit data set in (Shang et al., 2010), this research makes use of an implicit data set;
3. The persuasion profile information is used in collaborative filtering in order to aim for better recommendations than traditional collaborative filtering;
4. Since collaborative filtering suffers from sparsity, it has been examined whether persuasion profile information can partly overcome this problem.
4
Methods
In this research, two different types of experi-ment are executed; however, both of them take the same approach with regard to making rec-ommendations based on collaborative filtering.
Firstly, the similarity between users is calculated after which the preference of a user on all un-collected items is obtained, using this similarity scores. Finally, a list of items is recommended to the user by sorting them by its preference scores in descending order.
4.1 Similarity measures
4.1.1 Diffusion-based similarity
As earlier mentioned, Shang et al. (2010) pro-pose a collaborative filtering algorithm based on a measure of user similarity which integrates user preferences of both collected objects and used tags. The tags T in the formula below can analogously be seen as the persuasive profile in-formation used in this research.
As described in Shang et al. (2010), to obtain similarities between users, two pairs of correla-tions are employed: user-object and user-tag. If user u has collected an item α, auα = 1,
other-wise auα = 0. The initial step of the diffusion
process is to allocate a unit of resource to the target user v, which will be distributed to other users, such that each user gets a specific degree (Shang et al., 2010).
As visualized in figure 4, in the first diffu-sion step, target user v distributes the resource equally to all the objects he has collected. Then, each object distributes its resource equally to all
the users having collected it. Similarity between u and v reads as follows:
suv= P α∈O
auα× rαv
k(α) (1)
Here k(α) is the degree of object α in the user-object bipartite graph, O is the set of user-objects, and rαv is equal to what the object α gets from
user v. In the same manner, a unit of resource is initially located on the target user v, which will be equally distributed to all tags he has used; then the received resource is distributed to all its neighbouring users. Tag-based similarity be-tween user u and v reads as follows:
s0uv= 1 k0(v) P t∈T a0ut× a0vt k0(t) (2)
where k’(t) and k’(v) are the degrees of tag t and user v in the user-tag bipartite graph.
The user-item correlation and the user-tag correlation can be combined in a linear way, as follows:
s∗uv= λsuv+ (1 − λ)s0uv (3)
where λ ∈ [0, 1] is a tuneable parameter. The bigger this λ becomes, the smaller the influence of the tag information is on the similarity be-tween two users.
4.1.2 Cosine-based similarity
Among others, one of the most popular similar-ity measures in collaborative filtering is cosine-based similarity (Isinkaye et al., 2015; Shang et al., 2010; Lee et al., 2007). Therefore it is used in this research as the implemented similarity mea-sure to compare the diffusion-based approach with.
Cosine similarity is characterized as follows (Lee et al., 2007): s(a, b) = P j (Paj) × (Pbj) rP j (Paj2)rP j (Pbj2) (4)
Here, Paj and Pbj are respectively the
prefer-ences of user a and user b for item j.The similar-ity between them is equal to the angle between the vectors of user a and user b; the smaller this angle is, the more similar these users are (fig-ure 5). The persuasive profile information is in-corporated in the same way as in the diffusion-based approach, i.e. by means of a linear for-mula.
Figure 5: The cosine similarity between users can be viewed as the angle between their vectors. Adapted from Lee et al. (2007)
.
4.2 Preference and recommendation
The preference of user v for an item α is:
pvα= P u6=v
s∗uvauα (5)
Finally, all objects that user v has not collected, are being sorted in descending order by their preference score, and will be recommended to the user (Shang et al., 2010).
4.3 Data set
In this research, the data consists of logged ac-tivities retrieved from an e-commerce website, which of is partly presented in figure 6 . The columns of interest to this research are id (user), itemId (item) and principleType (principle).
Figure 6: Part of the logfile used in this research
4.3.1 Principle type
The last column principleType originates from a persuasive system and represents a certain type of influence strategy - the six persuasive princi-ples of Cialdini (1987). These strategies are all implemented in a specific manner, characteristic for the principle in question. For instance, when an online store show to a user how many people already bought a specific item, the system is us-ing the principle consensus in order to persuade him (Kaptein, 2011).
In the rest of this thesis, when the concept ’persuasive profile information’ is used, these principle types are meant.
4.3.2 Implicit feedback
Most recommendation systems rely on explicit feedback, which includes explicit input by users regarding their interest in products (Hu et al., 2008). On the contrary, implicit feedback indi-rect reflects opinions by observation of user be-haviour; and is the type of feedback employed in this research, i.e. clicks by users on items.
Figure 7: Removing the color-code for itemId
.
4.3.3 Pre-processing
Before the logfile is converted any further, it is prepared to prevent any inconveniences while compiling the algorithms.
The itemId initially contained a colour code, which caused a wide range of items due to items, because certain items are available in more than one colour. To partly overcome a sparse matrix, the itemId is modified and afterwards only con-sists of the item code. Therefore, two users who clicked on the same item but with a different color now prefer the same item (figure 7).
Furthermore, the baseline (0) principle is re-moved from the principleType column. This principle type value is only used for testing and transparent to the user and therefore not rele-vant for this research. In addition, if present, NaN-values and double rows are removed from the logfile.
4.4 Preparing matrices
As earlier mentioned, collaborative filtering suf-fers from sparsity. To reduce this sparseness, users as well as items haves to meet certain con-ditions. Because there are a lot of items present in the logfile, items that only interacted with less than three users - that is to say only three users clicked on this item - are removed. In the same way, users who clicked on less than three items are taken out. Besides, to evaluate the
implemented algorithms, the data set is split in a train and a test set. If a user only has one interaction and is used as test example, there is no interaction left to calculate similarity for this user.
Figure 8: The problem with applying conditions for users and items
A footnote needs to be made concerning meeting these conditions (the minimal number of interactions for both users and items). As in the example in figure 8, it is not sufficient to sim-ply apsim-ply both conditions. If C1 is applied, I1 is
dropped. Subsequently, condition C2 is applied
whereby U1 and U2 are deleted. Even though it
seems that both conditions are obtained, after applying C2, condition C1 fails, and C10 has to
be applied again.
To prevent this iterative process, figure 9 shows how this problem is treated. By apply-ing the first condition twice and the second con-dition once, only the former concon-dition is met. However, to encounter the problem that a user with only one interaction is in the sample set, sample users are derived from the data using only condition C2 to determine a set of sample
users. The few users who do not meet the sec-ond csec-ondition and are still in the data set are taken for granted and used as train data.
Figure 9: Resolve the condition problem
.
After the pre-processing steps are executed, the logfile is converted to two matrices, a user-item matrix, and a user-principle matrix. In both matrices the rows corresponds with the user vectors, which means that for every user a one is present in the cells related to an item or principle which he interacted with, or a zero if there was no interaction (see figure 8 for an example of a user-item matrix).
In both experiments one matrix is taken into account, i.e. a binary one - a 1 is present when there is an interaction with this item, a 0 when there is no interaction -.
The number of unique users after pre-processing is 5409, and there are 939 items; and despite the six persuasive principles described earlier, this data set only contains five. Hence, the dimensions of the user-item matrix are 5409 by 939 with 29435 unique interactions, and the user-principle matrix is 5409 by 5 with as many interactions. The distribution of the number of interaction users had - the session length - can be seen in figure 10.
Figure 10: The number of users per session length
.
5
Experiments and results
5.1 ARC-score
In contemplation of comparing the different kinds of algorithm and therefore investigating the performance, the ARC-score is used.
The ARC-score stands for ’average rank for the correct recommendation’, and represents the rank of a sample item in the ordered list which is retrieved after calculating the possible pref-erences of a user for the uncollected items from the data set (Burke, 2004) (figure 11).
Figure 11: Example for retrieving the ARC-score
.
5.2 Experiment A
As mentioned in section 3, in this research the proposed algorithm in (Shang et al., 2010) is val-idated by a data set containing persuasive profile information and compared to traditional collab-orative filtering using cosine similarity. There-fore, the research design of this experiment is al-most the same. In addition to the comparison of
the similarity measure, the differences in perfor-mance of the algorithms with different degrees of persuasive profile information are examined.
To test the different similarity types for col-laborative filtering, the data set is split into a test set (5 %) and a train set (95%), as can be seen in algorithm 1. The samples in the test set are taken out of the matrix to be evaluated afterwards.
Next, a matrix is created for both user-item combination and user-principle combina-tion. Subsequently, by means of these matrices, the similarities between users are calculated. As in formula 4 both similarities are combined in a linear form, in which the parameter λ expresses the degrees of persuasion profile information in the final similarity values used for prediction.
5.2.1 Results
Figure 12 shows the outcome, the ARC-score as a result of different values for λ for both diffusion-based similarity and cosine-similarity applied on a matrix consisting only of ones and zeros. Compared to the diffusion-based ap-proach, the cosine variant performs worse for all values for λ. In addition, since the bigger the value of λ is the smaller is the influence of the persuasive profile information, and it can be seen that cosine similarity only performs better with a bigger lambda, hence performance is only worse with the extra information used.
The diffusion-based variant however, per-forms slightly better with a value of 0.8 than with λ = 1.0. This means that when there is a little persuasive profile information present in the merged linear formula, the diffusion-based approach takes advantage of it. After averaging over ten independent runs with different random samples, for the optimal value of λ mentioned
Algorithm 1 Experiment A
1: Split log in train and test set
2: Get sample indices from test set
3: Create User-Item matrix from train set
4: Create User-Principle matrix from train set
5: Calculate similarities for UI-matrix for all similarity measures
6: Calculate similarities for UP-matrix for all similarity measures
7: for λ in range do
8: Merge UI- and UP-similarities taken the value for λ into account
9: Predict preferences of users for items
10: Evaluate prediction for samples with ARC-score
11: end for
above, an increase of 0.93% is achieved.
Figure 12: Performance of both diffusion-based and cosine-based collaborative filtering for a binary ma-trix
.
5.3 Experiment B
In this experiment recommendations to users are made per session length by means of algorithm 2 to examine the effect of more data - a longer session length - on the ARC-score. The test set can be seen to be constructed depending on the session length; e.g. if session length equals four, the test set consists of samples which used to be the fifth activity session length equals five -to the corresponding users. This is achieved by sorting the session for every user on ts
(times-tamp) which results in a chronological list (see figure 13).
Figure 13: Log file sorted by ts
.
In experiment B, both diffusion-based sim-ilarity and cosine simsim-ilarity are executed with different values for λ. As mentioned in ex-periment A - and seen in figure 12 - the op-timal value for the diffusion-based approach is 0.8 which caused a minor improvement com-pared to λ = 1.0. For this reason it will be compared to diffusion-based similarity without persuasive profile information (λ = 1.0) to in-vestigate the effect on the ARC-score for differ-ent session lengths. To measure the difference with cosine-similarity properly, the latter uses the same λ values.
Just as experiment A, this experiment is run ten times using different (random) test sets.
Figure 14: Plot of both similarity measures with two different values for λ, per session length
5.3.1 Results
Figure 14 shows the performances of the two dif-ferent similarity measures for the first fourteen interactions of the sample users.
As can be seen in the figure, the diffusion-based approach is performing almost the same for both values of λ (the red and blue lines), which means that adding a degree of persuasive information makes hardly any difference.
Furthermore, the green line, which corre-sponds to the cosine-based approach without any degree of extra information has better ARC-scores for all session lengths compared to the other three lines. When any extra information is actually present (the yellow line), the cosine variant is performing a lot worse.
6
Conclusion
In this research persuasive profile information is introduced as a third component to the tradi-tional user-item correlation. This information is incorporated in the collaborative filtering ap-proach for recommendation systems. As can been seen in section 3, by taking this approach several sub-components are addressed in order to examine if the diffusion-based approach will
perform better than the traditional approach of collaborative filtering with cosine similarity.
From the results from experiment A it can be said that the diffusion-based approach with the persuasive profile information performed slightly better than the variant without the ex-tra information, as function of λ. Seen this result (figure 12), the method of Shang et al. (2010) worked in the same manner for an im-plicit data set - which means for data without any explicit feedback from the user; e.g. rat-ings. However, the difference in performance for the diffusion-based approach for all values of λ bigger than 0.4, is very small in this research. In addition, cosine similarity using the persua-sive profile information is performing a lot worse than the same similarity measure without mak-ing use of the information.
Furthermore, since collaborative filtering suffers from sparsity, it is examined if the diffusion-based similarity measure in combina-tion with the persuasive principle informacombina-tion could overcome this problem partially. First it can be seen from experiment B that the diffusion-similarity measure is almost perform-ing the same for both with and without the per-suasive principle information despite the session length. Therefore, the added information does not affect the diffusion-based approach if it is examined per session length. It could be said - seen this result - that the earlier mentioned difference evaluated from experiment A, can be neglected.
Comparing the collaborative filtering algo-rithm with the cosine-similarity measure to both diffusion-based variants, it can be said from the results from experiment B that the former is per-forming better than the latter per session length. Therefore it can be said that the diffusion-based similarity measure incorporated in collaborative
Algorithm 2 Experiment B
1: Split log in train and test set . All users in the test set have a minimum session length
2: for SessionLength in range do
3: Take for every sample user all entries corresponding to SessionLength out of the test set
4: Take for every sample user all entries corresponding to SessionLength + out of the test set
5: Get all samples which are not in SessionLength as sample for every sample user
6: All samples together are the test set for this session length . Rest is train datas
7: Get sample indices from test set
8: Create User-Item matrix from train set
9: Create User-Principle matrix from train set
10: Calculate similarities for UI-matrix for all similarity measures
11: Calculate similarities for UP-matrix for all similarity measures
12: Merge UI- and UP-similarities . Here lambda is chosen which performed best in Experiment A
13: Predict preferences of users for items
14: Evaluate prediction for samples with ARC-score
15: end for
filtering with persuasive principle information is not performing better than tradition collab-orative filtering using the cosine-similarity ap-proach without the extra information.
7
Discussion
Seen the outcome of this research, the following can be explicitly addressed.
The data set only consisted of 5409 users and 393 items, and most of the users only had relatively very few interactions (see figure 10). If a user-item matrix is derived from these fig-ures, this matrix will become very sparse. It can be suggested that both similarity approaches suffered from this sparseness, which generally makes it hard to make proper recommendations (Lika et al., 2014).
On the other hand, the user-principle ma-trix was less sparse and therefore it should be easier to find better similarities between users. However, this matrix only consist of five
princi-ple types, which makes if harder to differentiate among users; e.g. if several users interact with five items and all items are accompanied with all five different principle types, they are very close related to eachother according to the persuasive profile information.
Furthermore, the data set is reduced to ma-trices which contained only binary values (one and zero) which only represented that the users interacted with these items were a one was present, hence the number of interactions is not taken into account.
Seen the likely inability to positively affect the performance of the recommendations of the user-principle matrix, due to the few number of principles in combination with a relatively large number of users and the binary values, this could be a reason for the bad performance of the cosine-similarity approach and the small differ-ence between the two types of diffusion-based approaches.
7.1 Future work
In the future, this research can be expanded by using a bigger data set consisting of more users and interactions in order to have less sparse ma-trices.
Furthermore, instead of only using the infor-mation that a user actually interacted with an item, in future research more interactions with one item could be taken into account. Maybe this ensures a clearer picture of what kind of principle type a user is sensitive for and there-fore can help to improve the diffusion-based ap-proach due to a weighted variant. In addition, other theorist defined more than forty principle types which can be used instead of five to find better similarities among users.
At last, in this research diffusion-based col-laborative filtering was used to examine the ef-fect of the persuasive profile information on rec-ommendation, because a same sort of third com-ponent was used in (Shang et al., 2010), i.e. the tag information. Unfortunately the desired re-sult was not obtained. To further explore the effect of persuasive profile information on recom-mendation systems, other methods can be used, such as pre-filtering (Burke, 2002), or an item-based approach (Zhang et al., 2010).
8
References
Adomavicius, G., & Tuzhilin, A. (2015). Context-aware recommender systems. In Recommender systems handbook (pp. 191–226). Springer.
Burke, R. (2002). Hybrid recommender systems: Survey and experiments. User modeling and user-adapted interaction, 12 (4), 331–370.
Burke, R. (2004). Hybrid recommender systems with case-based components. Advances in Case-Based Reasoning , 1087–1138.
Burke, R. (2005). Hybrid systems for personalized recommendations. Intelligent Techniques for Web Personalization, 133–152.
Burke, R. (2007). Hybrid web recommender systems. The adaptive web, 377–408.
Candillier, L., Jack, K., Fessant, F., & Meyer, F. (2009). State-of-the-art recommender systems. Collaborative and Social Information Retrieval and AccessTechniques for Improved User Model-ing.
Cialdini, R. B. (1987). Influence (Vol. 3). A. Michel.
Fogg, B. J. (1999). Persuasive technologies. Communications of the ACM , 42 (5), 27–29.
Hu, Y., Koren, Y., & Volinsky, C. (2008). Collaborative filtering for implicit feedback datasets. In Data mining, 2008. icdm’08. eighth ieee international conference on (pp. 263–272).
Isinkaye, F., Folajimi, Y., & Ojokoh, B. (2015). Recommendation systems: Principles, methods and evaluation. Egyptian Informatics Journal , 16 (3), 261–273.
Kaptein, M. (2011). Adaptive persuasive messages in an e-commerce setting: the use of persuasion profiles. In Ecis.
Kaptein, M., & Eckles, D. (2010). Selecting effective means to any end: Futures and ethics of persuasion profiling. In Persuasive (pp. 82–93).
Lee, T., Park, Y., & Park, Y.-T. (2007). A similarity measure for collaborative filtering with implicit feedback. Advanced Intelligent Computing Theories and Applications. With Aspects of Artificial Intelligence, 385–397.
Lika, B., Kolomvatsos, K., & Hadjiefthymiades, S. (2014). Facing the cold start problem in recom-mender systems. Expert Systems with Applications, 41 (4), 2065–2073.
Sarwar, B., Karypis, G., Konstan, J., & Riedl, J. (2001). Item-based collaborative filtering rec-ommendation algorithms. In Proceedings of the 10th international conference on world wide web (pp. 285–295).
Schafer, J. B., Konstan, J., & Riedl, J. (1999). Recommender systems in e-commerce. In Proceedings of the 1st acm conference on electronic commerce (pp. 158–166).
Shang, M.-S., Zhang, Z.-K., Zhou, T., & Zhang, Y.-C. (2010). Collaborative filtering with diffusion-based similarity on tripartite graphs. Physica A: Statistical Mechanics and its Applications, 389 (6), 1259–1264.
Yu, K., Schwaighofer, A., Tresp, V., Xu, X., & Kriegel, H.-P. (2004). Probabilistic memory-based collaborative filtering. IEEE Transactions on Knowledge and Data Engineering, 16 (1), 56–69. Zhang, Z.-K., Zhou, T., & Zhang, Y.-C. (2010). Personalized recommendation via integrated
dif-fusion on user–item–tag tripartite graphs. Physica A: Statistical Mechanics and its Applications, 389 (1), 179–186.
