Hybrid recommender systems combine several algorithms based on their hybridization strategy. Prediction algorithm selection strategy directly in uence the accuracy of the hybrid recommenders. Recent research has mostly focused on static hybridization schemes which are designed as xed combinations of prediction algorithms and do not change at run-time. However, people's tastes and desires are temporary and gradually evolve. Moreover, each domain has unique characteristics, trends and unique user interests. In this paper, we propose an adaptive method for hybrid recommender systems, in which the combination of algorithms are learned and dynamically updated from the results of previous predictions. We describe our hybrid recommender system, called AdaRec, that uses domain attributes to understand the domain drifts and trends, and user feedback in order to change it's prediction strategy at run-time, and adapt the combination of content-based and collaborative algorithms to have better results. Experiment results with datasets show that our system outperforms naive hybrid recommender.
Copyright is held by the author/owner(s). Workshop on the Practical Use of
Recommender Systems, Algorithms and Technologies (PRSAT 2010), held
in conjunction with RecSys 2010. September 30, 2010, Barcelona, Spain.
Content-based and collaborative ltering are the two major
recommendation techniques that have come to dominate the
current recommender system area. Content-based
recommender system uses the descriptions about the content of the
items (such as meta-data of the item), whereas collaborative
ltering system tries to identify users whose tastes are
similar and recommends items they like. Other recommendation
technologies include hybrid techniques, knowledge-based
approaches etc [
Previous research in this area, has shown that these
techniques su er from various potential problems-such as,
sparsity, reduced coverage, scalability, and cold-start problems
[
Most of the recommender system implementations focuses
on hybrid systems that use mixture of recommendation
approaches [
In this paper, we describe an Adaptive Hybrid Recommender System, called AdaRec, that modi es its switching strategy according to the performance of prediction techniques. Our hybrid recommender approach uses adaptive prediction strategies that determine which prediction techniques (algorithms) should be used at the moment an actual prediction is required. Initially we used manually created rulebased strategies which are static. These static hybridization schemes have drawbacks. They require expert knowledge and they are unable to adapt to emerging trends in the domain. We now focus on prediction strategies that learn by themselves.
The paper is organized as follows. Related work is described in Section 2. In Section 3, we present the adaptive prediction strategy model for hybrid recommenders. We then describe our experimental recommender systems' architecture & learning module that dynamically adjusts recommendation strategy in response to the changes in domain. An initial evaluation of our approach, based on MovieLens dataset, is presented in Section 6.
Personalization techniques have been investigated extensively
in several areas of computer science. Especially in the
domains of recommender systems, personalization algorithms
have been developed and deployed on various types of
systems [
There have been several research e orts to combine di erent
recommendation technologies. The BellKor system [
The use of machine learning algorithms for user modeling
purposes has recently attracted much attention. In [
In our system, we choice the Duine Framework2 for our
recommendation engine component, which is an open-source
hybrid recommendation system [
Hybrid recommendation systems combine multiple algorithms and de ne a switching behavior (strategy) among them. This strategy decides which technique to choose under what circumstances for a given prediction request. Recommender system's behavior is directly in uenced by the prediction strategy. The construction of accurate strategy that suits in all circumstances is a di cult process. A well-designed adaptive prediction strategy o ers advantages over the traditional static one.
In our approach we use the switching hybridization in
order to decide which prediction technique is most suitable
to provide a prediction. Prediction techniques, also called
the predictors are combined into an upper prediction model
that is called prediction strategy. The central concept in
combining multiple predictors using the switching
hybridization method is the prediction strategy. Prediction Strategy,
which de nes the algorithm selection strategy, changes the
behavior of the recommender engines at run-time.
Most of the currently available personalized information
systems focus on the use of a single selection technique or a xed
combination of techniques [
Di erent design approaches might be used for the prediction strategy adaptation. Rule based, case based, arti cial neural networks or Bayesian are some of the learning techniques. Each technique has its own strengths and weaknesses. In this paper, we introduce self adaptive prediction strategy learning module which employs a strategy based on 2http://duineframework.org/ its attached learning technique. Learning module initializes prediction engine according to the speci ed machine learning technique. This prediction strategy adapts itself to the current context by using the previous performance results of the techniques. Di erent machine learning algorithms that induce decision trees or decision rules could be attached to our experimental design.
Duine Recommender o ers extensive options for con guring
various recommender algorithms. It provides a sample of
most common recommendation algorithms that can be
combined in algorithm strategies. In the Duine Recommender
the knowledge that designs the selection strategy is provided
manually by experts [
Hybridization of a recommender system employs using the best prediction technique from the available ones. The main purpose of a prediction strategy is to use the most appropriate prediction technique in a particular context. Adaptive prediction strategy depicts the selection rules of prediction techniques. Figure 1 shows a sample prediction strategy that decides when to use which predictors (gray nodes) by using the threshold values (arrows).
Prediction strategy employs the selection rules of the
available prediction techniques. Our experimental hybrid
recommender system has plenty of pre-de ned prediction
techniques. These prediction techniques are implemented by
using di erent paradigms. Content-based and collaborative
ltering are the two principal paradigms for computing
recommendations [
To make decisions about which predictor is suitable for the current context, threshold values, predictors' state and users feedback are used by the adaptive prediction strategy. The state of a predictor is described by the amount and at the quality of knowledge that is available to the predictor. In other words, the knowledge that is used by the prediction technique is the basis of its predictions. One of the objectives of the prediction strategy is to select the right prediction technique according to the current states of the predictors. The initial strategy is de ned by using the expert knowledge. System starts with an initial prediction strategy. Later on the Learning Module adjusts the prediction strategy to the current systems' domain.
Decision trees and decision rules are model based strategies.
Each node in a decision tree represents some attributes and
each branch from a node corresponds to a possible value for
that attribute. When trying to classify a certain instance,
as seen in the Figure 1, one starts at the root of the decision
tree and move down the branches of the tree according to
the values of the attributes until a leaf node is reached. The
leaf nodes represent the decisions of the tree [
Adaptive prediction strategy is in the form of a decision tree. Another way to represent decision tree is using decision rules. Decision rules generally take the form of IF . . . THEN . . . rules, i.e. IF attribute1 = value1 AND attribute2 = value2 THEN result2.
Depending on the nature of the domains (movie, music, book etc.) di erent attribute-value combinations can be used for prediction strategy design. In our proposed system, since we tested on MovieLens dataset, we choice these speci c attributes that have meaningful correlations between movie domain and prediction techniques. We believe that, by measuring the changes on these attributes, we can capture the domain drifts and trends 1. item ratings count, the number of ratings that the current item has. 2. item similar user count, similar users count that have already rated the current item. 3. similar item count, the number of similar items count according to similarity measures. 4. main genre interest, main genre interest certainty of the current item among the users items. 5. sub genre interest, sub genre interest certainty of the current item among the users items.
Decision trees/decision rules are constructed using the combination of the above ve attributes. As shown in Figure 1 at each node of the decision tree a attribute is compared with a value, which is called threshold value. These ve attributes are used to classify the prediction algorithms. The Recommender System needs to work with the best suited prediction technique for its domain and users. In our system, we used and tested di erent prediction techniques. These are; topNDeviation, userAverage, socialFiltering, CBR (Case Based Reasoning), mainGenreLMS, subGenreLMS, informationFiltering. Because of the dynamic nature of the domain, these attributes create di erent forms of decision trees. Each domain has unique characteristics including user behaviors, emerging trends etc. Recommender engine able to adapt itself to the changes in the domain by analyzing the changes. Also the system can capture the trends in the domain able to re-design its' attached prediction strategy. In our system, we used and tested di erent prediction techniques.
The quality of a decision tree depends on both the classi
cation accuracy and the size of the tree[
Figure 2 depicts the architectural overview of the proposed AdaRec system. Our experimental framework is an extension of the open-source Duine Framework. System consists 3Waikato Environment for Knowledge http://www.cs.waikato.ac.nz/ml/weka/ of two core parts, Recommender Engine and Learning Module.
Recommender Engine is responsible for generating the predictions of items based on the previous user pro les and item contents. It attempts to recommend information items, such as movies, music, books, that are likely to be of interest to the user. The recommender generate the predictions by using its attached prediction strategy. The implementation here uses the open source Duine Framework for the recommender engine.
Learning Module handles the new prediction strategy creation upon the previous instances and performance results of the prediction techniques on each learning cycle. It allows the building of new decision trees/decision rules based on the previous recorded instances.
Learning cycle is a logical concept that represents the
redesign frequency of the prediction strategy. Each instance,
based on the indicated count, and prediction algorithm's
performance results are collected between two learning
cycles. The learning module modi es the values of attributes
in decision rules, which is also called threshold values
according to the gathered results of the prediction techniques
performance. The old prediction strategy is modi ed by using
recommender engines' machine learning algorithm (rule
tuning, rule adaptation, decision tree induction etc.). The
modi cation of the threshold values allows recommender system
to analyze&adapt the nature of the users and the domain.
The learning module rst tests the accuracy of the each
predictor in the system. Than the prediction strategy is
redesigned by the learning module in order to improve proper
use of predictors. Adaptive prediction strategy improves its'
prediction accuracy by learning better when to use which
predictors. The learning module adapts the hybrid
recommender system to the current characteristics of domain.
Previous predictions and user feedbacks are fed to the
training set of the next learning cycle. Inductive learning is used
in learning from the training set. In our experiments we
tested di erent (1K, 1.5K, 2K and 3K) instance sizes for
training sets. The training set contains the instances from
the previous learning cycle results. There are quite a few
inductive learning techniques to choose from, including
information theoretic ones (e.g. Rocchio classi er), neural
networks (e.g. back-propagation), instance-based methods
(e.g. nearest neighbour), rule learners (e.g. RIPPER),
decision trees (e.g. C4.5) and probabilistic classi ers (e.g. naive
Bayes) [
The User Pro le, is the representation of the user in the system. For each active user a user model is stored in the user pro le. User pro le holds the knowledge (such as preferences, feedbacks, feature weights etc.) about users in a structured way. The Recommender Shell, encapsulates the Recommender Engine's interaction with other modules. The shell serves the created prediction strategies to the engine. The Prediction Parser, produces the performance results of the prediction algorithms based on the analyzing of the collected predictions & feedbacks. This module handles the decomposition of the prediction results and generates the training set of sample instances with current attributes. User feedbacks and MAE (Mean Absolute Error) are the main criteria, which describe the trends in the domain. Adaptive prediction strategy learns its domain trends over time via unobtrusive monitoring and relevance feedback. In our proposed system, we focused self adaptive prediction strategy that classi es according to its' attached machine learning technique. This prediction strategy adapts itself to the current context by using the previous performance results of the techniques. Di erent machine learning algorithms that induce decision trees or decision rule sets could be attached to our experimental design.The architecture is open and exible enough to attach di erent machine learning algorithms.
In this section we present a brief discussion of our experimental dataset, evaluation metric followed by the experimental results and discussion.
In order to assess the impact of our proposed adaptive recommender and di erent machine learning algorithms, we calculated prediction accuracy (MAE) of the system using di erent con gurations of the machine learning schemes. Di erent MovieLens datasets are examined during the experiments.
The datasets are divided into temporal subsets according to their time-stamp values. Natural domain trends and changes in user interests are handled by using the subsets of the dataset.
We used data for our recommender system from MovieLens4,
which is a web-based research recommender system that
debuted in Fall 1997 [
In our experiments MovieLens one million ratings dataset
is used, with 6040 users and 3900 movies pertaining to 19
genres. MovieLens dataset contains explicit ratings about
movies and has a very high density. In order to train the
recommender system, the MovieLens dataset is divided in to
di erent temporal sets based on their distribution in time.
When testing the ratings of rst sets are used for
recommender engine training [
Recommender systems researchers use a number of di
erent measures for evaluating the success of the
recommendation or prediction algorithms [
PN M AE = i=1 jpi; rij
N where pi is the predicted value for item i and ri is the user's rating. 4http://www.movielens.umn.edu The aim of the experiments is to examine how the recommendation quality is a ected by our proposed learning module. The present model of the Duine Framework is non adaptive but it supports predictor level learning. This original state of the framework is referred to baseline. As shown in the Figure 1, Duine recommender uses a static prediction strategy as its' hybridization scheme, which does not change at run-time. We want to compare the prediction quality obtained from the framework's baseline (non adaptive) to the quality obtained by our proposed experimental framework (adaptive). The approach will be considered useful if the prediction accuracy is better than the baseline. At the rst iterations both systems are initialized with the same strategy, which is the default strategy of the Duine Framework. The validation process is handled using the following procedure: 1. The ratings provided by the dataset are fed to the system one by one, in the logical order of the system (ordered by timestamps). 2. When a rating is provided during validation, prediction strategy is invoked to provide a prediction for the current user and the current item. The average prediction error can be used a performance indicator of the attached prediction strategy. 3. After the error has been calculated, the given rating is provided as feedback to the recommender system. The adaptive system collects the feedback as well as the current attributes of the system as instances. 4. Whenever the collected instances reached the learning cycle's instance count (1000 instances for example), the prediction strategy of the system will be redesigned by the adaptive system according to the instances. This way, when the next learning cycle is processed, the adaptive system has learned from all the previously processed ratings. This process is repeated for all ratings at both adaptive and non-adaptive (baseline) systems in the test set. At the end MAE is calculated by averaging absolute errors within the baseline and the adaptive system, as described above.
Each prediction provided by the two di erent systems are examined. The prediction accuracy and the prediction error (MAE) are recorded. In experimenting with the MovieLens dataset, we considered both the proposed method, called adaptive system, and the existing method, called baseline. In the experiments, di erent number of instances, such as 1K, 2K and 3K, are used. The purpose of di erent number of instances was to compare the in uence of the instance size on algorithms at the same domain. In the adaptive system, C4.5, BF-Tree and Conjunctive Rules classi er algorithms are attached to the learning module and its results are recorded. We tuned the algorithms to optimize and congured to deliver the highest quality prediction without concern for performance.
We also plot the result of the best MAE (less is better) of the hybrid recommender in the current context at each iteration. The best MAE is referred as best at the charts. Therefore it is possible to compare the performance of the algorithms and the best possible result.
Figure 6 presents the prediction quality (average MAE) results of our experiments for the adaptive system as well as the original system referred to baseline. J48 algorithm is used in these experiments. In this chart, prediction quality is plotted for each of the iterations. On each iteration adaptive system re-designs its prediction strategy according to the previous iteration's performance result (feedbacks and results). It can be inferred from chart that, the prediction quality of the adaptive system performs better than the baseline. It can also be observed from the chart that the adaptive system adapts itself to the changes in the domain and users.
Figure 4 presents the prediction accuracy results of the different instance sizes. In the gure, we also plot the overall performance of ML algorithms as average. It can also be observed from the charts that as we increase the instance size of algorithms the quality tends to be superior (decreased MAE). In case of other algorithms it is expected that increasing the number of instances would mean small MAE values. The same trend is observed in the case of 2K and 3K instances. In order to determine the impact of the instance size, we carried out an experiment where we varied the value of instance size (1K, 1.5K, 2K and 3K). For each of these training set/instance size values we run our experiments. Figure 6 presents the whole picture of the adaptive system's performance results. From the plot we observe that the quality of MAE increases as we increase the instance size.
The results also show that, when using well tuned algorithms, the adaptive system is stable (better than the baseline) in obtaining the average prediction accuracy. This durability, which can be called the impact of learning, is established by the learning module. In order to adapt recommender engine to the current trends, the learning module re-designs the prediction strategy. The learning ability supports the recommender system adaptation to the changes.
In this paper, we introduced an adaptive hybrid recommender system, called AdaRec, that combines several recommender algorithms in which the combination parameters are learned and dynamically updated from the results of previous predictions. Research study shows that traditional static hybrid recommender systems su er from changing user preferences. In order to improve the recommendation performance, we handle domain drifts in our approach. The Learning Module re-designs its prediction (switching) strategy according to the performance of prediction techniques based on user feedbacks. As a result, the system adapts to the application domain, and the performance of recommendation increases as more data are accumulated. In the MovieLens dataset, the proposed adaptive system outperforms the baseline (naive hybrid system).
Initial experimental results show its potential impacts. Therefore, for the next step, we plan to further testing the learning module with various heterogeneous datasets. It would be interesting to examine the di erent domains other than movie (such as music, book, news etc. ). Also, our future work will explore the e ectiveness of other machine learning techniques for use in learning module. In our experiments we xed the used attributes for domain monitoring. It would be also interesting to use dynamic attributes, which means to use di erent attributes on di erent iterations. We believe that with this adaptive learning module, a traditional hybrid recommender should have higher chance to allow its users to e ciently obtain an accurate and con dent decision.