Recommender Systems suggest items that are likely to be the most interesting for users, based on the feedback, i.e. ratings, they provided on items already experienced in the past. Time-aware Recommender Systems (TARS) focus on temporal context of ratings in order to track the evolution of user preferences and to adapt suggestions accordingly. In fact, some people's interests tend to persist for a long time, while others change more quickly, because they might be related to volatile information needs. In this paper, we focus on the problem of building an e ective pro le for short-term preferences. A simple approach is to learn the short-term model from the most recent ratings, discarding older data. It is based on the assumption that the more recent the data is, the more it contributes to nd items the user will shortly be interested in. We propose an improvement of this classical model, which tracks the evolution of user interests by exploiting the content of the items, besides time information on ratings. When a new item-rating pair comes, the replacement of an older one is performed by taking into account both a decay function for user interests and content similarity between items, computed by distributional semantics models. Experimental results con rm the e ectiveness of the proposed approach.
Recommender systems adopts information ltering algorithms to suggest items or information that might be interesting to users. In general, these systems analyze the past behavior of a user, build a model or pro le of her interests, and exploit that pro le to nd potentially interesting items. In collaborative approaches, the user pro le is usually the vector of ratings assigned to all items that they have accessed, viewed, or purchased [12]. Content-based approaches rely on item and user descriptions (content) to build item representations and user pro les that suggest items similar to those a target user already rated (and liked) in the past [17].
One limitation of these traditional approaches is that the temporal context
of ratings is not taken in account, but actually user preferences are likely to
change over time: long term interests stay stable for a long time, short term
preferences tend to vary with higher frequency. There are some domains, such
as news recommendation, in which retaining this temporal distinction among
user preferences has an impact on the accuracy of suggestions [
Time-aware Recommender Systems (TARS) fall in the more general category
of contex-aware ones, that exploit the context in which users express their
preferences (such as: location, time, weather, emotional state) in order to adapt rating
prediction depending on the situation in which they are experiencing an item.
TARS focus on temporal context of ratings to adapt the recommendation list
accordingly. Regarding the usage of time information, the literature distinguishes
two classes of methods [
We focus on time-adaptive approaches, speci cally on those adopting some heuristics to penalize older preferences that are presumed to be less valid at recommendation time. These methods could be considered as a particular case of time decay heuristics, but they do not target a speci c recommendation time (morning, week-end, etc.). Our investigation speci cally concerns approaches that adopt some time-based strategy to learn separate models for short-term preferences and for interests that persists for a long time.
In particular, in this paper, we face with the problem of building an e ective short-term preference model able to predict items that will shortly be consumed by the user. To address this issue, a simple and popular approach is the use of sliding windows, that learns the model by including in the training set only the most recent ratings, while older data are discarded or weighted so that they contribute to the model in a limited way. The literature reports controversial results about the adoption of time weights for ratings provided at di erent times. For example, in [10] the authors show that recommendation accuracy is improved by an exponential decay function, while an opposite conclusion is drawn by performing rating prediction on the Net ix Prize dataset [14].
We argue that recent ratings of users re ect their short-term preferences more than old ratings, but this is not true for all old ratings, but only for those provided on items which are di erent from the recently rated ones. In other words, the hypothesis is that older ratings do not correspond de nitely to older preferences, as assumed in the classical sliding window approach, but content of the items should be taken into account as well, in order to discard old interests which di er from new ones.
The research question we want to investigate in this work is: \Is the gradual decay of in uence of ratings, combined with content similarity between items, useful for modeling short-term preferences?"
We propose a time-aware distributional content-based recommender system which suggests items the user will shortly be interested in. It exploits both an exponential decay function and semantic similarity between item descriptions for regulating the item participation in building of the user pro le. Items are described in a WordSpace through the geometrical metaphora of meanings. Related words are represented as near points (i.e. vectors), while the semantics of item descriptions (i.e. text fragments) is computed by summing the vectors associated with their words.
The paper is organized as follows: in Section 2 we discuss some relevant literature and compare existing approaches to the proposed one, described in the following section. Section 4 analyzes the results of the experiments performed to validate our proposal, while conclusions are drawn in the last section. 2
Following the characterization given by [
In [10], the authors propose a novel item-based algorithm (that identi es the
similarity between two items by comparing users' ratings on them), enhanced
by a time-decay function in a way that the items rated recently contribute more
to the prediction of the recommendations. The underlying assumption is the
same as in the sliding window approach: latest ratings reveal latest interests.
The proposed approach showed the actual improvement of precision obtained
by using an exponential decay function to weight ratings. In the later work [11],
a recency-based approach is proposed, in which each rating of the target user
is assigned a weight that is computed according to its deviation from her most
recent ratings on similar items. We ground our approach on similar basis, but
a signi cant di erence is that we adopt a decay function to weigh content
similarity among items recently rated and older ones, in order to select those to be
included in the training set for our model. Time-adaptive heuristics have been
used also by Cao [
Another method that adopts simple heuristics to improve collaborative
ltering is proposed in [
Other approaches adapted factorization algorithms to face with temporal e ects. Remarkable examples of these methods are described in [14] and [21]. Koren [14] suggested that a mere decay of older instances or usage of separate models for tracking the evolution of preferences cause a loss of prediction accuracy. The proposed solution is to model the temporal dynamics along the whole time period, allowing to separate volatile factors from durable ones, in order to capture the way user and product characteristics change over time. Xiong et al. [21] proposed a factorization method based on probabilistic latent factor models. In addition to the factors that are used to characterize entities, the authors introduce another set of latent features for each di erent time period. These additional factors represent the preference of latent features at each particular time, so that they are able to capture the evolution of preferences.
Compared to these work which handled the temporal dynamics in di erent ways, we focus on explicitly modeling short-term preferences and their in uence on recommendation of items that will shortly be consumed.
Among approaches that propose di erent models for long-term and
shortterm preferences, Cantador et al. [
Di erently from previous methods, our short-term model tries to capture semantic similarity among items by looking at their content, rather than simply co-occurence within sessions or clicking history, with the hope that the semantic approach, combined with time information, helps to discover short-term relatedness among them. 3
Models of Distributional Semantics have drawn a lot of attention in recent years due to their capability of capturing semantics at a latent level, without the requirement of learning algorithms or human-edited resources. Such models build a vector space of meanings where concepts are represented through vectors and the relatedness between meanings is expressed through a proximity function of the points they are represented by. Usually these models are built by skimming a large corpus in order to gather information about distribution of words in a text. Indeed, statistics about word co-occurrences are useful to infer paradigmatic relationships among words, i.e. relations about words that can be used interchangeably. One of the commonest use of such a model is for computing the similarity between words, since the vector components grasp the semantic of word usage in context. Then, the vector addition between words belonging to a text is an easy way to extend to a whole sentence/paragraph/document such a similarity. This work exploits the idea of adapting Distributional Semantic Model (DSM) to item descriptions as a uni ed framework for both representing the semantic content of items and computing the similarity between them. 3.1
The distributional semantic-based recommender system relies on Random Indexing (RI) [13] for building up the semantic space. Given a text corpus, RI technique consists of the following two steps: 1. A random vector is assigned to a term in the corpus vocabulary. This vector is highly dimensional, with very few elements -called seed - that take values in f 1; 1g. The dimension of the reduced space corresponds to the random vector dimensionality; 2. The semantic vector representation for the term is built by analyzing the whole corpus and summing the random vectors of co-occurring terms in a given text window.
Mathematically, the sum over random vectors corresponds to multiply the original co-occurrence matrix by a projection operator, which preserves the distance proportion between points. The resulting space, called WordSpace has two strenght points: 1) the reduced dimension enables a quicker computation of similarities, and 2) likewise Latent Semantic Analysis, shrinking the number of components to a smaller set of contexts makes high order relationships more prominent.
The item space is built upon the previously computed WordSpace: the item
representation comes from the sum of semantic vectors associated to the item
textual content. The user pro le, in turn, is built on the basis of the item vector
representations she liked or disliked before. In particular, the model keeps trace
of both positively and negatively rated items and builds two di erent pro le
vectors, u+ and u , as the sum of positive and negative items, respectively.
In order to get a single pro le vector on which basis the model will compute
the recommendations, we exploit the orthogonal projection operator, which has
been successfully employed in both retrieval and recommendation scenarios [
A time-adaptive user pro le should be able to grasp changes in the user's
behaviour in order to re ect the latest user tastes. We tackle this problem by
pro ling the user preferences with respect to a sliding window of time: i.e. the
items which contribute to building the pro le are those occurred shortly before
the recommendation. This kind of short-term model was initially proposed by
Billsus and Pazzani [
Let I = i1; :::; ik be such a pro le set, every time the user rates a new item inew, this is automatically added to I, while an older item is removed. The element to be discarded is selected as follows: iold = argminfi 2 I; sim(i; inew) e (tnew ti) g (1) where t identi es the rating time, for both the new item and that under observation, sim is the similarity between two items computed in the item space described in the previous subsection, and is a decay factor. Equation 1 aims to eliminate the item most dissimilar from the newly introduced one. However, in doing so we try to keep coherent the user pro le by weighting this factor with the exponential function, whose role is to smooth similarity through time. Then, two possibilities may occur: 1. The new item completely diverges from the user history. In this case all items in I will take on a small similarity and the exponential function will contribute more to equation 1; 2. The new item has some degree of similarity with some items in I. In this case the similarities will be reduced by the exponential function which serves to mediate the contribute of those similar items rated a long time ago. Among these, the rst scenario is the most interesting, since it re ects a potential new trend in the user's preference. Under this condition, the e ect of equation 1 on a new incoming item would be that of consolidating this new trend in the user pro le, if a newly added item is similar to that latest one, or quickly discard the \exception", thank to the contribute of the exponential function. The parameter plays in this context an important role, since it regulates how fast the exponential function has to reduce its rate [10]. 4
The goal of the evaluation is to assess the capability of the proposed timeadaptive algorithm to re ect the recency of user interests without loosing information about consolidated long-lasting preferences. Then, we compare our proposed algorithm (sim) with respect to a simple sliding window strategy ( fo). In fo, the window of items is kept constant: as a new item is added, the older one is removed, thus following a rst-in rst-out strategy. This approach discards items on the basis of a mere time factor, and no information about user preferences in long-time is preserved. 4.1
The evaluation is performed on the same dataset proposed by Adomavicius et
al. [
The objective behind the use of such a corpus is to provide as wider as possible coverage of all word usages in a language. The WordSpace represents the top 150; 000 most frequent keywors, the vector dimension is set up to 400, the number of seeds is of 10 elements, while the window size for computing cooccurrences is of 5 words. The recommender system is implemented in Java, and relies on Lucene API3 for building both the WordSpace and the recommendation model. The factor in equation 1 is set to 0.01. 4.2
We evaluate the proposed model on a top-N recommendation task. The
evaluation is performed as a time-dependent cross-validation, based on increasing
time window [
We compared the two time-adaptive methods on a variable k1, i.e. the training set dimension. We set k2 = 5 and N = 3, while varing k1 2 f5; : : : ; 15g. We decided to use this strategy in order to assess the validity of the proposed method when a wider pro le set is available.
Tables 1a and 1b report the results of the evaluation with respect to MAP and NDCG metrics. Both tables show a similar trend. The best result is obtained on 2 http://www.ark.cs.cmu.edu/personas/ 3 http://lucene.apache.org/ k1 = 7 by sim method, which gives also better overall gures. There are only two exceptions to this trend, with k1 = 11 and k1 = 14, although the di erences in reported values are very small. Another common trend showed by results is that, as we increase the k1 dimension, the di erences between sim and fo become smaller. However, we ascribe such a trend to the fact that by increasing the dimension of the training set, the di erences between fo and sim strategies become smaller since the variability in the set decreases (i.e. an increasingly number of items from the past of the user become part of the training set). Generally, the behavior of a user may be determined by her long-term interests, but at any given time, she is also a ected by short-term preferences or information needs. In this paper, we argue that time is not the only factor to be taken into account in order to distinguish among short-time and long-time interests. We started from the classical sliding window approach and suggested that older ratings do not correspond de nitely to older preferences, but content of the items should be considered as well. We proposed an approach that models short-term preferences by adopting a content-based sliding window approach: when a new ratings comes into the system, the replacement of an older one is performed by taking into account both a decay function for user interests and content similarity between items on which ratings are provided, computed by distributional semantics models. We compared the proposed approach to the simple FIFO strategy (the new rating replaces the oldest one). Experimental results con rmed the hypothesis that the gradual decay of in uence of ratings, combined with content similarity between items, is actually useful for modeling short-term preferences, especially when a few ratings are available to train the system. As a future work, we plan to evaluate our short-term model on a wider dataset. Furthermore, we want to design a model for long-term preferences, as well as a way to integrate the two models in order to have an overall recommendation strategy.
This work ful ls the research objectives of the PON 02 00323 2938699 CUP B86D130000300007 project \EFFEDIL - SOLUZIONI INNOVATIVE PER L'EFFICIENZA ENERGETICA IN EDILIZIA" funded by the Italian Ministry of University and Research (MIUR). 10. Ding, Y., Li, X.: Time weight collaborative ltering. In: Proceedings of the 14th ACM International Conference on Information and Knowledge Management. pp. 485{492. CIKM '05, ACM, New York, NY, USA (2005), http://doi.acm.org/10. 1145/1099554.1099689 11. Ding, Y., Li, X., Orlowska, M.E.: Recency-based collaborative ltering. In: Proceedings of the 17th Australasian Database Conference - Volume 49. pp. 99{107. ADC '06, Australian Computer Society, Inc., Darlinghurst, Australia, Australia (2006), http://dl.acm.org/citation.cfm?id=1151736.1151747 12. Herlocker, J.L., Konstan, J.A., Borchers, A., Riedl, J.: An algorithmic framework for performing collaborative ltering. In: SIGIR '99: Proceedings of the 22nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, August 15-19, 1999, Berkeley, CA, USA. pp. 230{237. ACM (1999), http://doi.acm.org/10.1145/312624.312682 13. Kanerva, P.: Sparse Distributed Memory. MIT Press (1988) 14. Koren, Y.: Collaborative ltering with temporal dynamics. In: Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. pp. 447{456. KDD '09, ACM, New York, NY, USA (2009), http://doi. acm.org/10.1145/1557019.1557072 15. Koren, Y.: Collaborative ltering with temporal dynamics. Commun. ACM 53(4), 89{97 (Apr 2010), http://doi.acm.org/10.1145/1721654.1721677 16. Lathia, N., Hailes, S., Capra, L.: Temporal collaborative ltering with adaptive neighbourhoods. In: Proceedings of the 32Nd International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 796{797. SIGIR '09, ACM, New York, NY, USA (2009) 17. Lops, P., de Gemmis, M., Semeraro, G.: Content-based recommender systems: State of the art and trends. In: Ricci, F., Rokach, L., Shapira, B., Kantor, P.B. (eds.) Recommender Systems Handbook, pp. 73{105. Springer (2011), http://dx. doi.org/10.1007/978-0-387-85820-3_3 18. Musto, C., Semeraro, G., Lops, P., de Gemmis, M.: Random indexing and negative user preferences for enhancing content-based recommender systems. In: Huemer, C., Setzer, T. (eds.) E-Commerce and Web Technologies, Lecture Notes in Business Information Processing, vol. 85, pp. 270{281. Springer Berlin Heidelberg (2011) 19. Shani, G., Gunawardana, A.: Evaluating recommendation systems. In: Ricci, F., Rokach, L., Shapira, B., Kantor, P.B. (eds.) Recommender Systems Handbook, pp. 257{297. Springer US (2011) 20. Xiang, L., Yuan, Q., Zhao, S., Chen, L., Zhang, X., Yang, Q., Sun, J.: Temporal recommendation on graphs via long- and short-term preference fusion. In: Proceedings of the 16th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. pp. 723{732. KDD '10, ACM, New York, NY, USA (2010), http://doi.acm.org/10.1145/1835804.1835896 21. Xiong, L., Chen, X., Huang, T., Schneider, J.G., Carbonell, J.G.: Temporal collaborative ltering with bayesian probabilistic tensor factorization. In: Proceedings of the SIAM International Conference on Data Mining, SDM 2010, April 29 - May 1, 2010, Columbus, Ohio, USA. pp. 211{222. SIAM (2010), http: //dx.doi.org/10.1137/1.9781611972801.19 22. Zimdars, A., Chickering, D.M., Meek, C.: Using temporal data for making recommendations. In: Proceedings of the 17th Conference in Uncertainty in Arti cial Intelligence. pp. 580{588. UAI '01, Morgan Kaufmann Publishers Inc., San Francisco, CA, USA (2001), http://dl.acm.org/citation.cfm?id=647235.720264