Item-based collaborative filtering recommendation algorithms
Summary (6 min read)
1. INTRODUCTION
- The amount of information in the world is increasing far more quickly than their ability to process it.
- A new user, Neo, is matched against the database to discover neighbors, which are other users who have historically had similar taste to Neo.
- In some ways these two challenges are in con ict, since the less time an algorithm spends searching for neighbors, the more scalable it will be, and the worse its quality.
- The bottleneck in conventional collaborative ltering algorithms is the search for neighbors among a large user population of potential neighbors [12].
- Recommendations for users are computed by nding items that are similar to other items the user has liked.
1.3 Organization
- The next section provides a brief background in collaborative ltering algorithms.
- The authors rst formally describe the collaborative ltering process and then discuss its two variants memorybased and model-based approaches.
- The authors then present some challenges associated with the memory-based approach.
2. COLLABORATIVE FILTERING BASED RECOMMENDER SYSTEMS
- Recommender systems systems apply data analysis techniques to the problem of helping users nd the items they would like to purchase at E-Commerce sites by producing a predicted likeliness score or a list of top{N recommended items for a given user.
- 0.1 Overview of the Collaborative Filtering Process Opinions can be explicitly given by the user as a rating score, generally within a certain numerical scale, or can be implicitly derived from purchase records, by analyzing timing logs, by mining web hyperlinks and so on [28, 16].
- Researchers have devised a number of collaborative ltering algorithms that can be divided into two main categories|.
- Memory-based algorithms utilize the entire user-item database to generate a prediction.
3. ITEM-BASED COLLABORATIVE FILTERING ALGORITHM
- Unlike the user-based collaborative ltering algorithm discussed in Section 2, the item-based approach looks into the set of items the target user has rated and computes how similar they are to the target item i and then selects k most similar items fi1; i2; : : : ; ikg.
- At the same time their corresponding similarities fsi1; si2; : : : ; sikg are also computed.
- Once the most similar items are found, the prediction is then computed by taking a weighted average of the target user's ratings on these similar items.
- The authors describe these two aspects, namely, the similarity computation and the prediction generation in details here.
3.1 Item Similarity Computation
- One critical step in the item-based collaborative ltering algorithm is to compute the similarity between items and then to select the most similar items.
- The basic idea in similarity computation between two items i and j is to rst isolate the users who have rated both of these items and then to apply a similarity computation technique to determine the similarity si;j .
- Figure 2 illustrates this process; here the matrix rows represent users and the columns represent items.
- There are a number of di erent ways to compute the similarity between items.
- These are cosine-based similarity, correlation-based similarity and adjusted-cosine similarity.
3.1.3 Adjusted Cosine Similarity
- Computing similarity using basic cosine measure in item-based case has one important drawback|the di erences in rating scale between di erent users are not taken into account.
- The adjusted cosine similarity o sets this drawback by subtracting the corresponding user average from each co-rated pair.
- Formally, the similarity between items i and j using this scheme is given by sim(i; j) = P u2U(Ru;i Ru)(Ru;j Ru)qP u2U(Ru;i Ru)2 qP u2U (Ru;j Ru)2 : Here Item-item similarity is computed by looking into co-rated items only.
- Isolation of the co-rated items and similarity computation, also known as Figure 2.
3.2.1 Weighted Sum
- Each ratings is weighted by the corresponding similarity si;j between items i and j.
- Formally, using the notion shown in Figure 3 the authors can denote the prediction Pu;i as Pu;i = P all similar items, N (si;N Ru;N)P all similar items, N (jsi;N j) Basically, this approach tries to capture how the active user rates the similar items.
- The weighted sum is scaled by the sum of the similarity terms to make sure the prediction is within the prede ned range.
3.2.2 Regression
- This approach is similar to the weighted sum method but instead of directly using the ratings of similar items it uses an approximation of the ratings based on regression model.
- In practice, the similarities computed using cosine or correlation measures may be misleading in the sense that two rating vectors may be distant (in Euclidean sense) yet may have very high similarity.
- In that case using the raw ratings of the \so-called" similar item may result in poor prediction.
- The basic idea is to use the same formula as the weighted sum technique, but instead of using the similar item N 's \raw" ratings values Ru;N 's, this model uses their approximated values R 0 u;N based on a linear regression model.
- The regression model parameters and are determined by going over both of the rating vectors.
3.3 Performance Implications
- The largest E-Commerce sites operate at a scale that stresses the direct implementation of collaborative ltering.
- For each item j the authors compute the k most similar items, where k n and record these item numbers and their similarities with j.
- Based on this model building step, their prediction generation algorithm works as follows.
- The authors observe a quality-performance trade-o here: to ensure good quality they must have a large model size, which leads to the performance problems discussed above.
- Then the authors perform experiments to compute prediction and response-time to determine the impact of the model size on quality and performance of the whole system.
4.1 Data set
- The authors used experimental data from their research website to evaluate di erent variants of item-based recommendation algorithms.
- Each week hundreds of users visit MovieLens to rate and receive recommendations for movies.
- For this purpose, the authors introduced a variable that determines what percentage of data is used as training and test sets; they call this variable x.
- For their experiments, the authors also take another factor into consideration, sparsity level of data sets.
4.2 Evaluation Metrics
- Recommender systems research has used several types of measures for evaluating the quality of a recommender system.
- They can be mainly categorized into two classes: Statistical accuracy metrics evaluate the accuracy of a system by comparing the numerical recommendation scores against the actual user ratings for the user-item pairs in the test dataset.
- Mean Absolute Error (MAE) between ratings and predictions is a widely used metric.
- The MAE is computed by rst summing these absolute errors of the N corresponding ratings-prediction pairs and then computing the average.
- The most commonly used decision support accuracy metrics are reversal rate, weighted errors and ROC sensitivity [23].
4.2.1 Experimental Procedure
- The authors started their experiments by rst dividing the data set into a training and a test portion.
- Before starting full experimental evaluation of di erent algorithms the authors determined the sensitivity of di erent parameters to di erent algorithms and from the sensitivity plots they xed the optimum values of these parameters and used them for the rest of the experiments.
- For conducted a 10-fold cross validation of their experiments by randomly choosing di erent training and test sets each time and taking the average of the MAE values.
- To compare the performance of item-based prediction the authors also entered the training ratings set into a collaborative ltering recommendation engine that employs the Pearson nearest neighbor algorithm (user-user).
- All their experiments were implemented using C and compiled using optimization ag 06.
4.3 Experimental Results
- In this section the authors present their experimental results of applying item-based collaborative ltering techniques for generating predictions.
- The authors results are mainly divided into two parts|quality results and performance results.
- In assessing the quality of recommendations, the authors rst determined the sensitivity of some parameters before running the main experiment.
- These parameters include the neighborhood size, the value of the training/test ratio x, and e ects of di erent similarity measures.
- For determining the sensitivity of various parameters, the authors focused only on the training data set and further divided it into a training and a test portion and used them to learn the parameters.
4.3.1 Effect of Similarity Algorithms
- The authors implemented three di erent similarity algorithms basic cosine, adjusted cosine and correlation as described in Section 3.1 and tested them on their data sets.
- For each simi- Relative performance of different similarity measures M A E larity algorithms, the authors implemented the algorithm to compute the neighborhood and used weighted sum algorithm to generate the prediction.
- The authors ran these experiments on their training data and used test set to compute Mean Absolute Error (MAE).
- It can be observed from the results that o setting the user-average for cosine similarity computation has a clear advantage, as the MAE is signi cantly lower in this case.
- Hence, the authors select the adjusted cosine similarity for the rest of their experiments.
4.3.2 Sensitivity of Training/Test Ratio
- For each of these training/test ratio values the authors ran their experiments using the two prediction generation techniques{basic weighted sum and regression based approach.
- The regression-based approach shows better results than the basic scheme for low values of x but as the authors increase x the quality tends to fall below the basic scheme.
4.3.3 Experiments with neighborhood size
- The size of the neighborhood has signi cant impact on the prediction quality [12].
- The authors can observe that the size of neighborhood does a ect the quality of prediction.
- But the two methods show di erent types of sensitivity.
- The basic item-item algorithm improves as the authors increase the neighborhood size from 10 to 30, after that the rate of increase diminishes and the curve tends to be at.
- On the other hand, the regression-based algorithm shows decrease in prediction quality with increased number of neighbors.
4.3.4 Quality Experiments
- Once the authors obtain the optimal values of the parameters, they compare both of their item-based approaches with the benchmark user-based algorithm.
- It can be observed from the charts that the basic Sensitivity of the parameter x M A E Sensitivity of the Neighborhood Size item-item algorithm out performs the user based algorithm at all values of x (neighborhood size = 30) and all values of neighborhood size (x = 0:8).
- Similarly at a neighborhood size of 60 user-user and item-item schemes show MAE of 0:732 and 0:726 respectively.
- The authors draw two conclusions from these results.
- Second, regression-based algorithms perform better with very sparse data set, but as the authors add more data the quality goes down.
4.3.5 Performance Results
- After showing that the item-based algorithm provides better quality of prediction than the user-based algorithm, the authors focus on the scalability issues.
- As discussed earlier, itembased similarity is more static and allows us to precompute the item neighborhood.
- This precomputation of the model has certain performance bene ts.
- To make the system even more scalable the authors looked into the sensitivity of the model size and then looked into the impact of model size on the response time and throughput.
4.4 Sensitivity of the Model Size
- Using the training data set the authors precomputed the item similarity using di erent model sizes and then used only the weighted sum prediction generation technique to provide the predictions.
- The authors then used the test data Sensitivity of the model size (at selected train/test ratio) M A E set to compute MAE and plotted the values.
- The authors repeated the entire process for three di erent x values (training/test ratios).
- The most important observation from these plots is the high accuracy that can be achieved using only a fraction of items.
- It appears from the plots that it is useful to precompute the item similarities using only a fraction of items and yet possible to obtain good prediction quality.
4.4.1 Impact of the model size on run-time and throughput
- Given the quality of prediction is reasonably good with small model size, the authors focus on the run-time and throughput of the system.
- This di erence is even more prominent with x = 0:8 where a model size of 200 requires only 1:292 seconds and the basic item-item case requires 36:34 seconds.
- To make the numbers comparable the authors compute the throughput (predictions generated per second) for the model based and basic item-item schemes.
4.5 Discussion
- From the experimental evaluation of the item-item collaborative ltering scheme the authors make some important observations.
- First, the item-item scheme provides better quality of predictions than the use-user (k-nearest neighbor) scheme.
- The improvement in quality is consistent over di erent neighborhood size and training/test ratio.
- The improvement is not signi cantly large.
- The authors experimental results support that claim.
5. CONCLUSION
- Recommender systems are a powerful new technology for extracting additional value for a business from its user databases.
- These systems help users nd items they want to buy from a business.
- Conversely, they help the business by generating more sales.
- Recommender systems are being stressed by the huge volume of user data in existing corporate databases, and will be stressed even more by the increasing volume of user data available on the Web.
- In this paper the authors presented and experimentally evaluated a new algorithm for CF-based recommender systems.
Did you find this useful? Give us your feedback
Citations
9,873 citations
Cites background or methods from "Item-based collaborative filtering ..."
...Other collaborative filtering methods include a Bayesian model [20], a probabilistic relational model [37], a linear regression [91], and a maximum entropy model [75]....
[...]
...[91] proposed using the same correlation-based and cosinebased techniques to compute similarities between items instead and obtain the ratings from them....
[...]
...Other important research issues that have been explored in the recommender systems literature include explainability [12], [42], trustworthiness [28], scalability [4], [39], [91], [93], and privacy [82], [93] issues of recommender systems....
[...]
...In the cosine-based approach [15], [91], the two users x and y are treated as two vectors in m-dimensional space, where m 1⁄4 jSxyj....
[...]
...In addition, [29], [91] present empirical evidence that item-based algorithms can provide better computational performance than traditional user-based collaborative methods while, at the same time, providing comparable or better quality than the best available userbased algorithms....
[...]
5,686 citations
Cites background or methods from "Item-based collaborative filtering ..."
...Several researchers have studied novelty and serendipity in the context of collaborative .ltering systems [Sarwar et al. 2001]....
[...]
...Shardanand and Maes [1995] measured “reversals”—large errors between the predicted and actual rating; we have used the signal-processing measure of the Receiver Operating Characteristic curve [Swets 1963] to measure a recommender’s potential as a filter [Konstan et al....
[...]
...Several researchers have studied novelty and serendipity in the context of collaborative filtering systems [Sarwar et al. 2001]....
[...]
4,788 citations
4,419 citations
Cites background or methods from "Item-based collaborative filtering ..."
...While early literature on recommendation has largely focused on explicit feedback [30, 31], recent attention is increasingly shifting towards implicit data [1, 14, 23]....
[...]
...In terms of user personalization, this approach shares a similar spirit as the item–item model [31, 25] that represents a user as her rated items....
[...]
..., ratings and clicks), known as collaborative filtering [31, 46]....
[...]
4,372 citations
References
12,443 citations
5,915 citations
5,644 citations
"Item-based collaborative filtering ..." refers background or methods in this paper
...One of the most promising such technologies is collaborative ltering [19, 27, 14, 16]....
[...]
...Collaborative Filtering (CF) [19, 27] is the most successful recommendation technique to date....
[...]
...The GroupLens research system [19, 16] provides a pseudonymous collaborative ltering solution for Usenet news and movies....
[...]
4,883 citations
4,557 citations
"Item-based collaborative filtering ..." refers background or methods in this paper
...schemes have been proposed to compute the association between items ranging from probabilistic approach [6] to more traditional item-item correlations [15, 13]....
[...]
...Clustering model treats collaborative ltering as a classi cation problem [2, 6, 29] and works by clustering similar users in same class and estimating the probability that a particular user is in a particular class C, and from there computes the conditional probability of ratings....
[...]
...The resulting model is very small, very fast, and essentially as accurate as nearest neighbor methods [6]....
[...]
...The Bayesian network model [6] formulates a probabilistic model for collaborative ltering problem....
[...]
...Clustering techniques usually produce less-personal recommendations than other methods, and in some cases, the clusters have worse accuracy than nearest neighbor algorithms [6]....
[...]