Mahout: adjusted cosine similarity for item based recommender

问题

For an assignment I'm supposed to test different types of recommenders, which I have to implement first. I've been looking around for a good library to do that (I had thought about Weka at first) and stumbled upon Mahout.

I must therefore put forward that: a) I'm completely new to Mahout b) I do not have a strong background in recommenders nor their algorithms (otherwise I wouldn't be doing this class...) and c) sorry but I'm far from being the best developper in the world ==> I'd appreciate if you could use layman terms (as far as possible...) :)

I've been following some tutorials (e.g. this, as well as part2) and got some preliminary results on item-based and user-based recommenders.

However, I'm not very happy with the item-based prediction. So far, I've only found similarity functions that do not take into consideration the users' rating-biases. I was wondering if there is something like adjusted cosine similarity. Any hints?

回答1:

Here is a sample of the AdjustedCosineSimilarity I created. You must remember that this will be slower than PearsonCorrelationSimilarity because of the sqrt computations, but will produce better results. At least for my dataset results were much better. But you should make a trade off, quality/performance, and depending of your needs you should use the implementation you want.

/**
 * Custom implementation of {@link AdjustedCosineSimilarity}
 * 
 * @author dmilchevski
 *
 */
public class AdjustedCosineSimilarity extends AbstractSimilarity {

  /**
   * Creates new {@link AdjustedCosineSimilarity}
   * 
   * @param dataModel
   * @throws TasteException
   */
    public AdjustedCosineSimilarity(DataModel dataModel)
            throws TasteException {
        this(dataModel, Weighting.UNWEIGHTED);
    }

    /**
     * Creates new {@link AdjustedCosineSimilarity}
     * 
     * @param dataModel
     * @param weighting
     * @throws TasteException
     */
    public AdjustedCosineSimilarity(DataModel dataModel, Weighting weighting)
            throws TasteException {
        super(dataModel, weighting, true);
        Preconditions.checkArgument(dataModel.hasPreferenceValues(),
                "DataModel doesn't have preference values");
    }

    /**
     * Compute the result
     */
    @Override
    double computeResult(int n, double sumXY, double sumX2, double sumY2, double sumXYdiff2) {
        if (n == 0) {
            return Double.NaN;
        }
        // Note that sum of X and sum of Y don't appear here since they are
        // assumed to be 0;
        // the data is assumed to be centered.
        double denominator = Math.sqrt(sumX2) * Math.sqrt(sumY2);
        if (denominator == 0.0) {
            // One or both parties has -all- the same ratings;
            // can't really say much similarity under this measure
            return Double.NaN;
        }
        return sumXY / denominator;
    }

    /**
     * Gets the average preference
     * @param prefs
     * @return
     */
    private double averagePreference(PreferenceArray prefs){
        double sum = 0.0;
        int n = prefs.length();
        for(int i=0; i<n; i++){
            sum+=prefs.getValue(i);
        }
        if(n>0){
            return sum/n;
        }
        return 0.0d;
    }

    /**
     * Compute the item similarity between two items
     */
    @Override
    public double itemSimilarity(long itemID1, long itemID2) throws TasteException {
        DataModel dataModel = getDataModel();
        PreferenceArray xPrefs = dataModel.getPreferencesForItem(itemID1);
        PreferenceArray yPrefs = dataModel.getPreferencesForItem(itemID2);
        int xLength = xPrefs.length();
        int yLength = yPrefs.length();

        if (xLength == 0 || yLength == 0) {
            return Double.NaN;
        }

        long xIndex = xPrefs.getUserID(0);
        long yIndex = yPrefs.getUserID(0);
        int xPrefIndex = 0;
        int yPrefIndex = 0;

        double sumX = 0.0;
        double sumX2 = 0.0;
        double sumY = 0.0;
        double sumY2 = 0.0;
        double sumXY = 0.0;
        double sumXYdiff2 = 0.0;
        int count = 0;

        // No, pref inferrers and transforms don't appy here. I think.

        while (true) {
            int compare = xIndex < yIndex ? -1 : xIndex > yIndex ? 1 : 0;
            if (compare == 0) {
                // Both users expressed a preference for the item
                double x = xPrefs.getValue(xPrefIndex);
                double y = yPrefs.getValue(yPrefIndex);
                long xUserId = xPrefs.getUserID(xPrefIndex);
                long yUserId = yPrefs.getUserID(yPrefIndex);

                double xMean = averagePreference(dataModel.getPreferencesFromUser(xUserId));
                double yMean = averagePreference(dataModel.getPreferencesFromUser(yUserId));

                sumXY += (x - xMean) * (y - yMean);
                sumX += x;
                sumX2 += (x - xMean) * (x - xMean);
                sumY += y;
                sumY2 += (y - yMean) * (y - yMean);
                double diff = x - y;
                sumXYdiff2 += diff * diff;
                count++;
            }
            if (compare <= 0) {
                if (++xPrefIndex == xLength) {
                    break;
                }
                xIndex = xPrefs.getUserID(xPrefIndex);
            }
            if (compare >= 0) {
                if (++yPrefIndex == yLength) {
                    break;
                }
                yIndex = yPrefs.getUserID(yPrefIndex);
            }
        }

        double result;

        // See comments above on these computations
        double n = (double) count;
        double meanX = sumX / n;
        double meanY = sumY / n;
        // double centeredSumXY = sumXY - meanY * sumX - meanX * sumY + n *
        // meanX * meanY;
        double centeredSumXY = sumXY - meanY * sumX;
        // double centeredSumX2 = sumX2 - 2.0 * meanX * sumX + n * meanX *
        // meanX;
        double centeredSumX2 = sumX2 - meanX * sumX;
        // double centeredSumY2 = sumY2 - 2.0 * meanY * sumY + n * meanY *
        // meanY;
        double centeredSumY2 = sumY2 - meanY * sumY;
//      result = computeResult(count, centeredSumXY, centeredSumX2,
//              centeredSumY2, sumXYdiff2);

        result = computeResult(count, sumXY, sumX2, sumY2, sumXYdiff2);

        if (!Double.isNaN(result)) {
            result = normalizeWeightResult(result, count,
                    dataModel.getNumUsers());
        }
        return result;
    }

}

来源：https://stackoverflow.com/questions/29419222/mahout-adjusted-cosine-similarity-for-item-based-recommender